[Paper Review] Enhancing Relation Extraction Using Syntactic Indicators and Sentential Contexts

Qiongxing Tao, Xiangfeng Luo+1 more|arXiv (Cornell University)|Dec 4, 2019

Topic Modeling28 references4 citations

TL;DR

This paper proposes an indicator-aware neural model that enhances relation extraction by integrating syntactic indicators—key phrases like prepositions that signal semantic relations—alongside full sentential context. Using BERT-based encoding with constraints from manually extracted syntactic indicators, the model improves relation classification by reducing noise and capturing more informative features, achieving an F1-score of 90.36% on the SemEval-2010 benchmark, outperforming state-of-the-art methods.

ABSTRACT

State-of-the-art methods for relation extraction consider the sentential context by modeling the entire sentence. However, syntactic indicators, certain phrases or words like prepositions that are more informative than other words and may be beneficial for identifying semantic relations. Other approaches using fixed text triggers capture such information but ignore the lexical diversity. To leverage both syntactic indicators and sentential contexts, we propose an indicator-aware approach for relation extraction. Firstly, we extract syntactic indicators under the guidance of syntactic knowledge. Then we construct a neural network to incorporate both syntactic indicators and the entire sentences into better relation representations. By this way, the proposed model alleviates the impact of noisy information from entire sentences and breaks the limit of text triggers. Experiments on the SemEval-2010 Task 8 benchmark dataset show that our model significantly outperforms the state-of-the-art methods.

Motivation & Objective

• To address the limitations of existing relation extraction models that treat all words equally and ignore informative syntactic cues.
• To reduce noise from irrelevant words in sentences that hinder accurate relation classification.
• To improve relation extraction performance by identifying and leveraging syntactic indicators—phrases like 'moved into' that signal semantic relations.
• To overcome the reliance on fixed lexical triggers, which fail when no explicit trigger is present.
• To demonstrate that combining syntactic indicators with full sentence context leads to superior performance in relation extraction.

Proposed method

• Syntactic indicators are extracted using syntactic knowledge, including entity disambiguation, principal component extraction, and unrelated entity removal.
• The model employs a BERT-based contextual encoder to represent both the full sentence and the extracted syntactic indicators.
• The syntactic indicator is used as a constraint to guide the contextual representation, focusing the model on relation-relevant features.
• A two-stream encoding mechanism processes the entire sentence and the syntactic indicator sequence separately before fusing their representations.
• The final classification is performed using a classifier head over the fused representation, with the indicator acting as a supervisory signal.
• The approach is validated using both BERT-based and non-BERT models, including CNN and Bi-LSTM components for ablation studies.

Experimental results

Research questions

• RQ1 Can syntactic indicators—phrases like prepositions or directional terms—improve relation extraction beyond standard sentence-level modeling?
• RQ2 How does incorporating syntactic indicators affect the model's ability to reduce noise from irrelevant words in sentences?
• RQ3 To what extent do syntactic indicators enhance performance compared to models relying only on entity proximity or fixed lexical triggers?
• RQ4 Does the constraint of syntactic indicators on contextual representations lead to more robust and accurate relation classification?
• RQ5 Can the proposed indicator-aware approach generalize across diverse relation types, especially those with high noise in the span between entities?

Key findings

• The proposed BERT-based model achieves an F1-score of 90.36% on the SemEval-2010 Task 8 benchmark, surpassing state-of-the-art methods.
• The inclusion of syntactic indicators increases the F1-score across all relation categories, with the most significant gains in Instrument-Agency (+3.13 pp), Member-Collection (+3.06 pp), Message-Topic (+2.61 pp), and Product-Producer (+5.95 pp).
• The non-BERT model, combining CNN and Bi-LSTM modules, achieves an F1-score of 85.9%, outperforming other CNN- and RNN-based models and those using high-level features like WordNet or NER tags.
• Removing the syntactic indicator constraint leads to a drop in performance, confirming that the indicator acts as a meaningful constraint that improves representation learning.
• Fusing features from both the full sentence and the indicator sequence yields better results than using either source alone, demonstrating the complementary value of both components.
• Excessive use of full-sentence features without filtering leads to degraded performance, confirming the presence of noise that must be mitigated.