Group Title: Department of Computer and Information Science and Engineering Technical Reports
Title: A Mapping repository for meta-querier evolution
Full Citation
Permanent Link:
 Material Information
Title: A Mapping repository for meta-querier evolution
Alternate Title: Department of Computer and Information Science and Engineering Technical Report
Physical Description: Book
Language: English
Creator: Li, Xiao
Cow, Randy
Publisher: Department of Computer and Information Science and Engineering, University of Florida
Place of Publication: Gainesville, Fla.
Copyright Date: 2008
 Record Information
Bibliographic ID: UF00095725
Volume ID: VID00001
Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.


This item has the following downloads:

2008459 ( PDF )

Full Text

A Mapping Repository

for Meta-querier Evolution

Xiao Li and Randy C'l,..-

CISE Technical Report 08-459
Department of CISE, University of Florida,
Gainesville, FL, 32611, USA
{xl, chow}

Abstract. There exist an increasing number of online databases search-
able only through query interfaces. Meta-querier with an integrated query
interface is a practical solution to effective searching of information from
multiple data sources. The multiplicity of data sources and their frequent
changes necessitate the need for maintenance to support the evolution
of meta-queriers. This paper presents a novel framework called MEQE
to address this unique issue of meta-querier evolution. Central to the
framework is the concept of a mapping repository, which is a domain-
based information base. This information base stores all current and
past mapping information, and their evolution orders, and functions as
a mapping storage for query translation from a meta-querier to the un-
derlying queriers. Moreover, it serves as a domain-based knowledge base
for (semi-)automatic schema matching and merging.
The innovation in the design of the mapping repository is a concep-
tual graph model, which integrates both mapping and version manage-
ment. In addition, the proposed framework can facilitate four advanced
features: self-validation, self-configuration, self-maintenance, and virtu-
alization. The paper describes details of the model and outlines an im-
plementation architecture. Finally, it characterizes classes of mapping
repositories and examines the potential use of the repository-based solu-
tions for schema matching and merging.

1 Introduction

Internet has been a great source of information for our daily life. However, the
explosion of information has made it difficult to efficiently retrieve useful infor-
mation, especially those hidden under the deep webs. Deep webs refer to the
on-line database searchable only through HTML query interfaces. Even in 2004,
a survey estimated the scale of on-line database had already reached 450,000
sites [11]. Various researchers have attempted to develop techniques to expose
these normally unreachable databases.
Studies have shown that building meta-queriers is the only practical solution
to searching effectively the contents from multiple on-line databases [8, 14, 23].
A meta-querier refers to a system that provides a unified access (i.e., a global

2 Xiao Li, Randy Chow

query interface) to multiple existing online databases (i.e., local queriers). With
a meta-querier, users only need to input a query in a global query form, which is
respectively translated to multiple local queries with the same semantics (called
query translation), and then the combined local query results are returned. In
addition, meta-queriers can also be customized (called customized meta-queriers)
[29], that is, the users are allowed to choose local queriers based on their interests
and habits.
In the Web environment, online databases often change not only their data
contents and formats, but also their semantics, query interfaces, and presenta-
tion I 1. [i.] And customized meta-queriers also can be changed because of the
insertion or deletion of local queriers by users. Sometimes, a slight change in
the bottom queriers could lead to a global effect and even jeopardize the whole
system. In addition, changes of query forms could be frequent and unpredictable.
Meta-queriers have to be incrementally evolved to adapt to these changes in or-
der to meet the 24/7 operating requirement. Therefore, meta-querier evolution
is more practical and considered a more critical issue than reconstruction for
long-running meta-queriers.
In this paper, we propose a novel framework called MEQE to address this
unique issue of meta-querier. Central to the framework is the concept of a map-
ping repository, which is the focus of this paper. The mapping repository is
a domain-based information base for meta-queriers' daily operations. Its main
functionalities include three points:
1) It supports mapping retrieval and management for query translation. First,
the core of query translation is the mappings from query inputs in a global
interface to inputs in local interfaces. Thus, mapping storage and management
are fundamental issues for meta-queriers. In addition, our proposed mapping
repository is shared by multiple meta-queriers to decrease storage space and
prevent update anomalies, as shown in Figure 1. The intended customized meta-
queriers could be numerous. Even for a specific meta-querier, the number of
alternative local queriers could be large [11]. Both scalability issues indicate
that the number of mappings in an integrated meta-querier system is potentially
overwhelming. Thus, mapping storage and management is an important issue.
2) It provides version storage and management resulting from meta-querier
evolution. Evolution implies the reuse of previous information to maintain the
inoperable meta-queriers rather than rebuilding the meta-queriers from scratch.
First, the unchanged self mappings can be directly reused. In addition, the devel-
opment history of self and peer queriers carries important information for finding
the changed or newly inserted mappings. Thus, version storage and management
are added into our proposed mapping repository.
3) Since a predefined and preconfigured knowledge cannot adapt well to dy-
namic changing environments, the design of the proposed mapping repository
addresses this shortfall by considering the following additional features: self-
configuration, self-maintenance, and self-validation.
The presentation of the proposed mapping repository for meta-querier evo-
lution is organized as follows: Section 2 provides a background of the related

A Mapping Repository for Meta-querier Evolution

meta queriers _

local queriers

." 1


Fig. 1. A mapping repository shared by multiple meta-queriers.

work. Section 3 describes the modeling of query interfaces we developed. Sec-
tion 4 gives an overview of the proposed framework MEQE. In Section 5 and
6, the prototypes and functionalities of the mapping repository are discussed in
more details. Section 7 characterizes classes of mapping repositories and exam-
ines the potential applications of the mapping repository. Finally, we conclude
and discuss future work in Section 8.

