在基于条件随机场的中文命名实体识别任务中,现有表示学习方法学习到的特征存在语义表示偏差,给中文命名实体识别带来噪音。针对此问题,提出了一种基于位置敏感Embedding的中文命名实体识别方法。该方法将上下文位置信息融入到现有的Embedding模型中,采用多尺度聚类方法抽取不同粒度的Embedding特征,通过条件随机场来识别中文命名实体。实验证明,该方法学习到的特征缓解了语义表示偏差,进一步提高了现有系统的性能,与传统方法进行相比,F值提高了2.85%。
In the task of Chinese named entity recognition based on conditional random fields, there is semantic bias on fea- tures learned by present representation learning. This paper presented a Chinese named entity recognition method based on po- sition-sensitive embedding model. This method applied the position information to the embedding model and used multi-scale word clustering to extract different size features. And then it recognized Chinese named entity with conditional random fields. The experiment shows that, this method improves the F-score by 2.85, compared to traditional methods.