Extending the Use of Previous Relevant Utterances for Response Ranking in Conversational Search

This technical report describes the approach of the Università della Svizzera italiana and the University of Amsterdam to TREC CAsT 2020. TREC CAsT provides a reusable benchmark for open-domain conversational information-seeking dialogues. Our system first performs query expansion by concatenating raw, relevant previous utterances, as predicted by an independent model trained on CAsTUR, with the current utterance. Initial ranking is performed by BM25, followed by ALBERT re-ranker trained on MS MARCO passage ranking task. Modifications of the approach include two different methods for utilising context: i) feeding the previous utterance and its top response to the model alongside the current one; ii) feeding up to 3 relevant utterances to the model and performing an attentive-sum to aggregate context information. Our last run uses automatically rewritten queries without context utilisation.