PaperlayerView exact revision on arXiv

Source-linked AI summary

UniRE: A Unified Label Space for Entity Relation Extraction

Yijun Wang, Changzhi Sun, Yuanbin Wu, Hao Zhou, Lei Li, Junchi Yan

arXiv:2107.04292v1cs.CLcs.LG

TL;DR

Joint entity-relation models may retain separate label spaces, and prior work describes benefits to separate encoders from capturing distinct contextual information. UniRE labels word-pair cells from a unified entity-relation label space; it performs better on ACE04 and SciERC, competitively on ACE05, and decodes 10x faster than the exact pipeline implementation.

  • Problem

    Joint models may share encoders yet retain separate entity and relation label spaces, while separate-label models are argued to capture distinct contextual information.

    Source: §1 · p. 2

  • Method

    UniRE assigns each word-pair cell a label from the union of entity and relation types, representing entities as squares and relations as rectangles.

    Source: §1 · p. 2

  • Results

    UniRE is better on ACE04 and SciERC and competitive on ACE05; decoding is 10x faster than the exact pipeline implementation.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The table formulation retains expressiveness for overlapping, directed, and undirected relations while unifying entity and relation labeling.

    Source: §1 · p. 2

    The task formulation assumes there are no overlapping entities in a sentence.

    Source: §3.1 · p. 3

Abstract

from arXiv · show

Many joint entity relation extraction models setup two separated label spaces for the two sub-tasks (i.e., entity detection and relation classification). We argue that this setting may hinder the information interaction between entities and relations. In this work, we propose to eliminate the different treatment on the two sub-tasks' label spaces. The input of our model is a table containing all word pairs from a sentence. Entities and relations are represented by squares and rectangles in the table. We apply a unified classifier to predict each cell's label, which unifies the learning of two sub-tasks. For testing, an effective (yet fast) approximate decoder is proposed for finding squares and rectangles from tables. Experiments on three benchmarks (ACE04, ACE05, SciERC) show that, using only half the number of parameters, our model achieves competitive accuracy with the best extractor, and is faster.

1 Introduction

Joint entity-relation models may still separate the subtasks’ label spaces; UniRE assigns labels from a unified space to word-pair cells. It reports better performance on ACE04 and SciERC, competitive performance on ACE05, and faster decoding.

1 Introduction

  • Even joint models commonly retain separate entity and relation label spaces, prompting the question of whether joint encoders and decoders warrant joint label spaces.Zhong and Chen (2020) argue that separate label spaces can benefit from separate encoders that capture distinct contextual information.Source: §1 · p. 2
  • A prior compound-label sequence model cannot detect overlapping relations or isolated entities.Overlapping relations involve an entity participating in multiple relations.Source: §1 · p. 2
  • UniRE labels each word-pair cell from the union of entity and relation types, representing entities as diagonal squares and relations as off-diagonal rectangles.The formulation is described as retaining expressiveness for overlapping, directed, and undirected relations.Source: §1 · p. 2
  • The extractor fills the table with biaffine predictions and structural regularizations, then approximately decodes squares and rectangles by finding split points.The proposed decoder differs from table-filling approaches that decode sequentially and fill tables incrementally.Source: §1 · p. 2
  • UniRE is better on ACE04 and SciERC and competitive on ACE05; decoding is 10x faster than the exact pipeline implementation.Its shared encoder uses half the parameters of the separate encoder in Zhong and Chen (2020).Source: §1 · p. 2

2 Task Definition

The task is to extract typed entity spans and typed relations between entities from a sentence. UniRE formulates this as labeling every word-pair cell in a table, then decoding entities and relations as rectangles.

2 Task Definition

  • An entity is a continuous word span with a predefined type, while a relation is a typed triplet linking two entities.The possible entity and relation types are denoted Ye and Yr, respectively.Source: §2 · p. 2
  • For each sentence, the table assigns every word-pair cell a label from entity types, relation types, or the no-relation label.Cells spanning an entity receive its type; cells spanning the two entities in a relation receive that relation type.Source: §2 · p. 3
  • At test time, decoding becomes a rectangle-finding problem, which the paper addresses with a joint decoding algorithm.The table formulation covers the sentence’s word pairs and labels cells corresponding to entity spans and relation pairs.Source: §2 · p. 3

3 Approach

The approach combines PLM-based biaffine table filling, structural training constraints, and a staged decoder for extracting entities and relations.

3.1 Biaffine Model

  • PLM contextualizes words, optionally extending the sentence to a fixed window W=200 for long-range dependencies; head/tail MLP projections support biaffine word-pair scoring.The biaffine-model formulation assumes no overlapping entities within a sentence.Source: §3.1 · p. 3

3.2 Table Filling

  • For each word pair, its score vector is softmaxed into a categorical distribution over Y, and the annotated cell label supplies the training target.Source: §3.2 · p. 3

3.3 Constraints

  • Training supplements independent cell-label likelihood with symmetry for entity and symmetric-relation tables and implication that relation probabilities cannot exceed argument-entity probabilities; all three losses are jointly optimized.Source: §3.3 · pp. 3–4

3.4 Decoding

  • Adjacent-row and column probability-vector distances are averaged, and positions exceeding threshold α=1.4 become span boundaries, yielding O(|s|) span decoding.Source: §3.4 · p. 5
  • After span proposals, entity types come from diagonal-square averages, then relation types from pairwise-entity rectangle averages, each selecting the highest-scoring label or no-label.Source: §3.4 · p. 5

4 Experiments

Experiments evaluate UniRE on three benchmarks and report benchmark comparisons alongside ablations, speed, threshold, context-window, and error analyses.

