An Intelligent Natural Language Conversational System for Academic Advising


Material Information

An Intelligent Natural Language Conversational System for Academic Advising
Physical Description:
1 online resource (107 p.)
Latorre, Edward M
University of Florida
Place of Publication:
Gainesville, Fla.
Publication Date:

Thesis/Dissertation Information

Doctorate ( Ph.D.)
Degree Grantor:
University of Florida
Degree Disciplines:
Electrical and Computer Engineering
Committee Chair:
Committee Co-Chair:
Committee Members:


Subjects / Keywords:
academic-advising -- advising -- conversational-agent -- dialog-system -- dialogue-manager -- natural-language-processing -- natural-language-understanding -- nlp
Electrical and Computer Engineering -- Dissertations, Academic -- UF
Electrical and Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation


Academic advisors assist students in academic, professional, social and personal matters. Successful advising systems increase student retention, improve graduation rates and help students meet educational goals. Research in academic advising shows a trend of integrating electronic technologies that advisors must adopt to increase their roles in student education, while resources are limited and the proportion of students increases. The main objective of this work is to deploy an expert-based natural language academic advising system, to manage multiple straightforward advising tasks and allow advisors to engage in non-mundane educative tasks. This system features a conversational agent as the user interface, includes modules for data collection, user data management, an academic advising knowledge base and a web design structure for the implementation platform. This system also includes an expert-based system to manage academic tasks such as course sequence planning, and a method to allow the users to contribute to the academic knowledge base. The system is operational for several hundred students of the Electrical and Computer Engineering Department of the University of Florida. This works covers the design, development, deployment and evaluation of this system. The students used the system during three experimental phases. For the third phase, the system performed well, obtaining close to 80%, on the traditional language processing measures of precision, recall, accuracy and F1 score. Assessment from the constituencies showed positive and assuring reviews. The development of this innovative application involves three research contributions to the field. Firstly, a comprehensive literature search and assessment of the academic advising field with recommended solutions based on the latest technological innovations. Secondly, the design of the first-known academic advising multi-task conversational system, using a robust and scalable linguistically inspired algorithm to resolve ambiguous references, suitable for conditions where large corpora of data are not available. The system contains measures for self-evaluation and for upgrading the knowledge base without involvement from the system developers. Thirdly, the implementation and evaluation of the system in a real-world scenario, which shows the viability of the application and initiated the development of a corpus for academic advising, valuable for the language processing and academic research communities.
General Note:
In the series University of Florida Digital Collections.
General Note:
Includes vita.
Includes bibliographical references.
Source of Description:
Description based on online resource; title from PDF title page.
Source of Description:
This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility:
by Edward M Latorre.
Thesis (Ph.D.)--University of Florida, 2014.
Co-adviser: LI,XIAOLIN.

Record Information

Source Institution:
Rights Management:
Applicable rights reserved.
lcc - LD1780 2014
System ID:

This item is only available as the following downloads:

Full Text




2014 Edwa r d M. Latorre Navarro


To my wife, family and friends


4 ACKNOWLEDGMENTS I am thankful to Dr. John G Harris, who encouraged my desire to become a Gator since I first approached him. Thanks to his vision and confidence in me, Albert is a smart, highly valu e d innovation that has been very rewarding to create I am appreciative of the professors on my PhD committee, whose courses directly influenced my path throughout this work. I thank Dr. Wind Cowles and her t eachings in psycholinguistics and biological real time language processing. I thank Dr. A. Antonio Arroyo for the introduction to NLP and machine intelligence. I thank Dr. Xiaolin Li for the experience in designing user friendly cyber physical systems and mobile applications. I express my gratitude for the support of the Electrical and Computer Engineering Department of the University of Florida during my studies. I thank the staff of the ECE Student Services Office for their kind service working with Alber t. I also thank my fellow comrades throughout all these years from the Hybrid Computational Group and Computational NeuroEngineering Laboratory. They have always offered a helping hand. I thank my family and friends throughout Puerto Rico and the United S tates, for always believing in me and being there for me. I thank my neighbor Dr. Jos M.A. Almodvar Fara. I could not ha ve asked for a better teammate. I thank, with all my heart, my mother, my father, my brothers, sister, grandparents, uncles, aunts, c ousins and in laws. My greatest inspirations to be here, Ana I. Navarro Vzquez, Pedro Navarro Barros, Pedro L. Navarro Vzquez, who now have a doctor in the family. The person who is always by my side, with whom I share every experience and makes my world go round, I am grateful for my wife, Karen Vanessa Muratti Rodrguez.


5 TABLE OF CONTENTS P age ACKNOWLEDGMENTS ................................ ................................ ................................ 4 LIST OF TABLES ................................ ................................ ................................ ........... 7 LIST OF FIGURE S ................................ ................................ ................................ ........ 8 LIST OF ABBREVIATIONS ................................ ................................ .......................... 10 ABSTRACT ................................ ................................ ................................ .................. 12 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ ... 14 Academic Advising ................................ ................................ ................................ 14 Natural Language Processing ................................ ................................ ............... 16 Natural Language Understanding ................................ ................................ .... 17 Conversational Agents ................................ ................................ .................... 19 Questi on Answering Systems ................................ ................................ .......... 23 Dissertation Topics ................................ ................................ ................................ 24 2 A CONVERSATIONAL SYSTEM FOR ACADEMIC ADVISING ............................. 26 Academic Advising ................................ ................................ ................................ 26 The Role o f the Academic Advisor ................................ ................................ .. 26 Academic Advising in ECE UF ................................ ................................ ........ 27 Fundamental Requirements for the Advising System ................................ ............ 28 Albert The Intelligent Academic Advising Conversational System ....................... 32 Advising Information ................................ ................................ ........................ 33 System Components ................................ ................................ ....................... 34 Advising Dialogue Manager ................................ ................................ ............ 37 Academic Planning System ................................ ................................ ............. 38 Syst em Evaluation ................................ ................................ .......................... 39 3 THE ADVISING DIALOGUE MANAGER ................................ ............................... 41 Natural Language Processing ................................ ................................ ............... 41 Natural Language Generation ................................ ................................ ......... 42 Natural Language Understanding ................................ ................................ .... 42 NLP Algorithm ................................ ................................ ................................ 43 Task Manager ................................ ................................ ................................ ....... 49 Spell Checking ................................ ................................ ................................ 49 Automatic Updates ................................ ................................ .......................... 52 Scaling Up the Natural Language Scripts ................................ ........................ 52


6 Statistical Data Collection ................................ ................................ ................ 54 4 THE ACADEMIC PLANNING SYSTEM ................................ ................................ 56 The Course Plan Analyzer ................................ ................................ ..................... 57 Obta ining the Academic Information ................................ ................................ 57 User Interface ................................ ................................ ................................ 58 The CoPA System ................................ ................................ ........................... 58 ................................ ................................ ........................ 64 Eval uating CoPA ................................ ................................ ................................ ... 65 5 EXPERIMENTS AND ASSESSMENT ................................ ................................ ... 66 Experiments ................................ ................................ ................................ .......... 66 Phase One ................................ ................................ ................................ ...... 68 Phase Two ................................ ................................ ................................ ...... 69 Phase Three ................................ ................................ ................................ ... 71 Analysis Summary ................................ ................................ ................................ 84 Current State of Albert ................................ ................................ ........................... 85 6 CONCLUSION AND FUTURE WORK ................................ ................................ ... 86 Conclusion ................................ ................................ ................................ ............. 86 Future Work ................................ ................................ ................................ ........... 87 APPENDIX A TOPICS COVERED BY ALBERT ................................ ................................ .......... 88 B SCREENSHOTS OF THE ALBERT WEBSITE ................................ ...................... 89 C SAMPLES FROM USER LOG FILES ................................ ................................ .... 94 D ADVISOR SURVEY QUOTES ................................ ................................ ............... 99 LIST OF REFERENCES ................................ ................................ ............................ 101 BIOGRAPHICAL SKETCH ................................ ................................ ......................... 107


7 LIST OF TABLES Table page 5 1 Results from the survey at the end of the CoPA process ................................ ... 76 5 2 ................................ 79 5 3 Results from the outcomes error estimation ................................ ...................... 82 5 4 System performance estimation results ................................ ............................. 83


8 LIST OF FIGURES Figure page 1 1 Basic architecture of a dialogue system ................................ ............................ 21 2 1 A model of the main components of Albert. ................................ ....................... 35 2 2 A screenshot of the website for Albert in its initial state. ................................ .... 36 3 1 Example of a procedure to match an input. ................................ ....................... 45 3 2 Example of a procedure for an input ................................ ................................ .. 51 3 3 Algorithm for course name submission ................................ .............................. 53 3 4 Questionnaire required to submit the academic planning process ..................... 55 4 1 A screenshot of Albert with the right side showing the first page of CoPA opened inside the frame. ................................ ................................ ................... 59 4 2 A segment of the template for entering the academic record in CoPA. .............. 60 4 3 A screenshot of Albert with the right side showing the second part of CoPA ..... 61 4 4 The template for choosing a course plan in CoPA, with an example of courses available to choose. ................................ ................................ ............. 62 4 5 Front page of the advisor website. ................................ ................................ ..... 64 4 6 The advisor website, with an example student report. ................................ ....... 6 4 5 1 Input statement histogram for phase one. ................................ ......................... 68 5 2 A screenshot of Albert, du ring phase two, showing the FAQ list on the right. .... 70 5 3 Input statement histogram for phase two. ................................ .......................... 70 5 4 St atistics for the video tutorial ................................ ................................ ........... 73 5 5 Input statement histogram for phase three. ................................ ....................... 74 5 6 Catalog year of each user who filled an academic record with CoPA. ............... 75 5 7 Survey results with a standard error bar for each question. ............................... 76 5 8 Distribution of ON input statements. ................................ ................................ .. 80 B 1 Main webpage for Albert ................................ ................................ .................... 89


9 B 2 Main FAQ webpage. ................................ ................................ .......................... 90 B 3 Main webpage for Albert showing second FAQ page. ................................ ....... 90 B 4 Full list of questions in the second FAQ webpage. ................................ ............ 91 B 5 Main webpage for Albert showing page one of CoPA. ................................ ....... 92 B 6 Main webpage for Albert showing an example page two of CoPA. .................... 92 B 7 Main webpage for Albert showing an example of the speech API. ..................... 93


10 LIST OF ABBREVIATIONS AI Artificial Intelligence AIML Artificial Intelligence Markup Language API Application Programming Interface BSCEE Bachelor of Science in Computer Engineering Hardware emphasis BSEE Bachelor of Science in Electrical Engineering CA Conversational Agent CoPA Course Plan Analyzer CS Chatscript DM Dialogue Manager DM Dialogue Manager ECE SSO Electrical and Computer Engineering Student Services Office ECE UF Electrical and Computer Engineering University of Florida FAQ Frequently Asked Questions FN False Negative FP False Positive HCI Human computer interaction IEEE Institute of Electrical and Electronics Engineers IR Information Retrieval ISIS Integrated Student Information System KB Knowledge Base KE Keyword Extraction LD Lexical Disambiguation LM Literal Match


11 LU Lexical Unit MT Machine Translation NLG Natural Language Generation NLP Natural Language Processing NLTK Natural Language Toolkit NLU Natural Language Understanding ON Outcome Negative OP Outcome Positive OS Operating System PM Partial Match POS Part of Speech POS Parts of Speech QA Question Answering TN True Negative TP True Positive WSD Word Sense Disambiguation


12 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy AN INTELLIGENT NATURAL LANGUAGE CONVERSATIONAL SYSTEM FOR ACADEMIC ADVISING By Edward M. Latorre Navarro May 2014 Chair: John G. Harris Major: Electrical and Computer Engineering Academic advisors assist students in academic, professional, social and personal matters. S uccessful advising system s increase student retention, improve graduation rates and help students meet educational goals. Resear ch in academic advising shows a trend of integrating electronic technologies that advisors must adopt to increase their roles in student education, while resources are limited and the proportion of students increases. The main objective of this work is to deploy an expert based natural language academic advising system, to manage multiple straightforward advising tasks and allow advisors to engage in non mu ndane educative tasks. This system features a conversational agent as the user interface, includes modules for data collection, user data management, an academic advising knowledge base and a web design structure for the implementation platform. This syst em also includes an expert based system to manage academic tasks such as course sequence planning, and a method to allow the users to contribute to the academic knowledge base.


