Re-ranking of Yahoo snippets with the JIRS passage retrieval system

Abstract

Passage Retrieval (PR) systems are used as first step of the actual Question Answering (QA) systems. Usually, PR systems are traditional information retrieval systems which are not oriented to the specific problem of QA. In fact, these systems only search for the question keywords. JIRS Distance Density n-gram system is a QA-oriented PR system which has given good results in QA tasks when this is applied over static document collections. JIRS is able to search for the question structure in the document collection in order to find the passages with the greatest probability to contain the answer. JIRS is a language-independent PR system which has been already adapted to a few non-agglutinative European languages (such as Spanish, Italian, English and French) as well as to the Arabic language. A first attempt to adapt it to the Urdu Indian language was also made. In this paper, we investigate the possibility of basing on the web the JIRS retrieval of passages. The experiments we carried out show that JIRS allow to improve the coverage of the correct answers re-ranking the snippets obtained with Yahoo search engine.