A new binary tagging-based model integrated with unsupervised tree hierarchy for relational triple extraction

Relational triple extraction (RTE) extracts entities and relations from unstructured text, serving as a crucial task for various NLP applications. Traditional pipeline approaches often face error propagation issues. The cascade binary tagging (CBT) method was introduced to mitigate this by linking entity recognition and relation extraction through shared parameters. However, CBT struggles with long-distance dependencies between subject and object entities, weakening performance. To address this, the COnRel model is proposed, integrating shallow and deep hierarchical information into the CBT framework. An unsupervised hierarchy parser generates multi-level tree structures, and a weight-transform method assigns higher weights to words closer in hierarchy to subject entities. This improves semantic representation of the subjects. In experiments, COnRel with shallow hierarchy outperforms the baseline model CasRel on the WebNLG dataset, and the full model, including deep hierarchy, excels on both WebNLG and NYT datasets, particularly for sentences 20-50 words in length.