Exploiting Document-Based Features for Clarification in Conversational Search

Asking clarifying questions in order to elicit user's information need is becoming an integral part of modern conversational search systems. Current work heavily relies on pre-collected clarifying questions or large-scale query logs. However, such work is very limited given that collecting all possible clarifying questions on a collection is not feasible. Moreover, modeling clarification based on query reformulation limits a model only to head queries with several occurrences in the log. In this work, we aim to address these limitations by exploiting several document-and ranking-based features to generate clarifying questions. We hypothesise that we can acquire enough evidence about different aspects of a query and extract useful facets to generate clarifying questions about. Specifically, we utilise Part-Of-Speech tagging, entity linking, and topic modelling in order to extract features from the ranked list of documents. Among the extracted features, we then extract potentially useful facets based on three different strategies, aimed to capture feature distinctiveness across documents. We then construct clarifying questions based on the extracted facets that are given to crowdsourcing workers to be evaluated in terms of usefulness. Moreover, our findings show significant improvements (+38% nDCGO3) in document retrieval performance with facet-expanded queries.