Cross-lingual Aspect-Based Sentiment Analysis with Large Language Models

Nancy

Cross-lingual Aspect-Based Sentiment Analysis (ABSA) aims to perform fine-grained sentiment analysis in a target language by transferring knowledge from a resource-rich source language, typically English, to a low-resource language. This task is crucial for expanding sentiment analysis capabilities to low-resource languages, where annotated data is scarce. Traditional cross-lingual ABSA methods often rely on external machine translation tools, which introduce noise and limit the scope of analysis. Recently, large language models (LLMs) have emerged as state-of-the-art tools in natural
language processing, offering efficient solutions to these challenges. In this work, we propose advanced LLM-based approaches for cross-lingual ABSA that eliminate the dependency on translation systems. We present two effective methods: a sequence-to-sequence framework with constrained decoding, and a pseudo-labelling technique where LLMs generate natural, high-quality target-language data for model refinement. Both approaches demonstrate significant performance improvements across multiple languages and models, with fine-tuned multilingual LLMs outperforming both English-centric LLMs and traditional translation-based baselines. Our findings highlight the effectiveness of LLMs in enhancing cross-lingual ABSA, paving the way for more robust and translation-free sentiment analysis across diverse linguistic contexts. Part of this research has already attracted attention at ACL 2025; in this talk, we go further and present the latest result.