Adaptive utterance rewriting for conversational search

摘要

In a conversational context, a user converses with a system through a sequence of natural-language questions, i.e., utterances. Starting from a given subject, the conversation evolves through sequences of user utterances and system replies. The retrieval of documents relevant to an utterance is difficult due to informal use of natural language in speech and the complexity of understanding the semantic context coming from previous utterances. We adopt the 2019 TREC Conversational Assistant Track (CAsT) framework to experiment with a modular architecture performing in order: (ⅰ) automatic utterance understanding and rewriting, (ⅱ) first-stage retrieval of candidate passages for the rewritten utterances, and (ⅲ) neural re-ranking of candidate passages. By understanding the conversational context, we propose adaptive utterance rewriting strategies based on the current utterance and the dialogue evolution of the user with the system. A classifier identifies those utterances lacking context information as well as the dependencies on the previous utterances. Experimentally, we evaluate the proposed architecture in terms of traditional information retrieval metrics at small cutoffs. Results demonstrate the effectiveness of our techniques, achieving an improvement up to 0.6512 (+201%) for P@l and 0.4484 (+214%) for nDCG@3 w.r.t. the CAsT baseline.