Web Services Discovery and Recommendation Based on Information Extraction and Symbolic Reputation

International Journal on Web Service Computing (IJWSC), Vol.4, No.1, March 2013

WEB SERVICES DISCOVERY AND RECOMMENDATION
BASED ON INFORMATION EXTRACTION AND SYMBOLIC
REPUTATION

Mustapha AZNAG1, Mohamed QUAFAFOU1, Nicolas DURAND1 and Zahi
JARIR2
1
Laboratory of Information and Systems Sciences, Aix-Marseille University, France
{mustapha.aznag,mohamed.quafafou,nicolas.durand}@univ-amu.fr
2
LISI Laboratory, FSSM, University of Cadi Ayyad, Morocco
zahijarir@ucam.ac.ma

ABSTRACT
This paper shows that the problem of web services representation is crucial and analyzes the various
factors that influence on it. It presents the traditional representation of web services considering traditional
textual descriptions based on the information contained in WSDL files. Unfortunately, textual web services
descriptions are dirty and need significant cleaning to keep only useful information. To deal with this
problem, we introduce rules based text tagging method, which allows filtering web service description to
keep only significant information. A new representation based on such filtered data is then introduced.
Many web services have empty descriptions. Also, we consider web services representations based on the
WSDL file structure (types, attributes, etc.). Alternatively, we introduce a new representation called
symbolic reputation, which is computed from relationships between web services. The impact of the use of
these representations on web service discovery and recommendation is studied and discussed in the
experimentation using real world web services.

KEYWORDS
Web services, WSDL file, data representation, information extraction, semantic tagging, symbolic
reputation, discovery and recommendation.

1. INTRODUCTION

Web services1 [1] are defined as a software systems designed to support interoperable machine-
to-machine interaction over a network. They are loosely coupled reusable software components
that encapsulate discrete functionality and are distributed and programmatically accessible over
the Internet. They are self contain, modular business applications that have open, internet-
oriented, standards based interfaces [7]. Different tasks like matching, ranking, discovery and
composition have been intensively studied to improve the general web services management
process. Thus, web services community has proposed different approaches and methods to deal
with these tasks. Empirical evaluations are generally proposed considering different simulation
scenarios. Nowadays, we are moving from web of data to web of services as the number of UDDI
Business Registries (URBs) is increasing. Moreover, the number of host that offer available web
services, through specific engines like Axis2, is significantly increasing. Consequently, the WSDL
files describing services can be crawled and parsed automatically. Such files contain different

1
http://www.w3.org/standards/webofservices/
2
http://axis.apache.org/axis/
DOI : 10.5121/ijwsc.2013.4101 1


kind of information like textual descriptions, simple and/or complex types, attributes, etc. (see
Section 3 for more details).Contrary to simulation based approaches, real world services inherit
the complexity of the real world. With the increasing number of published Web services
providing similar functionalities, it’s very tedious for a service consumer to make decision to
select the appropriate one according to her/his needs. Therefore mechanisms and techniques are
required to help consumers to make the best choice. In addition, web services use natural
language descriptions (ambiguity of sentences), they are multi-languages and cross-domains. In
this context, the representation of web service becomes a major challenge. Moreover, the quality
of representations has an important impact on services registries that provide mechanisms for web
services discovery, composition and recommendation. These tasks have been intensively studied
in the literature [4, 16, 17, 24, 9, 2, 12]

This paper studies different representations of web services and compares them considering the
following important tasks: discovery and recommendation. The first representation is centred on
textual descriptions of services and their available functions. It is produced from the web service
descriptions and enriched by integrating the descriptions of operations offered by services and the
structural elements of WSDL, for instance simple and/or complex types, attributes. The previous
representation is traditional and well known by the community of Web services. We consider this
representation as a baseline representation(B)for a web service. We introduce two advanced
representations that refine B considering semantic aspect of text and we show their interest inthe
experiments (see Section 7):

• Rules based text tagging (RBTT): a lot of services have very detailed descriptions,
especially when they offer several operations with their own descriptions. The main
question is how to recognize significant parts or entities in the text description and how to
use the filtered information to describe the web service?

All the previous representations are produced from the web service description, i.e, WSDL
content, which represents the provider point of view. Our goal is to infer a web service
representation from the description of its context (neighbour services).

• Symbolic reputation (SR): web services reputation is generally a numeric quantity
computed from user feedback [17]. The representation we consider here is qualitative as
it is produced from the symbolic descriptions of other web services. Consequently, such a
description describes the relationships of a service with others and does not describe the
service itself.

We introduce in this paper methods that compute the different representations to evaluate their
interest for discovery and recommendation of web services. We also describe a general
architecture of a System developed and used in the experimentation.
This paper is organized as follows. Section 2 provides an overview of related work. Section 3
details the WSDL parsing and the information extraction. Section 4 presents the different
representations of web services. The new proposed representations web services (RBTT and SR)
are discussed respectively in Section 4.2 and 4.3. Section 5 presents the general architecture of
our system. Section 6 is dedicated to the web service discovery and recommendation system.
Finally the experimental results are discussed in Section 7 before concluding.

2. RELATED WORK

In this section, we briefly discuss some of the research works related to discovering Web services.
Although various approaches can be used to locate and discover Web services on the web, we
have focused our research on the service reputation and discovery problems. Every web service
2


associates with a WSDL document that contains the description of the service. A lot of research
efforts have been devoted in utilizing WSDL documents [16, 28, 29, 25, 12, 26].Dong et al. [16]
proposed the Web services search engine Woogle that is capable of providing Web services
similarity search. However, their engine does not adequately consider data types, which usually
reveal important information about the functionalities of Web services [27]. Liu and Wong [25]
apply text mining techniques to extract features such as service content, context, host name, and
name, from Web service description files in order to cluster Web services. They proposed an
integrated feature mining and clustering approach for Web services as a predecessor to discovery,
hoping to help in building a search engine to crawl and cluster non-semantic Web services.
Elgazzar et al. [12] proposed a similar approach, which clusters WSDL documents to improve the
non-semantic web service discovery. They take the elements in WSDL documents as their
feature, and cluster web services into functionality based clusters. The clustering results can be
used to improve the quality of web service search results.

