Chinese text relation extraction with multi-instance multi-label BLSTM neural networks

Abstract

Recently, deep learning models have emerged as powerful tools for relation extraction. However, little work has been done on relation extraction for the Chinese language. One major challenge for relation extraction in Chinese texts is that Chinese sentences have no obvious word segmentation. This ambiguity increases the possibility of word segmentation errors. Another challenge is the lack of broad-scale Chinese text datasets. In this paper, we propose an attention-based multi-instance multi-label bidirectional long short-term memory network for distantly supervised Chinese relation extraction. Our model takes Chinese character embeddings and position embeddings as input without Chinese word segmentation errors. Then, the attention mechanism is used to extract richer Chinese character and sentence features. Finally, we handle the multi-label nature of relation extraction by using multi-label loss functions in the neural network classifier. Based on the idea of distant supervision, we constructed a new dataset for relation extraction in Chinese texts. Experiments on this dataset show that our method has achieved relatively high performance, and that the proposed network architecture is suitable for Chinese relation extraction. Furthermore, we also ran experiments on a popular English benchmark dataset, and the results show that our method is superior to some existing methods.