During the last years we are witnessing how the use of keywords has become the standard input when searching the Web. As opposite to the syntactic searches performed by traditional web search engines, the current research challenge is a semantics-guided information retrieval. The increasing pools of ontologies available on the Web can help to discover the semantics of user keywords and this information is priceless for many tasks, including new semantic search engines. In this paper we propose a system that takes as input a list of keywords provided by the user and discovers their possible meanings by consulting the knowledge represented by many (heterogeneous and distributed) ontologies. These keyword senses are semantically enriched with the synonym terms found during the ontology matching process: A synonymy measure based on statistics techniques and ontological similarity is used to integrate senses that are similar enough.
Although keyword-based search is a widely used technique for information
retrieval, traditional techniques do not consider the speci¯c semantics assigned by
the user: the same keywords can be used by di®erent users with the purpose
of accessing to di®erent information. Furthermore, the syntactic-based search
engines are very in°uenced by the enormous amount of information about
popular issues on the Web, i.e., the keyword \java": Java as programming language
eclipses the rest of possible senses (the Indonesian island, a co®ee plant, di®erent
US cities, etc). However, ontologies (which o®er a formal, explicit speci¯cation
of a shared conceptualization [
In this paper, we propose a system that takes as input a list of plain keywords provided by the user, discovers their semantics in run-time and obtains a ? This work is supported by the CICYT project TIN2004-07999-C02-02. ?? Work supported by a grant of Santander Central Hispano & University of Zaragoza. list of senses extracted from di®erent ontology pools; it deals with the possible overlapping among senses. The main steps of our approach are summarized in the following: 1. Extraction of Keyword Senses. First, the user keywords are normalized by a preprocessing step (e.g., rewriting them in lowercase, removing hyphens, etc.), and in order to discover the semantics of the user keywords, the system accesses to the shared knowledge stored in di®erent ontology pools available on the Web. The extracted senses are semantically enriched with the ontological senses of their synonyms (which are obtained from the ontology pool), whenever the system evaluates that the synonym senses matches to the semantics of the corresponding keyword sense. 2. Alignment of Senses. This process uses an incremental algorithm for the alignment of the di®erent keyword senses in order to remove the possible semantics redundancy among them. Senses are merged when the estimated synonymy probability between them is above a certain threshold. The synonymy measure combines a standard string distance metric with a structural similarity measure that is based on vector space techniques. Thus the result is a set of di®erent possible senses for each user keyword.
For e±ciency purposes, the system uses sampling and other statistic
techniques, as well as parallel processing, whenever possible. The output of our
system can be the input for a disambiguation process across keywords [
The rest of this paper is as follows. In Section 2 we show how the possible senses of each keyword are obtained and semantically enriched with their synonym senses. In Section 3 we describe the algorithm that computes the synonymy probability in order to integrate senses when a certain threshold is achieved. Finally, conclusions and future work appear in Section 4. 2
In this section we provide the details that show the contribution of this paper in
the task of automatically retrieving the possible senses for a set of user keywords.
In order to ¯nd the ontological terms that match those keywords, the system
accesses to Swoogle [
The system builds a sense for each URI obtained with the information
retrieved from matching terms in the ontology pool [
As matching terms could be ontology classes, properties or individuals, three lists of possible senses are associated with each keyword k: Sclass, Sprop and k k Sindv. In Figure 1 we show an example of some senses found in the ontology k pool for the user keyword \star". The system ¯nds in WordNet two matchings of keyword \star" as concept/class (s1 and s2), and one matching in the Travel Ontology2 as property of class \hotel" (s3). Notice that each sense is initialized with a popularity=1 and a synonymy degree=1.
celestial body binary star ... supernova class s1star= < {WN#star}, star "(astronomy) a celestial body of hot gases that...", 1, 1 > s2sctlaars=s < {WN#star, WN#lead, WN#principal}, star , "an actor who plays a principal role", 1, 1 > actor co−star ... film star domain (hotel) prop s3star= < {TravelOntology#star}, star , "", 1, 1 > Each keyword sense is enhanced incrementally with the synonyms terms extracted from the ontology pool. Therefore our system takes advantage of the shared ontologies available on the Web and semantically enriches the keyword senses with senses extracted from their synonyms. The synonym names are stored in the sense structure shown before, which gets upgraded everytime the sense is integrated with a (very similar) sense coming from other ontology. In order to evaluate the semantic similarity between the sense of a keyword and their synonyms, the system performs a sense alignment process (detailed in Section 3) which determines whether the semantics of the keyword sense and each synonym sense found represent the same semantic or not. After discarding the synonym senses that do not enrich the corresponding keyword sense, the result is a list of di®erent possible senses for each keyword.
This process can be limited in time; obtaining the senses is executed in parallel for each keyword; within that task, the semantic enrichment of each keyword sense with its synonym senses is performed in parallel too. 2 http://learn.tsinghua.edu.cn:8080/2003214945/travelontology.owl
We explain in this section the sense alignment process, which is used in two situations by our system: 1) to check which synonym senses represent the same semantics as their keyword senses, and 2) to avoid redundancy in the list of possible senses of each user keyword. However both tasks share a common goal: to ¯nd when two given senses represent very similar semantics; in that case they will be considered synonyms and both senses will be integrated3.
In order to decide if two senses must be integrated (as a single sense) or
not, the system computes their synonymy probability. Thus the system avoid
redundancy among the possible senses of a keyword. At present, several solutions
to determine the matching among ontological terms of distinct ontologies have
been proposed, see [
The synonymy measure used relies on both linguistic and structural characteristics of ontologies. The following steps are performed: 1) an initial computation using linguistic similarity, which consider labels as strings, 2) a recursive computation using structural similarity, which exploits the semantics of terms (ontological context) until a certain depth, and 3) the above values are combined to obtain the resultant synonymy measure.
Our proposal for sense alignment is not just a comparison between two senses but an iterative process, which improves the quality and e±ciency of ontology matching and enables the reuse of new discovered senses. In other words, each new integrated sense must be considered as candidate to integrate with the rest. For the same reason, new senses that do not integrate are stored because they could become the missing semantic gap between two senses. Although this method is costly (we limit its execution time), it performs a much better ontology alignment among senses. Due to space limitations, we do not detail this process.
In a variety of approaches, the similarity measure is only calculated among
ontological terms that plays as classes. However, unlike another works, we
propose a way to obtain the synonymy probability according to the type of senses
that we compare. Details about this process is available in [
In this paper we have presented a semantics-guided approach to discover the
possible senses for a set of keywords, by searching and extracting relevant knowledge
from di®erent ontology pools; ontology matching and synonymy estimation
techniques are used to merge senses considered similar enough. The main features
of our proposal are the following:
3 The integration process that we propose can be found in [
We believe that this technique to ¯nd out the semantic di®erent between senses (subsets of ontologies) can be applied to many ¯elds. As example, we are currently working on using the retrieved senses to generate queries expressed in a knowledge representation language to retrieve data corresponding to the user keywords.