Enhancing Pre-trained Chinese Character Representation with Word-aligned Attention

Most Chinese pre-trained encoders take a character as a basic unit and learn representations according to character's external contexts, ignoring the semantics expressed in the word, which is the smallest meaningful unit in Chinese. Hence, we propose a novel word aligned attention to incorporate word segmentation information, which is complementary to various Chinese pre-trained language models. Specifically, we devise a mixed-pooling strategy to align the character level attention to the word level, and propose an effective fusion method to solve the potential issue of segmentation error propagation. As a result, word and character information are explicitly integrated at the fine-tuning procedure. Experimental results on various Chinese NLP benchmarks demonstrate that our model could bring another significant gain over several pre-trained models.