13 The system is operational for several hundred students of the Electrical and C omputer Engineering Department of the University of Florida. This works covers the design, development, deployment and evaluation of this system. The students used the system during three experimental phases. For the th ird phase, the system perform e d well, obtaining close to 80%, on the traditional language processing measures of precision, recall, accuracy and F 1 score. Assessment from the constituencies showed positive and assuring reviews. The development of this innovative application involves three research contr ibutions to the field. Firstly, a comprehensive literature search and assessment of the academic advising field with recommended solutions based on the latest technological innovations. Secondly, the design of the first known academic advisin g multi task conversational system, using a robust and scalable linguistically inspired algorithm to resolve ambiguous references, suitable for conditions where large corpora of data are not available. The system contain s measures for self evaluation and f or upgrading the knowledge base without involvement from the system developers. Thirdly, the implementation and evaluation of the system in a real world scenario which shows the viability of the applicat ion and initiated the development of a corpus for ac ademic advising, valuable for the language processing and academic research communities


14 CHAPTER 1 INTRODUCTION Academic Advising Since the late 1800s, higher education institutions have employed academic advisors to assist students in academic, professional, social and personal matters [1] Academic advising was initially a response for a need to guide students in selecting their course sequence and later evolved to include assistance in personal matters. Later, in the 1970s, the advising practice became an examined activity as education institutions began comparing the methods and results of their advising systems [2] Today, academic advising still serves the same basic objectives and peer reviewed research is still very active given the many important implications of a successful ad vising system such as student retention, graduation rates and student educational goals including academic engagement and performance, and career planning [1] [3] Therefore, it is essential for academic programs to offer students an effective advising experience, which requires advisors to innovate at the speed of their stu dents, who each year tend to have higher expectations from their education institutions and a stronger synergy with digital technology. Thus we see the innovation trend in advising has been towards the use of communication technologies such as email, insta nt messaging, social networking, course management systems, podcasts, mobile applications, online videos and blogs [4] [6] A quick survey of university websites shows that many have introduced the concept of eAdvising, that is, utilizing electronic means, usually web based, to offer advising to students [7] [9] In 2003, the authors of [10] suggested that interactive advising modules for personalized discussions could be created using artificial intelligence (AI) techniques. In


15 2008 Leonard [4] detailed the profound effect of technology use by advisors and referenced the idea of an expert based interactive advising module as a possible future trend, but few institutions have shown intere st in developing such a system. A fully automated system is a better solution for the innovation trend in advising and addresses the concern of not enough advisors, for too many students, in an economically challenged environment [1] Such a system will lessen the burden of academic advisors from several mundane tasks and free up more of their time for the deeper aspects of advising such as career planning or managing extraordinary personal situations. There are several research publi cations describing expert based advising systems for helping students with straightforward repetitive tasks such as choosing majors, adhering to an appropriate curriculum sequence or accessing degree audits [11] [15] This work is inspired by the belief that, following current technological trends, new interactive advising s ystems should also include a natural language interface to allow students to communicate as freely and openly as with their actual advisors. Such an innovative system could be much more attractive to students as it would allow them to easily ask a wider ra nge of questions than those in previous expert based systems and obtain immediate responses to these instead of waiting for peers or advisors to read and reply, as with current eAdvising methods. knowledge, two publications exist that pres ent advising systems with an expert based system combined with natural language processing techniques for communication [16] [17] The first system was limited to only yes/no type questions and a few phrases to manage state transitions [16] The other system features an ontology based information retrieval engine to guide students in searching for answers after


16 entering keywords [17] The system developed for this research work allows unrestricted natural language communication utilizing state of the art natural language processing techniques. Natural Language Processing As the name implies, natural language processing (NLP) is the prime modern software technology for analyzing, understanding and interacting with humans in their natural language. The NLP field includes major research topics such as automatic summarization, information retrieval (IR), machine translation (MT), part of speech (POS) tagging, parsing, question answering (QA), sentiment analysis, and word sense disambiguation (WSD). In general, research in NLP comprises knowledge about linguistics, artificial in telligence, machine learning and to some extent, the psycholinguistic and psychological aspects of human language processing. Applications that use NLP involve human computer interaction (HCI) through either computer natural language generation or computer natural language understanding [18] Research in NLP is enjoying a peak in popularity after a widespread release of consumer and commercial HCI applications. Many well recognized HCI NLP applications are classified as artificial Conversational Agents (CA). In principle, these HCI NLP applications are currently challenged in their ability to understand the intention of users within a wide range of demographics, rega rdless of their diversity in syntax, semantics and pragmatics. This challenge is described among fundamental problems in NLP and linguistics such as reference resolution, ellipsis, modifier attachment, conjunction and disjunction, and word sense disambigua tion. In NLP, these problems are within the sub field of Natural Language Understanding (NLU). In a CA, the NLU is referred to as the unit where the system extracts meaning from the input text.


17 Natural Language Understanding Natural Language Understanding (NLU) comprises everything related to machine reading comprehension. In HCI applications, the NLU component is where the natural language input is transformed into a semantic representation, such as first order predicate calculus and lambda calculus, whic h allows the system to execute the required tasks. The NLU is where all the domain, discourse and world knowledge combines to determine the requested answer. Within NLU, a fundamental research topic is resolving ambiguity, that is, disambiguating multiple alternative meanings from a linguistic expression. Resolving ambiguity is usually divided into two specific problems, lexical disambiguation, a difference in meaning, and syntactic disambiguation, a difference in syntactic category as e.g. determining if a word is used as a noun or a verb [18] [19] Lexical or semantic disa mbiguation is prevalent in natural language and very much in English. In the WordNet 3.0 database [20] the average noun has 1.23 senses and the average verb 2.16 senses, where in WordNet sample texts the number of senses is roughly proportional to word frequency. Lexical disambiguation (LD) has always been a broad and challenging problem within NLP, cogni tive science and the psycholinguistics fields [18] LD tasks include the well studied task of POS tagging, choosing the proper tag within context, and the more chal lenging WSD, selecting the correct sense for a word within context. The term WSD denotes only when the actual sense or meaning of a word is ambiguous, e.g. bass as in fish versus bass as in music. The WSD problem is viewed as a classification task, with re search focused on methods known as lexical sample, using pre selected target words with defined senses, and all words, where the entire text is disambiguated using a full lexicon. Most proposals now


18 use unsupervised or semi supervised learning techniques, after years of mostly supervised training methods using lexical samples [21] Other LD tasks include: Reference resolution determine the entity to which a lingui stic expression is referring. Research for this task is divided between two subtopics: [18] i) Anaphora resolution finding the entity to which an expression refers. The expression is usually a pronoun and thus the subtask is referred to as pronominal anaphora resolution. Anaphora also includes use of possessive determiners and noun phrases. ii) Coreference resolution finding referring expressions in a text that refer t o the same entity. This is the harder problem of the two. Ellipsis identify from the conversation context an omitted verbal phrase from a sentence [22] Modifier attachment identify the entity to which each modifier has to be attached [23] Conjunction and disjunction when the word and denotes disjunction rather than conjunction [23] Named entity recognition detect and classify proper names. Relation detection and classification detect and classify semantic relations among entities. Resolving the lexical ambiguity difficulties will contribute to NLP research areas such as IR, MT, QA, sentiment analysis and text classification. Syntactic disambiguation is the process of determining how to cor rectly parse a sentence from multiple alternatives, also called the process of resolving structural ambiguity. Structural ambiguity is less common than semantic ambiguity and the two are orthogonal, as a word can have either related or unrelated meanings w hile in


19 of the art syntactic parsing systems using dynamic programming are effective at dealing with this task [18] Further complications with ambiguity in NLU may occur from user input of: Non standard language where the formal rules of capita lization, spelling and usage of characters are not followed. This predominantly occurs in applications where the users are allowed to write the input in natural language, such as text messages, blogs, discussion groups and social media websites. Idioms expressions consisting of sequences of words that have a combina torial literal meaning distinct from the figurative meaning of the word combination Neologisms, jargon and slang newly coined terms, words, or phrases that have not been accepted into formal language. B ound to popularity, sometimes these terms are eventually accepted into the formal language. Segmentations in English, segmentation issues may arise when a user inserts a hyphen between words to enhance meaning. They may also occur when a user invents abb reviations and uses entity names. In any case, most segmentation issues in English can be attributed to the previous mention of non standard language. These ambiguity difficulties occur often in HCI applications such as conversational agents, where the hu man is allowed to freely enter input into the system using natural language. The general concept is that people do not intentionally use ambiguity in every day language. Psycholinguistic research suggests that when people perceive ambiguous words, they mom entarily access their mental lexicons to rule out the unintended meanings [24] [25] Conversational Agents A conversational agent (CA) or dialogue system is a computer program designed to carry out an open and coherent conversation with a human [18] Designing a CA that publicly recognized NLP application after years of fictional AI characters made f amous through movies, novels and television. In addition to computer science and engineering,


20 designing a CA requires studies in linguistics and psychology to create a realistic human interaction experience. Present day computational hardware has allowed t his field to explode in a wide variety of commercial applications such as: 1. Personal assistants 2. Help desk agents and customer service representatives 3. Travel agents 4. Technical support 5. Automatic call routing 6. Automatic tutoring systems 7. Gaming characters 8. Virt ual friends and companions Many of these applications are enhanced with additional technology at the input and output of the system, such as speech processing and avatars, and are known as spoken dialogue systems and embodied conversational agents respect ively. Deployment environments of commercial applications include application programming interfaces (API), webpages, robots, in car systems and telephone based systems. Figure 1 1 shows the typical architecture of a basic dialogue system, which comprises of an NLU to receive the system input, a Dialogue Manager (DM) and a task manager to control the conversation, and a natural language generator to produce the output. The DM controls the architecture and structure of the dialogue, i.e., determines the outp ut of the system given the input. It also keeps record of the state of the system and serves as interface with the task manager. The DM, NLU and natural language language and kno wledge about the domain of the application. The task manager represents the link between the DM the domain knowledge base, and any other domain specific components of the dialogue system.


21 Figure 1 1 Basic architecture of a dialogue system The two well studied methods used in NLU systems are identified as pattern matching and statistical learning. Most practical applications utilize corpus based methods while most research developments are on machine learning techniques, tha nks in part to the availability of large corpora of data. Chapter 2 explains the motivation and requirements for choosing the method for the advising system. Dialogue management is a decision task, typically modeled in state space and implemented using kno wledge based or machine learning techniques. Two common advanced architectures for dialogue management are information state and plan based [18] [26] The term information state architecture covers a wide range of knowledge based and statistical methods that extract features from discourse context to determine the output. Pla n based methods use logic calculus to analyze within the knowledge base, input extracted features known as beliefs, desires and intentions. While distinct in principle, the two architectures require careful design for application constraints, maximizing fe ature extraction from the context of the input and many cycles of training and testing.


22 The NLG component hosts the process through which the output language is composed. Two well known basic methods to generate language are template based and generative based [18] For the first method, the simplest and thus most common, the dialogue designer preassembles sentences or prompts. The prompts usually contain parameters that the dialogue manager defines. The term generative based refers to a process where an output senten ce is constructed from words or even morphemes. Research in this area is focused on knowledge based and statistical approaches that use the Internet and other large corpora as sources [27] The evaluation of dialogue systems is an open research problem, with most proposals requiring carefully weighted evaluation metrics and taxonomies [18] [28] [30] Nonetheless, dialogue systems are ultimately evaluated on the satisfaction and efficiency costs for the users. For this reason, the Turing test is still a highly regarded evaluation for any CA [31] The Loeb ner Prize is the first and most renowned instantiation of a Turing Test [32] Though not a standard within the academic research field, the open competition usually takes place at well recognized universities and with participation by members of the academic and research field. The CAs that compete in the Loebner prize are commonly known as chatterbots and are full production systems viable for commercial applicatio ns. Two recent multiple year winners have released the main dialogue systems as the open source languages Artificial Intelligence Markup Language (AIML) [33] and ChatScript [34] Both systems utilize a template based approach for the NLU and NLG, and a knowledge based approach for the DM.


23 Released in 2001, AIML enjoys g reat success in the chatterbox community with thousands of members contributing to the knowledge base, releasing interpreters for multiple programming languages, and developing commercial releases and supplementary packages with advanced language features [35] [37] The advising system described previously, utilized AIML for its CA implementation [16] ChatScript is an open source scripting language for a rule based dialogue system. ChatScript was released in 2011 with significant features not available in AIML, e.g., a knowledge based information state architecture for the NLU and NLG, WordNet integration, a query based system for a dynamic database of tables and triples data structures, and analysis tools for user log files. Anoth er well known development platform for NLP applications is the Natural Language Toolkit (NLTK), an open source NLP platform written for the Python programming language [38] Released in 2002, NLTK features WordNet integration and a strong NLU component. Whereas NLTK is used mostly for educational and research purposes, a few recent publications have used NLTK when developing dialogue systems [39] [40] Question Answering S ystems When instead of a conversation, the system only responds to single queries independently, it is known as a QA system. Given strong commercial demand, QA systems ar e one of the oldest but also most active research topics in NLP. Because of the nature of many CA, these dialogue systems also contain a QA component. The two basic QA models are IR, where answers are obtained through a literal search, usually from the we b, and knowledge based, where answers are built after a successful analysis of the parsed input. In knowledge based systems, the information


24 for building the answers is collected from multiple documents. QA systems are usually designed to manage specific t ypes of questions and answers, e.g., factoid questions and complex narrative questions. Most QA commercial applications deal with factoid questions, while narrative questions are common in research systems [18] A system that uses a CA with a strong QA component must also manage elliptical questions [41] Therefore, the advis ing CA must manage all three types of questions, given that its main task is to serve as a QA application. Dissertation Topics The main objective of this work is to construct and deploy a real world expert based academic advising system that allows the stu dents to communicate freely in natural language. The system was designed to serve students of the Department of Electrical and Computer Engineering in the University of Florida (ECE UF). The system is accessible online for public use at any time. After com pletion, this system can serve as a basis for deployment in other academic programs. This system allows for training and testing of the algorithms developed in a well defined, domain specific, conversational question answering application, similar to othe r well known commercial systems such as SIRI with Apple, Watson with IBM and Wolfram Alpha. This work describes the production of a CA for academic advising, with the research focused on the NLU unit to develop a linguistically oriented algorithm suitable for this application, where there are no available corpora for machine training. The CA designed utilizes state of the art NLP algorithms and methods to enhance the system for utmost effectiveness and usability as evaluated by its users. The robustness of the algorithm allows for scaling up the input range of the system as the amount and


25 diversity of user interaction grows. Therefore, this work also introduces a methodology for allowing the system to scale up as new data and queries are obtained. This disse rtation is organized into the following chapters. Chapter 2 describes the task of academic advising, the user base for the system, introduces the conversational system for academic advising and specifies the components of the advising system. Chapter 3 des cribes the dialogue manager of the system, including the NLU, NLG and task managing components. Chapter 4 describes the academic planning system and the components for communicating information between the students and the advisors. Chapter 5 describes the field tests and the analysis of the results. Finally, Chapter 6 contains the conclusion, the contributions to the field and the plan of work for the future of the system. The appendix section includes a list of topics this application contains, the screen shots of the web interface, performance evaluations by the ECE UF advising personnel and examples of user log files from the experiments.


