Chinese Relation Extraction with Bi-directional Context-Based Lattice LSTM

Chinese entity relation extraction (Chinese RE) is a crucial task for various NLP applications. It aims to automatically extract relationships between entities in Chinese texts, thereby enhancing the accuracy of natural language understanding. Although existing hybrid methods can overcome some of the shortcomings of character-based and word-based methods, they still suffer from polysemy ambiguity, which results in inaccuracy when representing the relationships between entities in text. To address the issue, we propose a Bi-directional Contextbased Lattice (BC-Lattice) model for Chinese RE task. In detail, our BC-Lattice consists of: (1) A context-based polysemy weighting (CPW) module allocates weights to multiple senses of polysemous words from external knowledge base by modeling context-level information, thus obtaining more accurate representations of polysemous words; (2) A cross-attention semantic interaction-enhanced (CSI) classifier promotes exchange of semantic information between hidden states from forward and backward perspectives for more comprehensive representations of relation types. In experiments conducted on two public datasets from distinct domains, our method yields improved F1 score by up to 3.17%.