In [30], the authors proposed an architecture for Web services filtering and clustering. The service
filtering mechanism is based on user and application profiles that are described using OWL-S
(Web Ontology Language for Services). The effectiveness of the filters is based on a clustering
analysis that compares services related clusters. The objectives of this matchmaking process are
to save execution time and to improve the refinement of the stored data. Another similar approach
[31] concentrates on Web service discovery with OWL-S. The OWLS is first combined with
WSDL to represent service semantics before using a clustering algorithm to group the collections
of heterogeneous services together. Finally, a user query is matched against the clusters, in order
to return the suitable services. Nevertheless, the creation and maintenance of ontologies may be
difficult and involve a huge amount of human effort [28, 29].

Web service reputation represents a mechanism relying on feedbacks provided by
consumers/software agent to measure Web service trustworthiness. It is modelled as a vector of
aggregate consumers/software agent’s ratings for a web service. Also various rating feedbacks are
aggregated to derive a service provider’s reputation. Generally feedback is composed from
collected quality of service (QoS) information acquired from execution monitoring and those that
require consumer’s intervention like price or accuracy that cannot be monitored. According to the
QoS information published and a consumer preference including required QoS metrics, the QoS
registry will calculate an overall rating for each web service that matches the consumer’s search
request. Then the consumer will select the web service with the highest rating [17]. Most of
proposed reputation approaches use a central registry to collect and share consumer’s feedbacks.
Since this central architecture is subject of failure, other works based on peer-to-peer web
services are proposed to deal with decentralized reputation mechanism [32]. A service provider
that provides satisfactory service may get incorrect or false ratings from unfair or malicious
raters. One of challenging problem is protecting web service reputation from these incorrect
inputs. Some mechanisms have been proposed to detected and deal with dishonest feedbacks by
using dedicated monitoring agents to filter consumers evaluation [33], or collaborative filtering
techniques based on peer to peer solution [34]. In the literature, a rating of a service is a vector of
attribute values. The computed reputation rating may be a binary value (trusted or untrusted), a
scaled integer (e.g. 1-10), or on a continuous scale (e.g., [0, 1]). Therefore the satisfaction level of
web services is generally a normalized numerical value, representing quantitative reputation, used
for dynamic services ranking and selection.

Maximilien and Singh [3, 35] proposed a multi-agent based architecture where agents assist in
quality-based service selection using an agency to disseminate reputation and endorsement
information. Each proxy agent is autonomous but also collaborates with other agents to collect
other opinions and therefore maximizes its information to improve its decision-making. Liu, Ngu
and Zeng [36] proposed an algorithm about how to combine different QoS metrics to get a fair
overall rating for a web service. The proposed reputation can be defined as the average ranking
3


given to the service by end users. Majithia et al. [5] consider ratings of services in different
contexts and a coefficient (weight) is attached to each particular context. This coefficient reflects
its importance to a particular set of users. Based on this coefficient, they propose a method to
compute the reputation score as weighted sum of ratings for a service. Wish art et al. [37]
introduce an aging factor for the reputation score, which is applied to each of the ratings for a
service. Based on these two works, Xu et al. [17] propose to calculate reputation score as the
weighted average of all ratings of a service received from consumers, including an inclusion
factor representing the weight attached to each of the ratings for the service. Trust and reputation
mechanisms are closely related. Web service reputation can be considered as an aggregation of
evaluation for a service from consumers/software agent, while web service trust represents a
personalized and subjective opinion reflecting a web service [38]. Currently, several researches in
the area of trust and reputation topic are considered. We cite for instance the work [33] that
proposes an algorithm to analyse the trustworthiness for each consumer, which ultimately
facilitates the web services selection process considering feedbacks reported by trusted users than
others.

Web service community has been introduced to easy-of-use web service discovery. It can be
viewed as a mediator that holds the meta-data and registry information about its member services
and represents domain-specific knowledge [39]. Among Community-based approach, we present
for instance the work [34] that has proposed a community-based service selection approach based
on super agents. These agents share their information about services they have interacted with,
which is useful for other agents to make effective selection of services. This in order to maintain
communities and build community-based reputation for a service based on the opinions from all
community members that have similar interests and judgement criteria.

These efforts have an implicit hypothesis, which considers that web service reputation is a
numerical quantity. As far as we know, it’s the first time we break the quantitative reputation
hypothesis. This paper avoids this hypothesis considering the qualitative aspect of the reputation
and describes how to compute and use it.

3. WSDL DOCUMENT PARSING AND INFORMATION EXTRACTION

The Web Service Description Language (WSDL) is an XML-based language, designed according
to standards specified by the W3C that provides a model for describing web services. It describes
one or more services as collections of network endpoints, or ports. It provides the specifications
necessary to use the web service by describing the communication protocol, the message format
required to communicate with the service, the operations that the client can invoke and the service
location. Two versions of WSDL recommendation exist: the 1.13version, which is used in almost
all existing systems, and the 2.04 version which is intended to replace 1.1. These two versions are
functionally quite similar but have substantial differences in XML structure.