26 CHAPTER 2 A CONVERSATIONAL SYSTEM FOR ACADEMIC ADVISING This chapter begins with a brief narrative of academic advising tasks, followed by a description of the advising situation in the academic department where the system is installed, in view of the motivation for the academic advising system developed T he second section detail s the objectives and speci fications of the framework of this system. The third section introduces and describe s the main components of this system. Academic Advising The R ole of the A cademic A dvisor As described in Chapter 1, advising is an integral part of a successful academic d epartment of an educational institution. Academic advising seeks to encourage student success through teaching and learning that lead s to productive career and life planning skills [1] The tasks advisors engage in are traditionally identified as prescriptive and developmental advising. As described by Appleby, the prescriptive advisor dispenses expert advice, whereas the developmental advisor engages in a mutual lea rning making and evaluations skills [42] A study by Brown shows how the role of the advisor transitions from prescriptive to developmental throughout the academic years of the student [43] The study also describes how adv isors should be proactive with entry level students. Other studies support the idea of educative advising, where advisors are the teachers of the philosophy of the curriculum and the principles of how students learn [44] As teachers, advisors need tools to encourage students to seek the information they need, so they can consult their findings and goals with them. An increase in


27 developmental and educative tasks w ill require an increase in resources for advising. The system developed for this dissertation provide s advisors with an effective tool to streamline many prescriptive and educative tasks, thus allowing further developmental and educative tasks during their limited face to face time. With academic offices repeatedly struggling for resources, this automatic advising system is a favorable solution for the students, the advisors and the academic institutions [1] Academic A dvising in ECE UF The advising system was developed for s tudent s of the Department of Electrical and Computer Engineering of the University of Florida (ECE UF). Undergraduate students of ECE UF visit th eir academic advisor either by request of the advisor or to freely discuss a topic relevant to their studies. The first case is the most common, as these meetings enrollment in the stu dy pro gram. During these meetings most topics involve queries with accessible answers, i.e., prescriptive advising. These answers are usually obtained directly from facts or through logic analysis, and thus may be solved algorithmically through search algorithm s and logic calculus respectively. The other reason why students contact advisors is for the personal and professional issues that advisors must answer within their best judgment, i.e., developmental advising. The challenge for an automatic system lies in that any possible answer is greatly influenced by the personal circumstances of the student and very sensible to the context of the given situation. Since many of these questions arguably lack a definitive answer, it is reasonable to state that an advising system either will be biased when referring to such queries or must state t hat an answer is not practical for the system. The system either must anticipate every possible scenario or must


28 incorporate a decision based generative method to create the best possible judgment for the scenario. Clearly, developmental tasks are better l eft for the human advisors. Nevertheless while the main objective for this system does not include developmental advising, the system does manage a certain amount of the related topics. A dvisors and students of ECE UF do not have a centralized database of academic advising information This challenge affects training for the advisors, alternative resources of knowledge for the students and bookkeeping of all rules and guidelines. This is also a shortcoming in the effort to create a comprehensive knowledge base (KB) for the system; therefore, the knowledge for the system was obtained through meetings with the academic advisors and surveying multiple ECE UF and UF documents. Results from this work will provide ECE UF with a digital and easy to search through, advising KB The data from the system will help co nstruct an advising data corpus, currently unavailable in the research field. Following Chapter 1 and the conditions just described, the design choice for the advising system was a template based NLG and a knowledge based QA model Fundamental Requirements for the Advising System From a teaching perspective, some advising tasks of this system are similar to features available in conversational agent (CA) systems for e learning environments, such as those ex amined by Kerley et al. Gandhe et al and Mori et al [45] [47] The evolution of these CA applications shows an example of initially using pattern matching techniques and then statistical methods, as sufficient data becomes available. The success of many of these systems in education suggests the viability of the advising system here proposed. These systems also encourage additional non NLP features for the a dvising system such as, an escalation mechanism to forward selected


29 conversations to advisors and adding a limited amount of non topic conversations, or small talk, to inspire user rapport with the CA. In contrast e learning systems have fundamental diffe rences from the advising system here proposed, for instance e learning systems control the conversation with the users to allow for effective teaching the co nversatio n at all times. Kerley et al. presents a series of questions that developers of e learning systems must face [45] To resolve the fundamental requirements for the development of the conversational advising system, these questions are grouped into the following nine major prerequisites What approach to NLU must the system adopt? The enclosed domain of the main topics for this application and the lack of any availabl e corpus favor an approach to pattern matching techniques. How does the developer select the software to implement the NLU techniques? A solution for rapid deployment involves selecting the most advanced available software within the free and open source alternatives and building upon it. A l l components of the system are also built using open source programming software. Who controls the direction of the interaction, user or conversational agent? This design is meant to allow students to address their advi sing needs, therefore all the conversations are initiated and controlled by the users. What freedoms do users have through the interface? User input determines the states of the CA. The system must recognize the situations that call for state transitions. For example, the user will access and manage the course planning system upon request, at any moment of a conversation. To make the most out of the


30 with additional software an d communication platforms. Does the system structure the path along which the user must proceed? For the advising application, traditional dialogue managing methods such as dialogue trees do not work because the intention is not to engage and expand a conv ersation, the information solicited and stay or leave the topic only upon user request. This requires very good reference resolution to easily determine when a ling uistic expression is related to the same topic or starting anew. The refore, the system must allow the user to explore the conversation along any dimension they choose, including some off topic dialog, usage of follow up questions and contextual references. For off topic dialog, the system must quickly encourage returning to relevant tasks. As a service agent, the system must understand and help the users as efficiently and effectively as possible, that is, engage in the minimum possible dialog to understand the users and offer a straight to the point easy to understand response. The system output must be succinct, while preventing unnecessary follow up questions. The system shall not initiate off topic dialog or return answers that could potentially misinfor m the user. What will users say? Since there is no database of academic advising, the data was obtained through researching UF documents, advising documents from the web and through consultation s with ECE UF advisors. A llow ing students to communicate natu rally is akin to how users behave in the o nline and mobile communication media that are pervasive in our society This behavior includes informal conversational natural language intricacies such as implicitly conveyed information, semantics, pragmatics,


31 am biguity contextual references and neologisms which by social expectations users presume NLP applications will understand. T he system must also manage the input data without formal writing features such as punctuation and capitalization. It must also con tain a spell checking method for the unique vocabulary and non stan dard language of the user field, as well as the technical terms unique to each academic program, e.g. COMP1234 (course code), C+ (grade), C++ (either an unofficial course name or a programm ing language) and gened (general education course) Many of these unique terminologies must be collected from the user field, thus data collection from the users is crucial for the success of this system Many CA developers initiate the data collection pro cess using the Wizard of Oz technique, where a human, or wizard, simulates the behavior of the final system without [18 ] For this work, time and user availability hinder a chance for such a study. To maximize input data and productivity, the best option is developing and deploying a preliminary version of the system and gathering the data through the promotion of this beta system. How does the system handle unexpected responses elegantly? For unrecognized input, the syste m can suggest possible recognitions based on partial matches, request rephr asing or explicitly state that a response is not available. Another approach involves asking users to define the unrecognizable concepts and through semantic relations, map these definitions to the active KB. These alternatives are analyzed further in Chap ter 5, where we examine a methodology to scale up the CA. How does the system handle synonymous expressions? In traditional pattern matching techniques, developers address this challenge by predefining a


32 mapping of synonymous expressions. For everyday Engl ish, the lexical database WordNet is a well known resource [20] The advising CA system employs a software platform that includes WordNet along with many additiona l predefined concepts, which allows easily deal ing with this concern Conversely the definitions for the technical terminologies the users will use must be hand code d. How much testing will be required? C ommercial CA systems in e learning often undergo ov er two thousand conversations before reaching operational state [45] With an additional data collection phase, the advising system should require a greater amount of conversations to obtain a comparable performance. Three additional fundamental requirements f or this system include, first, that the system be available to the students at all times through a comfortable instrument such as a webpage or mobile application. Secondly, t he dialog ue system must systematically manage the complexities of multiple course sequences within an academic program and the multiple amendments that often occur in the academic field. Third, the system must obtain and store student academic data without comprom ising their private information. Specifically, in UF, student academic rec ords are protected by the Family Educ ational Rights and Privacy Act FERPA 20 U.S.C. § 1232g; 34 CFR Part 99, a U.S. f ederal law and by the Florid a Statute Section 1002.22 2007, a state law. Albert The Intelligent Academic Advising Conversational Syste m The system developed provides students with an academic advising service that reflects a human interaction experience through an online text application. The system does not require student training or additional human resources from the academic departm ents. This system enhance s the academic a dvising experience by offering students a service that is available at any time. Albert includes multiple advising


33 services accessible from any device with web access. Albert respects the privacy of its users and en courages the students to become independent and take responsibility for making decisions. Students who effectively utilize Albert will have the awareness to focus their meetings with the advisors in the subjects of developmental and educative advising. Advising I nformation Albert include s knowledge about the academic programs and policies, answers to a wide range of frequently asked questions (FAQ) in academic advising, it offers recommendations for the development of a course plan that leads to degree completion and referrals to other academic services. Appendix A contains the list of academic topics included with Albert. Students can search all this available information using natural language for a comparable experience to communicating with their adv isors. Albert contains the course scheduling information for all ECE UF courses, as information and a cron job updates the KB daily. The information is stored indefini tely and always available for access when queried by date. This process assures the information is always up to date without dependency of human maintenance. For energy consumption reasons, updates for the information of each term are suspended two weeks a fter each term commences. The easy access to this information benefits all students from ECE UF and other departments who want to enroll in courses from the department, faculty members, advisors and administrative personnel. All the other information in Al bert is hand scripted in the KB of the system. A s a courteous advisor employed by an academic department, the system output reflects an advisor who is polite, friendly maintain s a positive affective state with its


34 user s and demonstrate s a personality tra it intended to sympathize with the students of ECE UF. For this development of Albert, knowledge about degree programs and advising FAQ is mostly relevant to undergraduate students. While these students are the targets of all the advising challenges and be nefits discussed previously, future developments could include information about graduate degrees. Albert does contain all ECE UF undergraduate and graduate course information including the course schedules, pre requisites, descriptions, etc. The target us ers are students from the ECE UF two undergraduate degrees, Bachelor of Science in Electrical Engineering (BSEE) and Bachelor of Science in Computer Engineering Hardware emphasis (BSCEE). The target users also include the undergraduates who are enrolled in the combined Bachelor and Master of Science program, where students get an early start on their System C omponents The main script of Albert is written in the Python programming language, version 2.7.5. This script controls all the func tions of Albert and communication with each module including the web interface, the gateway interface for web communications, the user login routine the dialogue manager, the expert system for academic planning and the database. Figure 2 1 shows a model o f the Albert system. The dialogue manager refers to the complete dialogue system described in Chapter 1, namely, the NLU, NLG, DM and task manager.


35 Figure 2 1. A model of the main components of Albert. For students to effectively adopt the system, continual availability of Albert is required. This feature is accomplished by making Albert accessible through the web, specifically as a web browser graphical user interface. This website is hosted in a deskto p computer at ECE UF facilities and accessible via the Internet address The computer has an Intel Pentium 4 processor, 1.8 GB of RAM and is running the operating system (OS) Red Hat E nterprise Linux 6.4 The website was built with scripts written in Python, PHP, JavaScript, HTML and CSS programming languages, in addition to Unix scripts to manage the daily tasks. base d clients, is done through socket technology, which is widely used in web based software. The website was designed for simplicity and speed, with a load time of approximately one second on contemporary versions of the web browsers Google Chrome, Mozilla Fi refox and Apple Safari Users who access Albert using the Google Chrome web browser can also send messages using speech recognition software. The website includes an open source speech recognition API that allows users with the required hardware and Chrom e to input data using speech [48] User Input Output Web Interface Dialog Manager Academic Planner Database


36 Figure 2 2 shows a screenshot of Albert in its initial state. The mid upper left area shows the forms for logging into the system. Below these forms, the main window for communicating with Albert shows the instructions for connecting with the system. On the right of the main window, an independent web frame shows examples of FAQs the system contains. Screenshots showing additi onal states of the website are available in Appendix B, including a screenshot of the webpage using Chrome with the speech API. Figure 2 2. A screenshot of the website for Albert in its initial state. To protect the information of the students the sy stem requires users to register an account using an anonymous username and password combination This data is stored encrypted in the Albert web server using the standard secured hash algorithm


37 SHA 512 64 bits. Usernames and passwords are case sensitive; furthermore the system does not allow repeated usernames entered in different letter case. These security features allow students to safely share their usernames with their advisors, yet keep their information secured through the password. Anonymous acco unts also encourage students to communicate freely without repercussions, which is the best way to obtain sincere feedback about the system and the academic services. The account in Albert also allows storage of user information for recurrent advising sess ions. Advising Dialogue Manager The heart of Albert is the NLP system that drives all the input, output and states of the system, i.e., the a dv ising Dialogue Manager (DM ). Following the objectives and constraints listed above and the options reviewed in C hapter 1, the best option to develop the DM is to build it around the ChatScript (CS) scripting language. This work uses CS version 2.0. In the DM, CS is the main component and it includes most of the NLU and NLG structures The design of the DM allows a s traightforward method for redrafting and distributing the system for other academic programs, by using variables in the input patterns for the proper nouns and the contents of the well defined knowledge base. Specifically the advising FAQ corpus for the i nput patterns is generic enough such that deploying in other academic departments requires editing the KB, but not reworking the existing input templates. Since CS runs as a client of the DM, the DM controls communication of CS with additional task applica tions integrate d with Albert CS can also communicate directly with the OS through terminal commands at any moment. CS can read files from external applications via its internal database, subject to these applications creating the files in the required for mat. With these features, the DM integrates functions written in


