TriG-NER: Triplet-Grid Framework for Discontinuous Named Entity Recognition

TL;DR

TriG-NER tackles discontinuous named entity recognition by combining a grid-based word-pair tagging scheme with a token-level triplet loss that pulls together within-entity word-pairs and pushes apart those separated by non-entity tokens. The approach operates on a word-pair relationship grid produced from encoder representations, with decoding that constructs non-contiguous entity spans. Across three biomedical DNER benchmarks, TriG-NER consistently outperforms state-of-the-art grid-based baselines, especially for discontinuous elements, and shows robustness to tagging scheme variations. The work demonstrates the practical value of learning fine-grained token-level relations in a grid framework and suggests strong potential for generalisation to other structured prediction tasks, while also highlighting limitations of current LLM prompting for discontinuous NER.

Abstract

Discontinuous Named Entity Recognition (DNER) presents a challenging problem where entities may be scattered across multiple non-adjacent tokens, making traditional sequence labelling approaches inadequate. Existing methods predominantly rely on custom tagging schemes to handle these discontinuous entities, resulting in models tightly coupled to specific tagging strategies and lacking generalisability across diverse datasets. To address these challenges, we propose TriG-NER, a novel Triplet-Grid Framework that introduces a generalisable approach to learning robust token-level representations for discontinuous entity extraction. Our framework applies triplet loss at the token level, where similarity is defined by word pairs existing within the same entity, effectively pulling together similar and pushing apart dissimilar ones. This approach enhances entity boundary detection and reduces the dependency on specific tagging schemes by focusing on word-pair relationships within a flexible grid structure. We evaluate TriG-NER on three benchmark DNER datasets and demonstrate significant improvements over existing grid-based architectures. These results underscore our framework's effectiveness in capturing complex entity structures and its adaptability to various tagging schemes, setting a new benchmark for discontinuous entity extraction.

Figures (10)

• Figure 1: A case example involving discontinuous mentions. TriG-NER (a) perfectly extracts the DNE. LLMs (b,c,d), primarily trained to capture continuous sequences of text, face challenges in recognising entities split across discontinuous regions while maintaining coherence in prediction.
• Figure 2: Overall framework of the proposed TriG-NER
• Figure 3: Example of positive and negative candidates based on the anchor ("joint", "in") with a candidate window of 3.
• Figure 4: Example of positive and negative candidates for one-word entities (left) and one-word samples (right).
• Figure 5: Triplet Mining Methods