To manage efficiently web services descriptions, we extract all features that describe a web
service from the WSDL document and store them into a relational database. We recognize both
WSDL versions (1.1 and 2.0). During this process, we proceed in three steps (see Figure 1). The
first step consists of checking availability of web service and validating the content of WSDL
document. The second step is to get the WSDL document and read it directly from the WSDLURI
to extract all information of the document. In this step we describe the features to extract from the
WSDL document: (1) the name, the documentation and the version of the WSDL, (2) WSDL
types used by messages to transmit information between web services. Data types are often

3
http://www.w3.org/TR/wsdl
4
http://www.w3.org/TR/wsdl20/
4


specified using a XML Schema Definition (XSD). We extract all kind of elements and types that
can be as simple or complex types as a set of elements and/or attributes, and (3)a set of services
declared in the WSDL document. For each service we extract the name, the documentation and a
set of endpoints. Then, for each endpoint we extract the name, the address (which defines the
connection point to web service. It is typically represented by a simple HTTP URL) and the
binding(Name, Type, Style, Transport protocol). The binding specifies the interface as well as
defining the SOAP binding style (RPC/Document) and transport SOAP protocol. The interface
defines the operations to be performed for a web service, and the messages that are used to
perform the operation. We also extract for each operation the name, the documentation, the input
and the output parameters. The input/output parameters can be referred to the previously
extracted types/elements. Finally, the third step is dedicated to save the extracted WSDL features
into a database. The extracted information will be used during the generation of
representations(B, RBTT and SR). Before presenting the methods for calculating representations,
we discuss some text-processing standard used thereafter.

1. Tag removal: This step removes all HTML tags, CSS components, symbols(punctuation,
etc.).
2. Splitting and remove a stop words: Some terms are composed by several words
separated by a capital letter; we use therefore regular expression to extract these words.
To illustrate, the application of the regular expression ([A-Z][a-z]+) on this string "GetAll
Country Currencies Response" produces’ Get’, ’All’, ’Country’, ’Currencies’ and
’Response’. Furthermore, to extract the potential content words, we remove all the stop
words. Finally, the potential content words for the previous example are ’Country’ and
’Currencies’.
3. Word Stemming: In this step we use the Porter Stemmer [21] to remove words, which
have the same stem. Words with the same stem will usually have the same meaning. For
example, ’computer’, ’computing’ and ’compute’ have the stem ’comput’.

Figure 1. WSDL document parsing and information extraction

5


4. REPRESENTATIONS OF WEB SERVICES

In this section, we present a generation method of the traditional representation of web service
and we introduce two new representations. Let us note that the generated representations are
vectorial.

4.1. Baseline representation (B)

A web service can be described by a textual description extracted from WSDL document or given
by its provider when publishing in the UDDI. The current UDDI registry only allows searching
web services by their textual description. The first representation is centred on textual
descriptions of services and their offered functions. This is produced from the web service
descriptions and enriched by integrating the descriptions of operations offered by services. Let us
remark that the major disadvantage is that most web services have a poor or an empty description.
To complete this representation, we added all information described by WSDL types. The types
are used by messages to transmit information between web services. Consequently, WSDL types
are good features to describe the functionality of a service and are the most informative element
in WSDL document. For this reason, we extract all type names (elements, complex types, simple
types, attributes, enumerations) and we apply the textual processing (see Section 3 -steps 2 and 3)
to produce a set of words. Thus, we use the obtained set of words to construct the new
representation and we consider it as a baseline representation (B) for a web service.

4.2. Rules Based Text Tagging of web services descriptions (RBTT)

A lot of services have very detailed descriptions, especially when they offer several operations
with their own descriptions. The main question is how to recognize significant parts or entities in
the text description and how to use the filtered information to describe the web service? Our
approach consists in the definition of extraction rules to identify, extract and annotate relevant
multi-word terms from web service descriptions. This approach has been already used for
biological data [19]. The processing steps(Tokenization, Part-Of-Speech tagging, Extraction and
output generation) of Information Extraction have been implemented as modules using the
Lingua Stream platform5. Lingua Stream [20] is an integrated experimentation environment
targeted to researchers in natural language processing (NLP). Let us note that we used Tree
Tagger6 for the Part-Of-Speech tagging step. The extracted information is given in a form of a
XML file. In the context of the web service descriptions, we have defined a set of rules into
Prolog reflecting the Definite Clause Grammar (DCG) to recognize 3types of information: web
service names (namews), purposes of the web services(purpose), and the domain of utilization
(domain). Let us remark that we do not use patterns in the sense of Information Extraction that is
without a prior on the form of the expressions.

Figure 2. Structure of rules

Figure 2 presents the structure of the rules. From a ’context’, an expression(generally a multi-
word term, a nominal phrase) is recognized until a stop phrase is encountered. The context is a set

5
http://www.linguastream.org/
6
http://www.ims.uni-stuttgart.de/projekte/corplex/TreeTagger/DecisionTreeTagger.html
6


of ’trigger’ words. The stop phrases can be words, symbols, verbs, punctuation, etc. They depend
on the rule type.

Currently, the identification of the context has been done manually, however an automatic
learning of the context can also be considered. We have manually detected special phrases for
each type of information on an excerpt of the corpus. We have noted the corresponding trigger
words and the stop phrases. The common stop phrases are the trigger words, some punctuation
symbols and the relative pronouns. The current rule base consists of 35 rules – 13 for namews,16
for purpose, and 6 for domain. Let us note that different cases can be expressed by the same rule.
Let us take some examples.

Example 1:The web service description is ”CAPTCHA-image web service for serving and
validating CAPTCHA-images ”. We obtain this result:

<ws><id>WSID 172</id>
<namews>CAPTCHA-image</namews></ws>
Example 2: The web service description is ”The ICD9 coding system is an international
classification system which groups related disease entities and procedures for the purpose of
reporting statistical information. The system is widely used to for medical billing ”. We obtain
this result:

<ws><id>WSID 29</id>
<domain>medical billing</domain></ws>

In the example 1, the term ”CAPTCHA-image” is extracted by using the following set of rules:
Webservname (type:namews..name:N) --> start, nws(N).
start --> @lemma:’.’ ; ls_lookupToken(_,_,startPara).
nws(N) -->ls_token(N,tag:X,_), end, {not(member(X,[dt]))}.
nws(N) -->ls_token(N1,_), nws(N2), {concat(N1,N2,N)}.
end --> @lemma:web, @lemma:service ; @lemma:webservice.