38 CS, Python and PHP. For example, to manage the unique technical terms and proper nouns, an external spell checker in Python was constructed u sing methods surveyed from the literature [18] [38] [49] When CS finds a candidate for the spell checker, e.g. a noun term similar to a course name, it sends the term to the Python spell checker, which returns the possible corrections to CS to determine the system response. Chapter 3 covers the detail s abo ut the DM and processing the natural language input and output using th e CS system. Regarding computational performance, as CS is a production grade chat system, messages through the website took approximately one second when tested from computers connecte d through a local area network. Academic Planning S ystem Albert includes an expert system designed to offer students guidance and recommendations when preparing their course plans for each term. This expert system was developed in Python. This system allow s students to enter their academic record and receive recommendations on how to develop their course plan, based on the courses completed, the prerequisites approved, the courses offered each semester and related department and university rules. With this academic planning system, students can create their course plan for the next semester and send it electronically to their advisor for review. With course planning as the main feature, we refer to this system as the Course Plan Analyzer (CoPA). CoPA runs i ndependently from the DM allowing users to work on both systems simultaneously, analogous controls communication between both systems and automatically sends the data to an allowing advisors access to student submissions. Development of this procedure required building an independent website for advisors to access this


39 database. This process is part of the objective of allowing advisors to integrate Albert into their daily ad vising practice Chapter 4 covers the detail s about the academic planning system and CoPA. Regarding computational performance, the webpages for the CoPA system are also within the approximately one second delay between events the rest of Albert offers, w hen tested from computers connected through a local area network. System E valuation T o evaluate Albert, following the academic advising guidelines of the National Academic Advising Association [6] and the Council for the Advancement of Standards in Higher Education [50] assessment must include direct and indirect evaluation use qualitative and quantitative methodologies and d ata collected must include responses from students and other constituencies For quantitative analysis, Albert includes a numerical survey for the students built in the CoPA process and through CS, keeps a record of log files from all users. The log files contain the information of the input output messages, login events and message ti mestamps. Within the CS dialogue scripts, an original development allows to assess the number of questions that did not have an available answer, the questions responded to with nonspecific answers and other statistics that provide a measure for accuracy a nd task completion success. Chapter 3 covers additional details about the log files and their usage, and Chapter 5 shows the results obtained from these log files. For qualitative analysis, the measures include collecting feedback from the students and hav ing consultations with the ECE UF academic advisors for assessment reports. The data from the log files is also used to compose the academic advising FAQ corpus currently under development. As the data is processed, the system continues to


40 grow with additi onal input patterns. This data also offers assessment to advisors with information and statistics about the user data, to learn the areas of advising that most concern students. As reviewed by Soria in [3] adviso rs need detailed assessments of d retention, and student satisfaction with their academic institution.


41 CHAPTER 3 THE ADVISING DIALOGUE MANAGER Chapter 3 describes the advising dialogue manager (DM), namely, the Natural Language Understanding (NLU) system, the Natural Language Generation (NLG) system and the task manager. Natural Language Processing The NLU component of the DM produces a semantic representation of the natural language input that is appropriate for the dialogue tasks. The NLG consists of templates containing text, pointers, variables and other control functions. The main objective in the DM design was to r espond to questions from academic advising topics. The NLU contains several templates and measures for responding to secondary topics such as information about the Albert system, control of the system, the persona of the Albert advisor, conversation al syst ems, web browsers, mobile devices, time, dates, movies and other small talk. The structures in the NLU for these secondary topics use the same principles as the main topic, and the design calls for off topic dialog to promptly converge to on topic dialog. Accordingly, the focus of this section refers to development of the NLU for the advising topic only. The advising knowledge base (KB) in Albert contains a finite amount of answers that the system should return in a conversation. In contrast the system in put domain comprises infinitely possible styles in which users can request each answer. To create the KB, the output information was obtained through university documents and interviews with the academic advisors. T he input templates were derive d from a St andard English subset of the input domain. As designed, after deploying the system, the KB was continuously updated through analyzing the user log files The KB has the


42 information organized as a list of triples contain ing a question, its answer and the to pic to which the answer was classified Natural Language Generation As a template based system, the NLG mostly responds with strings of answers taken straight from the KB. The answers are constructed using as tools logic functions, random generators, contr ol structures and queries Queries include course scheduling saved with the course plan analyzer (CoPA). The system also contains a trivial amount grammar rules, mostly for handling unrecognized input and extending off topic dialog. For example, when responding to unrecognized questions, the NLG may either change the verb tense or rearrange the POS to let the user know the answer to that question is not available. Natural Language Understanding The system input is an informal conversational natural language with intricacies such as implicitly conveyed information, semantics, pragmatics, ambiguity and contextual references Therefore, the system requires an effectiv e reference resolution process to determine the intention of the user from the linguistic expression and its context T he NLU must classify every significant lexical unit (LU) and their relationships within the context of the conversation. A practical solu tion is to build an effective keyword extraction (KE) mechanism with context based ambiguity management Most current KE systems use similarity measure s with features such as term frequency, inverse document frequency, relative position of the first occurr ence and caps first r atio [51] [54] To be effective, these measures need either sufficient training data or data with formal grammar, none of which are available for Albert. Moreover, even with


43 sufficient data, most input quer ies for this system are too short for most f requency analysis techniques Nevertheless, the work of Hulth and Bellotti et al. shows that parts of speech ( POS ) tags, noun phrase chunks and lexical relations are significant features for KE algorithms, indep endent of the usual statistical term selection methods [45] [51] To s uccessfully classify LUs, the NLU must resolve the ambiguity problems identified in Chapter 1 namely, word disambiguation, reference resolution, modifier attachment, named entity recognition, relation detection and classification, the use of non standard E nglish, and elliptical questions NLP applications usually address ambiguity by adding probabilities to the grammar; however, this requires more data than is available. In addition, data from the initial experiments showed that students expect the system t o understand their messages written with informal grammar, incorrect sentence structure, e.g. fragments similar to computer commands, and without the formal writing features of punctuation and capitalization. The overall initial reaction of users was to re quest information akin to using a typical online search engine, i.e., entering isolated keywords. Appendix C shows user data from the latest log files. NLP A lgorithm The NLU and NLG components are built with the ChatScript (CS) scripting language CS resembles a knowledge based information state architecture features WordNet integration, a query based system with a dynamic database of tables and triples data structures, and a client server architecture that communicates using sockets. CS was creat ed on the basis that understanding natural language requires pattern matching and a method to infer from facts. Similar to semantic grammars, a script in CS comprises topics, templates concepts, facts and logic functions.


44 A topic is a collection of templ ates for input output matching. Each template contains the keywords, logic functions and control structures for input matching and the corresponding output message A concept is defined as any collection of LU s such as a sentence, phrase, word or parts of words. CS includes predefined concepts with WordNet synsets which are sets of cognitive synonyms Facts in CS are the elements of the query based database, with field values written as triples contain ing LUs or other facts. Facts are useful to build reco rds, arbitrary graphs, data arrays, tables and similar database components The input template design involved identifying the keywords, POS tags, noun phrase chunks and lexical relations of each input statement and selecting the most significant features, to define the keywords of the template and any respective topic keywords. For example, the topic course_schedule_ information includes as keywords, the list of all course names the list of all professor names professor and the teach The keywords for each input template, the words that convey the message, are extracted using the KE features referred to previously. For example, in W ho will teach C++ during the next semester the keywords are who teach C++ and next semester where who and teach must appear in that order. Using CS who teach word course name and term phrase where Teach word is any word that refers to a course being taught, e.g. teach, teaches, lectures a nd give. Course name is a name from the list of all courses. Term phrase is a word or phrase that refers to an academic term, e.g. next semester, next term, Fall 2012 and summer.


45 Figure 3 1 shows the algorithm for the example above. Clearly, this example i s not the only way to ask who will teach a specified course. For example, the user could ask W ho is the professor of C++ or if the request is within the context of the previous Who is teaching it The first case requires building a new template, which can be mapped to the output to the template defined above. The second case requires adding a new template that includes a method to determine the context of the previous input The system identifies th e context using features such as the current and previous input keywords, the keywords for the topics matched, the tense of any verb and the state of the variables representing potentially missing keywords. Algorithm: Respond to user request Who teaches course Input string S : W ho teaches C++ next semester? Desired output : C++ is taught next semester by Name of whom teaches C++ the next semester or C++ is not offered next semester If S contains a keyword of the topic course_schedule_information If S matches with a Who Teaches pattern If S contains a Term Phrase keyword Calculate the Term value Else Use currently stored Term value If the course C++ exists in the schedule of the Term Find the corresponding data element Instructor Return C++ is taught by Instructor in Term Else Return C++ is not offered in Term Figure 3 1. Example of a procedure to match an input requesting who teaches a specified course during the next semester. R ecognizing context is also a requirement for resolving lexical ambiguity. Resolving the ambiguity problems required a combination of solutions. For reference resolution and elliptical questions, in addition to the method described abov e for


46 recognizing context, the system uses the CS feature of rejoinders. Rejoinders are input templates, which follow parent templates that elicit some expected user response. Rejoinders also allowed some input templates at the end of topics to assume cert ain keywords were implied. The entity recognition problem is managed by defining case insensitive CS concepts with pre classified POS tags. In addition, the system has concepts defined for all the technical terms, neologisms, slang and significant LUs not available in the CS or WordNet dictionaries. The system processes modifier attachment, semantic relations and entries in nonstandard English, on a per template basis using logic functions and control structures. The system also has measures to deal with no nlinguistic ambiguity that the users inadvertently convey. In some cases, the best solution was to return all the probable Who is teaching EEL4995? collection of courses with individual instructors. In other cases, the best solution was to request more information from the user. N onetheless in most cases it is still preferable to convey some relevant information, and allow the user to either read it or try a different option. This user preference was palpable during the field tests; see Chapter 5. For instance Who teaches Circuits in the next term? the meaning of next term depends on the current date and during the spring semester, it could refer to either the summer or the fall term. For this case, the system will decide on a value for the term variable, determine the response and return the response including a quick method to obtain the value for the other term value. As evident in the algorithm of figure 3 1, once the user states an academic term, the system stores this value in case the user does not provide it in a future request.


47 When writing the input templates, the key tradeoff is between over fitting and not generalizing well, thus increasing missed inputs, o r under fitting and causing false positives. In this work, the precision of the template is inversely proportional to the rate of occurrence of the template. That is, the responses that users most seek have a lower accuracy and higher coverage. In contrast the precision of the template is proportional to the intricacy of the response, i.e., very specific answers have templates with higher accuracy. Within each topic, the templates are organized from complex to general, similar to an inverted decision tree. This approach minimizes false positives for questions with many specifics, a design constraint to prevent misleading the students, while causing most false positives errors in responses with broad information. When an input statement does not match with any template, the system will respond with an estimated match or request a new entry. The last t hree topics of the NLU script respond to these statements. The antepenultimate topic contains several on topic templates with either, partial answers because the full statement was not recognized or with suggestions to help the user obtain the anticipated response. This process is alike typical communications where if the receiver is doubtful of the input it notifies the sender of the issue or requests a confirmation of the received message. This topic also contains several templates for pronoun resolution, where the system will either redirect the user to a specific topic for the response or alert the us er that the statement was not recognized as written, to encourage the user to not repeat such statements. In general, this topic is designed to respond to any input statement that contains a significant keyword, such as a topic keyword, and return the most likely match based on a score of the significant


48 keywords. This topic also contains templates that call external spell checking procedures, which details are covered in the next section. The penultimate topic contains a small collection of less significan t keywords, such as verbs and common nouns, and some off topic templates. Most of the templates in this topic are produced after analyzing the user log files and finding various users who did not understand how to use the system. For example, many users wo uld expect the system to have knowledge about popular culture or to successfully respond to one word statements such as course enroll and hold Ideally, this topic should not respond to users, however, it proved useful in the preliminary tests. The last t opic responds to statements where the system cannot immediately help the user, either because it does not know to which topic to redirect the statement or because nothing was recognized. The templates are mostly for catching single keywords and responding with answers that explain how to obtain information related to the keywords. Some templates contain POS tags to generate responses to wh questions with a proper grammar. The topic ends with a collection of responses, which acknowledge that no part of the i nput was recognized. For evaluation purposes, the system classifies responses generated from these last two topics as indeterminate While these features enhance the template creating process, all template based systems are limited by the amount of pattern s for input matching. The next section describes a method to involve the users and experts in scaling up the advising KB. Chapter 5 includes the numerical specifications of the DM scripts and the results of the experiments.


49 Task Manager The task manager re presents all the functions Albert executes to complement the dialogue system. These functions include user account management, database management, the course plan analyzer, input validation, spell checking, statistical data collection, automatic updates, a scaling up routine and all communication for CS. The task manager is built with Python. This section closes the chapter with the description of the major functions, whereas the next chapter covers the course plan analyzer. Spell C hecking Albert is a text based system, thus spelling and typographical errors are very common, yet users expect the system to handle them effectively. CS includes a list of common spelling errors and the correct spelling that the system substitutes automatically o n inputs. This list, plus a corresponding list for the technical and local terms common to the Albert users, was continuously updated throughout the experimental phases. CS also has a built in spell checking method, although during the period of this work it was still a work in progress. The CS spell checking methodology was improved for this work by adding the capability to handle the required technical terms and unique expressions The CS spell checking procedure was also enhanced such to allow the system to return the results of the spell checking process, to alert the users that a correction was made and encourage them to verify the new input. While spell checking in general corrects typographical errors, sometimes the correction can change th e meaning of the question and cause the system to return an unexpected response. Since users are mostly expecting the answer to their problem, it is very important to let the user know if the response is the answer to the question they asked. This procedur e is akin to the


50 objective described for the non linguistic ambiguity resolution, where when in doubt, always present the user with a solution and allow them to choose a different option if the solution was not correct. As mentioned in Chapter 2, to manage the unique technical terms and proper nouns, a supplemental spell checker was constructed in Python u sing methods surveyed from the literature [18] [38] [49] including the measures: Deletion item contains characters in excess T ra nsposition item contains characters in misplaced locations Alt eration item contains incorrect characters Insertion item is missing characters As described in the previous section, the antepenultimate topic in CS contains the templates that call the external spell checking procedure. These templates are very similar to the counterpart templates meant to respond to the input, except for a placeholder where a missing term is expected, e.g. a noun term similar to a course name. The term in the location of the placeholder is evaluated to determine its lexical distance to the missing term. If the lexical distance is within a predetermined value, this candidate term is sent to the Python sp ell checker. This spell checking routine uses a similar lexical distance measure to determine the likelihood of the possible corrections. If a correction candidate has a significant likelihood versus the other candidates, this value is replaced in the ori ginal input statement and this new statement is sent to CS for evaluation. If there are candidates, but none of them is relatively significant, the system returns all the candidates to the user and prompts the user to choose a selection or rephrase the sta tement. If no candidates are found, the system terminates execution and responds to the user that statement was not recognized.