2 Related Work

2.1 Meta-querier
The most significant benefit of a meta-querier is its ability to utilize multiple local
search engines to query multiple different but related objects at the same time.
Many such commercial meta-queriers, such as Addall[1], Smartertravel[2] and
Kayak[3], have implemented similar but simplified functions. Their techniques
are proprietary and were often not reported.
In the research arena, multiple groups, such as WISF[_ I], MetaQuerier[12]
and DEQ UE [36], have attempted to analyze and attack this important problem
from various perspectives. These systems can be viewed as an automatic inte-
gration of web databases to facilitate efficient web exploration[12, 24, 36]. Each
complete system generally consists of the following components: source discov-
ery and selection[22, 26, 38], interface extraction[23, 47], schema matching[19, 21,
41, 43, 44], schema merging[16, 34, 45], query translation[25, 46], results extrac-
tion[27, 48], result in,. i -ii, [2" 37], and system maintenance.
However, these frameworks only focus on the initial construction of meta-
querier, but ignore the maintenance issue, which is considered more important
than initial construction. In addition, they do not describe any mechanism for the
storage of mappings, which can be used for the constructions and operations of
meta-queriers, such as schema matching, schema merging, and query translation.

2.2 Mapping Repository
For storing mappings, most -i. i,-[: 39,49] simply employ a large table. Its
columns represent different properties of mapping, such as element names, map-
ping expressions, and similarity values. Others use graph models in which nodes

4 Xiao Li, Randy Chow

denote schemas or fragments of schemas and edges represent mappings[9, 15].
Another approach[7, 32] is a combination of both. It models the mapping repos-
itory using graph representations with a finer granularity. That is, nodes in the
graph denotes elements in the table-based implementation rather than schemas
in the graph-based implementation.
Our mapping repository adopts the third approach. In the proposed graph
model, nodes and edges have multiple type attributes to assist subsequent main-
tenance and validation. The design of this model considers both mapping storage
and evolution, and thus mapping management and version management are in-
tegrated into this model.

3 Query Form Modeling

The proposed mapping repository requires a uniform representation of query
forms. This section introduces an undirected graph model for representing query
Current languages used to represent query forms include HTML, CSS, and
some scripting languages. Structures of query forms can be classified into two
groups: one-step forms and multi-step forms. Theoretically, the multi-step forms
can be decomposed to multiple one-step forms based on their appearance depen-
dence[36]. Therefore, this paper considers only the modeling of one-step forms.
For one-step form, the W3C's HTML specification[5] defines it as "a section of
a document containing normal content, markup, special elements called controls
(checkboxes, radio buttons, menus, etc.), and labels on those controls." The
returned contents of user requests are normally based on the modification of
controls (clicking radio buttons, entering text, etc.)

Select how you would like to search for a hotel
Ci A A C~d, M IfCaILdk
s_, FFIMt. -- I Mapping
Stale Pr re 4 Ip' o C t Repository
Nearby City instance Within 15miI25mi
Select dates of travel

Check-out date MM
Adults (ajelt) il Children C per room Graph
j FCi zP,

Fig. 2. A query interface and its graph models.

A Mapping Repository for Meta-querier Evolution

Typically, Interface Extractor (IE) can extract useful attributes such as con-
trol type, name/label, descriptive text, instances, data domain, default value,
scale/unit (e.g., kg, million, dollar), and data/value types (e.g., date type, time
format, char type, etc.) for each control[23,47]. The controls and their associated
attributes are referred to as Functional Components (FC) in our system. Two
FCs are called syntactically equivalent iff all the attributes of two FCs are
the same.
For instance, in the bottom left of Figure 2, FC1 is a functional compo-
nent extracted from a query interface. It has a control type "menu", a label
name "price", a descriptive text "How Much?", an instance set (from 211- 1 i
to :"-.- -i-," and "All"), data domain (from 20 to 400), default value ("All"),
unit (dollar), data type (integer domain).
To capture the structural semantics among different FCs, we translate query
forms from the native format into undirected graphs. A vertex in the graph
corresponds to a FC in one form. If two FCs are vertically or horizontally adjacent
to each other, an edge is used to connect them. Each edge is associated with a
boolean property to represent its adjacency type, vertical or horizontal. Each
maximum connected subgraph in the graph corresponds to a semantic block
with a descriptive text D (if available). Each block is assigned a unique identifier
called BlockID. In addition, each query form is identified by a uniform resource
locator (URL) and a version number (denoted by the last valid accessed time
The left portion of Figure 2 shows a query interface for a hotel booking
system. It requires users to input three categories of information. The middle
portion of Figure 2 shows its corresponding graph model. This graph is composed
of three corresponding semantic blocks. In each block, the vertical relations be-
tween FCs are denoted by solid lines, whereas the horizontal ones are denoted
by dotted lines. In the proposed framework, Graph Repository stores this FC
graph and Mapping Repository records a corresponding entity for each FC.

4 A framework of Meta-querier Evolution

This section gives a brief overview of the proposed maintenance framework called
MEQE for MEta-Querier Evolution. MEQE is to maintain inoperable meta-
queriers by mainly leveraging the previous learned information.
The maintenance of meta-queriers can be characterized in three categories:
a) Deleting unneeded queriers; b) Merging newly inserted on-line databases into
the current meta-querier; c) Repairing broken connections caused by the changes
of existing query forms. For the first situation, it can be easily implemented
by removing unneeded query forms' relations to the global query form. Thus,
MEQE only focuses on the last two situations. The input to the system is a
newly inserted query interface or an inoperable local query interface and the
output is a fixed meta-querier.
Figure 3 is a high-level flowchart that illustrates the steps involved in main-
taining a meta-querier. First, Interface Extractor (cf.3) extracts FCs from the

6 Xiao Li, Randy Chow

OI Newlylnserted Inoperable
Interfaces Interface

(I Interface Extractor

( Change Detector Graph

Schema Matcher
R X d Mapping
Schema Merger

) A fixed

Fig. 3. A mapping repository.

newly inserted (cf.l) or inoperable (cf.2) query forms. For inoperable query
forms, ( I-.. ..; Detector (cf.4) [30] begins to syntactically compare the current
FC graph with the last version, stored in Graph Repository (cf.8), and passes the
difference to Schema Matcher (cf.5). Then, for each modified/new FC, Schema
Matcher utilizes its structure information (stored in Graph Repository) and his-
tory information (stored in Mapping Repository cf.9) to obtain the best global-to-
local mapping correspondence and expression. If one is found, Schema Matcher
inserts a new mapping to Mapping Repository. Otherwise, Schema Merger (cf.6)
attempts to merge it into the global interface and updates Mapping Repository.
If the fixed meta-querier (cf.7) does not trigger any exception, the maintenance
is successful. Otherwise it evokes human intervention.

