HIN: Hierarchical Inference Network for Document-Level Relation Extraction

Hengzhu Tang, Yanan Cao, Zhenyu Zhang, Jiangxia Cao, Fang Fang, Shi Wang, Pengfei Yin

Abstract

Document-level Relation Extraction (RE) requires identifying relationships between entities that may span multiple sentences, necessitating complex reasoning and information aggregation. We present the Hierarchical Inference Network (HIN), a model designed to capture relational facts across three granularities: entity, sentence, and document levels. By explicitly modeling this hierarchy, HIN achieves state-of-the-art performance on the DocRED dataset, effectively handling long-range dependencies and multi-hop reasoning.

Introduction

Traditional Relation Extraction (RE) models primarily focus on sentence-level tasks, where the goal is to identify relations between entities within a single sentence. However, real-world documents often distribute evidence across multiple sentences. For instance, determining a "creator" relationship might require linking a show's name in one sentence to a writer's name several sentences later.

Document-level RE introduces significant challenges, including the need for robust coreference resolution, multi-hop reasoning to connect disparate facts, and the ability to filter noise from irrelevant sentences. To address these, we evaluate models on the DocRED dataset, which contains over 5,000 documents where approximately 40.7% of relational facts require cross-sentence reasoning and 61.1% require some form of complex inference.

Methods

The Hierarchical Inference Network (HIN) processes document information through a bottom-up architecture consisting of three primary layers:

  1. Entity-Level Inference: Entities are represented by averaging their mentions. Inspired by knowledge graph embeddings, HIN applies a "translation constraint" ($e_{subject} + relation \approx e_{object}$) and a bi-affine layer to model entity interactions in latent subspaces.
  2. Sentence-Level Inference: A BiLSTM with attention generates sentence representations. A semantic matching mechanism compares entity-level information against each sentence to identify relevant evidence.
  3. Document-Level Inference: A hierarchical BiLSTM aggregates sentence-level inferences, using attention to focus on supporting sentences while discarding noise.
Architecture of the Hierarchical Inference Network (HIN).

The final prediction is made by concatenating entity-level and document-level representations, allowing for multi-label classification of entity pairs.

Key Results

HIN demonstrates superior performance on the DocRED benchmark. Notably, HIN-BERT achieves an F1 score of 55.60% on the test set, outperforming previous baselines such as Context-Aware (50.70%) and standard BERT-RE (53.20%).

Performance comparison of HIN against baseline models on the DocRED dataset.

Key insights from the evaluation include:

  • Hierarchical Advantage: Ablation studies show that removing the hierarchical aggregation leads to a 2.81% drop in F1, confirming the importance of multi-granularity modeling.
  • Long-Range Robustness: Unlike many models that struggle as the distance between supporting sentences increases, HIN maintains effectiveness even when evidence is spread across up to 7 sentences.
  • BERT Integration: While HIN performs well with GloVe, the transition to BERT embeddings provides a significant boost, highlighting the model's ability to leverage deep contextual representations.

Conclusion

The Hierarchical Inference Network provides a structured approach to document-level RE by mimicking human-like reading strategies. While HIN sets a new standard on DocRED, future work remains in incorporating external common-sense reasoning to handle cases where relationships are implied rather than explicitly stated. Additionally, improving performance on extremely sparse evidence (8+ sentences) remains an open challenge for the field.

Summary updated

BibTeX
@inproceedings{tang2020hin,
  author = {Hengzhu Tang and Yanan Cao and Zhenyu Zhang and Jiangxia Cao and Fang Fang and Shi Wang and Pengfei Yin},
  title = {{HIN: Hierarchical Inference Network for Document-Level Relation Extraction}},
  booktitle = {PAKDD},
  year = {2020},
  doi = {10.1007/978-3-030-47426-3\_16},
}
Figure