Abstract:
Conversational retrieval is a process where a user posts a series of
queries/questions to a retrieval system. Our work addresses two challenges unique to this setting. First, we designed a novel retrieval framework for conversational retrieval. We combine multiple representations for a given query and multiple representations of the document to improve retrieval effectiveness. Second, we addressed the novel
challenge of predicting the effectiveness of retrieval performed for
a query in the conversation; i.e., estimating retrieval effectiveness with no relevance judgments. We present two prediction methods. The first utilizes information induced from previous queries in the conversation. The second prediction method uses information induced from alternative forms of representation for the given query.
Empirical evaluation attests to the merits of our retrieval and prediction methods.
https://technion.zoom.us/j/91808480009
