Improving Autoregressive Training with Dynamic Oracles
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2024
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866910485684682752 |
|---|---|
| author | Yang, Jianing Visvanathan, Harshine Wang, Yilin Hu, Xinyi Gormley, Matthew |
| author_facet | Yang, Jianing Visvanathan, Harshine Wang, Yilin Hu, Xinyi Gormley, Matthew |
| contents | Many tasks within NLP can be framed as sequential decision problems, ranging from sequence tagging to text generation. However, for many tasks, the standard training methods, including maximum likelihood (teacher forcing) and scheduled sampling, suffer from exposure bias and a mismatch between metrics employed during training and inference. DAgger provides a solution to mitigate these problems, yet it requires a metric-specific dynamic oracle algorithm, which does not exist for many common metrics like span-based F1, ROUGE, and BLEU. In this paper, we develop these novel dynamic oracles and show they maintain DAgger's no-regret guarantee for decomposable metrics like span-based F1. We evaluate the algorithm's performance on named entity recognition (NER), text summarization, and machine translation (MT). While DAgger with dynamic oracle yields less favorable results in our MT experiments, it outperforms the baseline techniques in NER and text summarization. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2406_09393 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Improving Autoregressive Training with Dynamic Oracles Yang, Jianing Visvanathan, Harshine Wang, Yilin Hu, Xinyi Gormley, Matthew Computation and Language Artificial Intelligence Machine Learning Many tasks within NLP can be framed as sequential decision problems, ranging from sequence tagging to text generation. However, for many tasks, the standard training methods, including maximum likelihood (teacher forcing) and scheduled sampling, suffer from exposure bias and a mismatch between metrics employed during training and inference. DAgger provides a solution to mitigate these problems, yet it requires a metric-specific dynamic oracle algorithm, which does not exist for many common metrics like span-based F1, ROUGE, and BLEU. In this paper, we develop these novel dynamic oracles and show they maintain DAgger's no-regret guarantee for decomposable metrics like span-based F1. We evaluate the algorithm's performance on named entity recognition (NER), text summarization, and machine translation (MT). While DAgger with dynamic oracle yields less favorable results in our MT experiments, it outperforms the baseline techniques in NER and text summarization. |
| title | Improving Autoregressive Training with Dynamic Oracles |
| topic | Computation and Language Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2406.09393 |