PLM$^{3}$RC: A PLM-Based Method for Multilingual Relation Classification Without Using Parallel Corpora

Abstract

When building knowledge bases using relation classification (RC) techniques, it is essential to handle data from various languages and domains. However, most existing multilingual RC methods require parallel corpora and some may lead to obviously increased computation overhead. To address these challenges, we propose a multilingual RC method based on PLMs, called PLM$^{3}$RC. Our method first introduce entity prompts and multilingual PLMs to construct a shared encoder network, and capture shared language-invariant knowledge of relation classification. Then multiple domain-specific classifiers are introduced to capture language- and domain-specific knowledge. For realizing multilingual knowledge transfer, we further introduce adversarial learning and design linear discriminators. We conducted a series of experiments to verify the effectiveness of our method. Our experiments show that PLM$^{3}$RC outperform other mainstream RC method in multilingual scenarios, with regular computing resources. Besides, the results clearly demonstrate that our method can effectively leverages multilingual data for knowledge transfer, without using parallel corpora.