The trigger phrase is the start of a paragraph (”startPara ”) or the start of a new sentence. The
rules ”nws ” allow the system to recognize the multiword terms. The end phrase is here Web
Service or Web Service (with or with out capitals). Let us remark that the rule ”nws ” avoid the
recognition of a single determiner (”dt ”) just before the end phrase.

Our system differs from the classical approaches. From Information Extraction method point of
view, simple declarative extraction rules are designed, making the implementation process ”light
and quick”. The rules are domain-specific, but by no means corpus-specific. Moreover, our
method is endogenous: no resources such as knowledge base or dictionary are needed at the
beginning. There sources are constructed on the fly – the system learns new terms (which can be
new terms in the domain or missing) to be used later or in other corpora and/or in other text
mining applications.

All the previous representations are produced from the web service description, i.e, WSDL
content, which represents the provider point of view. Our goal is to infer a web service
representation from the description of its context (neighbour services).

4.3. Symbolic reputation model (SR)

The reputation of a web service reflects a common perception of other web services or customers
towards that service. In other words, it aggregates the web service ratings given by consumers.

7


Typically, a reputation would be built from a history of ratings. Several reputation systems have
been proposed in the literature [17, 3, 5, 13]. However, feedbacks from users are not easily
available and Hidden Markov Models (HMM) were used to predict the reputation of a service
provider [10]. In [8], the authors propose an algorithm for generating locally calculated reputation
ratings from a semantic network. This effort has an implicit hypothesis, which considers that web
service reputation is a numerical quantity. As far as we know, it is the first time we go beyond the
quantitative reputation by proposing a symbolic reputation notion. In fact, our work considers the
qualitative aspect of the reputation and describes how to compute and use it. In this paper, we
consider that the numeric reputation of a service s , denoted Rnum (s), results from customer’s
feedback and computed as the weighted average of all ratings the service received from
customers [17]:
Ns

∑s i
∗ l di
R num ( s ) = i=1
Ns
(1)
∑l di

i=1

Where Ns is the number of ratings for the service s , s i is the ithservice rating, l ∈ [0,1] is the
inclusion factor tacking into account the dimension time (smaller l means that the more recent
ratings have a larger impact in the reputation score), and di is the age of the ithservice rating.

Before computing symbolic reputation, we explain the structure of the web service space by
modelling relationships between services. Web services dependency graph was used to represent
the structure of the web services space, i.e. web services relationships, by generating a network
depending on the Input / Output of each web service. This induced network is then used for web
service composition [23,2]. The general approach consists of searching first for web services
operations dependencies using the I/O parameters of their operations then the web services
dependency graph is induced. For seek of simplicity, we consider here that a dependency occurs
between two services when they offer at least two dependent operations: a service requires (resp.
provides) data from (resp. to) another service. of course, other dependency definitions can be used
without affecting our general symbolic reputation computing approach.

Definition 4.3.1 Operations dependency:
Let F = { f1, f2,..., fn } be the set of operations offered by all web services published in the UDDI.
InF( fi ) , OutF( fi ) represent respectively the inputs and the outputs of the operation fi ,
a ∈[0,1] a given threshold and sim the similarity function.
The operation f j depends on the operation fi , denoted fi → f j , if and only if
∀p ∈ InF( f j ), ∃q ∈ OutF( fi ) : sim(p, q) ≥ a . sim(a, b) =1− NGD(a, b) is the featureless
similarity factor computed between words a and b using Normalized Google Distance (NGD)
[22] as a featureless distance measure between words.

Definition 4.3.2 Web services dependency graph:
Let S = {s1, s2,..., sn } be the web services space containing all services published in the UDDI,
we define the dependency graph as the directed graph G = (S,V ) where
V = {(si , s j ) ∈ S × S,∃fi ∈ si , ∃f j ∈ s j : fi → f j } . We note f ∈ s when the service s offers the
operation f .

8


This construction of the direct web services dependency graph is needed to compute the symbolic
reputation. Our model of symbolic reputation is based on random walks models. A random walk
is a sequence of nodes in a graph constructed by the following process: select a random starting
node in the graph, then move to select a neighbour of this node at random and so forth. The
random walk analysis has been applied to different fields as a fundamental model for random
processes in time [11]. More particularly, a formal model was proposed in [6] to allow the
computation of the reputation of a web page using the web hyperlink structure. We have adapted
this model to compute the symbolic reputation of web services.

We consider the constructed dependency graph. The probability of visiting a service s for term t
at step n of the walk, denoted P n (s, t), is defined as follows:

d
P n−1 (q, t)  N if t ∈ VR(q)
P n (s, t) = (1− d)∑ + t (2)
q→s O(q) 
0 Otherwise

The previous probability is the core of the algorithm that computes the symbolic reputation. In
fact, suppose that with the probability d the random walker jumps into a service uniformly
chosen at random from the set of services that contains the term t . In this context, the probability
d
that a random walker visits a service s in a random jump is ( Nt denotes the total number of
Nt
services on the UDDI that contains the term t ) if the vectorial representation VR(q) of service q
contains the term t and it is zero otherwise. The probability that the walker visits the service s at
1− d n−1
step n after visiting one among its parent service q is P (q, t) where P n−1 (q, t) denotes
O(q)
the probability that the walker visits the service q for term t at step n −1 and O(q) denotes the
number of services that require data from q (parents of s in the dependency graph). Let us
denote VR(s) , the vectorial representation of service s . Algorithm 1 gives the details on the
computation of the symbolic reputation of a given service.