5 Mapping Repository

In the core of MEQE, Mapping Repository stores all the mapping information,
including the current and past mappings and their evolution order. Such in-
formation can be employed by Schema Matcher and Schema Merger to fix the
broken or newly inserted local queriers. Mapping Repository is also used for
mapping retrieval during query translation.
We start from the representation of mappings. Having a semantically rich rep-
resentation of mappings is particularly important. In a meta-querier, a mapping
can be defined as Mapping (Expression, ListFCL, ListFCc), where ListFCL
and ListFCG are two ordered list of FCs' inputs in a local query form and a
global form, respectively, and Erpression denotes a high-level declarative ex-
pression that specifies the transformation from ListFCG to ListFCL. Expres-
sions can be list-oriented functions (e.g. equivalence, concatenation, mathematic
expressions) or other more complex statements (e.g. if-else, while-loop). And
mapping cardinality can be 1:1, l:n and n:m (n > 1 and m > 1).

A Mapping Repository for Meta-querier Evolution

In addition, meta-queriers are not static. The evolution of meta-queriers com-
plicates the mapping representation:
a) A local querier often develops multiple versions over time. Each version
corresponds to a specific query form. Meta-queriers are required to update the
relations between this altered local query form and the global form. Thus we
need to seek and update the corresponding FC mappings. The previous self
mapping information facilitates finding the proper mappings. And more recent
versions carry more weight in understanding the current active version. Thus,
development history needs to be stored with each mapping.
b) The collection of local queriers for a meta-querier can change, especially for
a customized meta-querier. When a new querier is added, the corresponding new
mappings need to be inserted into Mapping Repository. New mappings generated
by matchers are individually attached with a similarity value (ranging between
0 and 1), indicating how well the match is.
Given the above discussion of mapping representation, we are ready to present
a graph model for Mapping Repository and its unique managing capabilities in
the following subsections.

5.1 Conceptual Graph Model

Mapping Repository is modeled using a directed graph, with three kinds of nodes
(regular nodes, super nodes, and object nodes) and two kinds of edges (mapping
edges and version edges), as illustrated in Figure 4.
As described in Section 3 on query form modeling, a query form is represented
by an undirected FC graph, where each FC refers to a control and its associated
attributes. Undirected FC graphs are stored in Graph Repository separated from
Mapping Repository.
In Mapping Repository, a regular node (shown as a solid-line round) is in-
tended to represent a FC. Regular nodes are the most basic building block in this
model. Super nodes (shown as a solid-line rectangle) are the clusters of regular
nodes whose FC inputs can be transformed to each other by an 1' i. "
mapping. Mapping edges (shown as solid arrowed lines) connecting super nodes
are employed to store the corresponding mapping information, such as mapping
expressions, validation statuses and similarities values.
Furthermore, each querier has various different versions. Each version is a
specific query form modeled in an undirected FC graph. We observe that most
FCs in two subsequent versions of a querier do not change[6]. And thus, the
concept of a semantic object is introduced to denote a set of FCs satisfying:
1) syntactically different (defined in Section 3); 2) with the same semantics;
3) from the same querier. Each FC is called a version of its semantic object.
Given the above definition of a semantic object, it is easy to deduce: 1) Two
syntactically equivalent FCs from different versions of a querier belong to the
same version of a semantic object; 2) There exist multiple pairs of FCs which are
the same version of a semantic object but respectively in two different versions
of a query form.

8 Xiao Li, Randy Chow

In Mapping Repository, each object node (shown as a dotted line circle) refers
to a semantic object, including all its versions (regular nodes). Different versions
of the same semantic object are connected by version edges (shown as dotted
arrowed lines) based on the order that they were developed.

Motivating Scenario To motivate the conceptual model, we use a scenario
that is illustrated in Figure 4. Consider two customized meta-queriers, MQ1 and

Fig. 4. A mapping repository.

MQ2, which are respectively built on two groups of local queriers, {LQ1, LQ3}
and {LQ1,LQ2}.
Initially (illustrated in Figure 4(a)), each of these three local queries has
only one version (i.e., the currently active query form). Each version has only
one FC, whose type is a text-entry box. MQ1 and MQ2 share the same mapping
repository. In this mapping repository, three FCs in three local queriers and two
FCs in two meta-queriers are represented by five regular nodes. They are divided
into a global super node {C} and a local super node {A}, respectively, based on
their sources. A mapping edge el records a 1:1 mapping expression from {C}
to {A}. Additionally, each querier has a object node containing its own regular
As time passes, LQ2 and LQ3 respectively develop one and four new versions,
as shown in Figure 4(b). In this example, the change happened in LQ2 does not
break the current mapping, and thus, only a new regular node representing the
FC (Text-entryC) is substituted for the FC (Text-entryB) to be the active
regular node for LQ2. And LQs's four new versions use different representation
methods from the original one, such as, different control types. That is to say,

A Mapping Repository for Meta-querier Evolution

Text-entry D's inputs cannot be transformed by an "equivalence" mapping to
Text-entry E, F, G, H's inputs. And thus, a new super node B is created and
connected by a mapping edge 62, which stores a new mapping expression, from
a global super node C.
This example assumes that all these FCs have the same semantics, and thus
each querier maintains a single semantic object. In each object node, version
edges are used to denote the development history.
To adapt to the local evolution, MQ1 wants to update its own global query
form by replacing its currently inaccurate FC (Text-entry I) with the FC (Text-
entry J). But this change does not affect the mappings from local queriers, and
thus a new regular node is added to super node C.
In the above discussion, several concepts and operations are introduced infor-
mally. To make these concepts effective and usable by developers, their semantics
are specified precisely in the following subsections.

Regular nodes Regular nodes represent the most fundamental components
FCs in the Mapping Repository.

