Named Entity Recognition For Chinese Social Media With Domain Adversarial Training And Language Modeling

Recent years have seen a surge of interest in natural language processing (NLP) for social media because the massive unstructured data from social media provide valuable information. However, natural language processing in this domain often suffers from the lack of large scale labeled data used for building models. In this paper, we focus specifically on the task of named entity recognition (NER) for Chinese social media. We propose a neural network model for domain adaptation which builds on domain-adversarial training and language modeling. The model is capable of learning from multiple sources of training data: labeled indomain data, labeled out-of-domain data, as well as (large-scale) unlabeled in-domain data. To demonstrate the effectiveness of our approach, we experiment on an enlarged Chinese social media corpus. Results show that the approach outperforms baselines significantly.