Yeon Seonwoo

Ph.D.

Semantic matching has been a core task in many NLP fields, such as machine reading comprehension (MRC), document retrieval, and semantic textual similarity. However, this task has been defined differently in each field depending on the textual sources' complexity and scale. In this talk, I will introduce 1) how semantic matching has been defined in each NLP sub-field, 2) what problems have occurred, and 3) how these problems have been alleviated. I will first talk about a weakly supervised method that enhances the context prediction performance of MRC models. Then, I will talk about a weakly supervised multi-hop document retriever that correctly predicts a combination of documents that semantically aligns with a given question. Finally, I will talk about an unsupervised sentence embedding method that leverages a corpus-level context to alleviate the imprecise sentence embedding problem in semantic textual similarity.