GCRE-GPT: A Generative Model for Comparative Relation Extraction

TL;DR

GCRE-GPT reframes comparative relation extraction as a generation task using a GPT-2–based architecture with an encoder augmented by prompt words and a comparative-aware decoder. A filter layer ensures that generated relation components originate from the input, and autoregressive cross-entropy trains the model to produce relation texts like "t1 vs. t2 in a". The method achieves state-of-the-art F1 on two benchmarks, especially for sentences with a single relation, and can handle multiple relations after data augmentation. Limitations include small dataset sizes and prompt-design constraints, pointing to future work on missing comparative elements and improved multi-relation handling. Overall, the work demonstrates the viability and advantages of generation-based, prompt-guided extraction for complex relational information in text.

Abstract

Given comparative text, comparative relation extraction aims to extract two targets (\eg two cameras) in comparison and the aspect they are compared for (\eg image quality). The extracted comparative relations form the basis of further opinion analysis.Existing solutions formulate this task as a sequence labeling task, to extract targets and aspects. However, they cannot directly extract comparative relation(s) from text. In this paper, we show that comparative relations can be directly extracted with high accuracy, by generative model. Based on GPT-2, we propose a Generation-based Comparative Relation Extractor (GCRE-GPT). Experiment results show that \modelname achieves state-of-the-art accuracy on two datasets.

Figures (1)

• Figure 1: The architecture of GCRE-GPT.