Retrieval using Passage Retrieval , Connectivity

This report describes the participation at the Web track of the TREC-(Italy). TREC-9 has been our rst participation to TREC and, then, to the Web track. In the following, we describe the experimental approach we have chosen, the research hypotheses and questions, the problems we encountered, the results we reached and our conclusions. We consider this experience as the rst step towards the participation to the next Web tracks. The approach we have taken to address the problems and the research questions regards both the scientiic side and the implementation side. As regards to the scientiic side, we employed an experimental approach that mixes both classical advanced information retrieval (IR) techniques, and connectivity-based algorithms for IR on the Web. Figure 1 depicts the whole process being described below. Speciically, we have chosen those classical IR techniques, i.e. passage retrieval and blind relevance feedback, which have proven to be eeective to produce good retrieval results 1]. Moreover, we are interested to test whether the connectivity-based algorithms, which have been proposed in diierent Web contexts, are eeective tools to improve classical techniques. As regards to the implementation side, we developed in-house software and employed other software modules that are publicly available. 1 passage extraction baseline connectivity based algorithms similarity computation based algorithms using similarity connectivity connectivity based result baseline result weighted connectivity based result documents passages ranked documents link data hubs and authorities similarity data weighted hubs and authorities Web Interdata passages robot Tidy Figure 1: The experimental process. Bold text refers to the submitted runs. Baseline. First 10 passages { title and paragraphs { are extracted from each document and indexed using a stop-list augmented with Web stopwords, the Porter's stemming algorithm, and by keeping non-stemmed words; for example, the word \White" has been stored together with \white". Title-only and title-description queries are automatically generated, and indexed as passages did. For each query, top 10; 000 passages are retrieved and ranked by F-4 2]. The lists of retrieved passages are reweighted through blind relevance feedback by considering top 100 passages as relevant. The lists of newly 10; 000 retrieved passages are mapped to retrieved documents. The document score is the sum is the of the scores of the mapped passages. Connectivity-based algorithm. A modiied version of the HITS (Hyper-link Induced Topic Search) algorithm is applied on the provided link les, where the link weight is the baseline score; Similarity-based algorithm. In-and …