5. GENERAL ARCHITECTURE AND UDDI EXTENSION

The aim of our implementation is to extend a UDDI by introducing the different approaches
previously discussed. Our system supports service consumers to discover web services that meet
their requirements. It also recommends other services to customers using symbolic reputation. In
addition, our platform allows enriching the UDDI register automatically by collecting Web
services published in others UDDI registries. In our implementation, services or agents (dynamic
Web services) of our system reside in a remote server that consumers know a priori. Our system
enables service providers to publish their services with QoS (price, availability, response time, . .
.) and consumers to find web services that meet their functional requirements and/or QoS
requirements. Figure 3 gives an overview of the general architecture by outlining the different
components and their relationships. The general architecture of our system consists of the
following components:

9


ALGORITHM 1: SYMBOLIC REPUTATION COMPUTING

REQUIRE:

• d: The random surfer jump.
• k: Maximum number of iteration.
• s: A given service.
ENSURE: SR(s): Symbolic reputation of the service s.

1: SR(s) = ∅
2: VR(s) = get Vectorial Representation(s)
3: FORALL term t ∈ VR(s)DO
4: P(s,t) = d / Nt
5: ENDFOR
6: FOR l =1, 2,.., k DO
7: IF l < k THEN
8: d' = d
9: ELSE
10: d ' =1
11: ENDIF
12: FOR every path ql →... → q1 → s of length l and every term t in VR(ql )DO
13: P(s, t) = 0 if term t has not been seen before
P(s, t) = P(s, t)+[(1− d)l / ∏ O(qi )(d ' / Nt )
l
14:
i=1
15: ENDFOR
16: ENDFOR
17: FOR every term t with P(s,t) >1/ Nt DO
18: SR(s) = SR(s)∪t
19: ENDFOR
20: RETURN SR(s)

• UDDI Registry: It allows service providers to publish their web services.
• UDDI Manager: It contains two interfaces Service Publisher and Service Finder.
Service Publisher offers all the necessary operations to service providers to manage their
services in the UDDI registry. This is the only possible way to interact with the UDDI
registry. Service Finder allows finding services according to functional requirements. It
can also provide information that describes a web service published in the UDDI registry.
• Discovery Agent: It receives consumer’s requests, finds the web services that meet their
requirements and sends the response to the customer. It also collects feedback rating from
customers expressed as a score reflecting their level of satisfaction after interacting with
the discovered service.
• Reputation Manager: It allows calculating the symbolic reputation of web services.
• Clustering Manager: It allows creating the clusters of web services.
• WSDL Manager: It allows parsing, extracting features from WSDL documents and save
them in the WSDB database.
• WVMC rawler: It is a dynamic web service that runs automatically to collect Web
services published in other UDDI registries.
• WSDB Manager: It allows managing the WSDB database (CRUD -
Create/Read/Update/Delete operations).
10


Figure 3. General architecture and UDDI extension

(a). Services register.
(b). Register / unregister services and business on UDDI Registry.
(c). Save the QoS advertisements (price, availability, response time, . . .) if are specified for
services to register on database WSDB.
(d). CRUD (Create/Read/Update/Delete) operations on database WSDB.
(e). Request/Response.
(f). Find services from UDDI by functional requirements.
(g). (1) Get Qos and reputation advertisements for services that functional requirements
match with user’s query. (2) Send user’s feedback score for web service.
(h). Bind a web service.
(i). Get services from UDDI and calculate the Quantitative/Qualitative reputation.
(j). Save the computed reputation for each service on WSDB.
(k). Get services from UDDI and construct the clusters of Web services.
(l). Save the constructed clusters on WSDB.
(m). Get the new WSDL URI published on UDDI Registry and parse it to extract all
information of WSDL document.
(n). Save the features extracted by WSDL Parser on WSDB.
(o). Crawler WSDLs URI from others UDDI registries.
(p). Register the new crawled web services on UDDI registry.

6. WEB SERVICE DISCOVERY AND RECOMMENDATION SYSTEM

The proposed web service discovery and selection algorithm is based on the three following
operators: Matching, Ranking and Selection (see Algorithm 2):
11


• Matching: Find services that meet the consumer’s functional requirements. Then
maintain services that meet only the non-functional requirements (QoS and/or
reputation).
• Ranking: Calculate an overall score that combines the quality of service(availability,
response time, . . .) and/or reputation score and use it to classify the web services,
• Selection: Select only the best services according to the classification results and the
maximum number of services specified by consumer, and return them.
Algorithm 2 shows the details of the discovery process and selection of our system. The
functional matching process is based on the consumer’s functional requirements. The Qos
matching process is based on the algorithm proposed by Maximilien and Singh [4] and improved
by Xu et al. [17].A service consumer sends a service discovery request to the Discovery Agent
(see Figure 3), which then contacts the UDDI Manager and WSDB Manager to find the services
that meet the functional requirements (func Matching() - line 1).If no matched service is found,
the Discovery Agent returns an empty result to the consumer. If multiple services match the
functional requirements the discovery service contacts the WSDB Manager to retrieve the Qos
attributes for candidates services previously selected. If multiple services match the Qos
requirements (Qos Matching() - line 5) the discovery agent calculates the Qos score for each
service. If the consumer’s reputation requirement is specified, the discovery agent contacts the
WSDB Manager to retrieve the reputation score for each candidate services. Then the ranking
process (Ranking() - line 8) calculates the overall score and uses it to rank the candidates
services. If no reputation requirement is specified the ranking process (Ranking - line 12) uses
only the computed Qos score. Finally, the selection process (Select) selects nbMax services with
the highest scores and returns them to the consumer (nb Maxdenotes the maximum number of
service to be returned as specified by the consumer). If nbMax is not specified, one service is
randomly selected from services whose Qos score (or overall score) is greater than the given
threshold.

We also propose an algorithm for web services recommendation using the symbolic reputation
(SR). Algorithm 3 shows the details of the symbolic reputation based recommendation. The
general method is based on the following steps: (1) we use Algorithm 2 to discover services that
match consumer’s requirements.(2) We retrieve the symbolic reputation SR(s) for each
discovered services. Then, we find all services q that have a vectorial representation VR(q) such
that SR(s)∩VR(q) ≠ ∅ and its cardinality is greater than a given threshold a . Then, the ranking
process (Ranking - line 9) uses the computed Qos score and the reputation score to rank the
candidates services previously selected. Finally, the selection process (Select - line 10) selects
the services with the highest scores and returns them to the consumer. Only the best services
according to the ranking result are recommended.

