Dual Interactive Attention Network for Joint Entity and Relation Extraction

The joint entity and relation extraction method establishes a bond between tasks, surpassing sequential extraction based on the pipeline method. Many joint works focus on learning a unified representation for both tasks to explore the correlations between Named Entity Recognition (NER) and Relation Extraction (RE). However, they suffer from the feature confusion that features extracted from one task may conflict with those from the other. To address this issue, we propose a novel Dual Interactive Attention Network to learn independent representations and meanwhile guarantee bidirectional and fine-grained interaction between NER and RE. Specifically, we propose a Fine-grained Attention Cross-Unit to model interaction at the token level, which fully explores the correlation between entity and relation. To obtain task-specific representation, we introduce a novel attention mechanism that can capture the correlations among multiple sequences from the specific task and performs better than the traditional self-attention network. We conduct extensive experiments on five standard benchmarks (ACE04, ACE05, ADE, CoNLL04, SciERC) and achieve state-of-the-art performance, demonstrating the effectiveness of our approach in joint entity and relation extraction.