Semantic annotation is commonly recognized as one of the cornerstones of the semantic Web. In the context of Web services, semantic annotations can support e ective and e cient discovery of services, and guide their composition into work ows. Because semantic annotation is a time consuming and expensive task, (semi-)automatic approaches for semantic annotation extraction are required. In this paper, we propose a semi-automatic extraction approach of lightweight semantic annotations from textual description of Web services. In contrast with most of the existing semi-automatic approaches for semantic annotations of Web services which rely on a prede ned domain ontology, we investigate the use of NLP techniques to derive service properties given a corpus of textual description of bioinformatics services. We evaluate the performance of the annotation extraction method and the importance of lightweight annotations to classify bioinformatics Web services in order to bootstrap the service discovery process. Our framework relies an unsupervised clustering approach based on a simultaneous clustering algorithm that enables to determine biclusters of Web services and semantic annotations highly correlated.
During the last decade, semantic Web services (SWS) [
In this paper, we propose a bottom-up approach to extract domain-dependant lightweight semantic annotation from textual description of Web services. Such annotations of Web services aims to capture static (i.e., domain concepts) and procedural knowledge (i.e., tasks) of a domain. Despite their importance, few domain ontologies exist for the purpose of Web services annotation, and thus, building such ontologies is a challenging task. Natural language documentations of Web services are short textual descriptions intended to close the "semantic gap" between low-level technical features of Web services (e.g., data types, port types, or data formats) and the high-level, meaning-bearing features a user is interested in and refers to when discovering a Web service. Hence, our semi-automatic approach combines di erent extraction patterns to generate lighweight annotations describing service properties such as inputs, outputs, or functionnalities. We notice that our extraction method provides a good starting point for ontology building.
Therefore, we rely on a simultaneous clustering algorithm, namely CROKI2
[
The paper is organized as follows. The section 2 reviews related work conducted in the elds of automatic annotation of Web services and block clustering. Section 3 presents our framework for semantic annotation and clustering of Web services. In the section 4, we present and discuss the results of our experimentations. Section 5 concludes the paper and outlines our future work.
2.1
Converting an existing Web service into a semantic Web service requires signi
cant e ort and must be repeated for each new Web service. We review in this
section research work that focus on learning semantic annotations by exploiting
textual descriptions, WSDL les or even Web forms. Hess and al. proposes ASSAM
(Automated Semantic Annotation with Machine Learning), a semi-automatic
WSDL annotator application. ASSAM [
Also, within the bioinformatics space, Afzal et al. [
Ambite and al. [
With the expectable growth of the number of available Web services and service
repositories, the need for mechanisms that enable the automatic organization
and discovery of services becomes increasingly important. In this context, most
of the existing research rely on a one-way clustering, either annotations clustering
[
The proposed framework is comprised of two main steps. The rst one aims to
perform a semi-automatic semantic annotation extraction from Web services
textual documentations. Semantic annotations enables to describe service properties
such as functionalities, inputs, outputs, and other domain-dependant features.
One particluarity of textual Web service description is that they employ natural
language in a speci c way. In fact, such texts belong to what was de ned as
sublanguages [
The semantic annotation extraction phase allows to identify two types of knowledge: domain concepts and procedural knowledge describing services tasks. First, a morphosyntactic analysis of textual description of Web services is performed. In this step, a sentence splitter and a tokeniser components are used to extract sentences and basic linguistic entities. Then, a POS (Part-Of-Speech) Tagger is performed to associate to each word (token) a grammatical category and thus distinguish the morphology of various entities. For example, the sentence below, the tagger identify a verb (i.e., compute), three nouns (i.e., structure, RNA, sequence), an adjective (i.e., secondary ), and a preposition (i.e., for ). compute (VB) Secondary (JJ) Structure (NN) for (Prep) RNA (NN) sequence (NN).