ALGORITHM 2: WEB SERVICE DISCOVERY AND SELECTION
REQUIRE:

• fRq: Functional requirements
• qosRq: Qos requirements
• repRq: Reputation requirements
• nbMax: Maximum number of services to be returned
ENSURE: select: A set of discovered services

12


1: Service[] f Match= func Matching(fRq)
2: select = Service[]
3: qMatch = Service[]
4: IFqosRq != null THEN
5: qMatch= Qos Matching(f Match,qosRq)
6: sRank = Service[]
7: IFrepRq != null THEN
8: sRank= Ranking(qMatch, qosRq, repRq)
9: select = Select(sRank,nbMax,"QosAndRep")
10: RETURN select
11: ELSE
12: sRank= Ranking (qos Match, qosRq)
13: select= Select (sRank, nbMax, "Qos")
14: RETURN select
15: ENDIF
16: ELSE
17: select= Select(fMatch, nbMax, "Functional")
18: RETURN select
19: ENDIF

ALGORITHM 3: SYMBOLIC REPUTATION BASED RECOMMENDATION
REQUIRE: ds: Discovered web service. S: set of all services in the UDDI

ENSURE: L: List of web services to recommend

1: L ≠ ∅
2: SR(ds) = get Symbolic Reputation(ds)
3: FORALL q ∈ S DO
4: VR(q) = get Vectorial Representation (q)
5: IF SR(ds)∩VR(q) ≥ a THEN
6: L = L ∪q
7: ENDIF
8: ENDFOR
9: L = Ranking( L )
10: L = Select( L )
11: RETURN L

7. EVALUATION
We have considered different web service sources like WebservicesX.net7, xMethods.net8and
seekda.com9. Our general UDDI architecture allows using different UDDI registries, which are
used through specific wrappers. We have collected around 8,500 services. However, these
services are not classified into categories. This is a real problem when we want to evaluate the
usefulness of our representations using precision and recall measures. For this reason, we have

7
http://www.webservicex.net/ws/default.aspx
8
http://www.xmethods.net/ve2/index.po
9
http://www.seekda.com
13


considered a new web services source, i.e. service-finder.eu10, which classifies its services using
an ontology [18]. The number of services, which are classified into a category by service-finder is
only reduced to only 1,647 web services. 465 services among them are not available: they are
discarded. The collected web services are multi-languages. We have automated the recognition of
the language and keep only those using English language. Finally, 993 web services are used.

Let us denote ’Collected WS’ the number of collected web services for each category,’ Available
WS’ the number of available services and ’Used WS’ the number of used services for each
category after the process of recognition of the language. See Table 1 for more details.

Table 1. Classification of the used web services according to the category

Categories Collected WS Available WS Used WS
Weather 57 39 35
Address Information 433 316 301
SMS 45 36 20
Currency Exchange 27 23 18
Stock 163 153 142
Payment 47 23 15
Converters 23 15 9
News 37 29 25
Jobs 18 13 12
Travel 653 433 331
Logistics Shipping 14 12 6
Multimedia Video 10 10 8
Identity Verification 22 13 11
Government 11 7 7
Mathematics 36 26 24
Translation 51 34 29
Total 1647 1182 993

In the rest of this section, we discuss the results of the web services discovery considering the
different presented representations. Finally, we discuss the web services recommendation results.

We use the two classical measures, Precision and Recall, to evaluate the performance of our
approach. Precision and Recall are often used to evaluate information retrieval schemes [14].
Precision can be seen as a measure of exactness or fidelity, whereas Recall is a measure of
completeness [14]. In our context, the Precision and Recall are defined as follows:
A
Pr ecision = (3)
B
A
Recall = (4)
C
Where A is the number of relevant services retrieved, B is the total number of irrelevant and
relevant services retrieved, and C is the total number of relevant services in the whole
collection.

In order to evaluate the interest of the evoked representations, i.e., B, RBTT and SR , we have
considered the following experimental protocol :

10
http://www.service-finder.eu
14


• Discover the web services that belong to a given category. To achieve this goal, we
generate a query reqc defined by the name of category c , for example "Weather",
"Address Information", "SMS", "Currency Exchange" and "Stock".
• Compute the precision and the recall for each representation. These measures compare
the returned result with the services belonging to the category (see Table 1).
• Evaluate the global precision and recall for each representation.

The results, given in Table 2, show that the categories of web services are not represented at the
same level of precision. In general, the representation RBTT is the best. Then we have B and in
last SR. Note that the overall recall value is medium about all representations. In fact, the RBTT
is the best representation for the category ’Weather’. For the other categories, we note that only
the RBTT has obtained the medium values. Let us remark that B and SR representations give the
lowest precisions for respectively "Currency" and "Jobs"(see Table 2). Finally, we have the
category "SMS" for which we obtain medium results except for SR that gives the lowest
precisions. In conclusion, this is difficult to find an efficient representation for all categories as
the description of web services are generally heterogeneous. The lessons we have learned from
our experimental results are:

1. The traditional representation B may be efficient for some categories but is not robust and
may lead to very low precisions,
2. The symbolic reputation (SR) is not efficient for web services discovery, except for one
category (i.e. ’Weather’), and consequently it will not be used for web services discovery,
3. The rules based text tagging (RBTT) is robust (i.e. the results do not vary according to
the category) and gives correct precision results for all categories.

The previous results show that the symbolic reputation is not appropriate for web service
discovery. In fact, it does not represent the description of a service itself, but it represents the
description of the relation the service has with others. Thus, we have used the SR representation
for web services recommendation (see Algorithm 3). The basic idea behind this task is to enrich
the returned results during the web services discovery. For each discovered service, we search the
set of web services to be recommended with these discovered services. The result of a query is
not reduced only to the services that match the query, but it also contains recommended services
for each discovered service (Target service).

