UQ IElab at TREC 2020 CAsT Track

Abstract

This paper describes the work done by the IELAB for the TREC Conversational Assistance Track (CAsT) 2020. IELAB investigated two methods to improve both the retrieval and re-ranking stages of a conversational IR system. The first method used an Adapted Query (AQ), which extracted context from the first utterance only to expand all subsequent queries for a conversational session. The second method utilized a query likelihood model (QLM) which used a pre-trained deep language model to estimate the likelihood that a query could be generated by a document. Specifically, we used the text-to-text transfer transformer (T5) as a scoring functions for re-ranking passages.