We distinguish di erent types of syntactic patterns depending on the semantic annotation type. Syntactic patterns describe selectional constraints that exploit sublanguages particularities. We distinguish syntactic patterns that allow to extract inputs and outputs of services, services tasks, and domain-dependant features which are strongly related to the bioinformatics domain:
composition issue. We observed that, in majority of textual descriptions of Web services, verbs identify the functionnality performed by a Web service. In our work, we consider di erent classes of verbs which inform on the service task. For example, VBRetrieval is the class of verbs that indicates a retrieval process (e.g., get, retrieve, fetch, search, nd, return, query ). A frequently occuring pattern which involves this verbs class and the preposition from can be used to easily determine the output and the retrieved resource as described by the following selectional pattern:
VBRetrieval <Output> from <Source>.
Other verb classes were recognized, such as VBExtraction which is a class of
verbs denoting an extraction process, VBExtraction=fextract, scan, identify,
locate, analyseg.
2. Identifying inputs and outputs of Web services. Inputs and outputs
of Web services denote domain concepts which are generally depicted by
nouns in the corpus. However, to get high-quality annotations, we create a
list of biological terms comprised by a set of single word terms. When two
or more biological concepts are used together, we interpret them as a
single biological concept and update the list by adding it, i.e., gene expression,
transcription factors, protein structure, tertiary protein structure, amino acid
sequence, chromosome segment, etc. We de ne di erent heuristics that
identify the roles of concepts (input or output) depending on the structure of
the sentence. Some extraction patterns are presented in Table 2. Therefore,
our extraction patterns identi es cases when several concepts are related via
logical operators such as "and ", "or ". In this case, the same role is assigned
to each concept.
3. Identifying domain-dependant features. We de ne a set of extraction
patterns that focus on bioinformatics-dependant features. For example, we
propose patterns to identify data formats (e.g., FASTA, GFF, GIF, etc.)
related to inputs/outputs formats. An example of such patterns is described as
follows: % computes <OutputService> for % <InputService> described
with <dataFormat> %.
We propose to use a simultaneous clustering approach to classify Web services
in terms of semantic annotations. Our approach aims to nd biclusters of Web
services and annotations by applying CROKI2 algorithm [
Croki2 algorithm. The Croki2 algorithm is applied to the contingency table composed of services and annotations to identify a row partition P = (P1; :::; PK ) composed of K clusters and a column partition Q = (Q1; :::; QL) composed of L clusters that maximizes X 2 value of the new contingency table (P,Q) obtained by regrouping rows and columns in respectively K and L clusters. Croki2 consists in applying K-means algorithm on rows and on columns alternatively to construct a series of couples of partitions (P n; Qn) that optimizes Chi2 value of the new contingency table T1(P; Q) de ned by this expression: k 2 [1; :::; K] and l 2 [1; :::; L].
Marginal frequencies in table T1 are :
T1(k; l) = X X
aij i2Pk j2Ql fkl = X X
fij
i2Pk j2Ql
fk: = X
f:l = X
i2Pk
j2Ql
fi:
f:j
Biclusters validity. The application of Croki2 algorithm leads to an exhaustive
enumeration of biclusters. It is possible to select only biclusters satisfying certain
criteria such as a user-speci ed bicluster size, bicluster homogeneity and bicluster
relevancy [
{ Homogeneity H is the inertia conserved by the bicluster divided by the initial inertia. and
H = Bkl=Tkl Tkl = X X fi:f:j (fij =fi:f:j
1)2 i2Pk j2Ql Bkl = gk:g:l(gkl=gk:g:l 1)2 The value of this ratio is between 0 and 1. A high value of this ratio indicates that the bicluster is homogenous. { Relevancy R is the inertia conserved by the bicluster divided by the global inertia.
R = Bkl=B
Our experimental corpus consists of 100 bioinformatics services descriptions from
the biocatalogue2, a new curated life science Web services repository. The
development of Biocatalogue shows the dramatic increase of bioinformatics Web
services and tools with 2053 services and 148 providers3. Biocatalogue allows users
to discover Web services through keyword-based retrieval or category browsing.
Annotations manually attached to Web services are either textual descriptions or
lists of tags. Tagging Web services with a set of lexical tokens de ned by users
is not a perfect way to enable an e cient service discovery. Manual resource
tagging is an error prone and time consuming task. Figure 1 shows the top-20
tags used on biocatalogue. In total, 951 tags were created by users to describe
services. The use of tags to describe Web services raises several issues such as
the ambiguity of their signi cance (e.g., BioMoby or soaplab in Figure 1), the
variability of the spelling for several tags that may refer to the same concept.