51 Figure 3 2 summarizes this algorithm with an example similar to that in figure 3 1, where the incorrect entry is the name of th e course from which the user is requesting Who teaches EEL3101 however that course does not exist. The spell EEL3101 is not an EE course I recognize. If I a ssume you meant EEL3105, then my answer is: For Spring 2014 we have EEL 3105 by Latchman, Haniph A Algorithm: Respond to a user request with the spelling error Who teaches a course Input string S : W ho teaches EEL3101? Desired output : EEL3101 is taught by Name of lecturer. As EEL3101 does not exist, the desired output is to infer the intention of the user and respond with the expected response. If S matches with the Who Teaches pattern with a missing term Determine if S contains a term X si milar to any missing term candidate. If X is within a predefined distance Send X to the Python spell checking process Else End execution of this template. CS continues iteration input matching. Algorithm: Python spell checking process Input string X : EEL3101, an incorrect value for a course name Desired output : Most likely course name Find all terms Y similar to X, within a lexical distance of 2. Assign a score K to each term. If a value Y K is statistically significant versus all other Y K Return Y K to the task manager, to re form the input and send to CS. Else if any value Y K exists Send all Y K to the task manager, to return these options to the user. Else End execution. The task manager responds to the user that the input statement was not recognized. Figure 3 2. Example of a procedure for an input requesting who teaches a specified course, where the name of the course is incorrect


52 Automatic U pdates Albert has two main procedures for automatic updates; one procedure updates the schedule of courses database, the other updates the CS scripts. Both procedures are written using Python and Linux OS scripts. The course schedule information is available to t The system includes the Python library Beautiful Soup version 4.0, to extract the information from the webpage [54] The script then parses the data and saves it in a format that is readable by CS. A Linux cron job information is updated. A separate cron job runs a daily procedure of shell scripts that updates all the CS content. This scheduled task has multiple purposes. First, to copy and store all the data on a remote drive for backup purposes. The second purpose is to update the CS knowledge base with the course schedule information. The third purpose is for programmer ease, as this procedure allows the developer to work upgrades to the system at any time and know that the changes will be available on the next day. As detailed below, the system contains methods for automatic scaling up, which does not involve programmer involvement. Therefore, this procedure will also automatically update the system with the changes made by the user community and thus achieve the automatic scaling up. Scaling Up the N atural L anguage S cripts As discussed throughout this work, the productivity of NLU sy stems depends on the amount of data available for the design. Ideally, the NLP system will include methods to allow the system to scale up, with minimal involvement from the developers.


53 For this reason the design of Albert has measures in place to allow s uch improvements to the system. This work includes the design of one scaling up method, specifically a procedure to allow users to suggest unofficial names for courses. Collecting the academic advising data includes gathering the official names for all t he courses and a list of nicknames, abbreviations and acronyms that the ECE UF community commonly uses for these courses. Reasonably it is not possible to obtain every possible moniker for each course; therefore, a viable solution is to have the users sub mit new course names to the system. Users have tw o methods to submit course names to Albert. The first method is through a direct request, i.e., they implicitly state that they want to submit a course name. The second method occurs when the system recogniz es a request for course information, e.g. who teaches the course, when is the course or where is the course taught, but the name of the course is not recognized and no spelling correction was obtained. Figure 3 3 shows the algorithm for submitting a course name. Algorithm: Allow users to submit course names Input : System recognizes user intent to submit a course name OR course information was requested and the system did not recognize the name Desired output : Acknowledgment of successful name submission System requests the official title of the course or the course code If course is recognized Request from user the new course name. Send course code and new name to the ECE UF advisor for approval. Else Inform user that name was not recognized and to try again if preferred. Figure 3 3. Algorithm for course name submission


54 The algorithm obtains from the user the official name of the implied course and the recommended course moniker. This 2 tuple is automatically sent to the ECE UF advisor database, which the advisors can access online through an independent webpage developed in this work for the task. The second phase of the scaling up procedure involves the ECE UF ad visors using the online system to accept or reject the user proposed course name. If the name is rejected, the process ends. If the name is accepted, the 2 tuple is added to a table in CS, which was designed in this work for this purpose. With the automati c update described in the previous section, this data pair is available for users by the next calendar day. This section of the advisor online system was not published due to an interruption in the BSEE advising service and low user turnout. Statistical D a ta C ollection As explained previously, the DM classifies input statements according to the information returned to the user. Hence the system produces a computation of the indeterminate responses. However, the system does not provide an automatic estimat e of false positives outcomes that is, input statements incorrectly matched to a determinate response, or false negatives outcomes, i.e., input that should have matched. Therefore, the analysis of the false positives and false negatives outcomes is done through an estimation of the data. The results from the estimation allow computing the standard statistical evaluation metrics from i nformation retrieval, POS tagging, parsing systems, named entity recognition systems and WSD systems, namely, field such as precision recall, accuracy and the F 1 measure [18] [56]


55 Additional scripts developed for the system evaluation include procedures to count the number of logins p er user, the number of input statements per user and how many suggestions each user made through the scaling up process. To submit the results from the academic planning process, users must answer a questionnaire of three Likert items, as shown in figure 3 4. The figure also shows an optional write in text area where students can leave feedback. The results of the questionnaire and comments section are not sent to the ECE UF advisors. Figure 3 4. Questionnaire required to submit the academic planning pr ocess


56 CHAPTER 4 THE ACADEMIC PLANNING SYSTEM The fundamental task of academic advisors is helping students complete the requisites for graduation. For the typical student, the academic advisor is someone who helps them choose which courses they should take and provides solutions for any issue concerning academic advancement. Students expect personal attention with precise answers for their unique questions. The theme throughout this work is that Albert provides widespread solutions for FAQs, to allow students and advisors to focus their limited face to face time on personal discussions, developmental and educative advising. Notwithstanding, an account with Albert allows students to manage data stored during their conversat ions, as detailed in Chapter 3, and from their academic records. Using this data, Albert offers individualized recommendations on course enrollment and a service for students to send to their advisor, their choice of courses for enrollment. This chapter de scribes the academic planning system. Integrated Student Information System (ISIS), with a degree auditing component that is available for all of its students. ISIS allows students to study their a cademic record, complete course enrolment and to process official UF documentation. Experiential data from the ECE UF advisors indicate presentation and student reluctance to dealing with it. Therefore, most students repeatedly show up for their advising appointments unprepared to discuss their resolve this issue.


57 The Course Plan Analyzer The Course Plan Analyze r (CoPA) is a curriculum management and degree auditing service, which contains an expert system that analyze s record to provide students their options for selecting a course plan for the ensuing academic term. Obtaining the A cadem ic I nformation efficient step is to link Albert to ISIS to obtain the necessary information. Ideally, this would be through a secured private connection that would allow students to download identity. However, federal and state laws limit access to this information only to authorized personnel. Therefore, the practical options for this work wer e to use synthetic data or to develop a method to allow students to enter the minimal information necessary to evaluate their academic progress, without compromising their privacy. The most useful option is to request students to manually enter their data The decision is based on student benefit, the value for this research work, design cost and time constraints. Ultimately having students examine and enter their grades is a favorable self assessment exercise for planning their course plan. At the implementation time of CoPA, the ECE UF BSCEE degree was undergoing a renovation. Therefore, the published version of CoPA contains the BSEE program only. The rules and information of this academic degree were obtained by surveying the ECE UF academic advi sors, the UF website and the official UF catalog. CoPA contains all the rules and requisites for taking each course in the BSEE curriculum, plus every course from the ECE department.


58 User I nterface The main expert system and the database for CoPA are built with Python. The user interface of CoPA is built on PHP, a popular server side scripting language that is well suited for web development. These independent modules of Albert, in Python and PHP, communicate via sockets. The user interface of CoPA runs inside the Albert messaging system and CoPA With this feature, students can ask Albert a bout academic records, academic rules, the official course schedule, details about the courses and related information while they enter the forms in the CoPA webpages. Figure 4 1 shows a screenshot of Albert, where the web frame on the right shows the init ial webpage for CoPA. This frame shows the upper contents of the CoPA webpage, with the additional information accessible by scrolling down on the frame. Appendix B shows additional screenshots of Albert and the CoPA webpages. The CoPA S ystem Students can initialize CoPA by requesting it, with expressions similar to I want to update my degree audit and I want to enter my grades To prevent an unintended initialization, Albert will ask users to confirm the request for CoPA before presenting the initial CoPA webpage. An example of this process is evident in figure 4 1. Upon to how it appears in the UF catalog. The student can enter the following academic information in thi s template. Catalog year, the year of admission into the academic program Grades in completed courses Courses currently enrolled in Courses dropped


59 Figure 4 1. A screenshot of Albert with the right side showing the first page of CoPA opened inside the f rame. The left side, the main window, shows the conversational exchange that took place to initiate the CoPA system. Figure 4 2 shows a segment of the template for entering an academic record in CoPA. Students choose their status in each course using the drop down menu and the checkbox. The default value for each course grade is not taken so students need only to edit what they have taken. Students who have special cases, such as courses taken in other institutions, will have their course equivalencies is In the template, elective courses have a text box where students can enter the name or code of the course. The BSEE program has a strict list of courses from which students must choose their technical and specialization ele ctives. The template requires students to enter the correct code, for every course in this list that they have taken. Other elective courses do not include validation for the manual entry box.


60 Figure 4 2. A segment of the template for entering the academic record in CoPA.


61 After students submit their academic record template, CoPA will determine: 1. The courses remaining in the plan of study to complete the degree requirements. 2. The courses the student can take based on the requisites and availability for the semester, along with a separate list for the technical and specialization electives. 3. The courses the student can repeat, if desired. Once CoPA completes this analysis, Albert will invite students to prepare a course plan for the next term. When students accept to complete this process, Albert opens the second CoPA webpage, which contains the course information listed above. Figure 4 3 shows a screenshot of Albert with the mess age exchange that took place for Albert to open the CoPA webpage, after the student submitted the academic record. The frame on the right side of the screenshot, shows the uppermost section of the CoPA webpage. Figure 4 3. A screenshot of Albert with t he right side showing the second part of CoPA


62 Students who have submitted their grades to the system, can ask for this webpage at any moment by stating, for example, I want to choose courses for next semester or I want to submit my courses Figure 4 4 show s this second CoPA webpage, with an example of courses available to choose. Figure 4 4. The template for choosing a course plan in CoPA, with an example of courses available to choose.


63 As seen in figure 4 4, each course listed in this CoPA webpage has a checkbox to the left, and for elective courses, a text box to the right. Students will select each course they wish to take during the next term and optionally enter the code or names of any elective course. T he t echnical and specialization EE elect ives do not have an accompanying text box, given these are listed separately with their titles. In the most recent release, the special topic courses, which use the same numeration for multiple titles, also have a text box for entering the name. This separ ate list ends with a text box for students to enter special case courses that are not on the pre approved list. The courses listed in the Plan of Study section are shown in the same order as in the template in figure 4 2, which is the order suggested in the BSEE curriculum. Students who a re g raduation c andidate s can select that option at the top part of the CoPA webpage; see the upper left part of figure 4 4. Students can submit additional courses that are not part of their major, by using the text area f orm at the end of the last list of courses. The last portion of figure 4 4 shows the three Likert feedback questions that students must complete in order to submit their courses. The write in text area at the end, where students can leave additional feedba ck, is not required for submission. When students complete the course selection process, they submit the academic submission, thus students can repeat the process a s many times as they prefer. The When a student is prepared to discuss the course plan with the advisor, the


64 results from the online database. Students do not need to share their passwords with the advisor, thus keeping all their other data anonymous. Albert website was built using PHP. Figure 4 5 shows the front page of the website, which contains a simple login form. To register, advisors will obtain the credentia ls from the develope r. Once logged in, advisors will see a form to search for the usernames, and to the right, the text pad area that they can use to take notes during their advising sessions. Figure 4 6 shows this webpage, after a s uccessful search of the user ed Figure 4 5. Front page of the advisor website. Figure 4 6. The advisor website, with an example student report.


65 Upon a successful username search, the system will return the information submitted by the student divided in up to 4 lists, i .e. the courses chosen from the Plan of Study section, the courses chosen from the EE electives section, any course the student chose to repeat and the manually written entries. The results will also show if the student selected the graduation candidate o ption. Evaluating CoPA Albert, with the addition of the course plan system, was officially published for students of the BSEE program during enrollment period in the Fall 2013 semester. Students of the BSEE program were invited to use Albert during this p eriod and the results obtained were used to evaluate the effectiveness of the system. As a direct measurement of CoPA, the system records the number of users who submitted an academic record, the number of users who submitted a course plan to the advisor database and the catalog year of each user The ECE advisor provided the number of user s who requested an evaluation of their course plan submission in Albert. Indirect evaluation is done through the questionnaire at the end of the CoPA process, which contains one question that directly relates to CoPA, the user comments and surveys from the advisors.


