PLP-RC:Point–Line–Plane Fusion for Discriminative Relation Classification with LLMs

Pengjiu Xia, Wu Yuan, Yidian Huang, Wen Yuan, Wandong Zhang

Keywords: LLM, Embedding, Relation Classification

TL;DR: PLP-RC is a discriminative LLM-based framework that fuses point, line, and plane representations, avoiding hallucinations and achieving SOTA on TACRED, TACREV, and RE-TACRED.

Abstract: Relation classification is a fundamental NLP task that involves identifying the semantic relations between entity pairs in a given text. While pre-trained language models have advanced this area, effectively integrating local entity information with global context remains a key challenge. Large Language Models offer rich world knowledge, but their generative use often suffers from hallucinations, limiting reliability. To address these issues, we propose a Point–Line–Plane fusion framework for discriminative relation classification with LLM embeddings. Entity spans are modeled as local point representations, the end of sequence token provides a global plane representation, and an attention-based line representation aligns the two. This discriminative paradigm avoids hallucinations while fully exploiting LLM representations. Our method achieves new SOTA performance on TACRED, TACREV, and RE-TACRED benchmarks, outperforming both discriminative and generative baselines. Ablation studies provide further evidence for the effectiveness of our design in achieving context-aware relation classification.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 7445