A novel chinese relation extraction method using polysemy rethinking mechanism

Qihui Zhao, Tianhan Gao, Nan Guo

Abstract: The methods of Chinese relation extraction(CRE) based on the neural network can be divided into two categories according to the input mode(word-based and character-based). The performance of word-based models depends on the accuracy of word segmentation. Unfortunately, there are still errors in existing word segmentation tools (methods). Among the character-based models, Lattice LSTM-based models have been successful in CRE. However, such RNN-based models cannot meet the requirements of parallel computing and thus have natural drawbacks in model training and inference. There is much word polysemy in Chinese that constrains the performance of CRE. Most CRE models are built on English datasets, which often perform poorly on Chinese datasets. To address the above issues, we propose a method for CRE with the Polysemy Rethinking Mechanism. In this method, (1) we use a CNN-based architecture in which input is characters. It can incorporate word-level information through the lexicon to correct the error caused by word segmentation. (2) We propose a Polysemy Rethinking Mechanism, which can alleviate the problems caused by multiple meanings of one word by adding multiple sense information to the model. (3) Compared with the Lattice LSTM-based model, our model improves computational efficiency to gain results. We conduct experiments on two real-world datasets of CRE. The results show that our method achieves better performance than the state-of-the-art ones.