Semantic Re-ranking in Ad-hoc Robust Retrieval

This paper proposes an investigation about a re-ranking strategy presented at SIGIR 2010. In that work we describe a re-ranking strategy in which the output of a semantic based IR system is used to re-weigh documents by exploiting inter-document similarities computed on a vector space. The space is built using the Random Indexing technique. The effectiveness of the strategy has been evaluated in the context of the CLEF Ad-Hoc Robust-WSD Task, while in this paper we propose new experiments in the TREC Ad-Hoc Robust Track 2004. 1 Background and Motivation A general approach to overcome the word ambiguity problem in IR involves the representation of documents by word meanings. Among the most investigated techniques are those that rely on WordNet synsets through which groups of synonym words are uniquely identified and linked to each other by semantic relations. The Robust-WSD task at Cross Language Evaluation Forum (CLEF) [1] has shown that results improve when aggregation strategies are exploited. The method proposed in [6] describes a different approach to document aggregation based on a variation of the “inter-document similarities” [8] idea. The method combines two retrieval strategies that work at two different representation levels: keyword and synset. The ranked list of documents retrieved using the synsetbased representation (synset list) is exploited to re-rank the list of documents retrieved using the keyword-based one (keyword list). The insight of this method is that documents in the keyword list with the highest number of similar documents in the synset list should climb in the result set. The approach tries to re-weigh documents in response to a query by promoting those documents with the highest number of supporters. In this context, a supporter is a document with content similar to the target one. Inter-document similarities is computed relying on the Random Index technique to build a vector space in which similar documents are represented close. Let us denote by Lk and Ls the ranked lists of documents retrieved using keywords and synsets representation, respectively. The idea behind our re-ranking method is to give more evidence to the documents in Lk that are widely supported by similar documents occurring in both lists. The method requires the following steps: 1 A semantic lexicon for the English language. 2 Pierpaolo Basile, Annalina Caputo, and Giovanni Semeraro 1. For each document di ∈ Lk compute the supporters(di, α), which is the set of α documents {d1, ...dα} ⊂ Lk with the highest inter-document similarity to di. 2. Get the overlap supporters = {dj ∈ Ls : dj ∈ supporters(di, α)} which is the set of documents occurring in both Ls and supporters. 3. Assign to di a new score S(di) taking into account supporting documents computed in the step 2. Formally: S(di) = θ ∗ Ssupporters + (1− θ) ∗ Sk(di) (1) where Ssupporters = ∑ dj∈overlap supporters Sk(dj) ∗ Ss(dj) (2) and Sk(dj) is the score of dj in Lk, while Ss(dj) is the score of dj in Ls, and θ is a free parameter used to smooth Ssupporters, which denotes the scores combination of supporting documents.