Article Preview
TopIntroduction
The classical Information Retrieval Systems (IRSs) mainly use words to represent the content of documents and queries, hence, the matching is carried out by using these words on lexical level rather than on semantic level, in other words, when the IR system takes a query, it simply retrieves documents that contain the query words without considering the semantics behind them. Furthermore, many users give different words to refer to the same concept or they only provide few words and do not explicitly describe the information need. Therefore, Users’ query words may be quite different and not specific enough to the ones used in the documents for describing the same semantics. The user may then get many of irrelevant documents in the result set. Clearly, such vocabulary gaps make the retrieval effectiveness non-optimal and decrease the IRS performance.
Query expansion (QE) is one of the proposed approaches used to solve problems mentioned above, based on the following principle; more the number of keywords in the query is greater, more the information need is well described, since it certainly includes a greater number of index keywords that represent relevant documents (Xu & Croft, 2000). Indeed, it consists of adding new words or terms into the original query in order to improve the retrieval performance.
In the past two decades, several query expansion techniques were developed, i.e. the strategy of the direct reformulation, it consists by adding new terms to the initial query that are extracted from external resources such as, ontologies using concepts (Bhogal, Macfarlane, & Smith, 2007) or thesauri as Wikipedia (Li, Luk, Ho, & Chung, 2007). In the strategy of indirect reformulation, this consists by adding new terms extracted from a list of documents already considered relevant, called Relevance Feedback when it is supervised and documents are chosen by the user, or Pseudo Relevance Feedback (PRF) when documents are automatically selected by the system (Xu & Croft, 1996).
Although the traditional methods for query expansion have got many improvements, however, they have to be used carefully, because, it might degrade the performance of the IRS, This is due to the incorrect choice of terms in which implies the divergence to user's information needs (Cronen-Townsend, Zhou, & Croft, 2004). In addition, existing approaches mainly focus on isolated keywords without regard for the query's semantics to expand it integrally, i.e. they do not exploit semantic relations between keywords. They also consist by adding directly terms from the external resource, for which results are bound to be less subjective.
This paper presents a new hybrid approach of query expansion (QE) for Arabic Information Retrieval (IR); in which the authors demonstrate how the PRF expansion can be combined to an external resource such as the Arabic WordNet (AWN) to enhance QE process, in selection and weighting of expansion terms. Thus, the combined approach aims to represent the user's information need by a semantic tree whereby each node is an expansion concept including one or several candidate terms. In other words, is to build a weighted concept hierarchy following the user’s information need and semantic relations of keywords.
The proposed approach is subdivided into three major stages. First, the query is pre-processed, which usually includes normalization, elimination of stopwords, and stemming, etc. Whereby, simple and compound terms are extract. These terms are then matched on the AWN to extract and disambiguate appropriate concepts. The concept list is also given to the IR system to return a set of documents ranked according to relevance criterion of the system. Thereby, the most important terms are extracted from the top-R documents of the Pseudo Relevance Feedback (PRF) process. Second, the query expansion procedure starts by building a set of initial semantic trees, in which roots are the original keywords and their synonyms found in the AWN resource, all these trees are integrated and consolidated to one big tree; by adding relations to concepts that are semantically related, and deletion of concepts that are semantically distant.
Finally, a weight is assigned to each node of the integrated tree. These weights allow selecting candidate concepts according a given threshold in order to generate the best candidate terms for the QE process.