Reasoning with Latent Structure Refinement for Document-Level Relation Extraction

Guoshun Nan, Zhijiang Guo, Ivan Sekulic, Wei Lu

Abstract

Document-level relation extraction requires integrating information within and across multiple sentences of a document and capturing complex interactions between inter-sentence entities. However, effective aggregation of relevant information in the document remains a challenging research question. Existing approaches construct static document-level graphs based on syntactic trees, co-references or heuristics from the unstructured text to model the dependencies. Unlike previous methods that may not be able to capture rich non-local interactions for inference, we propose a novel model that empowers the relational reasoning across sentences by automatically inducing the latent document-level graph. We further develop a refinement strategy, which enables the model to incrementally aggregate relevant information for multi-hop reasoning. Specifically, our model achieves an F1 score of 59.05 on a large-scale document-level dataset (DocRED), significantly improving over the previous results, and also yields new state-of-the-art results on the CDR and GDA dataset. Furthermore, extensive analyses show that the model is able to discover more accurate inter-sentence relations.

Introduction

Extracting relationships between entities across an entire document is a significant challenge in Natural Language Processing because the "reasoning path" often spans multiple sentences. Traditional Relation Extraction (RE) systems typically operate at the sentence level, identifying relationships within a single string of text. However, real-world information is often scattered across a document. For instance, to conclude a relationship between two entities, a system might need to link a mention in the first sentence to a pronoun in the second, which then connects to a third entity in the final sentence.

Existing document-level RE models usually rely on static graphs built using "hard" rules, such as syntactic dependency trees or co-reference links. These human-defined heuristics are often incomplete or noisy, failing to capture the complex, non-local interactions required for sophisticated inference. This paper introduces Latent Structure Refinement (LSR), a model that treats the document structure as a latent variable, automatically inducing and iteratively refining a document-level graph to perform multi-hop reasoning.

Methods

The LSR model moves away from fixed structures in favor of a dynamic, learnable approach. The architecture consists of three main stages:

  1. Node Construction: The model extracts three types of nodes: Mention Nodes (individual entity occurrences), Entity Nodes (global representations created by averaging mentions), and Meta Dependency Path (MDP) Nodes (tokens on the shortest dependency path between mentions within a sentence).
  2. Dynamic Reasoner: Instead of a pre-defined adjacency matrix, a Structure Induction module based on a variant of the Matrix-Tree Theorem calculates edge probabilities between nodes, effectively "learning" the document's internal logic.
  3. Iterative Refinement: The model stacks $N$ reasoning blocks. In each block, node representations are updated using a Densely Connected Graph Convolutional Network (GCN), and these updated features are used to refine the latent graph for the next step.
LSR Architecture and Iterative Refinement

This iterative process allows the model to incrementally aggregate information, moving from shallow local connections to complex non-local interactions.

Results and Discussion

The model was evaluated on three major benchmarks: DocRED (large-scale general domain), CDR (Chemical-Disease Reactions), and GDA (Gene-Disease Associations).

Key Findings

  • Latent beats Static: Inducing a graph structure is more effective than providing a fixed one. By allowing the model to learn relevant connections, it discovers reasoning paths that manual heuristics might miss.
  • The Power of Iteration: The first iteration typically captures local connections, while subsequent refinements capture "non-local" interactions where distant entities are linked.
  • SOTA Performance: LSR achieved an F1 score of 59.05 on DocRED when paired with BERT, setting a new state-of-the-art at the time.
  • Inter-sentence Superiority: The model showed significant gains on relations spanning multiple sentences, outperforming previous graph-based models like EoG.
Performance Comparison and Refinement Analysis

As shown in Figure 6, LSR consistently outperforms baseline models (QAGCN, EoG, AGGCN) across Overall, Intra-sentence, and Inter-sentence F1 scores. The refinement blocks contribute to performance stability and peak around 2-3 iterations for inter-sentence reasoning.

Conclusion

LSR demonstrates that learning document structure as a latent variable is a powerful approach for document-level relation extraction. While the model currently relies on external parsers for node construction (specifically MDP nodes), which can introduce noise in specialized domains like biomedicine, it significantly advances the ability to perform multi-hop reasoning across sentences. Future work will focus on inducing nodes directly from text to create a truly end-to-end, structure-free reasoning system.

Summary updated

Abstract

Document-level relation extraction requires integrating information within and across multiple sentences of a document and capturing complex interactions between inter-sentence entities. However, effective aggregation of relevant information in the document remains a challenging research question. Existing approaches construct static document-level graphs based on syntactic trees, co-references or heuristics from the unstructured text to model the dependencies. Unlike previous methods that may not be able to capture rich non-local interactions for inference, we propose a novel model that empowers the relational reasoning across sentences by automatically inducing the latent document-level graph. We further develop a refinement strategy, which enables the model to incrementally aggregate relevant information for multi-hop reasoning. Specifically, our model achieves an F1 score of 59.05 on a large-scale document-level dataset (DocRED), significantly improving over the previous results, and also yields new state-of-the-art results on the CDR and GDA dataset. Furthermore, extensive analyses show that the model is able to discover more accurate inter-sentence relations.
BibTeX
@inproceedings{nan2020reasoning,
  author = {Guoshun Nan and Zhijiang Guo and Ivan Sekulic and Wei Lu},
  title = {{Reasoning with Latent Structure Refinement for Document-Level Relation Extraction}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
  url = {https://openreview.net/forum?id=BWeCjrGW4Q2},
}
Figure