Source-linked AI summary
Joint Extraction of Entities and Relations Based on a Novel Decomposition Strategy
Bowen Yu, Zhenyu Zhang, Xiaobo Shu, Yubin Wang, Tingwen Liu, Bin Wang, Sujian Li
TL;DR
Joint extraction methods either create redundant entity pairs or overlook the internal dependencies among entities and relations. The paper decomposes the task into head-entity and tail-entity/relation extraction with span-based sequence labeling, achieving state-of-the-art F1 scores across three public datasets. The model is reported to improve normal, overlapping, and multiple relation extraction, although it does not solve entity-pair overlap and uses a training-time gold distance signal.
Problem
Joint extraction must identify related entity pairs while avoiding redundant candidate pairs and preserving dependencies among head-entities, tail-entities, and relations.
Method
The method decomposes extraction into HE and TER subtasks, then models them as span-based sequence-labeling problems with hierarchical boundary tagging and multi-span decoding.
Results
The method achieves state-of-the-art F1 scores of 59.0% (+5.2%), 78.0% (+5.9%) and 83.1% (+21.5%) on normal, overlapping and multiple relation extraction, respectively.
Takeaways & Limitations
Functional decomposition simplifies learning and supports strong performance across normal, overlapping, and multiple relation extraction on three public datasets.
Takeaways & Limitations
ETL-Span cannot solve entity-pair overlapping, and training uses the gold distance to the correct nearest start position to prevent error propagation.
Abstract
from arXiv · showhide
Joint extraction of entities and relations aims to detect entity pairs along with their relations using a single model. Prior work typically solves this task in the extract-then-classify or unified labeling manner. However, these methods either suffer from the redundant entity pairs, or ignore the important inner structure in the process of extracting entities and relations. To address these limitations, in this paper, we first decompose the joint extraction task into two interrelated subtasks, namely HE extraction and TER extraction. The former subtask is to distinguish all head-entities that may be involved with target relations, and the latter is to identify corresponding tail-entities and relations for each extracted head-entity. Next, these two subtasks are further deconstructed into several sequence labeling problems based on our proposed span-based tagging scheme, which are conveniently solved by a hierarchical boundary tagger and a multi-span decoding algorithm. Owing to the reasonable decomposition strategy, our model can fully capture the semantic interdependency between different steps, as well as reduce noise from irrelevant entity pairs. Experimental results show that our method outperforms previous work by 5.2%, 5.9% and 21.5% (F1 score), achieving a new state-of-the-art on three public datasets
1 INTRODUCTION
The paper reframes joint entity-relation extraction as interrelated head-entity and tail-entity/relation subtasks, then solves them through span-based sequence labeling. This decomposition targets redundant entity pairs and preserves dependencies among extraction steps.
- Joint extraction supports knowledge-base construction but must handle overlapping relations, where multiple relations share an entity.
- Prior extract-then-classify methods generate roughly N^2 entity pairs, most labeled non-relation, potentially misleading relation classifiers.
- Unified labeling methods can identify overlapping relations poorly and overlook dependencies among head-entities, tail-entities, and relations.
- The proposed ETL strategy first extracts candidate head-entities, then predicts each head-entity’s corresponding tail-entities and relations.It decomposes p(h, r, t|S) into p(h|S)p(r, t|h, S), allowing TER extraction to use the given head-entity’s semantic and positional information.
- Span-based tagging labels entity or relation types at start and end positions, using a hierarchical boundary tagger and multi-span decoding.For m head-entities, the task becomes 2 + 2m sequence-labeling subtasks.
- 59.0% (+5.2%), 78.0% (+5.9%) and 83.1% (+21.5%) F1 scores establish new state-of-the-art results on normal, overlapping and multiple relation extraction.These results are reported on NYT-single, NYT-multi and WebNLG, respectively.
2 METHODOLOGY
The method decomposes joint extraction into head-entity and tail-entity/relation extraction, then solves these subtasks with span-based boundary tagging and hierarchical decoding. A shared encoder supplies head-entity information to tail-entity and relation extraction, while multi-span decoding handles overlapping outputs.
- Tagging Scheme: The tagging scheme converts head-entity and tail-entity/relation extraction into separate start- and end-position sequence-labeling subtasks.Head entities receive entity-type labels at their boundaries, while tail-entity boundaries receive relation-type labels for a given head entity.
- Tagging Scheme: The model tags each sentence 2 + 2 × m times, where m is the number of head-entities, rather than using n query-position sequences.This design is intended to recognize overlapping relations while reducing the number of tag sequences when m << n.
- Extraction System: The hierarchical boundary tagger provides a unified extraction module for head and tail entities, differing mainly in prior knowledge and output label sets.The same general architecture is applied to entity-type labeling for head entities and relation-type labeling for tail entities.
