Robust triple extraction with cascade bidirectional capsule network

Recent approaches have witnessed the success of neural models for triple extraction. However, we empirically observe that previous approaches may fail for those disambiguate text expressed in a similar context and generate triples that contradict the commonsense. Such issues severely hinder the generalization of triple extraction in real-world applications. Motivated by the capsule networks' power of modeling latent structures and the implicit entity-relation schema, we propose a novel Cascade Bidirectional Capsule Network (CBCapsule) to address those issues. We firstly introduce a cascade capsule network to dynamically aggregate context representations and then propose a bidirectional routing mechanism to encourage interaction between the high level (e.g., relations) and low level (e.g., entities) capsules. Experimental results on three benchmarks show that our proposed approach is more efficient than baselines and has a more robust generalization ability with complex surface forms.

Automated summary

Recent research has shown that previous approaches may fail in handling text ambiguities in similar contexts and lead to contradictions in commonsense. Inspired by capsule networks and implicit entity-relation schema, a novel Cascade Bidirectional Capsule Network is proposed to address these issues. The experimental results demonstrate that the proposed approach is more efficient and has a stronger generalization ability to handle complex surface forms.