Definition 1. A regular node represents a FC in a four-tuple {FCContent, URL, BlocklD, LS},
where FCContent refers to this FC's associated attributes, URL and BlocklD re-
spectively identify the query form and the semantic block that this FC is located,
and LS is the life span, i.e., the active period of this FC.
Regular nodes can be identified by a triple, { URL, LS, BlocklD}. First, dif-
ferent queriers could contain syntactically equivalent FCs, but these FCs are
represented by different regular nodes in Mapping Repository. In order to rec-
ognize which querier the regular node belongs to, the querier's unique identifier,
URL, is necessary. Second, a querier often develop multiple versions over time.
Normally, the changes of a querier only influence a small part of its FCs, thus
syntactically equivalent FCs possibly appear in multiple versions of a querier. To
avoid storing repetitive regular nodes, Mapping Repository employs time spans,
instead of a time point, as the version identifier of a FC. Third, it is observed
that syntactically equivalent FCs does not co-exist in the same semantic block[6].
And it is possible that two syntactically equivalent FCs with different purposes
exist in two different semantic blocks of a query form. Thus, BlockID is also
included. In sum, a specific FC can be fetched from Mapping Repository by its
triple identifier, and we also can use this triple to obtain a set of its source forms
(different versions of a query form) from Graph Repository.
Regular nodes can be further classified into local or global based on the source
of query forms.

Super Nodes and Mapping Edges The relations among FCs can be dis-
tinguished to three categories: 1)global-to-local: the mappings from global FCs
to local FCs; 2)local-to-local: the mappings among local FCs; 3)global-to-global:
the mappings among global FCs.

10 Xiao Li, Randy Chow

All these three types of mappings are required for the proposed framework
MEQE. In Mapping Repository, super nodes are intended for storing local-
to-local or global-to-global mappings whose relations are one-to-one semantic-
equivalent. Mapping edges represent the remaining ones.

Definition 2. A super node is an unordered collection of regular nodes sat-
isfying the following two requirements:
1) All regular nodes in this collection must be either all local or all global.
2) For each pair of distinct regular nodes, their FC inputs can be transformed
to each other by an 1'. d1' -. In mapping.
Depending on the constitution of super nodes, Mapping Repository distin-
guishes local super nodes from global super nodes. Additionally, since the inputs
to regular nodes in the same super node are totally same, they always share
the same mappings to other super nodes, i.e., the same correspondences and

Definition 3. A mapping edge is a directed edge from a source super-node
list NodeList1 to a destination super-node list NodeList2, where NodeListl and
NodeList2 are two ordered lists of super-nodes. Each mapping edge has three
attributes {ME, SimV, VS}. ME denotes a mapping expression specifying a
directional function operating on the inputs of NodeList1 to produce the inputs
of NodeList2. SimV and VS are two matrixes which respectively express the
similarities degree (called similarity matrix) and the status (called status matrix)
of each mapping denoted by this mapping edge.
The concept of ordered node list is proposed for supporting the storage of
n : m mappings. The node numbers of sources and destination lists depend on
the mapping cardinality, that is, 1:1, l:n, or n:m. ME is based on the specific
sequence of source and target nodes, that is, different orders of super nodes in
the source or destination list might produce different MEs.
Each mapping edge represents a set of mappings (called a mapping set)
1) Its mapping expression is the same as ME;
2) Its input consists of a list of regular nodes, each of which is chosen from
a unique source super node. The regular nodes in the output list are similarly
chosen from the destination super nodes.
3) All its inputs must come from the same query form. It is similarly true
for the outputs.
The third requirement is required; otherwise, the semantics of a global query
input is separated into multiple parts of different local query forms.
The validation state and similarity value of a mapping in such mapping set
are respectively stored in the corresponding element in VS and SimV matrixes.
Its status can be in one of the three states: human-validated, computer-validated
and not-validated (default). Its similarity value ranges from 0 (minimum) to 1.0.
Its initial value is equal to the similarity degrees with all the other regular nodes
in the super node it is clustered into. This value will fluctuate under the control
of self-validation mechanism (discussed in Section 5.2.1).

A Mapping Repository for Meta-querier Evolution

Digging a little deeper into the definitions of super nodes and mapping edges,
we can see an important consideration embedded within it: A super node bundles
the regular nodes into a single logical entity. This allows for coarser granularity
of presentation and explicit modeling of relationships among the super nodes. As
long as a regular node is assigned to a super node, all the pre-configured map-
pings associated with this super node are also applied on this regular node. Thus,
schema matching and merging can be transformed to reused-oriented strategies
in the context of Mapping Repository.
Definition 4. A mapping graph is a directional graph G =< V, E >, where
V is a set of all super nodes in Mapping Repository and E is a set of all mapping
edges in Mapping Repository.
A mapping graph contains complete mapping information to support map-
ping management operations such as insertion and modification of mappings
which might involve changes of expressions and similarity values.
For example, Figure 5 (a) illustrates a mapping graph and Figure 5 (b)
depicts the composition of the super node {H}. This mapping graph covers all

local I global

(a) (b)

Fig. 5. A mapping graph.

three mapping types: a global-to-global one (E A), five global-to-local ones
(e.g., A H), and various local-to-local one (e.g., G K and 1 2). There
exists a 3:2 complex mapping from the super-node list {A, B, C} to the super-
node list {F, G}. The global super node {D} is mapped to three different super
nodes, {H}, {I}, {J}.

Object Nodes and Version Edges Mapping Repository serves as not only
the repository for mapping retrieval in query translation, but also a knowledge
base for (semi-)automatic Schema Matcher and Schema Merger in the proposed
MEQE framework. Besides self and peer mapping information[20, 44], their de-
velopment history can also facilitates finding the proper current mapping. De-
velopment history denotes all the mapping changes and their sequences during
a period. Object nodes and version edges are employed to store development
As declared in the previous section, a mapping is defined as a triple (Expres-
sion, ListFCL, ListFCG). In Mapping Repository, a mapping is expressed by a

12 Xiao Li, Randy Chow

mapping edge and two regular node lists. We assume that changes on Expression
is for correcting the previous one, so its evolution history is ignored. The changes
on a mapping can be represented by the evolutions of its two regular node lists.

