Source-linked AI summary
RESIDE: Improving Distantly-Supervised Neural Relation Extraction using Side Information
Shikhar Vashishth, Rishabh Joshi, Sai Suman Prayaga, Chiranjib Bhattacharyya, Partha Talukdar
TL;DR
Distant supervision scales relation-extraction training but can violate its labeling assumption, while Knowledge Base side information remains underused. RESIDE integrates entity types, relation aliases, and syntactic GCN encodings into a neural extractor. On benchmark datasets, it achieves higher precision across the recall range and remains competitive with limited alias information.
Problem
Distant supervision supplies large training sets but can assign incorrect labels, while Knowledge Base entity types and relation aliases are not fully exploited.
Method
RESIDE combines entity type and relation alias information with syntactic information encoded by Graph Convolution Networks in a neural relation extractor.
Results
RESIDE achieves higher precision over the entire recall range on the Riedel and GDS datasets and outperforms PCNN+ATT and BGWA.
Takeaways & Limitations
RESIDE remains competitive with very limited relation alias information, while performance improves when more aliases are available.
Abstract
from arXiv · showhide
Distantly-supervised Relation Extraction (RE) methods train an extractor by automatically aligning relation instances in a Knowledge Base (KB) with unstructured text. In addition to relation instances, KBs often contain other relevant side information, such as aliases of relations (e.g., founded and co-founded are aliases for the relation founderOfCompany). RE models usually ignore such readily available side information. In this paper, we propose RESIDE, a distantly-supervised neural relation extraction method which utilizes additional side information from KBs for improved relation extraction. It uses entity type and relation alias information for imposing soft constraints while predicting relations. RESIDE employs Graph Convolution Networks (GCN) to encode syntactic information from text and improves performance even when limited side information is available. Through extensive experiments on benchmark datasets, we demonstrate RESIDE's effectiveness. We have made RESIDE's source code available to encourage reproducible research.
1 Introduction
Relation extraction addresses incomplete Knowledge Bases by extracting entity relationships from text, but distant supervision can introduce noisy labels. RESIDE uses entity types, relation aliases, and syntactic GCN representations as additional supervision for improved distant-supervised extraction.
- Motivation: Distant supervision automatically creates large relation-extraction datasets by aligning text with Knowledge Base relationships.Its assumption that every sentence mentioning related entities expresses that relation does not always hold.
- Motivation: Multi-instance learning relaxes the distant-supervision assumption, but earlier approaches rely on potentially noisy NLP-tool features.
- Motivation: Entity types and relation aliases provide relevant Knowledge Base side information for predicting relations such as founderOfCompany.For example, person and organization types, or phrases such as “was started by,” can constrain relation prediction.
- Contributions: RESIDE imposes soft constraints using entity type and relation alias information while encoding sentence syntax with Graph Convolution Networks.
- Contributions: RESIDE uses GCN-based syntactic modeling and performs competitively even with limited side information.
- Contributions: RESIDE demonstrates effectiveness over state-of-the-art baselines on benchmark datasets.
2 Related Work
Related work develops distant-supervision, neural, syntactic, and side-information approaches for relation extraction. These methods address limited labeled data, noisy distant-supervision assumptions, feature engineering, and the use of Knowledge Base information.
- Distant supervision: Distant supervision heuristically aligns text with Knowledge Base relations to create large datasets, but can produce wrongly labeled sentences.
- Distant supervision: Multi-instance single-label and multi-instance multi-label learning relax distant-supervision assumptions for valid or overlapping relations.
- Neural relation extraction: Neural relation-extraction models reduce reliance on hand-engineered features using CNNs, piecewise pooling, and attention over instances or sentences.
- Syntactic modeling: Dependency-tree features capture syntactic information, while GCNs have been used to model this information effectively.
- Side information: Prior side-information methods use entity descriptions or types, while Knowledge Bases can provide reliable type information directly.Entity descriptions are not available for all entities.
3 Background: Graph Convolution Networks (GCN)
GCNs update each node using representations from neighboring nodes in a directed, labeled graph. RESIDE augments the graph with reverse edges and self-loops, supports stacked layers, and can suppress noisy edges through learned gating.
- GCN on labeled directed graphs: A directed labeled graph represents each edge as (u, v, l_uv), with vertices and edges defining the graph structure.
- GCN on labeled directed graphs: RESIDE augments directed graphs with reverse edges and self-loops because information need not propagate only along the original direction.
- GCN on labeled directed graphs: Each GCN update produces a node representation from its immediate neighbors using label-dependent parameters and a nonlinear activation.
- GCN on labeled directed graphs: Stacking multiple GCN layers captures multi-hop neighborhoods.
- Integrating edge importance: Edgewise gating assigns relevance scores to graph edges to subdue noisy edges in automatically constructed graphs.The gated embedding is computed using learned edge parameters and a sigmoid function.