66 CHAPTER 5 EXPERIMENTS AND ASSESSMENT Albert was designed for rapid deployment to maximize the data collected. User data is crucial for the success of the system and the best available method to create the advising corpus. User data analysis was a major factor triggered approximately seventy percent of the input templates produced Albert went through three incremental phases of data collection and testing While data collected during the first two phases was not enough for significant statistical analysis, it did provide user b ehavioral analytics and significant design decisions such as the need for a course planning system and a broad FAQ list on the webpage. This chapter describes the experimental phases a nd the major results obtained in each phase. The chapter closes with an overall assessment of the system and the current state of operation. Experiments All the experiments were designed for the undergraduate students of the ECE UF programs BSEE and BSCEE. The ECE UF department has approximately 950 students in these programs, with approximately 500 in the BSEE program. The experiments took place between October 2012 and December 2013. Each testing phase began when the ECE UF advisors would inform the stude nts of the availability of Albert and provided the web address for the system. The web address was not available in any other medium, including search engines. Students visit their advisors in the Electrical and Computer Engineering Student Services Offic e (ECE SSO). To encourage students waiting to meet the advisors inside the ECE SSO, a desktop computer was deployed for the exclusive usage of Albert. This


67 computer was available during normal operation hours of the ECE SSO, i.e. business days from 8:00AM website were available 24/7 during the entire fifteen month period, however, stude nt use was minimal outside each delimited phase and thus not included in the results. Undergraduate students at UF must visit their advisors before enrollment period, to discuss their projected course plan and obtain authorization to enroll for courses the publication of required to bring a draft of the course plan form they must turn in during the meeting; however, most students arrive unprepared for this task. In addition, this period is the busiest for advi sors each semester, as many students arrive at the same time, usually shortly before their scheduled date for enrollment. Each experimental phase for Albert refers to this advising period during the Fall 2012, Spring 2013 and Fall 2013 semesters respective ly. Since the objective with Albert is to provide an advising experience that is as close as possible to traditional human interaction, advisors did not provide students any would provide the login instructions, example questions and for phase three, a video tutorial. Results refer to users as unique accounts created. For fairness, identifiable log files were removed, including those from faculty and advisors. Users who made two or fewer statements were also removed. The remaining user files were included in the results, even when the user never meant to converse about academic advising


68 Phase O ne The first experimental phase took place during a six week period between Octobe r 2012 and November 2012. For this implementation, the system did not require a password and CoPA was not available. The webpage showed some instructions and a handful of example FAQ. The main NLU script had approximately 160 input templates, for roughly 1 00 unique responses. Log files from 5 2 user accounts revealed t he average number of logins per user account at 1.5 and t he average number of input statements per user at 16. The input count histogram in figure 5 1 shows that 50% of users made 10 or fewer e ntries. Figure 5 1. Input statement histogram for phase one. The data shows that most users initiated the dialog with a predetermined question. In general, most users experimented with random statements, with many using the e xample FAQs and the help command. Therefore, to teach students about the information in Albert, the FAQ in the webpage display was increased in the subsequent


69 implementation. Users with over fifteen statements showed enthusiasm for the chat system by askin g from a wide range of topics. The FAQ list would also influence students to focus on these topics. The statements that were not correctly recognized had three significant reasons. As expected, the first reason was that the information was not included in the design. The data collected provided information for the subsequent designs. The second reason was that users would approach Albert as if it were a standard search engine, that is they used mostly one word entries and fragmented statements that did not follow natural language. This result prompted adding to Albert the ability to recognize keywords of interest to inform the user about the information available on the topic. This allowed users to navigate towards the solution they seek through natural con versation. The third most common reason for errors was that students requested answers that required knowledge of their academic records. While the system stated that accounts were anonymous and student registration only required choosing a random username to login, students still reques ted answers that required knowledge of their academic records. Some students explicitly stated that the system needed access to their academic records to provide a service they would appreciate. The log files also showed that the system needed a robust spe ll checking process. Phase T wo The second phase of experimentation took place during a four week period between March and April 2013 For this implementation, the system did not require a password and CoPA was not available. The website had 14 questions on the right side help frame, as seen in figure 5 2. The main NLU script had close to 300 input templates, for roughly 160 unique responses.


70 Figure 5 2. A screenshot of Albert, during phase two, showing the FAQ list on the right. Figure 5 3. Input statement histogram for phase two.


71 Log f iles from 103 user accounts indicated the average number of logins per user account was 1.9 and the average number of input entries per user was 11.4. The input count histogram in figure 5 3 sh ows that 62% of the users made less than ten statements. For clarity in figure 5 3, the values for input statements are discontinuous after twenty The log files showed that the most requested feature was personalized recommendations for course enrollment Other highly requested features were also related to course enrollment planning, e.g., the ability to track their progress through the degree and obtain recommendations for course sequence planning. These requests prompted the development of CoPA. With r espect to recognition errors, the data was inline with the results obtained in phase one. Correspondingly the data provided advising topics for the ensuing phase. While Albert contain ed close to 160 unique answers to academic FAQs, most students did not s eek many of these answers, probably due to not having the need for them at the time of the conversation or because they lacked the academic experience to necessitate finding out the information. Many students opted to start the conversation with one of the FAQs listed on the website and then ask similar and follow up questions as they thought of them. For many of these students, this exercise proved to be a useful academic informative session. To instigate further dialog and display the system capabilities the FAQ list was extend ed for the subsequent experimental phase. Phase T hree The third experimental phase took place during an eight week period between October 4, 2013 and November 27, 2013. The system for this implementation contained all the features described in Chapter 3, Chapter 4 and the screenshots in Appendix B.


72 The main NLU script had approximately 415 input templates, with over 200 unique responses. For this phase, only the BSEE advisor sent out an announcement to the students. Therefore, the a mount of BSCEE students who communicated with Albert was minimal. The system compiled data from 387 users during this phase. The plan for CoPA was to complement the advisors during the pre enrollment task However, shortly before this task period began the BSEE advising position became vacant. As a response, Albert was assigned the full task of advising BSEE students during the first half of the period. Albert was initially announced to students who had completed at least two years of the degree and to all students afterwards In order to receive authorization to enroll for courses, students would complete a course plan using CoPA and then e mail the request along with their Albert username to the ECE UF advisor. The name CoPA was not revealed to the student s. The instructions specified that students should submit their course plan using Albert. While the interruption in the advising services prevented testing the s econd phase of the s caling up process, the first half of the process was operating The system recorded submissions for course names from six users. Three of these users made useful contributions. As an official ECE UF service, students were offered four sources for help with using Albert and completing the CoPA process. The first was an email addre ss exclusive for this service, which nobody used. As the second source, the developer was available for tutoring in the ECE UF student computer laboratory during the first three week period. About twenty students attended these sessions. The third source w as the ECE UF Student Branch of the Institute of Electrical and Electronics Engineers (IEEE).


73 Two weeks before phase three began, the developer presented a workshop to approximately twenty chair members of the ECE IEEE The ECE IEEE estimates that ten stud ents visited them for help with Albert during this period As the fourth source for help with completing the CoPA process with minimal effort, Albert featured a link for an online video tutorial hosted in the YouTube website. The tutorial video length is 11 minutes, with about five minutes describing an example of using CoPA. The rest of the video describes additional features and the usage of Albert. Figure 5 4 shows a graph with viewer data provided by YouTube Figure 5 4. Statistics for the video tutorial. Data obtained through YouTube Analytics. Though video views dropped to zero by the end of November, the YouTube results do not necessarily represent unique ECE students using Albert. Nonetheless, the plot shows that 19 computers with unique cookies viewed the video on October 4, the day the advisor made the announcement of Albert. These 19 views had an average duration of approximately five minutes. On the next business day, Monday October 7,


74 the video had 18 unique c ookie viewers with an average of five minutes per view. With these numbers reflecting peak views, clearly the tutorial is too long. The plot also shows students tend to deal with advising topics mostly at the beginning of the week. Views picked up slightly after November 4, the day that enrollment began, though by then, the BSEE advisor was available for service as well. Figure 5 5 shows the input count histogram for the 387 users during this phase. The histogram shows that 53% of users made less than ten e ntries. Many users used the system explicitly for the CoPA process. Registering and completing the CoPA process required a minimum of four statements. The average l ogin per user was two, with 78% of users having two or less. Figure 5 5. Input statement histogram for phase three. From the user total, 292 entered their grades. Figure 5 6 shows the catalog year, that is, the year of admission into the BSEE program, of these users. Clearly, many users did not write the correct response, as 2014 or more is not possible. Fifty five 0 5 10 15 20 25 30 35 40 45 50 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 47 49 54 60 75 106 Users Number of input statements Input Count Histogram


75 users asked Albert to explain the phrase catalog year. It is likely that many students entered either the current year or the expected graduation year, judging by those who entered a future date. The results for the years 2009 through 2012 should be accurate and the overall shape of their distribution is as expected, i.e., with most students in their second and third year of academic progress. This data will contribute to the assessment of the ECE SSO services. Figure 5 6. Catalog year of each user who filled an academic record with CoPA. The ECE SSO estimates that 240 s tudents requested and completed the course plan analysis using Albert, which is about 50% of the BSEE students and 65% of the user accounts. Albert coll ected survey results from 224 students. The discrepancy between the number of surveys and the ECE SSO estimate is mostly due to students who sent their course plan via email, while claiming to have completed CoPA. This 0 10 20 30 40 50 60 70 80 90 2000 2002 2007 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2029 Users Catalog year Catalog Year of Each User


76 outcome is expected when people working under time pressure integrate a new process into a known system. Table 5 1 and figure 5 7 show the results from the survey. Table 5 1. Results from the survey at the end of the CoPA process Question Is this process helpf ul for your academic planning? Is this system easy to use? What is your opinion on the natural language chat application? Average Value (1 5) 2.92 3.29 3.14 Figure 5 7. Survey results with a standard error bar for each question. The survey results reflect mostly positive reviews, as a majority of students gave values of three and four for how much they liked it and how easy it was. The rating of helpfulness is smeared over two, three and four. Conversational systems with similar surve ys obtained averages within the low threes, to four and a half [46] [47] However, 0% 5% 10% 15% 20% 25% 30% 35% 40% 1 2 3 4 5 Percentage of total Survey value Survey Results with Standard Error Bar Helpful Easy Like


77 these systems have a reduced input domain, as the system is who drives the dialog, and users know beforehand what to expect from the system. To help clarify the reasoning behind the survey results, a comment section was added to the survey on October 12, 2013. Of the 191 users who completed CoPA after this date, 61 wrote comments, of which eleven were positive feedback, 45 were negatives remarks and five were technical suggestions with neutral sentiment. The positive remarks were general compl iments and appreciating the course plan system. Of the negative comments, 47% criticized the system as not appropriate for substituting a human advisor, 42% criticized having to manually enter their academic record, while the remaining 11% reported technic al difficulties and unique situations. Therefore, a minimum of 89% of the negative comments was outside the scope of the implementation The disapproval of the method for entering the grades was completely understandable and expected. As described in Chap ter 4, this was the appropriate resolution for the privacy concern in this testing phase, as the academic department would manage the final product and thus have the resources to automatically transfer the information upon student request. To address the comments about substituting a human advisor, the main solution is to inform the students about the objectives for Albert and allow the advisor to integrate the system into the daily tasks. A basis throughout the design of Albert is that to appreciate the v alue of this service, students must understand that the objective is to assist and not replace. The fact that almost half of the negative comments concern this statement validates its significance Unfortunately, the disruption in the advising


78 services ham pered this guideline. Altogether 89% of the negative remarks require realistic provision s from the academic institution and thus easy to implement. Appendix C shows example dialog extracted from the user log files. Responses by the system are curtailed fo r reader ease while user grammatical and typographical errors are preserved for authenticity. The ECE advisor, who processed the student submissions from Albert, completed a review of the system and the course submission process. The review is available in Appendix D. In general the advisor was very encouraging of the system and service provided, with recommendations in line with student reviews. Phase described in Chapter 3. This measur e classifies each input in Albert as the following. Literal match (LM) are input statements that match a FAQ listed on the webpage. Partial match (PM) are statements that have partially matching templates. Outcome negative (ON) includes false negatives (FN), i.e., statements not recognized by the system and true negatives (TN), i.e., statements outside the design scope. Outcome positive (OP) includes correct responses or true positives (TP), and false positives (FP). The LM statements are 71 FAQs listed in the webpage. These include the examples for initiating CoPA, the help command and an example from each topic in Albert. The LM statements do not have any uncertainty for recognition; therefore, these are subtracted from the total input to evaluate the s ystem error. As CoPA was a main feature of phase three, eliminating these commands increases the estimated error. Most of the PM statements the system responds with are advice for the user to obtain the information of interest or incomplete answers This includes some statements


79 that are outside the scope of the system, to which the system responds with statements related to the topic and lets the user know that more information on the topic is not available. Although these templates were successful in the ir design, given the user did not directly receive the desired response, these are not classified as outcome positive. Results are available from 366 users between October 7 and November 27. Table 5 2 shows the results of users who made an expression under each classification and the amount of input statements under each class. Table 5 Input classification Total LM Total Original PM ON OP Users 366 60.49% 11.99% 55.04% 99.73% Input Statements 4952 12.52% 4332 1.69% 14.66% 83.52% The results in Table 5 2 show that 60% of the students copied an instruction exactly as written, which for all purposes is akin to making a selection from a menu. While this result suggests that the interface could benefit from menu selections, the objecti ve of this experiment is to encourage unrestricted expressions. Having approximately 80% of the students initiate the CoPA process and 60% using example input shows the student preference of using the quickest possible method to achieve a goal. Nevertheles s to serve as an educational tool, the previous data showed students benefit from a display of the FAQ. A solution is to include a display of topics with less example statements, while extending the NLP routines that allow users to access data by navigati ng through topics. While this approach is not within the scope of the vision for Albert, the data shows that including both approaches would increase the user base


80 and allow users to explore the capabilities of the system through a more familiar experience Table 5 2 shows that less than 15% of all input was not recognized, however, these statements came from 55% of the users. This result is expected from a system with a restricted domain that accepts all type of input. Any user who decides to test the boun daries of the system will contribute to this result. Figure 5 7 shows the distribution of the ON classified statements for these users. Figure 5 8. Distribution of ON input statements. To reduce the ON responses, as in the previous phases during revision of the log files, false negatives were continuously identified and updated in the NLU system. As the data collection increased, the number of false negatives decreased By the fifth week of the eight week period, the amount of false negativ es per user had dropped to almost zero. Approximately 25% of the users accessed the system during this 0 20 40 60 80 100 120 140 160 180 0 1 2 3 4 5 6 7 -17 Users Number of input statements not recognized ON Statements per User


