The paper proposes a novel Test-Time Code-Switching (TT-CSW) framework for cross-lingual Aspect Sentiment Triplet Extraction (ASTE).
It introduces a bilingual alignment-based code-switching technique that enhances test-time predictions, effectively addressing term boundary detection and out-of-dictionary challenges.
------
📌 TT-CSW enables test-time code-switching, fixing low-resource ASTE issues. It dynamically augments test data with bilingual signals, improving boundary detection and reducing segmentation errors in opinion-aspect mapping.
📌 TT-CSW beats GPT-4 by refining multilingual alignment at test time. It embeds structured bilingual knowledge into smaller models, achieving 14.2% higher weighted F1 than ChatGPT without massive pretraining.
📌 Heuristic phrase-switching with voting ensures robust multilingual extraction. TT-CSW generates multiple bilingual variants, aligns them, and selects the best prediction, reducing hallucinations and improving F1 by 3.7% across datasets.
---
https://arxiv.org/abs/2501.14144
Original Problem 🤔:
→ ASTE models perform well in high-resource languages but struggle in low-resource languages due to limited training data.
→ Existing cross-lingual transfer methods face term boundary detection issues and out-of-dictionary problems, leading to incorrect extractions.
→ Traditional code-switching techniques only work during training, limiting their effectiveness at test time.
---
Solution in this Paper 🔧:
→ TT-CSW introduces bilingual code-switching during both training and testing phases to bridge the gap between bilingual training data and monolingual test-time inference.
→ In training, a generative model learns to extract bilingual ASTE triplets using a boundary-aware code-switching approach that maintains aspect and opinion term integrity.
→ In testing, an alignment-based code-switching method augments monolingual sentences into bilingual variants and aligns extracted triplets back into the target language, improving prediction consistency.
→ A heuristic phrase-switching strategy generates multiple bilingual augmentations, and the final prediction is determined through alignment and voting mechanisms.
---
Key Insights 💡:
→ Incorporating bilingual information at test time significantly improves term boundary recognition and reduces translation inconsistencies.
→ Fine-tuned generative models using TT-CSW outperform larger LLMs (ChatGPT & GPT-4) in cross-lingual ASTE, demonstrating efficiency gains in smaller models.
→ The method provides a multilingual view of input text, leveraging both source and target language features for better extraction accuracy.
---
Results 📊:
→ Achieves an average 3.7% improvement in weighted F1-score across four datasets.
→ Outperforms ChatGPT by 14.2% and GPT-4 by 5.0% in weighted F1-score.
→ Enhances baseline models (mT5-base, m2m100) by 15.2% to 43.8% on specific datasets.
Share this post