Definition 5. A version edge is a directed edge from a regular node list
NodeList1 to another regular node list NodeList2, where NodeList2 is the next
version of NodeListl for the same semantic object. A version graph is a direc-
tional graph G =< V, E >, where V is a set of all regular nodes in Mapping
Repository and E is a set of all version edges in Mapping Repository. An object
node is a maximum connected subgraph in a version graph.
A version edge reflects the order of two subsequent versions. The different
versions of a semantic object indicate the same semantics in different formats.
In a global or local object node, two regular nodes (versions) may differ in
their representation formats, (i.e., FC's attributes) such as control type, de-
scriptive texts and default values. For instances, there are two regular nodes,
Ni{FCContenti, URL1, BlocklD1, LS1} and N2{FCContent2, URL2, Block-
ID2, LS2} If they satisfy: FCContentl : FCContent2, URL1 URL2, LS1 E
LS2 and BlockID1 = BlockID2, and both N1 and N2's super nodes correspond
to the same global nodes, we regard them as the different versions of one object.
Each semantic object corresponds to a maximum connected subgraph (de-
noted as an object node) in a version graph. Thus, an object node stores all the
versions of a semantic object and its development orders. Since version orders are
based on the development history of FCs, each object node should be a tree-type
regular node graph (e.g., a version graph illustrated in Figure 6).

local global

',,------ ,, '

-- L r '

Fig. 6. A version graph.

Given the concepts described above, the critical issue is to automatically
identify which regular nodes belong to the same semantic object and which
nodes are newly inserted.
For global queriers, it is easy to capture the contents and orders of changes
since MEQE tracks the whole evolution process. But for local queriers, all we
have are two subsequent versions of a querier (i.e., two query forms) without any
knowledge of the steps that led from one form to another. Thus we need to find
the correspondent FCs between two versions automatically.

A Mapping Repository for Meta-querier Evolution

We observe that the semantics of a FC remain unchanged if it stays in the
same semantic block in different versions [6]. Furthermore, since synonym FCs
rarely co-exist in the same semantic block [21,40], each global FC corresponds
to no more than one local FC in a semantic block.
Based on these two observations, our algorithm can be built upon the fol-
lowing three assumptions:
1) The relations between semantic objects can be expressed by the orders
and locations of FCs in the FC graph.
2) If a FC changes its position but remains in the same semantic block, the
semantic of this FC will not be changed.
3) In the same semantic block of each form, two FCs do not correspond to
the same global FCs.
Therefore, a local object consists of a set of local regular nodes satisfying
three requirements:
1) Corresponding to the same global super node list;
2) In the same query form;
3) In the same semantic block.
Based on the above criteria, local objects can be built and maintained auto-
matically by Algorithm 1 sketched below. The algorithm is for inserting a regular
node LR to a proper object node ON in Mapping Repository R. Notice that the
algorithm is only for 1 : n mappings, but it is easy to extent it to n : m.

Input: Local regular node LR, Mapping Repository R
Output: Object node ON
BF = NodesInSameBlockForm(LR, R);
GN = GlobalDestination(LR);
foreach Local regular node i in BF do
Ni GlobalDestination(i);
if N, n GN 1 0 then
ON Object node of i;
return ON;
ON new ObjectNode();
return ON;
Algorithm 1: Insert a regular node to a proper object node in Mapping

Algorithm 1 is composed of three steps:
1) Searching the whole Mapping Repository (R) to obtain all the regular
nodes (BF) in the same block and query form with LR.

14 Xiao Li, Randy Chow

2) Finding out all the possible global super nodes lists (GN) which can reach
LR through different directed mapping edges.
3) Comparing GN with all possible global super nodes lists (Ni) of each
regular node i in BF. If there exists the same global super node list, then the
object node of regular node i is the target ON. If there is no such regular node
i, a new local object node will be built to contain LR.

Mapping Repository integrates the version graph with the mapping graph.
Its basic building blocks are regular nodes, mapping edges and version edges.
Object nodes and super nodes are employed for coarser granularity based on
different purposes, version and mapping management.

5.2 Advanced Features

In order to accommodate the dynamic of web environment and the complexity
of query forms, Mapping Repository needs to provide advanced features such as
self-validation, self-configuration, self-maintenance, and virtualization.

Self-validation In meta-queriers, the mappings from global forms to local forms
are precisely engineered, and thus have one special property: high similarity
with a value close to 1.0. If human experts insert or modify a regular node,
the repository automatically assumes that experts have validated the mappings
related to this regular node, i.e., their validation statuses will be set to "human-
v IlI, I l I with the highest similarity values (1.0).
For the mappings derived by Schema Matcher (computers), their initial sim-
ilarity values depend on the algorithm implemented in the matcher. But these
values are not statistic. As time passes, the meta-queriers continuously examine
the usage of mappings, and adjust the similarity values based on the validation of
mappings. We propose the following similarity function to obtain the similarity
value of a mapping at the time point t2:

1 if E(n, m) + SimVtl > 1;
SimVt2 0 if E(n, m) + SimVtl < 0;
E(n,m) + SimVtl otherwise.

where n and m are the numbers of correct execution and improper execu-
tion between the time points tl and t2, respectively, SimVtl is the similarity
value at the time point t1, and E(n, m) is equal to the degrees of changes be-
tween tl and t2. In the initial implementation, we adapt the following definition:
E(n, m) a x n + b x m(b > a), where a and b are constants. The penalties
of improper executions are larger than the benefits from possibly proper execu-
tions, and thus the value of b is greater than a.

To address the features of self-configuration and self- maintenance, we need
to introduce the notion of mapping transitivity as illustrated in Figure 7.

A Mapping Repository for Meta-querier Evolution

Definition 6. Transitive: for any three super node lists A, B and C, if there
exist two mapping edges mi (from B to A) and m2 (from C to B), another new
mapping edge ms (from C to A) can be derived by combining mi and m2 in
order, denoted by mi m2.
Mapping composition is a challenging problem. Research results [10, 18, 33]
state that not all the mappings can be composed, that is, the intermediate states
cannot be circumvented. For instance, in Figure 7 (a), C has to be (partly)
converted to the B's format and then to the A's.



M4 (b)

(c) (d)

Fig. 7. Mapping transitivity.

