ENIAC-ML: Environmentally Interactive Active Meta-learning for Zero-Shot Relation Triplet Extraction

ACL ARR 2025 February Submission7408 Authors

Abstract: Zero-shot relation triplet extraction (ZeroRTE) task aims to extract unseen relations and corresponding entities from the text. Existing methods conflate the Relation Extraction (RE) and Named Entity Recognition (NER) subtasks. Moreover, some methods introduce synthetic data or information that contains noise, resulting in failures on ZeroRTE. We propose a novel meta-learning approach named Environmentally Interactive ACtive Meta-Learning (ENIAC-ML) that can mimic human processing on ZeroRTE. We decompose ZeroRTE into RE and NER subtasks and train the model using a pipelined approach. We further develop an active meta-learning approach that can acquire knowledge by interacting with an agent in the environment, autonomously determine the focus of learning, and mitigate the impact of noise in external information. The experimental results demonstrate that ENIAC-ML surpasses existing methods on Fewrel and Wiki-ZSL datasets. Our code is available at https://anonymous.4open.science/r/ENIAC-ML-E0FF.

Paper Type: Long

Research Area: Special Theme (conference specific)

Research Area Keywords: Relation triplet extraction, zero-shot learning, meta learning, active learning

Contribution Types: Model analysis & interpretability, Approaches to low-resource settings

Languages Studied: English

Submission Number: 7408