Research On The Fusion Of Character-level And Segment-level Model For Chinese NER

Named Entity Recognition is a fundamental task in Natural language processing.Its aim is to recognize named entities in texts.Unlike English texts which every word is divided by space tokens,the character sequence in Chinese texts is continuous.So,the Chinese NER task is generally divided into two categories: character-level models and word-level models.Most of the research work focuses on character-level models and treats NER as a character-level sequence labeling task.But as our research finds out that character representations cannot utilize word-level information to classify named entities.We propose several methods to tackle this problem.The work in this thesis is as follows:1.Based on the common way of Chinese representations,we propose a new way of Chinese representations method : combining character-level information with word-level information in Chinese NER task.By fully exploiting the character-level information and word-level semantic information,we observe large improvements against single character-based models in Chinese NER task.2.We propose a named entity recognition architecture based on searching entity segments.This model behaves differently than character-level models or word-level models.It can use the word-level information without the interference of segmentation error.The model tries to increase its ability to classify named entities by using the information from the boundary word and the inside feature of the segments in the representation of the segments.3.This thesis designs a fusion model which is based on the character-level model and the segment-level model.By using the segment-level model as the fundamental architecture,we try to integrate the segment-level information with character-level information.By the integration of the two ways of Chinese representations,the new model can distill complex and representative features which are crucial in Chinese NER task.Experiments show that the fusion model achieves better results than character-level models and segment-level models.