Exploring Query Reformulation for Conversational Information Seeking

Few tasks have been designed for conversational information seeking. To fill this gap, this year’s TREC Conversation Assistance Track (CAsT) is proposed to advance research on conversational search systems. We built a model that first read the dialogue context, then retrieved candidate response information from a large collection of paragraphs. In order to perform passage retrieval task, we first applied the coreference resolution method to format questions into queries, and we use Indri to retrieve top 100 relevant passages. During the second phrase, we applied fine-tuned BERT model to rerank retrieved passage.