My research interests include various aspects of information retrieval, recommender systems, and text mining. My major contributions and current research focus include:

Neural Information Retrieval: Following our initial study on distributed representations for information retrieval (ICTIR16a, ICTIR16b, CIKM16), we have developed multiple neural network models for the fundamental IR tasks. They include relevance-based word embedding (SIGIR17a), context-aware query representation (WWW17), and semi-structured document representation (WSDM18). We have also invented the first learning to rank model that can retrieve (as opposed to re-rank) documents from large-scale collections, called standalone neural ranking model (CIKM18a).

Weak Supervision for IR: We have developed a student-teacher learning model for training machine learning IR models with no labeled data, called weak supervision. We have empirically shown the effectiveness of weak supervision in various IR tasks, including ad-hoc retrieval (SIGIR17b), representation learning (SIGIR17a), and query performance prediction (SIGIR18a). We have also theoretically study weak supervision for information retrieval (ICTIR18). Check out our SIGIR 2018 workshop on learning from limited or noisy data (LND4IR).

Conversational Information Seeking: I have been working on conversational search and recommendation with a focus on mixed-initiative interactions, such as asking clarifying questions. We have designed and implemented an extensible platform for conversational information seeking research, called Macaw (see the GitHub repository). We have developed models for generating and selecting clarifying questions (WWW20, SIGIR20a), models for utilizing user response to clarifying questions (SIGIR20b), and offline evaluation methodologies for asking clarifying questions in conversational systems (SIGIR19a). I have recently started new directions related to multi-modal conversational information seeking (SIGIR21). We will participate in the Alexa Prize 2021 (see this news article). Related to conversational search, we have implemented a unified search framework for mobile devices (SIGIR18b, CIKM18b).

IR Meets RecSys: I have made efforts on bridging the gap between the IR and RecSys communities. We have modeled query expansion as collaborative filtering (CIKM16), proposed content-based recommendation models using IR techniques (RecSys17, IRJ18), and designed a joint search and recommendation model (DESIRES18, WSDM20). I co-organized the ACM Recommender Systems Challenge in 2018 on music playlist generation and continuation, which can be seen as a RecSys or IR task (RecSys18). Check out our visions on music recommendation (IJMIR18) and the overview of ACM RecSys Challenge 2018 (TIST19).