Semantically enhanced pseudo relevance feedback for Arabic information retrieval

Abstract

The conventional information retrieval (IR) framework consists of four primary phases, namely, pre-processing, indexing, querying and retrieving results. Some phases of the current Arabic IR (AIR) framework have several drawbacks. This research aims to enhance an AIR by improving the processes in a conventional IR framework. We introduce an enhanced stop-word list in the pre-processing level and investigate several Arabic stemmers. In addition, an Arabic WordNet was utilized in the corpus and query expansion levels. We also adopted semantic information for the Pseudo Relevance Feedback. The enhanced Arabic IR framework was built and evaluated using TREC 2001 data. The technique of using the Arabic WordNet to build a semantic relationship between query and corpus in two levels, that is, the corpus and query levels, is a new one. The enhanced AIR framework demonstrated an improvement by 49% in terms of mean average precision, with an increase of 7.3% in recall compared with the baseline framework.

Word-embedding-based pseudo-relevance feedback for Arabic information retrieval

“…The main contribution consists of using word’s part-of-speech to select the appropriate synonyms. More recently, Atwan et al [9] presented an automatic corpus-based expansion technique combining AWN and corpus-based semantic similarity to select expansion terms. The results showed that the automatic expansion technique enhances the accuracy of Arabic IR on TREC 2001 dataset.…”

“…Ideally, however, expansion terms should be selected based on their similarity to query terms as well as their distribution in the set of pseudo-relevant documents. Some studies have indeed proposed to do so with mutual information [8–10]. We propose here to use word embedding for this task and focus on the Arabic language.…”

Query Expansion for Arabic Information Retrieval Model: Performance Analysis and Modification

“…The preprocessing steps are done on the document terms before building the index and on the user query before matching process. The preprocessing should be done first to gain the benefit of speeding-up the retrieval time [18,19]. The preprocessing steps involve tokenization, removal of stop-words and stemming.…”

“…These words don't give any hint for the content of their documents. In information retrieval systems, stop-words should be eliminated (by referring to a stop-word list) from the query text and from the set of index terms [18,20]. Figure 3 shows the tokens of a document title after removing the stop-words.…”

“…To facilitate the matching process, a matching model is used. In this paper, the Vector Space Model (VSM) is used for the matching operation [18,20,[27][28][29][30][31][32]. The VSM is an algebraic model where it uses non-binary weights that are assigned to the index terms of documents and queries.…”