부산대학교 도서관

An important challenge in named entity recognition (NER) is how to distinguish the correct meanings of polysemy entities in specific contexts. With the emergence of a large number of emerging entities and various new surface forms of entities, this challenge becomes more and more prominent. Another important challenge is that collecting annotations of named entities is usually expensive and laborious, especially in technical domains such as biomedical publications, financial reports, academic literature, etc. The key to solving above two challenges is that the model should be able to recognize entities utilizing information beyond entities. Entity trigger refers to a group of words that can act as explanation to recognize the positions and categories of entities. In view of the above two challenges, we propose an enhanced machine reading comprehension (MRC) framework with Triggers as Explanations for NER (i.e. MRC-TE). Specifically, our framework combines the information contained in entities and triggers into the same model through a multi-task learning method. Experimental results on two dataset CONLL2003 and BC5CDR demonstrates the effectiveness of our framework.