LACA: Improving Cross-lingual Aspect-Based Sentiment Analysis with LLM Data Augmentation

2025-08-13

AI Summary

Cross-lingual aspect-based sentiment analysis (ABSA) faces challenges in transferring knowledge from source to target languages, with existing approaches relying on error-prone machine translation—introducing substantial pseudo-label noise and semantic distortion. This paper proposes a translation-free pseudo-label generation framework: leveraging large language models (LLMs) with semantics-preserving prompt engineering to directly generate high-quality natural sentences in the target language alongside corresponding fine-grained sentiment labels. Integrated with a two-stage self-training strategy, it constructs high-fidelity pseudo-labeled datasets without any target-language annotations. The method supports fine-tuning diverse backbone models—including generative ones—and achieves consistent improvements over translation-based state-of-the-art methods across six languages and five models. Results demonstrate strong generalizability and effectiveness: fine-tuned LLMs significantly outperform smaller multilingual models, confirming the superiority of direct, translation-free label generation.

Technology Category

• Natural Language Processing: Generation
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Language and Vision

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Cross-lingual aspect-based sentiment analysis (ABSA) involves detailed sentiment analysis in a target language by transferring knowledge from a source language with available annotated data. Most existing methods depend heavily on often unreliable translation tools to bridge the language gap. In this paper, we propose a new approach that leverages a large language model (LLM) to generate high-quality pseudo-labelled data in the target language without the need for translation tools. First, the framework trains an ABSA model to obtain predictions for unlabelled target language data. Next, LLM is prompted to generate natural sentences that better represent these noisy predictions than the original text. The ABSA model is then further fine-tuned on the resulting pseudo-labelled dataset. We demonstrate the effectiveness of this method across six languages and five backbone models, surpassing previous state-of-the-art translation-based approaches. The proposed framework also supports generative models, and we show that fine-tuned LLMs outperform smaller multilingual models.

Problem

• Enhancing cross-lingual ABSA without translation tools
• Generating pseudo-labelled data using LLMs for target language
• Improving sentiment analysis accuracy across multiple languages

Innovation

• LLM generates pseudo-labelled target language data
• No reliance on translation tools required
• Fine-tunes ABSA model with enhanced pseudo-data