Person Name Extraction from TV program Using Pre-trained Language Model and News Headline

In this study, we focus on extracting person names from text contained in from TV programs and TV-CMs broadcast in Japan. Person name extraction from TV programs is a challenging task because person names are often recognized with other entity names such as locations and organizations. To tackle this problem, we experiment with an existing named entity extraction method using a list of names extracted from Wikipedia. However, this method results in low precision in extracting the person names, which leads to the problem that the list of person names is not effective for TV program texts. In this paper, to solve this problem, we propose a person name extraction method using Conditional Random Fields (CRF) with ELMo pre-training language model. Also, we propose to use news headlines to construct the person name extraction model for effective extraction of person names performing on the TV programs. As a result of our experiments, the proposed model trained with training data and news headlines provides the highest precision. These results show that adding the news headlines as external information to TV program data is effective for person name extraction.