In the current Mapping Repository, we do not compose the mappings but
combine them in order to predict the new one, i.e., going through every inter-
mediate state in the mapping path. Its similarity value can be computed by
multiplying the similarity values of all the mappings involved in the mapping

Self-configuration Self-configuration is an automatic mechanism to generate
the unknown global-to-local mappings for each local super node. By exploiting
mapping transitivity, it calculates the similarity values for all the mapping paths.
We can obtain a new mapping by combining the mappings through the mapping
path with the highest similarity value.
For instance in the Figure 7 (b), there exist n mapping paths from C to A.
The path C B1 A consists of the mapping m2 from C to B1 and mn from
B1 to A. Its similarity value is equal to Sim[C, B1] x Sim[Bi, A]. Suppose this
path has the highest similarity value among all the paths, the new mapping ms
can be derived by the combination of mi and m2, i.e., mi m2.

Self-maintenance Self-maintenance is to facilitate efficient restructuring of
Mapping Repository upon changes, such as deletion, modification and insertion

16 Xiao Li, Randy Chow

of edges and nodes. We demonstrate three cases common in the evolution of
1) Removing a mapping edge: Improper mapping edges can be removed au-
tomatically or manually. As illustrated in the Figure 7 (b), if the mapping edge
ms is removed, Mapping Repository will attempt to find a new mapping edge
from the mapping paths from C to A, just like self-configuration.
2) Updating derived mappings: C'l i_, -. of any mapping edge will trigger the
update of its derived mappings. If this edge is deleted and cannot be replaced
by the combination of other edges, the derived mapping edges are also removed;
otherwise, they should be updated to the new ones. For instance in the Figure
7 (c), the mapping ms derived by mi and m2 is updated to m5 due to the fact
that m2 is changed to m4, where m5 = mi m4.
3) Replacing a global super node: In meta-queriers, human experts some-
times need to re-organize the global query form. Some global super nodes will
be replaced. It is a laborious, time-consuming and error-prone task to update
numerous mappings between the replaced global super nodes and local super
nodes. As long as experts specify the mapping from the new global super nodes
to the replaced, Mapping Repository can update all the related global-to-local
mappings automatically. For instance in the Figure 7 (d), given the mapping m2
from the new super node C to the replaced node B, a new mapping ms from C
to A can be derived by combining m2 and mi in order.

Virtualization Mapping Repository can be shared by many customized meta-
queriers, although they are built on different collections of local query engines.
A view is a virtual or logical mapping repository. Each view is a sub-graph of
the whole mapping and version graph. Each meta-querier or local querier owns
its own view, which includes all the related nodes and edges.
Views hide the complexity of graphs and make it easier to maintain the
mapping repository for human editors. Views can be categorized to updatable
or read-only views to increase the security of Mapping Repository.

6 System structure

Mapping Repository is implemented above the open-source system Alignmen-
tAPI[4]. Our proposed query form model is transformed to ontology format in
OWL. Each query form can be regarded as an instance of this form ontology.
Our system is organized in two layers, as shown in Figure 8. Layer 1 (storage
layer) is responsible for physical storage of mapping and version graphs. Layer
2 (logic layer) implements all the logics of mapping repository discussed in the
previous sections.
Storage layer stores mapping and version graphs in tables. It also provides the
storageAPI to translate the graph operations to the operations on the relational
schema, such as, seeking all the reachable super nodes from a super node N.
Roughly, logic layer can be divided into three levels. Each level offers a functional
interface the operations of which may be used by higher levels. Level 1 (node

A Mapping Repository for Meta-querier Evolution

(Schema Matching) (Mapping Retrieval (Human Interaction) ( Schema Merging)

Applications I I

Repository API
system level views complex operations Logic
(structure level mapping graphs version graphs)
(node level regular nodes )
Storage API
Relational Database) Storage

Fig. 8. The system structure.

level) implements the concept of regular nodes. Level 2 (structure level) and 3
(system level) implement the mechanisms for mapping management and version
Node level offers the basic operations on regular nodes. Regular nodes can
be viewed, inserted, edited or deleted. At this level, no relation between regular
nodes is considered. All the regular nodes are operated separately.
Structure level manages mapping graphs and version graphs. Super and ob-
ject nodes, mapping and version edges are defined and implemented in this level.
There are three kinds of operations: 1) Creation and deletion of super nodes and
object nodes; 2) Manipulation of mapping graphs and version graphs (Creation
and deletion of edges, modification of mapping edge's ME, mapping edge shift
between two super nodes, and the version edge shift between two regular nodes);
3) Search and comparison of graph elements including regular nodes, mapping
edges, version edges, super nodes, and object nodes.
In system level, the support of four advanced features (described in Section
5.2) is implemented, e.g., providing a mechanism to define customized views to
adapt customized meta-queriers. This level also provides some common com-
plex operations, a sequence of basic operations, such as, combining the mapping
expressions through a mapping path.
Mapping Repository can be employed by different applications such as map-
ping retrieval, schema matching and schema merging. In the next section, two
repository-based applications, schema matching and merging, are discussed in

7 Discussion

7.1 Design Issues
Mapping repositories function as knowledge bases for storing mapping informa-
tion. There are many dimensions by which mapping repositories can be classified.
Mapping repositories have the following four properties:

