Source-linked AI summary
Document-level Relation Extraction as Semantic Segmentation
Ningyu Zhang, Xiang Chen, Xin Xie, Shumin Deng, Chuanqi Tan, Mosha Chen, Fei Huang, Luo Si, Huajun Chen
TL;DR
Document-level relation extraction must classify multiple entity pairs, including relations spanning sentences, while prior approaches overlook interdependency among triples. DocuNet predicts an entity-level relation matrix as a semantic-segmentation mask, combining contextual encoding, U-shaped global modeling, and balanced softmax; it reports state-of-the-art performance on DocRED, CDR, and GDA.
Problem
Document-level RE must identify relations among multiple entity pairs and across sentences, but prior models often overlook interdependency among multiple relational triples.
Method
DocuNet reformulates document-level RE as semantic segmentation over an entity-level relation matrix, using an encoder, U-shaped segmentation module, and balanced softmax.
Results
DocuNet achieves state-of-the-art performance on the DocRED, CDR, and GDA benchmark datasets.
Takeaways & Limitations
The results support capturing both local context and global information among triples for document-level relation extraction.
Abstract
from arXiv · showhide
Document-level relation extraction aims to extract relations among multiple entity pairs from a document. Previously proposed graph-based or transformer-based models utilize the entities independently, regardless of global information among relational triples. This paper approaches the problem by predicting an entity-level relation matrix to capture local and global information, parallel to the semantic segmentation task in computer vision. Herein, we propose a Document U-shaped Network for document-level relation extraction. Specifically, we leverage an encoder module to capture the context information of entities and a U-shaped segmentation module over the image-style feature map to capture global interdependency among triples. Experimental results show that our approach can obtain state-of-the-art performance on three benchmark datasets DocRED, CDR, and GDA.
1 Introduction
Document-level relation extraction must identify relations across sentences and among multiple entity pairs, where global interdependencies can guide classification. DocuNet reformulates the task as semantic segmentation over an entity-level relation matrix and reports state-of-the-art performance on three benchmarks.
- Above 40.7% of relations can only be identified at the document level, motivating models that extract relations beyond individual sentences.
- Document-level RE classifies multiple entity pairs at once, including inter-sentence pairs with long-distance dependencies.Relations such as those between Eldersburg and U.S. or Baltimore and U.S. require information from different sentences.
- Interdependency among relation triples can guide classification, constrain implausible relations, and support inference across entities.The paper illustrates this with Maryland–U.S., Eldersburg–Maryland, and Eldersburg–U.S. relations.
- DocuNet formulates document-level RE as semantic segmentation over an entity-level relation matrix, using an encoder for context and U-shaped segmentation for global interdependency.Relevant entity-pair features are treated as an image, and each entity pair receives a relation-type mask.
- DocuNet captures local context and global triple interdependency, and achieves state-of-the-art performance on three benchmark datasets.
2 Related Work
Prior document-level RE work uses graph-based reasoning or pretrained transformers, but much of it emphasizes local entity representations over global connections among triples. This paper instead draws on semantic segmentation and U-Net to model those interactions.
- Document-level RE research has used graph-based models and transformer-based models to represent and reason over document information.Examples include graph-enhanced attention, heterogeneous graph models, BERT variants, hierarchical inference, and ATLOP.
- Graph-based approaches model interactions among text spans, sentences, and potential relation instances with graph structures and graph neural methods.
- Transformer-based approaches use pretrained language models to capture long-distance dependencies without explicitly generating document graphs.
- Most previous studies emphasized local entity representations and overlooked high-level global connections and interdependency among multiple relations.
- DocuNet is inspired by semantic segmentation and U-Net, using a contracting path for context and an expanding path with skip connections for localization.The paper presents this as the first formulation of relation extraction as a semantic segmentation task.
3 Methodology
DocuNet represents document-level relation extraction as prediction over an entity-level relation matrix, combining contextual entity encoding with U-shaped segmentation to model local and global information.
- 3.1 Preliminary: The task is formulated as an N × N entity-level relation matrix, where each entry represents the relation type between an entity pair.Entities are ordered by their first appearance in the document, and the matrix parallels the pixel-level mask in semantic segmentation.
