arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2103.07162cs.CLcs.LG
收藏

BERT是跨学科知识学习者吗?预训练模型可移植性的惊人发现

Is BERT a Cross-Disciplinary Knowledge Learner? A Surprising Finding of Pre-trained Models' Transferability

National Taiwan University(台湾大学)

Wei-Tsung Kao, Hung-Yi Lee

摘要:

本文研究了在文本数据上预训练的模型(如BERT)的能力是否可以转移到一般的令牌序列分类应用中。为了验证预训练模型的可移植性,我们在文本分类任务上测试了预训练模型,这些文本分类任务具有标记不匹配的含义,以及真实世界的非文本标记序列分类数据,包括氨基酸,DNA和音乐。我们发现,即使在非文本数据上,在文本上预训练的模型收敛得更快,比随机初始化的模型表现更好,并且只比使用特定任务知识的模型略差。我们还发现,文本和非文本预训练模型的表示具有非平凡的相似性。

英文摘要:

This paper investigates whether the power of the models pre-trained on text data, such as BERT, can be transferred to general token sequence classification applications. To verify pre-trained models' transferability, we test the pre-trained models on text classification tasks with meanings of tokens mismatches, and real-world non-text token sequence classification data, including amino acid, DNA, and music. We find that even on non-text data, the models pre-trained on text converge faster, perform better than the randomly initialized models, and only slightly worse than the models using task-specific knowledge. We also find that the representations of the text and non-text pre-trained models share non-trivial similarities.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