81 culminating period. During t hese last three weeks no additional measures were taken to minimize ON outcomes from statements that were outside the scope of the design. To determine the number of FP and FN, it is necessary to manually evaluate the OP and ON statements. For this evaluation, a FP outcome is a response that is not relevant to the input statement. This definition of FP error conditions the sys capabilities of responding to the statement with respect to the available information. For Who is my financial advisor For financial information please see the following website is relevant and it is the best response available in the system. If for the same input the Your BSEE advisor is Shannon M Chillingworth classified as FP, given it was not gener ated by the expected template by design. A FN outcome occurs when the input statement has an available answer, yet the system did not respond correctly. Following the same example as above, if a user asks, Who is the person in charge of financial advisin g I do not understand that question responded as stated previously. TN outcomes occur when the input statement is outside the design scope. The FN outcomes include cases caused by technical difficulties in other areas of the system; however, with respect to the overall system performance they are still incorrect responses. It is possible to estimate the FP and FN results by selecting a random sample of the OP and ON statements. To select the samples, a uniform random number is assigned to each of the 3618 OP statements and to each of the 641 ON statements. Each statement is tested for the binary outcome FP or not FP, and FN or not FN


82 respectively. By treating each statement as an independent random variable, the outcomes of the tests will follow a Bernoulli distribution. In the Bernoulli distribution, the maximum variance occurs when the mean is equal to one half, thus th e variance is equal to one fourth. For a given sample size by the central limit theorem, the distribution is closest to the standard normal distribution when is near one half. Under this scenario, a conservative estimate of the sample size needed to estimate with a confidence level and margin of error is given in equation 5.1 where the operator represents rounding to the next integer and is the estimated standard score for a given two sided confidence level [57] (5.1) For a maximum margin of error equal to 10% and a confidence interval of 95%, equation 5.1 returns a minimum size of 97 samples. Table 5 3 summarizes the result for each test. Table 5 3. Results from the outcomes error estimation Outcome Total Statements Confidence Interval Error Margin Sample Size Error Positive 3618 95% 10% 97 15.46% Negative 641 95% 10% 97 24.74% The tests returned 15 false positive statements, for 15.46% of the OP samples and 24 false negatives for 24.74% of the ON samples Using these results, it is possible to determine the statistical measures precision, recall, accuracy and the F 1 measure [18] Precision is defined as (5.2)


83 to measure the percentage of outcomes that were correct. Recall is defined as (5.3) to measure the percentage of positive outcomes given the statements with expected outcome positive. Accuracy is defined as (5.4) to measure the percentage of statements that obtained the expected outcome. The unbias ed F measure, or F 1 measure, is the harmonic mean of precision and recall. (5.5) Table 5 4 summarizes the result for each test. Table 5 4. System p erformance e stimation r esults Precision Recall Accuracy F 1 measure 84.54 % 77.36 % 79.90% 0.81 The results show the system performance metrics are all close to 80%. These results are estimated minimums; given the error is an estimated maximum. Recent studies for comparable e learning systems show Albert offers a high level performance for the complexity of the task [58] [59] Regarding the ma in NLP tasks in Albert, specifically keyword extraction, question answering and natural language interfaces, recently published systems that require data corpora for training, obtain results that show the performance of A lbert is highly competitive [53] [60] [63] Another possible performance measure is a study of human accuracy and response time, though such a study is not available. In any case, the results show Albert offers a competitive


84 performance versus the available technologi es, and much promise for advancement as data collection continues. The results for precision and recall are in line with the design goal of providing students with high precision answers and minimizing false positives. At the same time, the results show th at over 75% of the queries to the system, obtained a reply that is effective versus an input not recognized type answer. This result is valuable considering users did not have any training on how to use the system. Analysis Summary The enhancements developed throughout the experiments proved successful in increasing system performance and user interest for the system. The results of the third experimentation phase showed Albert has developed into a practical and valuable application for academic advising. Overall, ECE UF students and ECE SSO were appreciative of the services provided and supportive for future developments. The data shows that, as is customary in electronic text communications, students prefer to ask questions using th e minimum amount of words possible, instead of through a Standard English sentence. Therefore, Albert should include methods to allow this user preference, while concurrently inspiring students to use complete expressions. As a new service, students were initially reluctant to agree to use Albert for advising chores. Indeed, much effort is required in marketing the service as a useful option versus visiting a human advisor, who is usually a short walk away. However, during peak periods, when the waiting ti me is long, the predisposition to experiment with a new option is very high. Therefore, the implementation of CoPA, a tool for an essential process that all students must complete each academic semester, confirmed the viability of Albert as a real world so lution.


85 Current State of Albert Albert continues to serve the ECE UF department during the Spring 2014 semester. During this fourth experimental phase, Albert will continue growing as an effective advising solution. As in previous phases, Albert has been updated to reflect the most recent results, including a revision of the FP and FN statements from the third feature a twofold process for CoPA, that is, it will allow stud ents to simultaneously prepare a course plan for summer and fall. CoPA will provide students with a richer course planning experience by also providing information to plan for courses ahead of the ensuing semester. Through requests similar to when can I g raduate or in how many semesters can I finish my degree students will obtain a suggested academic track for the courses remaining to complete the degree, divided by semesters. write on the provided text pad area. Their notes are saved in the server and they retrieve them upon searching for the username. The ECE UF BSEE advisor requested this feature, as the system becomes an integral database of the daily advising tasks. The st atistics measured from the data in phase three will provide the baseline for the performance evaluation of phase four. In addition, the ECE SSO can compare the service provided between both semesters and provide additional evaluation of the convenience of this system with their tasks. The main webpage design was updated to meet most of the accessibility compliances of the Americans with Disabilities Act and Section 508, 29 U.S.C. § 794d, of the Workforce Rehabilitation Act of 1973


86 CHAPTER 6 CONCLUSION AND FUTURE WORK Conclusion The fundamental objective of this work is to provide students with an automated online advising experience that is as close as possible to traditional human interaction. This work presents the de sign, development, deployment and evaluation of an i ntelligent n atural l anguage c onversational s ystem for a cademic a dvising. The online system is available 24/7, it includes an advising dialog engine, an academic planning system that integrates the students with the advisors, and a method to allow the users base. The system offers academic departments an instrument to potentially increase student integration, retentio n, satisfaction and performance, by providi ng additional advising services without the need for additional personnel. This achievement is possible by allowing academic advisors to allocate resources to tasks typically classified as developmental and educative advising, instead of repetitive time in tensive tasks related to prescriptive advising. This work contributes real world solutions for the academic community, through a unique combination of software application s and advanced NLP techniques. Every resource used for this system is original or av ailable open sourced. With the experimental data collected, this work has originated a KB of academic advising FAQ s that will serve to build a corpus on the topic Throughout extended use of the system, Albert can potentially exhaust the scope of academic advising FAQs to allow development of systems for multiple programs through mostly answer editing.


87 This system is currently in operation for students of the BSEE and BSCEE degree programs of the University of Florida ECE Department. The online system Alb ert provide s instant service as an academic advising assistant for students, professors and additional members of the academic community. This system allows users to obtain academic information easily through communication analogous to how two people inter act, thus not requiring a user learning curve. After three major experimental evaluation periods the system has proved to be a practical and valuable application for academic advising, based on positive reviews from the users, the advising personnel and the statistical measures. Within a confidence interval of and a margin of error of Albert showed an estimated precision of a recall of an accuracy of and a F 1 measure of Overall, users were supportive and excited about future developments. Future Work The immediate developments for this system include expanding CoPA for multiple semesters and adding a computerized method for submitting the academic record. Using a full academic record from the students will increase the service s the system can implement for students and advisors. Already in development, the advising corpus will provide the data to integrate statistical methods into the dialogue manager, such as those described in Chapter 2. Additional future updates include an escalation mechanism to forward selected student conversations to the advisors adding course information from other academic departments and traditional advising tasks referred to in Chapter 2. For example, Albert could provide students the available path s towards graduation and suggestions for a balanced course plan


88 APPENDIX A T OPICS COVERED BY ALBERT 1. Topics related to managing CoPA 2. Determine the default academic term for student questions 3. Introductions 4. Information related to each curriculum 5. Academic rules from the university and academic department 6. Information related to classes in the students academic program 7. 8. Course prerequisites 9. EE technical el ectives and core courses 10. EE specialization courses 11. CEE technical electives 12. CEE major courses 13. Breadth electives 14. Tracking courses 15. Non ECE courses 16. Internships 17. BSMS program 18. Minors 19. Financial assistance 20. Faculty and staff office hours 21. UF academic calendar 22. Student organizations 23. Academic questions not captured by any previous topic 24. UF campus 25. Teach information to Albert 26. General and off topic questions and comments 27. Information about Albert 28. Keyword catching 29. No response matched; default answers


89 APPENDIX B S CREENSHOTS OF THE ALBERT WEBSITE Figure B 1. Main webpage for Albert


90 Figure B 2 Main FAQ webpage. Figure B 3 Main webpage for Albert showing second FAQ page.


91 Figure B 4. Full list of questions in the second FAQ webpage


92 Figure B 5. Main webpage for Albert showing page one of CoPA. Figure B 6. Main webpage for Albert showing an example page two of CoPA


93 Figure B 7. Main webpage for Albert showing an example of the speech API.


94 APPENDIX C SAMPLES FROM USER LOG FILES This appendix shows excerpts from the log files of six users, recorded during the third experimental phase. User statements are in bold font weight and their syntactic errors are preserved for authenticity. Italic type represen ts actions. Responses by the system are curtailed for clarity. Some html code is included to illustrate how the responses were presented, specifically for bold font and
for new lines. User 1 Welcome to the University of Florida Electrical and Computer Engineering Department's Natural Language Advising System 1.02 Caiman For help at any time type in "Help" and read the example questions. Please be advised that all curriculum information is for ECE students.
I'll remember our conversation wh en you return.

Hello Freezing7, my name is Albert. To better serve you please tell me, are you an EE or a CEE major? What courses do you recommend I take? ==> It seems I don't recognize your degree program. If you are in the B SEE program please enter EE. If you are in the BSCE "Computer Engineering Hardware" program please enter CEE. If neither applies to you please answer None . EE ==> Very well Freezing7, you are an Electrical Engineering student. W hat can I do for you? What courses do you recommend I take? ==> I will tell you which courses you could take next semester after you enter your grades. Do you want to enter your grades into the database? (yes | no) yes ==> No response. The first frame o f CoPA is launched at this time. confirm ==> Do you want to choose your courses for the next semester and submit the list to your advisor? yees ==> To better under stand you I assumed you meant: yes
Instead of: yees The second frame of CoPA is launched at this time. drop ==> Drop? I can't tell what you're saying with a one word sentence. Look at the example questions on the right of this webpage for ideas of dialog we can have.

All I can say about the word drop is
Noun drop: the act of dropping something. Verb drop: the act of dropping something. can i drop a class ==> So you want to drop a course? Are you in upper division or lower division? (You're in lower division if you haven't passed 60 credits yet.) lower ==> Students are given TWO drops prior to completing 60 credit hours. Also


95 know about the college of engineering drop rule; The College of Engineering .. User 2 OK F9788, you are an Electrical Engineering student. What can I do for you? What courses do you recomm end I take? ==> I will tell you which courses you could take next semester after you enter your grades. Do you want to enter your grades into the database? (yes | no) y ==> No response. The first frame of CoPA is launched at this time. confirm ==> Do you want to choose your courses for the next semester and submit the list to your advisor? what courses do you recommend I take? ==> Assuming your info is up to date, these are the courses you could take next semester.

EEL 3472 Electromagne tic Fields and Applications 1

College breadth elective User 3 Welcome back willwrdn I want to drop a course ==> So you want to drop a course? Are you in upper division or lower division? (You're in lower division if you haven't passed 60 cr edits yet.) upper ==> Students are given TWO drops after completing 60 credit hours. Any unused lower division drop is lost forever. what is the deadline for dropping a course ==> I'll tell you this, under extreme circumstances that are beyond your control, it is possible to drop a course after am I able to drop a course via ISIS? ==> So you want to drop a course? Are you in upper division or lower division? (You're in lower division if you h aven't passed 60 credits yet.) upper ==> Students are given TWO drops after okay ==> All right. so how do I drop a course? ==> You drop classes via ISIS. If the add/drop period has ended you can still do it, but you need to see your advisor so he can complete the process. alright! thank you! That's all the information I need today ==> Well, I am learning every day. What can I do for you? logout ==> Before you go, please share any feedback or suggestions to help you better next time. Perhaps it was just that my querie s were better and fewer this time, but the system seems to have improved. Good job! I hope it gets to access to more information soon so that it can answer more questions instead of some variation of I don't know ==> All right, thanks!
Later, Gator!


96 User 4 Hi Novae! show my schedule ==> You want me to show you the courses you chose for next semester programming classes ==> Electrical Engineering students must take one programming course in Java, C or C++ for their major. The following courses are recommended: COP3502 COP3504, EEL3834 or COP2271. digital logic prequisite ==> To better understand you I assumed you meant: Digital_Logic perquisite
Instead of: digital logic prequisite

I reco gnize Digital Logic is a course in our department, but could you be more specific about your request?
Maybe the example questions on the right of this webpage can help.
I certainly can tell you the number for Digital Logic, it's EEL 3701C. prere quisite for eel 3701c ==> The prerequisite for EEL 3701C is a Programming Course. is eel3834 being taught next semester? ==> The course EEL 3834 i s not offered in Spring 2014 what classes can i take ==> Assuming your info is up to date, these are the courses you could take next semester.

