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

This repository accompanies the paper
“Zero-Shot to Full-Resource: Cross-lingual Transfer Strategies for Aspect-Based Sentiment Analysis”,
which presents a comprehensive cross-lingual study of recent ABSA methods across eight languages and four subtasks and is under review for publication at LREC 2026.

---

📘 Overview

Aspect-Based Sentiment Analysis (ABSA) aims to extract fine-grained opinions by identifying sentiments toward specific aspects within text.
Despite major advances in transfer learning and large language models (LLMs), most research has remains English-focused, and cross-lingual generalization of ABSA methods is still work in progress.

This work systematically evaluates five representative ABSA paradigms — covering both traditional and modern LLM-based approaches — under multilingual, cross-lingual, and zero-resource conditions.

---

🧩 Key Contributions

• Comprehensive Multilingual Benchmarking:
  Evaluation of state-of-the-art ABSA methods across eight languages (English, French, Spanish, Dutch, Russian, Turkish, German, Czech) and four subtasks:

  • Aspect Category Detection (ACD)
  • Aspect Category Sentiment Analysis (ACSA)
  • Targeted Aspect Sentiment Detection (TASD)
  • Aspect Sentiment Quad Prediction (ASQP)

• Unified Experimental Framework:
  Comparison of three major modeling paradigms:

  1. Encoder-only (BERT-CLF, Hier-GCN)
  2. Sequence-to-Sequence (T5, DLO)
  3. Decoder-only LLMs (Gemma 3, LLaMA 3.1) for prompting and instruction tuning

• Resource-Level Analysis:
  Experiments conducted under three resource settings:

  • Zero-Resource: no training data or language-specific models
  • Data-Only: training data without language-specific models
  • Full-Resource: training data and language-specific models

• Cross-Lingual Adaptation Strategies:
  Evaluation of multilingual transfer through:

  • Machine translation
  • Code-switching
  • Balanced multilingual training across source languages

• New German ABSA Datasets:

  • GERestaurant [Hellwig et al., 2024]: A version of the dataset adapted to the SemEval Rest16 schema (Aspect Categories)
  • GERest: first German ASQP dataset for structured opinion extraction, fully aligned with the SemEval Rest16 schema

• Empirical Insights:

  • Instruction-tuned LLMs generally achieve the highest scores, especially when it comes to more complex tasks (TASD, ASQP).
  • Encoder-based models remain competitive for simpler classification tasks.
  • Code-switching provides consistent improvements in zero-resource settings.
  • Multilingual models remain competitive, yet fall behind language-specific models.

---

📊 GERest Dataset

The GERest dataset extends the TASD corpus GERestaurant and mirrors the structure of the English ASQP-Rest16 benchmark [Zhang et al., 2021].
It provides German training, validation, and test splits for aspect sentiment quad prediction.

Dataset composition:

• Training: 1,264 examples
• Validation: 316 examples
• Test: 544 examples
• Aspect categories: 13
• Polarity types: Positive, Negative, Neutral
• Reference types: Explicit, Implicit

---

🧠 Annotation Process

GERest was created following the ASQP annotation guidelines by [Zhang et al., 2021] and [Wan et al., 2020].

1. Annotator A (B.Sc. student) performed the initial labeling.
2. Annotator B (Ph.D. student with ABSA experience) reviewed and refined all examples.
3. Out of 2,124 sentences, 184 received suggested label changes; 179 (97%) were adopted.
4. Disagreements (n = 5) were resolved jointly.

This two-stage review ensured high annotation consistency and alignment with multilingual ABSA standards.

---

⚙️ Methods Overview

Language-specific pretrained models were used only when publicly available for a given language–architecture combination.
Otherwise, the corresponding multilingual model (e.g., mBERT, mT5) was retained.

---

📝 Citation

If you use this repository or the GERest dataset, please cite:

@inproceedings{fehle2026-lrec,
  title={Zero-Shot to Full-Resource: Cross-lingual Transfer Strategies for Aspect-Based Sentiment Analysis},
  author={Fehle, Jakob and Hellwig, Nils Constantin and Kruschwitz, Udo and Wolff, Christian},
  year={2026},
  booktitle={Proceedings of the Fifteenth Language Resources and Evaluation Conference},
}