By reformulating open-domain question answering as retrieving pre-indexed phrases, phrase retrieval (Seo et al., 2019) has shown great run-time efficiency over existing QA models while simplifying the process of retrieving fine-grained knowledge from unstructured text. In this talk, I'm going to introduce phrase retrieval and the technical challenges it poses, and how we can tackle them by learning dense representations of phrases at scale. We further show how we can use phrase representations as an unstructured knowledge base that goes beyond open-domain question answering and retrieves various text units (e.g., phrases, sentences, passages) for different downstream tasks such as slot filling and dialogue. I'll conclude the talk with the remaining challenges that we need to tackle in the future including domain adaptation and entity generalization.

온라인줌 링크: https://snu-ac-kr.zoom.us/j/86352991372

Jinhyuk Lee is a research professor at Korea University and was formerly a visiting postdoctoral research associate at Princeton University. His research area is based on natural language processing and deep learning. Specifically, he is interested in learning generalizable text retrieval with dense vectors (i.e., dense retrieval) and tackling challenges in biomedical NLP. His work on NLP and BioNLP has been published in top NLP conferences including ACL (2019, 2020, 2021) and EMNLP (2018, 2020, 2021), and bioinformatics journals including Bioinformatics (2019, 2020). He received his Ph.D. degree in Computer Science and Engineering at Korea University. Previously, he received a B.S. in Computer Science and Engineering at Korea University.