4 Experiments

  • Experiments use ACE04, ACE05, and SciERC, with micro-averaged precision, recall, and F1 under strict evaluation.Hyperparameters are tuned on ACE05 development data using averaged entity and relation F1, then held fixed on ACE04 and SciERC.Source: §4 · pp. 5–6

4.1 Performance Comparison

  • UniRE is best on ACE04 and SciERC and comparable on ACE05; versus the previous best joint model, entity F1 rises by +0.9 and +0.7 and relation F1 by +3.4 and +1.7 on ACE04 and ACE05, respectively.The overall evaluation compares prior work and UniRE across the three datasets.Source: Table 3 · p. 6, §4.1 · pp. 6–7

4.3 Inference Speed

  • With W = 200, UniRE is more than 10× faster than the pipeline method and has comparable or better relation performance; it is competitive with the faster approximate pipeline version but performs better.Using W = 100 accelerates inference further with slight performance drops, while hard decoding is much slower than UniRE.Source: Table 5 · p. 7, §4.3 · pp. 7–8

4.2 Ablation Study

  • Removing symmetrical loss decreases SciERC entity and relation performance by 1.1 and 1.4 points, respectively; removing implication loss harms ACE05 relation performance by 1.0 point, showing structural information is useful.Source: §4.2 · p. 7

4.4 Impact of Different Threshold α

  • At α = 1.4, the authors report a general best setting across ACE04, ACE05, and SciERC; span and entity performance sharply decline as α rises from 1.4 to 1.5.Relation performance begins declining slowly from α = 1.5; the authors attribute the higher relation threshold to sparse relations and entities that do not participate in any relation.Source: Figure 5 · p. 8, §4.4 · p. 8

4.5 Context Window and Logit Dropout Rate

  • Entity and relation performance improve significantly from W = 100 to W = 200, drop sharply from W = 200 to W = 300, and are best at dropout rate p = 0.2.The final model uses W = 200 and p = 0.2.Source: Table 6 · p. 8, §4.5 · p. 8

4.6 Error Analysis

  • Not-found errors exceed type errors for both entities and relations, while span-splitting errors are relatively few.The authors attribute the not-found errors to class imbalance in table filling and reserve this problem for future work; examples also show decoding correcting biaffine prediction errors using table structure and neighboring cells.Source: §4.6 · p. 8, Figure 6 · p. 9

5 Related Work

Related work is organized around separate versus unified label spaces for entity and relation extraction. The paper follows unified-label-space approaches with a two-dimensional table intended to handle overlapping relations.

5 Related Work

  • The related-work discussion divides existing entity-relation extraction methods into two categories according to their adopted label space.Source: §5 · p. 9
  • Separate-label-space methods treat entity recognition and relation classification as distinct tasks, ranging from independent pipeline models to shared-encoder models with separate decoders and joint decoding algorithms.Table filling is also described as parameter sharing in a table structure.Source: §5 · p. 9
  • Unified-label-space approaches cast extraction as tagging, transition-based parsing, or sequence-to-sequence generation, while the proposed 2D table targets overlapping relations without requiring expert-designed transition systems.The paper contrasts its approach with Zheng et al. (2017) and Wang et al. (2018).Source: §5 · p. 9

6 Conclusion

The paper presents a unified label space and a table-based model for extracting entities and relations. Experiments on three benchmarks are reported to show state-of-the-art performance and promising efficiency.

6 Conclusion

  • Experiments on three benchmarks show the proposed method achieves state-of-the-art performance and promising efficiency.Source: §6 · p. 9
  • The model represents entities as squares and relations as rectangles in a novel table, then fills it with a biaffine model and decodes the predictions jointly.Source: §6 · p. 9
  • The method extracts entities and relations in a unified label space to mine interaction between the two subtasks.Source: §6 · p. 9

A Decoding Algorithm

The appendix points to Algorithm 1 for a formal description of the decoding algorithm.

A Decoding Algorithm

B Datasets

The experiments use ACE04, ACE05, and SciERC, with dataset-specific splits and corrected annotations for undirected relations.

B Datasets

  • ACE04 uses 5-fold cross-validation, while ACE05 uses 351 training, 80 validation, and 80 test examples; ACE04 also reserves 10% of training data for development.Both ACE corpora annotate 7 entity types and 6 relation types.Source: B Datasets · p. 12
  • SciERC contains 500 scientific abstracts, annotates 6 entity types and 7 relation types, and uses 350/50/100 training, validation, and test examples.Source: B Datasets · p. 12
  • Undirected relations are corrected by adding the reverse triplet, so each corresponds to two rectangles symmetric about the diagonal.Source: B Datasets · p. 12
Loading 2107.04292v1…