The recommended services are ranked using the ranking process (Algorithm 3), which uses the
computed Qos score and the reputation scores. We select only the services with the highest scores
and return them to the consumer. Table3 shows the results for services belonging to the
"Weather" category. We have evaluated manually these recommended services. The result is
really significant as the recommended services generally perform the same task, i.e. belongs to
the same category of the target service or may be composed with it.

Table 2.Web service discovery evaluation (Precision (P) and Recall (R) in %).

B RBTT SR

Categories P% R% P% R% P% R%

Weather 68,97 57,14 81,82 51,43 68,97 57,14

SMS 48,48 80,00 55,56 50,00 12,12 80,00

15


Currency Exchange 16,67 16,67 58,33 77,78 8,51 22,22

Jobs 04,35 08,33 57,14 33,33 6,67 41.67

Average 34,61 40,53 63,21 53,13 24,06 50,27

Table 3. Recommended web services for the category "Weather"

Target service:

http://www.deeptraining.com/webservices/weather.asmx?wsdl
Recommended web services:

http://ws.cdyne.com/WeatherWS/Weather.asmx?wsdl
http://ws.soatrader.com/harbormist.com/0.1/WeatherService?WSDL
http://rightactionscript.com/webservices/nusoap/server.php?wsdl
http://www.deeptraining.com/webservices/weather.asmx?wsdl
http://ws.soatrader.com/weather.gov/0.1/ndfdXML?WSDL
http://ws.soatrader.com/cis.temple.edu/0.1/DOTSFastWeather?WSDL
http://ws2.serviceobjects.net/fw/fastweather.asmx?wsdl
http://ws.serviceobjects.com/fw/FastWeather.asmx?WSDL
http://ws.soatrader.com/webservicex.com/0.2/GlobalWeather?WSDL
http://ws.soatrader.com/ejse.com/0.1/Service?WSDL
http://ws.soatrader.com/deere.com/0.1/WeatherForecastServiceService?WSDL

8. CONCLUSION

This paper revisits the traditional representation of web services using its textual descriptions and
the elements of the WSDL structure as types, attributes, etc. We have proposed two new web
services representations. The first one is based on semantic tagging of web services descriptions
to keep only the most significant parts. Whereas, the second proposed representation, i.e.,
symbolic reputation is more contextual as it does not consider only the service itself, but its
neighbours. The main conclusions of the experimentation with real-world web services are:

• Traditional representation (B) behaves correctly but it is not robust when we consider
different concepts (categories).
• Rules based text tagging (RBTT) is more robust(i.e. the results do not vary according to
the category) even if its precision is lower than traditional representations.
• Symbolic reputation (SR) is not appropriate for the web services discovery tasks but it is
more efficient for the web services recommendation task.

In the future, we will continue our work on symbolic reputation for the recommendation task. We
will increase the number of web services used for the experimentation. Finally, we will make
available online the gate corresponding to the full implementation of our system.

16


REFERENCES
[1] W3C (2004).“Web services architecture”. Technical report, W3C Working Group Note 11 February
2004. Retrieved April 30, 2006 from http ://www.w3.org/TR/ws-arch.
[2] Abrehet Mohammed Omer, Alexander Schill, “Dependency Based Automatic Service Composition
Using Directed Graph”, Next Generation Web Services Practices, 2009 (NWESP ’09). Fifth
International Conference on, vol., no., pp.76-81.
[3] E. Michael Maximilien and Munindar P. Singh. (2002). “Reputation and Endorsement for Web
Services”. ACM SIGecom Exchanges, Vol. 3(1), pp.24-31.
[4] E. Michael Maximilien and Munindar P. Singh. (2004). “Toward autonomic web services trust and
selection”. ICSOC’04: Proceedings of the 2nd international conference on Service oriented
computing, pp. 212-221.
[5] Majithia, S.; Ali, A.S.; Rana, O.F.;Walker, D.W. (2004). “Reputation-based Semantic Service
Discovery”. In Proc. of the 13th IEEE Intl.Workshops on Enabling Technologies: Infrastructure for
Collaborative Enterprises (WETICE), pp.297-302, Modena, Italy.