Finally, the lack of explicit knowledge representations in folksonomies (a set of
tags) to express whenever the tag describes for example a service task, service
input or output which prevents their use towards a signi cant resource
discovery. In our work, Web services are semantically annotated based on their textual
descriptions. Extracted semantic annotations enable to automatically construct
a semantic service pro le. In following, we evaluate respectively the annotation
extraction module and the block clustering algorithm.
We designed an annotation extraction module using the GATE [
rules to input service descriptions in order to generate semantic annotations. Indeed, JAPE (Java Annotation Patterns) engine provides nite state transduction over annotations based on regular expressions. A JAPE grammar consists of a set pattern/action rules. A JAPE rule has a Left-Hand-Side (LHS) and a Right-hand-Side (RHS). The LHS speci es the annotation pattern that may contain regular expression operators (e.g., *, ?, +). The RHS consists of annotation manipulation statements. Annotations matched on the LHS of a rule are referred to on RHS by means of labels that are attached to patten elements. The gazetteer lookup modules, part of the JAPE engine, enable to identify domain concepts in the textual description based on a set of lists of tokens. We have created di erent lexicons lists containing bioconcepts, service tasks, dataformats and identi ers (e.g., EntrezGene ID, KEGG ID). Figure 2 illustrates an example of JAPE rule for input service annotation.
We evaluate the results of our experimentations in terms of three metrics: precision, recall and F-measure as depicted in Table 3. The three metrics are calculated as follows.
P recision =
Correct + 1=2P artial
Correct + Spurious + 1=2P artial Recall =
Correct + 1=2P artial
Correct + M issing + 1=2P artial F measure = ( 2 + 1)P
R 2R + P
GATE provides an automatic tool for automatic evaluation, named AnnotationDi to compare a set of annotations generated manually and the set of the annotations generated by our extraction method. To measure the performance of the extraction method, we manually identi ed semantic annotations from the service descriptions corpus. Then, using the AnnotationDi Tool, we compared this set of annotations with the ones that were extracted through extraction patterns.
The application of Croki2 algorithm leads to an exhaustive enumeration of
biclusters. The data used to evaluate the Croki2 algorithm consists on 98 services
and 78 annotations only. The choice of meaningful ones is based on homogeneity
and Relevancy as described in the previous section. Given that CROKI2
algorithm uses k-means to cluster rows and columns, the number of clusters needs
to be speci ed by user. Therefore, we extend the use of some validity indices,
namely BH [
Best biclusters have high values of homogeneity and relevancy ( g.3 and Table 4). For example, biclusters 2, 3, 4 and 6 are the most homogeneous (H=100%) and bicluster 5 is the most relevant (R=10%). Services and annotations that compose each selected bicluster are highly correlated. Each service in a bicluster is described by a subset of annotations and each annotation in a bicluster describe only services belonging to the same bicluster. All biclusters are signi cant from the bioinformatics view. For example, bicluster 1 is comprised by services related to pathway and protein interactions, bicluster 2 is composed of services related only to pairwise sequence alignment, in contrast with bicluster 5 which is comprised by services related to pairwise and multiple sequence alignment. 5
This work is part of our ongoing research work. We propose a semi-automatic approach to learn lightweight semantic annotations given a corpus of textual descriptions of Web services. The conducted experimentations show that the approach allows to generate high-quality annotations, mostly because of the ne-grained extraction rules of the approach and the regularity of the sublanguage used to describe Web services in the bioinformatics domain. Our approach consists on a good starting point towards building domain ontologies. As future work, we aim to develop a methodology of domain ontologies building devoted to semantic annotations of Web services by harvesting textual descriptions, WSDL les, and even existing domain ontologies. The main goal of the methodology would be the automatic construction of semantic Web services. Therefore, one motivation of this work is to facilitate the resource discovery within the bioinformatics domain. Thus, we rely on a block clustering algorithm to determine a set of biclusters of services coupled with a set of semantic annotations highly correlated. The results demonstrate the potential of block clustering to model the relatedness between both resources and annotations which is very prominent in the context of service discovery.