融合迁移学习的中文命名实体识别

Research on Chinese Named Entity Recognition Fusing Transfer Learning

命名实体识别是自然语言处理研究领域中的一项很重要的基础性任务,是实体关系抽取和事件抽取等高层任务重要基石.如何在缺乏标注语料或只有少量标注语料条件下,提高命名实体识别的性能是自然语言处理领域的一个重要研究方向.针对这一问题,提出一种基于实例的迁移学习算法——TLNER_AdaBoost.该方法通过自动调整训练集中实例样本的权重和计算辅助训练样本的迁移能力来提高训练语料质量,并选取采用不完全标注语料的自学习方法和采用完全标注语料的基于条件随机场的方法来对该方法进行实验对比分析.经实验对比分析得知,本文方法在提高命名实体识别的准确率、召回率和F值的同时,大大降低了人工标注语料的工作量.

Named entity recognition is a very important basic task in natural language processing research;it is the high-level task cornerstone of the entity relationship extraction and event extraction. When we are faced with the lack of annotation corpus, or only a small amount of tagging corpus, how to improve the performance of named entity recognition was an important research direction in natural language processing. In order to solve this problem, we present an instance-based transfer learning algorithm TLNER_Ada- Boost. This method improved the quality of training corpus by adjusting automatically weight of the whole training instances samples and calculated the transfer ability of assist training samples, then improved the performance of named entity recognition. We use the self-learning method with incomplete tagging corpus and conditional random field method with completely tagging corpus as the comparison method to do experiment contrast analysis. The experiment results show that our method can not only improve the rate of accuracy, recall and F-Measure of named entity recognition, but also decrease the workload of manual tagged corpus.