Leveraging knowledge graph for domain-specific Chinese named entity recognition via lexicon-based relational graph transformer

Leveraging knowledge graphs (KGs) has been an emerging direction to improve the performance of deep learning-based Chinese named entity recognition (CNER). Nevertheless, most existing methods directly inject correlated words into sentences but ignore word boundaries that are crucial for CNER. Conflicts among incorrect word segmentations may misguide models to predict incorrect labels. To solve this problem, this work investigates a novel lexicon-based relational graph transformer (LRGT), which combines relational graph-structured inputs and transformer tailored for lexicon-augmented CNER. In LRGT, characters and self-matched lexicon words are fully interacted through a two-phase relational graph softmax message passing mechanism. The finally enhanced character representation in LRGT dynamically integrates both lexical and relative positional information, which is distinguishable for the identification. Results on four benchmark datasets demonstrate that LRGT significantly outperforms several state-of-the-art methods. We further demonstrate that LRGT with KG achieves higher performance on two public specific-domain CNER datasets. LRGT performs up to 3.35 times faster than several typical baselines while achieving better F1-score by up to 1.92% and 2.24%, respectively.