Centrala begrepp
Prompt-based TOPRO method enhances token-level sequence labeling tasks for zero-shot cross-lingual transfer.
Sammanfattning
The article introduces TOPRO, a novel method for token-level sequence labeling tasks. It decomposes input sentences into tokens and applies prompts to each token, improving cross-lingual transfer performance. The study compares TOPRO with Vanilla Fine-Tuning and Prompt-Tuning on NER and POS tagging tasks across multiple languages.
Structure:
- Abstract: Introduces the TOPRO method for token-level sequence labeling tasks.
- Introduction: Discusses multilingual pretrained language models and zero-shot cross-lingual transfer methods.
- Data Extraction: Identifies key metrics supporting the effectiveness of TOPRO in cross-lingual transfer.
- Results and Analysis: Presents results showing TOPRO outperforming baselines in NER and POS tagging tasks.
- Error Analysis: Analyzes selected instances from the UDPOS task to highlight differences between Vanilla and TOPRO predictions.
- Exploratory Study in MLLMs: Compares the performance of TOPRO with existing benchmarking methods on MLLMs.
Statistik
"Our experiments show that TOPRO outperforms the baselines with the MPLMs."
"The improvements in languages such as Persian (fa), Gujarati (gu), Hebrew (he), Japanese (ja), Kazakh (kk), Burmese (my), Telugu (te), Thai (th), Urdu (ur), and Chinese (zh) are above average."
"The improvements over Vanilla in Chinese reach 44.3% and 38.53% for XLM-R and mT5, respectively."
Citat
"TOPRO outperforms Vanilla Fine-Tuning and Prompt-Tuning in zero-shot cross-lingual transfer."
"TOPRO shows a noticeable performance improvement and could serve as a potential benchmark for sequence labeling tasks."