- Hierarchical Boundary Tagger: The hierarchical boundary tagger predicts end positions using start-position tagging results and relative-position embeddings.The end-position layer receives lower-level tagging results, hidden states, and position embeddings based on the nearest preceding start position.
- Hierarchical Boundary Tagger: The multi-span decoding algorithm combines predicted start and end tags to extract multiple targets from each sentence.It tracks the nearest preceding start position and uses a constant C when no preceding start exists, constraining end positions relative to starts.
- Extraction System: A shared BiLSTM encoder feeds a HE extractor, after which each extracted head entity triggers TER extraction using its semantic and position information.The HE extractor uses token representations and a global sentence representation, while the TER extractor additionally models the head entity, context, and relative distance.
3.1 Experimental Settings
Experiments use three benchmark datasets, evaluate exact triplet extraction with micro-averaged metrics, and compare against established joint and unified-labeling baselines.
- Datasets: The experiments use NYT-single, NYT-multi, and WebNLG benchmark datasets.NYT-single uses 395 manually annotated test sentences; NYT-multi targets overlapping relation extraction with 56,195 training sentences.
- Datasets: The test data are divided into Normal, SingleEntityOverlap, and EntityPairOverlap categories to assess overlapping-relation extraction.Normal has no overlapping entities; EPO has identical entity pairs with different relations.
- Evaluation: A triplet is correct only when its relation type and both entity offsets are correct, using micro Precision, Recall, and F1.Both head and tail offsets must match for an entity to be considered correct.
- Baselines: Baselines include Cotype, NovelTagging, MultiDecoder, MultiHead, PA-LSTM, and GraphRel.These represent jointly trained, unified sequence-tagging, sequence-to-sequence, multi-relation, and extract-then-classify approaches.
- Model Variants: ETL-Span is evaluated alongside a BiLSTM-CRF baseline that replaces the proposed tagger while retaining the input features.The comparison baseline uses a BIES-based tagging scheme.
3.2 Experimental Results and Analyses
ETL-Span achieves state-of-the-art extraction performance across three datasets while improving efficiency and supporting varied relation patterns. Ablations attribute gains to decomposition, joint learning, hierarchical tagging, and span-based decoding, although EPO overlap remains a limitation.
- Main Results: ETL-Span improves over GraphRel by 16.1% and 40.2% F1 on NYT-multi and WebNLG, respectively.The paper attributes these gains to integrating tail-entity and relation extraction and excluding redundant entity pairs.
- Main Results: ETL-Span decodes up to 3.7 times faster than ETL-BIES while using about one-quarter the batch time and one-fifth the GPU memory during training.The paper links the efficiency advantage to avoiding CRF decoding overhead and reducing the tag space.
- Ablation Study: Removing hierarchical tagging reduces F1 by 2.5%, while replacing joint learning with a pipeline reduces F1 by 5.3%.The ablation results support using start-position information for end prediction and training the HE and TER extractors cooperatively.
- Analysis on Different Sentence Types: ETL-Span surpasses OrderRL by 6.1% on Normal, 6.9% on SEO, and 0.6% on EPO overlapping categories.The comparison is reported on the NYT-multi test set.
- Analysis on Different Sentence Types: ETL-Span outperforms baselines across sentences containing 1, 2, 3, 4, or ≥5 triplets, including an 8.8% gain over OrderRL for single-triplet sentences.These experiments evaluate performance by sentence triplet count on NYT-multi.
4 RELATED WORK
Related work approaches joint entity-relation extraction through pipelines, joint representations, unified tagging, or sequence-to-sequence generation. The paper positions its extract-then-label paradigm as combining selective entity identification with structured relation labeling.
- Traditional and Joint Methods: Traditional pipelined methods first extract entity spans and then discover relations between entity mentions.Entity recognition is commonly formulated as sequence labeling, while relation extraction is treated separately.
- Traditional and Joint Methods: Joint learning methods share representations or use graph and table structures to model entities and relations together.Some approaches require complicated feature engineering, while graph convolutional methods consider many word-pair candidates.
- Limitations of Prior Work: Extract-then-classify systems can overwhelm relation classifiers with redundant entity pairs, whereas sequence-to-sequence systems may fail on multi-word entities.Unified tagging methods are introduced as another line of work but are discussed alongside these limitations.
- Proposed Positioning: The proposed extract-then-label paradigm first identifies head-entities likely to participate in target triplets, then labels their tail-entities and relations.The design aims to reduce redundant pairs and better capture correlations between head-entities and tail-entities than unified labeling approaches.
5 CONCLUSIONS
The proposed end-to-end sequence labeling framework uses novel task decomposition for joint entity–relation extraction, achieving state-of-the-art results on three public datasets.
- The framework decomposes joint entity–relation extraction into sequence labeling subtasks.
- The functional decomposition simplifies learning and improves the overall learning outcome.
- Further analysis shows the model handles normal, overlapping, and multiple relation extraction.
- The method achieves a new state-of-the-art on three public datasets.