- 3.2 Encoder Module: A pretrained language-model encoder marks entity mentions and aggregates overlapping dynamic-window representations to produce entity embeddings.Mention markers identify entity positions, while logsumexp pooling accumulates signals across mentions.
- 3.2 Encoder Module: Entity-pair features combine similarity-based and context-based relevance strategies, using element-wise, cosine, and bilinear similarities or entity-aware attention.The context-based strategy uses attention weights, document embeddings, learnable matrices, and multiple transformer heads.
- 3.3 U-shaped Segmentation Module: The entity-level feature tensor is treated as a D-channel image, and a U-Net module performs semantic-segmentation-style relation prediction.The segmentation module is designed to capture global interdependency among entity-pair relations while distributing aggregated information across output positions.
- 3.4 Classification Module: Given entity-pair embeddings and matrix representations, a feedforward network and bilinear function produce relation probabilities.The probability computation uses the entity-pair representation from Y together with learnable relation-specific parameters.
- 3.4 Classification Module: Training uses balanced softmax to address the imbalanced relation distribution in which many entity pairs have the NA relation.The method adds a category 0 and separates target-category and non-target-category scores around a threshold.
4 Experiments
DocuNet is evaluated on DocRED, CDR, and GDA against graph-based and transformer-based baselines, with ablations and case studies examining its components and global-information modeling.
- 4.1 Dataset: DocuNet is evaluated on three document-level relation extraction datasets: DocRED, CDR, and GDA.DocRED is a crowdsourced general-domain dataset, while CDR and GDA target biomedical interactions.
- 4.3 Results on the DocRED Dataset: DocuNet-BERTbase outperforms ATLOP-BERTbase on DocRED, while DocuNet with RoBERTa-large achieves a new state-of-the-art result.The authors also report holding first position on the CodaLab scoreboard as of January 20, 2021, without external data.
- 4.4 Results on the Biomedical Datasets: DocuNet-SciBERTbase improves F1 by 6.9% on CDR and 1.4% on GDA compared with ATLOP-SciBERTbase.SciBERT is used because it is pretrained on scientific publication corpora.
- 4.5 Ablation Study: Removing U-shaped segmentation decreases dev F1 by 2.18%, while removing balanced softmax decreases it by 1.32%.The ablation results identify both modules as beneficial and particularly sensitive to performance.
- 4.5 Ablation Study: The context-based strategy exceeds the similarity-functions strategy by 0.84 F1 in the ablation study.The similarity-based variant directly computes entity correlation as the input matrix instead of using contextual information.
- 4.6 Case Study: In a case study, only DocuNet-BERTbase deduces shared performer and publication-date relations across “Without Me” and “The Eminem Show”.The inferred values are “Eminem” and “May 26, 2002”, respectively; the authors connect this behavior to cross-sentence relational reasoning.
- 4.7 Analysis: The U-shaped model consistently outperforms its ablated counterpart, with larger improvements as the number of entities increases.This analysis is performed on DocRED development groups divided by entity count.
5 Conclusion and Future Work
The paper formulates document-level relation extraction as semantic segmentation and introduces DocuNet to capture local and global information. It reports better performance than baselines and suggests convolution over entity-entity relation matrices may support relational reasoning.
- 5 Conclusion and Future Work: The paper formulates document-level relation extraction as semantic segmentation and introduces the Document U-shaped Network.This is presented as the study’s first step toward applying semantic-segmentation ideas to document-level relation extraction.
- 5 Conclusion and Future Work: DocuNet achieves better performance than baselines by capturing local and global information.The conclusion summarizes the model’s reported performance across the evaluated benchmarks.
- 5 Conclusion and Future Work: Convolution over the entity-entity relation matrix may implicitly conduct relational reasoning among entities.The authors identify this as an empirical observation rather than a formally established guarantee.
- 5 Conclusion and Future Work: Future work will apply the approach to other span-level classification tasks, including aspect-based sentiment analysis and nested named entity recognition.These applications are proposed as extensions of the current approach.