18 Xiao Li, Randy Chow

Single-domain v.s. cross-domain: whether mapping repositories can be from
the queriers in the same domain or multiple domains. The representation meth-
ods in the same domain are limited. For example, the vocabulary used in query
forms tends to converge at a relatively small size[ll]. Thus, the previous map-
pings in the same domain are very useful for (semi-)automatic construction and
maintenance of meta-queriers.
The current version of Mapping Repository mainly focuses on the storage of
mappings in a single domain. In the future, Mapping Repository needs to con-
sider more issues for supporting the meta-queriers that cross multiple domains,
such as, dealing with the contradicted and common mappings from different
Dynamic v.s. static: whether mapping evolution can be supported or not.
Mapping is inherently dynamic and abstract perception of relation between local
and global queriers. The proposed mapping repository can modify or add the
corresponding contents based on the changes in the local and global queriers.
Single-version v.s. multi-version: whether the previous versions are stored
or not. Mapping information is active and valid at different time intervals. Al-
though the previous versions become invalid, they are very precious resources for
understanding the current valid version of local queriers. The proposed Mapping
Repository stores all the versions of mappings.
Single-meta-querier v.s. multi-meta-querier: whether mapping repositories
only serve a specific meta-querier or multiple ones. A mapping repository ought
to be designed for the sharing by multiple meta-queriers if they have many com-
mon mapping information. Compared with the storage space used by multiple
individual mapping repositories, the space for the shared repositories is much
smaller. And the shared repository also prevents update anomalies. Multiple
customized meta-queriers can be simultaneously maintained by modifying the
common part in the mapping repository. Lastly, the more mapping information a
mapping repository owns, the better performance (semi-)automatic construction
and maintenance of meta-queriers have. Thus, this property is also supported
by the proposed Mapping Repository.

7.2 Applications

The designers of meta-queriers can implement their own operations on Map-
ping Repository based on the specific application (domain). For example, 1) a
repository-based mapping retrieval to utilize the storage of mappings to fetch
the required mappings for query translation; 2) supporting human interaction
for repository-based mapping validation and modification. In the following, two
application examples that use Mapping Repository are presented in details.

Repository-based schema matching is a reuse-oriented matching strategy.
It aims to find a global FC correspondence (FCG) and a mapping expression for
each pending local FC (FCL). It leverages the evolution history of self and peer
mappings to learn new mappings. This process is composed of three steps:

A Mapping Repository for Meta-querier Evolution

1) Scanning Mapping Repository to find the most similar (higher than a
threshold) local super node (SNL) for each pending FCL. Each super node is
indeed a cluster of highly similar regular nodes, and thus this search problem is
equivalent to the incremental clustering algorithms which is widely studied[13,
42]. And the similarity degrees between regular nodes are determined by various
associated information, such as their FC attributes, development history, and
their locations and adjacencies in FC graphs. These values can be calculated by
most existing schema or ontology matching strategies[17, 35], since the relations
between FCL and the FCs stored in SNL are totally equivalent (i.e., ") rather
than other mapping expressions.
2) Finding the most suitable mapping path (MP) which is from global su-
per nodes to the SNL (found in step 1). The possible paths from global super
nodes can be obtained by enumerating all the paths to SNL. Then the similarity
value of each path is calculated by exploiting mapping transitivity, just like the
example in Section 5.2.2. The path with the highest value is finally selected as
3) Choosing a FCG from the super node SNG which is the source of MP
(found in step 2). In most cases, the selection should best-effort maintain un-
changed with the subsequent version of the global interface. The other principles
of selection depend on the specific application. In addition, the corresponding
expression can be achieved by combining all the mappings involved in the MP.
As shown in Figure 7 (b), consider A and C are respectively a local super node
and a global super one. If A is the SNL and the path C -B B1 A is MP, C
is the SNG with a mapping expression ms.
Repository-based schema matching focuses on reusing previous matching re-
sults of self and peers. For providing better match candidates, this approach
should be combined with other schema-based or instance-based schema match-
ing techniques[17, 35].

Repository-based schema merging is a reused-oriented merging strategy. It
aims to integrate multiple local query forms into a unified global query form
by using the previous validated self and peer mappings. Local and global query
forms are made up of FCs, which are individually represented by a regular node.
The primary objective of this approach is to find a set of global regular nodes
(RNG) that can be mapped to all (or the most significant) local regular nodes
(RNL) of local query forms.
Since each RNL corresponds to a local super node (SNL), it is not hard to
obtain a set of global super nodes (SNSetc) which can reach each SNL through
directed mapping edges. The principles of choosing a set of RNG from all the
SNSetG are listed in the order of priority:
1) RNG should not conflict with each other, such as naming and structural
2) RNG should maximally satisfy the constraints of all the RNL.
3) The number of chosen RNG should be minimized.

20 Xiao Li, Randy Chow

In essence, it is a NP-hard optimization problem. Many approximation or
heuristic algorithms have been proposed in the previous research results[16, 34,

8 Conclusion and future work

In this paper, we propose a framework named MEQE to tackle a unique and im-
portant topic, meta-querier evolution. In the center of MEQE, Mapping Repos-
itory functions as a mapping storage for query translation and serves as a
domain-based knowledge base for Schema Matcher and Merger. Mapping Repos-
itory employs an innovative conceptual graph model which integrates both map-
ping and version management. In addition, it supports four advanced features:
self-validation, self-configuration, self-maintenance, and virtualization. Although
Mapping Repository is primarily designed for the meta-querier evolution, it
should be useful to other applications in the field of information integration.
Currently, we have implemented the basic Mapping Repository. Future work
in this direction includes:
1) Modification and improvement of Mapping Repository to support cross-
domain meta-queriers.
2) Implementation of other components of MEQE, such as Schema Matcher
and Merger.
3) Collection and analysis of the evolutions of real query forms.


4. Alignment API and Alignment Server.
5. W3C's HTML specification.
6. The UIUC web integration repository (TEL-8 query interfaces). Com-
puter Science Department, University of Illinois at Urbana-Champaign., 2003.
7. S. Bergamaschi, S. Castano, and M. Vincini. Semantic integration of semistructured
and structured data sources. SIGMOD Rec., 28(1):5459, 1999.
8. M. K. Bergman. The Deep Web: Surfacing hidden value. Journal of Electronic Pub-
lishing, 7(1), August 2001.
9. P. A. Bernstein. Applying model management to classical meta data problems. In
CIDR, 2003.
10. P. A. Bernstein, T. J. Green, S. Melnik, and A. Nash. Implementing mapping
composition. The VLDB Journal, 17(2):333353, 2008.
11. K. C.-C. ('l. I.. B. He, C. Li, M. Patel, and Z. Zhang. Structured databases on
the web: Observations and implications. SIGMOD Rec., 33(3):6170, 2004.
12. K. C.-C. C'l .... B. He, and Z. Zhang. Toward large scale integration: Building a
metaquerier over databases on the web. In CIDR, 2005.
13. M. Charikar, C. Chekuri, T. Feder, and R. Motwani. Incremental clustering and
dynamic information retrieval. SIAM J. Comput., 33(6):14171440, 2004.