[6] DavoodRafiei Alberto O. Mendelzon“What is this Page Known for? Computing Web Page
Reputations”, 9th InternationalWorld WideWeb Conference, Amsterdam, 2000.
[7] Gustavo Alonso, Fabio Casati, Harumi Kuno, Vijay Machiraju, “Web Services - Concepts,
Architectures and Applications”, Springer Verlag, Berlin Heidelberg, 2004.
[8] Jennifer Goldbeck, James Hendler, “Inferring Reputation on the Semantic Web”, WWW 2004, NY
USA.
[9] Petrova-Antonova, D. et S. Ilieva (2009). “Towards a Unifying View of QoS-Enhanced Web Service
Description and Discovery Approaches”. In YR-SOC, pp. 99-113.
[10] Zaki Malik, Ihsan Akbar, Athman Boughettaya, “Web service reputation Assessment Using a Hidden
Markov Model”, ICSOC-ServiceWave 2009, LNCS 5900, pp. 576-591, 2009.
[11] Laszlo Lova’sz, “Random Walks on Graphs: A Survey”, Bolyai society, Mathematical studies, 2,
1993.
[12] Khalid Elgazzar, Ahmed E. Hassan, Patrick Martin, “Clustering WSDL Documents to Bootstrap the
Discovery of Web Services”, IEEE International Conference on, pp. 147-154, 2010 IEEE
International Conference on Web Services, 2010.
[13] Surya Nepal and Zaki Malik and AthmanBouguettaya. “Reputation Management for Composite
Services in Service-Oriented Systems”, Int. J. Web Service Res. 8(2): 29-52 (2011).
[14] John Makhoul, Francis Kubala, Richard Schwartz, Ralph Weischedel, “Performance measures for
information extraction”, DARPA Broadcast News Workshop, Herndon, VA, February 1999.
[15] Shuiguang Deng, Zhaohui Wu, Jian Wu, Ying Li, Jianwei Yin, “An Efficient Service Discovery
Method and its Application”, International Journal of Web Services Research, Vol. 6, No. 4, pp. 94-
117, 2009.
[16] Xin Dong, Alon Halevy, Jayant Madhavan, Ema Nemes, Jun Zhang, “Similarity Search for Web
Services”, Proceedings of the 30th VLDB Conference, Toronto, Canada, pp. 372-383, 2004.
[17] ZiqiangXu, Patrick Martin, Wendy Powley and Farhana Zulkernine, 2007, “Reputation Enhanced
QoS-based Web services Discovery”, IEEE International Conference on Web Services (ICWS 2007).
[18] Adam Funk and KalinaBontcheva. “Ontology-Based Categorization of Web Services with Machine
Learning”, Proceedings of the Seventh conference on International Language Resources
andEvaluation (LREC’10), may 2010.
[19] Charnois, T. and Durand, N. and Klema, J. “Automated Information Extraction from Gene
Summaries, ECML/PKDD 2006 Workshop on Data and Text Mining for Integrative Biology
(DTMIB 2006), pages 4-15, Berlin, Germany, September 2006.
[20] Bilhaut, F. and Widlcher, A. “LinguaStream: An Integrated Environment for Computational
Linguistics Experimentation”, the European Chapter of the Association of Computational Linguistics,
2006, Trento, Italy
[21] M. F. Porter, “An Algorithm for Suffix Stripping”, Program, Vol. 14, No. 3, pp. 130-137, 1980.
[22] Cilibrasi, Rudi L, Vitnyi, Paul M. B., “The Google similarity distance”, IEEE Transactions on
Knowledge and Data Engineering, Vol. 19, No. 3, pp. 370-383, March 2007.
[23] Seyyed Vahid Hashemian, Farhad Mavaddat, “A Graph-Based Approach to Web Services
Composition”, Proceedings of the Symposium on Applications and the Internet, p.183-189, January
31-February 04, 2005

17


[24] Yan, J. and Piao, J. “Towards QoS-Based Web Services Discovery”. ICSOC’08: Proceedings of the
2nd international conference on Service oriented computing, pp. 200-210.
[25] Wei Liu, Wilson Wong, “Web service clustering using text mining techniques”, International Journal
of Agent- Oriented Software Engineering, Vol. 3, No. 1, pp. 6-26,2009.
[26] C. Platzer, F. Rosenberg and S. Dustdar: “Web service clustering using multi dimentionalangles as
proximity measures”. ACM Trans. Internet Technol. 9(3), pp. 1-26(2009).
[27] NatalliaKokash, “A Comparison of Web Service Interface Similarity Measures”. Frontiers in
Artificial Intelligence and Applications, Vol. 142, pp.220-231, 2006.
[28] C. Atkinson, P. Bostan, O. Hummel and D. Stoll. “A Practical Approach to Webservice Discovery
and Retrieval”. In 2007 IEEE International Conference on Web services (ICWS 2007),2007
[29] H. Lausen and T. Haselwanter. “Finding Web services”. In the 1st European Semantic Technology
Conference, Vienna, Austria,2007
[30] W. Abramowicz and K. Haniewicz and M. Kaczmarek and D. Zyskowski. “Architecture for Web
services filtering and clustering”. In Proceedings of ICIW’2007.
[31] R. Nayak and B. Lee. “Web service Discovery with Additional Semantics and Clustering”. In Web
Intelligence, IEEE/WIC/ACM International Conference, 2007
[32] Le-Hung Vu, Manfred Hauswirth and Karl Aberer, “QoS-based service selection and ranking with
trust and reputation management”, Proceedings of OTM’05, R.Meersman and Z. Tari (Eds.), LNCS
3760, p.p. 466-483, 2005.
[33] Z. Malik and A. Bouguettaya, “Evaluating Rater Credibility for Reputation Assessmentof Web
Services”, International Web Information Systems Engineering Conference (WISE 07), Nancy,
France, December 2007.
[34] Wang Y., Zhang J., Vassileva J. “Effective Web Service Selection via Communities Formed by
Super-Agents”, Proceedings of IEEE/WIC/ACM WI 2010, Toronto, August 2010, pp.549-556.
[35] E. M. Maximilien and M. P. Singh, “Multi agent System for Dynamic Web Services Selection”, In
Proceedings of 1st Workshop on Service-Oriented Computing and Agent-Based Engineering
(SOCABE at AAMAS), July 25-29, 2005, Utrecht, The Netherlands
[36] Y. Liu, S. Ngu, and L. Zeng. “QoS Computation and Policing in Dynamic Web Service Selection”.
WWW’2004, May 2004.
[37] Wishart, R., Robinson, R., Indulska, J., and Josang, A. “Superstring Rep: Reputation-enhanced
Service Discovery”, Australasian conference on Computer Science, Vol. 38: 49-57, 2005.
[38] Yao Wang, JulitaVassileva, “Towards Trust and Reputation Based Web ServiceSelection”, In Multi-
Agent and Grid Systems(MAGS) Journal, 2007.
[39] X. Z. Liu, G. Huang, and H. Mei, “A community-centric approach to auto mated service
composition”, Science in China, Series F, vol. 53, no. 1, pp. 50-63, 2010.

18

Web Services Discovery and Recommendation Based on Information Extraction and Symbolic Reputation

More Related Content

What's hot (11)

Similar to Web Services Discovery and Recommendation Based on Information Extraction and Symbolic Reputation (20)

Recently uploaded (20)

Web Services Discovery and Recommendation Based on Information Extraction and Symbolic Reputation