RUCIR at TREC 2019: Conversational Assistance Track

Abstract

In this paper we give a brief overview of the RUCIR group's participation in the TREC 2019 Conversational Assistance Track. All our runs for the Conversational Assistance Track are on the full MS MARCO Conversational Search Sessions dataset and use the online Indri retrieval system hosted at CMU. For the Conversational Assistance Track, our runs try to solve conversational retrieval problems from two directions: One is to improve the search results by modifying the user's current query, including query reference resolution and incorporate the information from user's history queries in the same session. Run 1, Run 2 and Run 4 use this method. The other direction is to design a neural network to model user's global search intent and current search intent to get the retrieval results and run3 uses this method.