Entangled Relations: Leveraging NLI and Meta-analysis to Enhance Biomedical Relation Extraction

TL;DR

This work reframes relation extraction as a natural language inference task by verbalizing relation classes into hypotheses and training with entailed/non-entailed labels. It introduces three key enhancements—meta-class analysis, feasible hypothesis filtering, and group-based prediction selection—yielding large F1 improvements on BioRED and ReTACRED, and demonstrating domain-agnostic applicability. The method integrates entity-type abstraction and LLM-generated templates to reduce data sparsity and enhance relational reasoning, with extensive ablations and cross-domain experiments supporting its effectiveness. Overall, MetaEntail-RE advances RE-to-NLI by leveraging mutual exclusivity in relation classes and structured hypothesis filtering, offering substantial practical impact for biomedical and broader information extraction tasks, along with open-source resources for further research.

Abstract

Recent research efforts have explored the potential of leveraging natural language inference (NLI) techniques to enhance relation extraction (RE). In this vein, we introduce MetaEntailRE, a novel adaptation method that harnesses NLI principles to enhance RE performance. Our approach follows past works by verbalizing relation classes into class-indicative hypotheses, aligning a traditionally multi-class classification task to one of textual entailment. We introduce three key enhancements: (1) Meta-class analysis which, instead of labeling non-entailed premise-hypothesis pairs with the less informative "neutral" entailment label, provides additional context by analyzing overarching meta-relationships between classes; (2) Feasible hypothesis filtering, which removes unlikely hypotheses from consideration based on domain knowledge derived from data; and (3) Group-based prediction selection, which further improves performance by selecting highly confident predictions. MetaEntailRE is conceptually simple and empirically powerful, yielding significant improvements over conventional relation extraction techniques and other NLI formulations. We observe surprisingly large F1 gains of 17.6 points on BioRED and 13.4 points on ReTACRED compared to conventional methods, underscoring the versatility of MetaEntailRE across both biomedical and general domains.

Figures (2)

• Figure 1: Data flow used for MetaEntail-RE. The original RE input instance (A) is converted into a premise where surface forms are masked with corresponding entity types (B). Each relation class is verbalized into a hypothesis (C), and a feasible hypothesis filter (D) removes infeasible hypotheses based on the pair of entity types. NLI labels are generated via meta-class analysis (E), which are the labels used to fine-tune an LLM via cross-entropy (F). Finally, we use softmax probabilities as a proxy for the model's confidence and select the most confident "entail" prediction among the group of predictions (G). Note that the model makes three predictions in this example—one for each feasible hypothesis. The second "entail" prediction is incorrect but the group-based prediction module selects the first and correct "entail" prediction by assessing the model's confidence.
• Figure 2: $\Delta$F1 per relation class when leveraging meta-class analysis to assign NLI labels.