4 RESIDE Overview
RESIDE represents an entity-pair bag through syntactic sentence encoding, side-information acquisition, and attention-based instance aggregation. The resulting bag representation is combined with entity types for relation prediction.
- RESIDE Overview: RESIDE comprises syntactic sentence encoding, side-information acquisition, and instance-set aggregation components.
- Syntactic Sentence Encoding: Syntactic sentence encoding combines Bi-GRU local-context representations with dependency-tree GCN encodings and token attention.
- Side Information Acquisition: Side-information acquisition obtains additional supervision from Knowledge Bases and relevant information through Open IE methods.
- Instance Set Aggregation: Instance-set aggregation concatenates sentence representations with matched relation embeddings, applies sentence attention, and adds entity-type embeddings before softmax classification.
5 RESIDE Details
RESIDE builds bag representations by combining token-level syntactic encoding with relation-alias and entity-type side information. Attention aggregates sentence evidence before classification with a softmax over relations.
- 5.1 Syntactic Sentence Encoding: RESIDE encodes tokens with Bi-GRU representations, dependency-tree GCN embeddings, and token attention to form sentence representations.GCN captures long-range syntactic dependencies, while attention downweights irrelevant tokens.
- 5.1 Syntactic Sentence Encoding: The syntactic GCN reduces 55 dependency edge labels to forward, backward, and self-loop directions to limit overparameterization.The resulting edge labels are used when updating token embeddings through GCN layers.
- 5.2 Relation Alias Side Information: Relation alias information matches extracted relation phrases with an embedding-space alias set and uses the closest alias embedding.Open IE extracts phrases between target entities, while dependency-path tokens at one hop extend the phrase set.
- 5.2 Relation Alias Side Information: When multiple phrases produce matches, RESIDE averages their relation embeddings and concatenates the result with the syntactic sentence representation.The model uses this combined representation as sentence-level input for bag aggregation.
- 5.2 Entity Type Side Information: Entity type embeddings are concatenated with the final bag representation, using 38 coarse FIGER types instead of all 112 fine-grained types.Multiple types for an entity are averaged before concatenation.
- 5.3 Instance Set Aggregation: Sentence attention weights representations augmented with matched relation embeddings; the weighted bag representation is concatenated with subject and object type embeddings before softmax classification.The classifier outputs a probability distribution over relations.
6 Experimental Setup
The experiments use the Riedel and GIDS distant-supervision datasets, held-out evaluation, and precision-recall and top-N precision metrics. RESIDE is compared with established relation-extraction baselines, with neural baselines used on GIDS.
- 6.1 Datasets: RESIDE is evaluated on the Riedel and GIDS distant-supervision datasets.Riedel aligns Freebase relations with New York Times sentences from separate training and test years; GIDS extends a Google relation-extraction corpus with additional entity-pair instances.
- 6.1 Datasets: The GIDS dataset is designed to satisfy the at-least-one multi-instance assumption, making automatic evaluation more reliable without manual verification.Its construction adds instances for each entity pair.
- 6.4 Results: RESIDE achieves higher precision across the entire recall range than all baselines on both datasets.The result is reported from the precision-recall curves in Figure 3.
- 6.2 Baselines: The comparison includes Mintz, MultiR, MIMLRE, PCNN, and PCNN+ATT baselines, with only neural baselines used on GIDS.These baselines span logistic-regression, graphical-model, and CNN-based approaches.
- 6.3 Evaluation: Evaluation uses a held-out scheme comparing relations discovered in test articles with Freebase relations, reporting precision-recall curves and top-N precision.The metrics are applied in the experiments across the benchmark datasets.
7 Results
RESIDE outperforms neural baselines across evaluated settings, while ablations show that GCNs and side information contribute to performance. The model remains competitive with limited relation-alias information and improves as more aliases become available.
- 7.1 Performance Comparison: RESIDE achieves higher precision across the recall range on both Riedel and GDS datasets.
- 7.1 Performance Comparison: RESIDE attains improved P@N precision with one, two, or all randomly selected sentences from each bag.
- 7.2 Ablation Results: Removing RESIDE components causes drastic performance degradation, validating the effectiveness of GCNs for encoding syntactic information.
- 7.2 Ablation Results: Side information improves performance and is complementary to features extracted from text.
- 7.3 Effect of Relation Alias Side Information: RESIDE performs competitively with very limited relation-alias information, performs best with KB-provided aliases, and improves further as alias information increases.
8 Conclusion
The conclusion presents RESIDE as a neural distant-supervision model that integrates KB side information and GCN-based syntactic encoding. Experiments show effectiveness over state-of-the-art baselines, robustness to limited side information, and publicly available source code supports reproducibility.
- RESIDE uses entity type and relation-alias information from Knowledge Bases to improve distant supervised relation extraction.
- RESIDE employs Graph Convolution Networks to encode sentence syntax and remains robust to limited side information.
- Experiments on benchmark datasets demonstrate RESIDE’s effectiveness over state-of-the-art baselines.
- RESIDE’s source code is publicly available to promote reproducible research.