Shared Representation Learning with Self-Attention for Cross-Domain Chinese Hedge Cue Recognition

摘要

Hedge is usually used to express uncertainty and possibility. Research on the Chinese hedge cue detection is of great significance to the Chinese factual information extraction. So far, existing approaches focus on learning a model for Chinese hedge detection based on domain-specific corpus. However, the difference between hedge cue distributions in various domains makes domain-specific models difficult to extend to other domains. To make full use of out-of-domain data and reduce the cost of manual data annotation, we propose a novel cross-domain hedge cue detection approach based on shared representations (domain-general representations). Our method uses stacked CNNs model with self-attention to learn the shared semantic representations across different domains through parameter sharing mechanism. Meanwhile, we introduce adversarial training to separate the private features of each domain from the shared representations by iteratively training the generator and the discriminator alternately. Experiments on Chinese Biomedical domain and Wikipedia domain show that our method gets a significant improvement on cross-domain Chinese hedge cue detection, compared with instance-based transfer learning and feature-based transfer learning methods.