Zero-Shot to Full-Resource: Cross-lingual Transfer Strategies for Aspect-Based Sentiment Analysis

2026-04-29 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A multilingual evaluation of Aspect-Based Sentiment Analysis (ABSA) approaches was conducted across seven languages: English, German, French, Dutch, Russian, Spanish, and Czech. The study compared various transformer architectures under zero-resource, data-only, and full-resource settings for four ABSA subtasks: Aspect Category Detection (ACD), Aspect-Category Sentiment Analysis (ACSA), Target Aspect Sentiment Detection (TASD), and Aspect Sentiment Quad Prediction (ASQP). Fine-tuned Large Language Models (LLMs) achieved the highest scores, especially in generative tasks, while few-shot LLMs and smaller encoder models were competitive in simpler setups. The research found that cross-lingual training on multiple non-target languages optimized transfer for fine-tuned LLMs, whereas code-switching was more effective for smaller encoder or sequence-to-sequence models. Additionally, two new German datasets, an adapted GERestaurant and the first German ASQP dataset (GERest), were introduced.

Key takeaway

For research scientists developing multilingual NLP systems, you should recognize that optimal cross-lingual transfer strategies for Aspect-Based Sentiment Analysis (ABSA) are architecture-dependent. If you are working with Large Language Models, prioritize cross-lingual training on diverse non-target languages. Conversely, for smaller encoder or sequence-to-sequence models, integrate code-switching techniques to maximize performance. Your choice of strategy directly impacts model efficacy across different resource settings and languages.

Key insights

Cross-lingual transfer strategies for ABSA vary in effectiveness based on model architecture and resource availability.

Principles

• LLMs excel in complex generative ABSA tasks.
• Cross-lingual training boosts LLM transfer.
• Code-switching benefits smaller encoder models.

Method

The study systematically compared transformer architectures using cross-lingual transfer, code-switching, and machine translation across zero-resource, data-only, and full-resource settings for four ABSA subtasks.

In practice

• Consider fine-tuned LLMs for generative ABSA.
• Utilize code-switching for smaller models.
• Explore new German ABSA datasets.

Topics

• Aspect-Based Sentiment Analysis
• Cross-lingual Transfer
• Large Language Models
• Code-switching
• Multilingual NLP

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• Beyond Bilingual Transfer: Multilingual Code-Switching in Instruction Tuning — Multilingual code-switching instruction tuning improves LLM performance beyond bilingual settings.
• Aspect-Based Sentiment Analysis in the Accommodation Domain Using the BERTimbau Model — BERTimbau model fine-tuning strategies for Portuguese ABSA show trade-offs between recall and precision.
• CLP-Transfer: Cross-Lingual and Progressive Transfer Learning — CLP-Transfer simplifies cross-lingual model transfer using token overlap and a small helper model, but shows limited downstream task performance.
• Datasets for Verb Alternations across Languages: BLM Templates and Data Augmentation Strategies — This work addresses the underexplored ability of large language models (LLMs) to capture cross-sentence paradigmatic patterns, specifically verb alternations.
• Authorship Attribution in Multilingual Machine-Generated Texts — Multilingual authorship attribution for LLM-generated text is complex, with current monolingual methods showing limited cross-lingual transferability across diverse language families.
• DiffuSent: Towards a Unified Diffusion Framework for Aspect-Based Sentiment Analysis — DiffuSent unifies ABSA subtasks via non-auto-regressive boundary denoising, improving accuracy and speed for multi-word sentiment extraction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.