A Mapping Repository for Meta-querier Evolution

14. S.-L. ('Cl,.... K. C.-C. Cl' .... and C. Zhai. Context-aware wrapping: Synchro-
nized data extraction. In VLDB, 2007.
15. H.-H. Do. Schema Matching and Mapping-based Data Integration. Ph. D. disser-
tation, Department of Computer Science, University of Leipzig, Germany, 2006.
16. E. Dragut, W. Wu, P. Sistla, C. Yu, and W. Meng. Merging source query interfaces
on web databases. In ICDE, 2006.
17. J. Euzenat and P. Shvaiko. Ontology Matching. Springer-Verlag New York, Inc.,
18. R. Fagin, P. G. Kolaitis, L. Popa, and W.-C. Tan. Composing schema map-
pings: Second-order dependencies to the rescue. ACM Trans. Database Syst.,
30(4):9941055, 2005.
19. B. He and K. C.-C. Chang. Statistical schema matching across web query interfaces.
In SIGMOD, 2003.
20. B. He and K. C.-C. Chang. A holistic paradigm for large scale schema matching.
SIGMOD Rec., 33(4):2025, Dec. 2004.
21. B. He, K. C.-C. C(' ...-i and J. Han. Discovering complex matching across web
query interfaces: A correlation mining approach. In KDD, 2004.
22. B. He, T. Tao, and K. C.-C. Chang. Organizing structured web sources by query
schemas: a clustering approach. In CIKM, 2004.
23. H. He, W. Meng, Y. Lu, C. Yu, and Z. Wu. Towards deeper understanding of the
search interfaces of the deep web. World Wide Web, 10(2):133155, Jun. 2007.
24. H. He, W. Meng, C. Yu, and Z. Wu. Wise-integrator: An automatic integrator of
web search interfaces for e-commerce. In VLDB, 2003.
25. G. Kabra, C. Li, and K. C.-C. Chang. Query routing: Finding ways in the maze
of the deep web. In ICDE-WIRI, 2005.
26. Y. Lu, H. He, Q. Peng, W. Meng, and C. Yu. Clustering e-commerce search en-
gines based on their search interface pages using wise-cluster. Data Knowl. Eng.,
59(2):231246, Nov. 2006.
27. Y. Lu, H. He, H. Zhao, W. Meng, and C. Yu. Annotating structured data of the
deep web. In ICDE, 2007.
28. Y. Lu, W. Meng, C. Y. L. Shu, and K. L. Liu. Evaluation of result merging strate-
gies for metasearch engines. In WISE, 2005.
29. Y. Lu, Z. Wu, H. Zhao, W. Meng, K. Liu, V. Raghavan, and C. T. Yu. My-
searchview: a customized metasearch engine generator. In SIGMOD, 2007.
30. R. McCann, B. A1Shebli, Q. Le, H. Nguyen, L. Vu, and A. Doan. Mapping main-
tenance for data integration systems. In VLDB, 2005.
31. S. Melnik, E. Rahm, and P. A. Bernstein. Rondo: a programming platform for
generic model management. In SIGMOD, 2003.
32. P. Mitra, G. Wiederhold, and M. L. Kersten. A graph-oriented model for articula-
tion of ontology interdependencies. In EDBT, 2000.
33. A. Nash, P. A. Bernstein, and S. Melnik. Composition of mappings given by em-
bedded dependencies. ACM Trans. Database Syst., 32(1):4, 2007.
34. R. Pottinger and P. A. Bernstein. Schema merging and mapping creation for rela-
tional sources. In EDBT, 2008.
35. E. Rahm and P. A. Bernstein. A survey of approaches to automatic schema match-
ing. The VLDB Journal, 10(4):334350, Dec. 2001.
36. D. Shestakov, S. S. Bhowmick, and E.-P. Lim. Deque: querying the deep web. Data
Knowl. Eng., 52(3):273311, 2005.
37. W. Su, J. Wang, Q. Huang, and F. Lochovsky. Query result ranking over e-
commerce web databases. In CIKM, 2006.

22 Xiao Li, Randy Chow

38. W. Su, J. Wang, and F. Lochovsky. Automatic hierarchical classification of struc-
tured deep web databases. In WISE, 2006.
39. A. Thor and E. Rahm. Moma a mapping-based object matching system. In CIDR,
40. F. L. W. Su, J. Wang. Holistic schema matching for web query interface. In EDBT,
41. J. Wang, F. J. Wen, Lochovsky, and W.-Y. Ma. Instance-based schema matching
for web databases by domain-specific query probing. In VLDB, 2004.
42. D. H. W,.1. -..1.... T. R. Ioerger, and J. Yen. An incremental approach to building
a cluster hierarchy. In ICDM, 2002.
43. W. Wu, A. Doan, and C. Yu. Webiq: Learning from the web to match deep-web
query interfaces. In ICDE, 2006.
44. W. Wu, C. Yu, A. Doan, and W. Meng. An interactive clustering-based approach
to integrating source query interfaces on the deep web. In SIGMOD, 2004.
45. W. Wu, A. Doan, and C. Yu. Merging interface schemas on the deep web via
clustering aggregation. In ICDM, 2005.
46. Z. Zhang, B. He, and K. C.-C. Chang. On-the-fly constraint mapping across web
query interfaces. In VLDB-IIWeb, 2004.
47. Z. Zhang, B. He, and K. C.-C. Chang. Understanding web query interfaces: Best-
effort parsing with hidden syntax. In SIGMOD, 2004.
48. H. Zhao, W. Meng, Z. Wu, V. Raghavan, and C. Yu. Fully automatic wrapper
generation for search engines. In WWW, 2005.
49. A. Zhdanova and P. Shvaiko. Community-driven ontology matching. In ESWC,

University of Florida Home Page
© 2004 - 2010 University of Florida George A. Smathers Libraries.
All rights reserved.

Acceptable Use, Copyright, and Disclaimer Statement
Last updated October 10, 2010 - - mvs