Professional Communication circuits 2 ==> I recognize CIRCUITS 2 is a course in our department, but could you be more specific about your request?
Maybe the example questions is eel3701c being o ffered spring ==> For Spring 2014 we have EEL 3701C is Who will teach EEE3308C? ==> For Spring 2014 we have
EEE 3308C by Yoon,Yong Kyu digital logic prereq ==> For Spring 2014 we have
EEL 3701C by Gugel,Karl S digital logic prerequisite == > For Spring 2014 we have EEL 3701C by Gugel,Karl S eel3701c any sections available ==> For Spring 2014 we have
EEL 3701C by Gugel,Karl S Is it ok if I take less than 12 credit hours? ==> UF policy states that the minimum full time load for undergr aduate students is 12 credits. The minimum full time load How many critical tracking courses are required for my major? ==> You need to pass 8 critical tracking courses. You need to pass all the listed tracking courses What courses did I select for n ext semester? ==> I was thinking about taking two lab classes. Is that a good idea? ==> These are the courses with labs included:
EEL3701C Where can I see my curriculum? ==> Here is your c urriculum. EE Curriculum


97 User 5 Log 1 Hello gil3167, my name is Albert. To better serve you please tell me, are you an EE or a CEE major? EE ==> OK gil3167, you are an Electrical Engineering s tudent. What can I do for you? I need to have my hold lifted ==> I can't know if your hold was removed because I am not authorized to access ISIS. If your academic hold was removed you should have gotten a confirmation email by the ECE advisor. I want to choose my courses for spring ==> You can do this if you have y our grades up to date in my system. To update your degree audit, just let me know yes ==> No response. The first frame of CoPA is launched at this time. what is catalog year? ==> Students graduate under the catalog in effect is ANT 2000 a humanities ==> Could you rephrase that please? ANT 2000 ==> I'm not following. are my grades analyzed? ==> Ok gil3167 I have a list of courses you could take next semester, assuming you pass those e nrolled in. After you submit your courses through that webpage, you can then email the EE advisor
Do you want to choose your courses for the next semester and su bmit the list to your advisor? yes ==> No response. The second frame of CoPA is launched at this time. i'm interested in a minor ==> You can find the general information regarding computer engineering minor ==> You can find the general information regarding bye ==> Before you go, please share any feedback or suggestions to help you better next time. User 5 Log 2 => Hi again gil3167 how many credits is digital design ==> EEL 4712C has 4 credits. how many credits is electromagnetic fields ==> EEL 3472 has 3 credits. does eel 3472 have a lab? ==> These are the courses with labs included:
is engineering statistics hard = => I only know information for courses of our department. is digital design hard ==> A good course is where you will learn the important who is teaching engineering statistics ==> Who teaches engineering? I can't tell what do you recommend that i take next ==> I recommend you study hard and


98 what courses should i take ==> Assuming your info is up to date, these are what are the college breadth electives ==> A College Breadth Elective can be User 5 Log 3 => Good to see you again gil3167 choose classes ==> I didn't quite understand your statement. Regarding EE what courses do you recommend i take? ==> Assuming your info is up to date i would like to register ==> You register for courses via ISIS. < br> You can ask removing my hold ==> As you may have read in your advisor's email, I will help Do you want to enter your grades into the database? (yes | no) yes ==> No response. The first frame of CoPA is launched at this time. done Do you want to choose your courses for the next semester and submit the list to your advisor? yes ==> No response. The second frame of CoPA is launched at this time. User 6 enter my grades Do you want to enter your grades into the database? yes ==> No response. The first frame of CoPA is launched at this time. confirmed my grades ==> Excuse me aradulescu1, I'm not sure what you mean. Student submits first CoPA page. confirmed my grades ==> Do you want to choose your courses for the next semester an d submit the list to your advisor? no ==> No problem. What can I do for you? I want to choose my courses for the next semester ==> No response. The second frame of CoPA is launched at this time. What courses do you recommend. ==> Ass uming your info I'm done. ==> Done with what? Talking to you. ==> Is that so? I need to learn about that. Log out. ==> Before you go, please share any feedback or suggestions to help you better next time. I did. ==> All right, thanks!
Later, Gator!



PAGE 100

10 0

PAGE 101

101 LIST OF REFERENCES [1] V. N. Gordon, W. R. Habley, T. J. Grites and National Academic Advising Association. Academic Advising: A Comprehensive Handbook, 2nd ed. San Francisco: Jossey Bass, 2008. [2] S.H. Frost, "Historical and Philosophical Foundations for Academic Advising", in Academic Advising: A Comprehensive Handbook, V. Gordon, W. R. Habley, T. J. Grites and National Academic Advising Association, Eds. San Francisco: Jossey Bass, pp. 3 16, 2008. [3] The Ment or: An Academic Advising J. Oct 2012. [Online]. Available: satisfaction/ Retrieved Feb 2013. [4] Academic Advising: A Comprehensive Handbook, V. Gordon, W. R. Habley, T. J. Grites and National Academic Advising Association, Eds. San Francisco: Jossey Bass, pp. 292 306, 2008. [5] L. Waldner, D. McDaniel, T. Esteves and T Anderson Generation eAdvising Tool to Build Community and R The Ment or: An Academic Advising J. Oct 2012. [Online]. Available: eadvising tool build com munity retain students/ Retrieved Feb 2013. [6] National Academic Advising Association, Advising Technology Innovation Awards. [Online]. Available: Awards/EPublication.htm Retrieved Feb 2013. [7] The Mentor: An Academic Advising Journal [Online]. Available: advising system/ Retrieved Feb 2013. [8] University Advising Center, The University of Texas at Arlington, Arlington, TX. [Online]. Available: planning/uac/index.php Retrieved Feb 2013. [9] Academic Advising and Career Center, eAdvising, University of Michigan Fl int, Flint, Michigan. [Online]. Available: Retrieved Feb 2013. [10] E. R., White and M. J. Leonard, Faculty Advising and Technology, in Faculty advising examined: Enhancing the potential of college faculty as advisors G. L. Kramer, Ed. Bolton, MA: Anker Publishing Company, 2003.

PAGE 102

102 [11] Sci. Coll., vol. 18, no. 3, pp.17 25, Feb 2003. [12] F. Albalooshi and S. Shatnawi, "HE Advisor: A multidisciplinary web based higher education advisory system", Global J. Computer Sci. and Tech., vol. 10, no. 7, pp. 37 49, Sept. 2010. [13] A. N. Nambiar and A. K. Int. Conf. Educational and Inform. Tech., vol. 1, pp. V1 31 V1 315, Sept. 2010. [14] T. Feghali, I. Zbib and S. Hallal, "A Web based Decision Support Tool for Academic Advising", J. Educational Technolog y & Society, vol. 14, no. 1, pp. 82 94, 2011. [15] E. Onyeka, D. Olawande and A. Charles, "CAES: A model of an RBR CBR course advisory expert system," Int. Conf. Inform. Soc., pp.37 42, June 2010. [16] tudent 2010. [17] 2010. [Online]. Av ailable: um/intelligentCounselingSystema24/219101 Retriev ed Feb 2013. [18] D. Jurasky and J. H. Martin, Speech and Language Processing An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition, 2nd Ed, New Jersey: Pearson Education, 2008. [19] clopedia of Philosophy Online. E. Craig General Ed., Taylor & Francis Group, 2012. [Online]. Available: Retrieved Feb 2013. [20] Princeton University, "About WordNet", Princeton University, 2010. [Online]. Available: Retrieved Feb 2013. [21] 41 no. 2, pp. 1 69, 2009. [22] Quarteroni, S. and Manandhar, S. "Designing an interactive open domain question answering system." J. Natural Language Eng, vol. 15, no. 1, 73 95, 2008. [23] es to Databases 81, 1995.

PAGE 103

103 [24] T. Harley, The Psychology of Language: From Data to Theory, 3rd Ed., Hove: Psychology Press, 2008. [25] olysemy 223, Elsevier Science, USA, 2002. [26] Conversational Agents and Natural Language Interaction: Techniques and Effective Practices D. Perez Marin and I. Pascual Nieto, Eds., USA: IGI Global, 2011. [27] E. Reiter and R. Dale, Building Natural Language Generation Systems, Cambridge University Press, 2000. [28] evaluating spoken Computational Linguistics, pp. 271 280, 1997. [29] V. Hung, M. Elvir, A. Gonzalez and R. Demara, "Towards a method for evaluating naturalness in conversational dialog systems," Syst. Manage. and Cybernet ics IEEE Int. Conf., pp.1236 1241, 2009. [30] D. Oelke, Ming Hao, C. Rohrdantz, D.A. Keim, U. Dayal, L. Haug and H. Janetzko, "Visual opinion analysis of customer feedback data," Visual Analytics Science and Technology. VAST 2009. IEEE Symposium on pp. 187 194 Oct. 2009. [31] Artif. Intell. Med. vol. 44, no. 2, pp. 83 89, Oct. 2008. [32] The Loebner Prize in Artificial Intelligence, [Online]. Available: prize.html Retrieved Feb 2013. [33] R.S. Wallace, Alice A.I. Foundation, [Online]. Available: Retrieved Feb 2013 [34] B. Wilcox, ChatScript project, [Online]. Available: Retrieved Feb 2013. [35] Pandorabots, the open source software robot hosting service, [Online]. Available: Retrieved Feb 2013. [36] N. Bush and N. Dubrovsky,, Open Source Community, [Online]. Available: Retrieved Feb 201 3. [37] F.A. Mikic, J.C. Burguillo, M. Llamas, D.A. Rodriguez and E. Rodriguez, "CHARLIE: An AIML based chatterbot which works as an interface among INES and humans," EAEEIE Annual Conference pp. 1 6, 2009.

PAGE 104

104 [38] S. Bird, E. Klein and E. Loper, Natural Language Proc essing with Python, O'Reilly Media, 2009. [Online]. Available: Retrieved Feb 2013. [39] P. Ljunglf, " An open source Python library for developing ISU based dialogue systems", Proc. IWSDS, Kloster Irsee, Germany, 2009 [40] K.E. Boyer, E.Y. Ha, R. Phillips, M.D. Wallis, M.A. Vouk and J.C. Lester, "Dialogue Act Modeling in a Complex Task oriented Domain", Proc 11 th Annu. Meeting Special Interest Group Discourse and Dialogue Association Comp utational Linguistics pp. 297 305, 2010. [41] S. Quarteroni and S. Manandhar, "Designing an interactive open domain question answering system" J. Natural Language Eng ., vol. 15, no. 1, pp. 73 95, Jan 2008. [42] n Academic Advising: A Comprehensive Handbook, V. Gordon, W. R. Habley, T. J. Grites and National Academic Advising Association, Eds. San Francisco: Jossey Bass, pp. 85 102, 2008. [43] Academ ic Advising: A Comprehensive Handbook, V. Gordon, W. R. Habley, T. J. Grites and National Academic Advising Association, Eds. San Francisco: Jossey Bass, pp. 309 322, 2008. [44] Acade mic Advising: A Comprehensive Handbook, V. Gordon, W. R. Habley, T. J. Grites and National Academic Advising Association, Eds. San Francisco: Jossey Bass, pp. 17 35, 2008. [45] Learning", Applicat. and Innovations in Intelligent Systems XVI pp. 169 182, 2009. [46] S. Gandhe, N. Wh itman, D. Traum and R. Artstein, An integrated authoring tool for tactic Proc. 6th IJCAI Workshop on Knowledge and Reasoning in Practical Dialogue Systems Association for the Advancement of Artificial Intelligence, 2009. [47] J. Comput. Cult. Herit. vol. 6, no. 2, pp. 10:1 10:15, May 2013. [48] Contributors to the Web Speech API Specification, Web Speech API, Speech API Community Group, 2012. [Online]. Available: api/raw file/tip/ speechapi.html Retrieved Feb 2013. [49] P. Norvig. How to Write a Spelling Corrector [Online]. Available: correct.html Retrieved Feb 2013.

PAGE 105

105 [50] Council for the Advancement of Standards in [51] Stockholm University, 200 4. [52] Conf. Advances Comput. En tertainment Technology, T Romo et al. Eds. ACM, New York, art. 17, ch. 2, 2011. [53] 2 nd Workshop Making Sense of Microposts 21st Int. Conf. World Wide Web, 2012. [54] D. Artif. Intell., P. Anthony, M. Ishizuka and D. Lukose, Eds. Springer Verlag, Berlin, Heidelberg, pp. 6 37 648, 2012. [55] L. Richardson, Beautiful Soup [Online]. Available: Retrieved Feb 2013. [56] R. A. Baeza Yates and B. Ribeiro Neto, Modern Information Retri eval Addison Wesley Longman Publishing, Boston, MA, 1999. [57] J. A. Gubner. Probability and Random Processes for Electrical and Computer Engineers, Cambridge University Press, Cambridge, United Kingdom, 2006. [58] A. Olney, M. Louwerse, E. Matthews, J. Marineau, H Hite Mitchell, and A. NAACL Building Educational Applicat. using NLP, Assoc for Computational Linguistics vol. 2, 1 8, USA, 2003. [59] W. Ward, R. Cole, D. Bolaos, C. Buchenroth Martin, E. Svirsky, S. V an Vuuren, T. Weston, J. Zheng and L. Becker, "My science tutor: A conversational multimedia virtual tutor for elementary school science", ACM Trans Speech and Language Process. TSLP, vol. 7, no. 4, 2011. [60] Specific Knowledge Representation for Web Self Proc 2nd Int Conf Human Syst Interaction IEEE, May 21 23, Catania, Italy, pp. 298 305, 2009 [61] M. Minock, "C Phrase: A system for building robust nat ural language interfaces to databases", Data & Knowledge Engineering vol. 69 no. 3, pp. 290 302, 2010. [62] M. Olvera Lobo and J. Gutirrez Artacho, "Open vs. Restricted Domain QA Systems in the Biomedical Field," J. Inf. Sci. vol. 37, no. 2, pp. 152 162, 201 1.

PAGE 106

106 [63] Web Semantics: Science, Services and Agents on the World Wide Web vol. 21, pp. 3 13, 2013.

PAGE 107

107 BIOGRAPHICAL SKETCH Edward is a Professional Engineer and a Doctor of Philosophy in electrical and computer engineering, whose research area comprises natural language processing, software application development, and computer science and engineering education. Edward obtain obtained the PhD and a Master of Science in electrical and computer engineering from the University of Florida. For his academic achievements and passion, he obtained a Google scholarship for the 2013 2014 academic year