BERT
基于变换器的双向编码器表示技术(英語:Bidirectional Encoder Representations from Transformers,BERT)是用于自然语言处理(NLP)的预训练技术,由Google提出。[1][2]2018年,雅各布·德夫林和同事创建并发布了BERT。Google正在利用BERT来更好地理解用户搜索语句的语义。[3] 2020年的一项文献调查得出结论:「在一年多一点的时间里,BERT已经成为NLP实验中无处不在的基线」,算上分析和改进模型的研究出版物超过150篇。[4]
最初的英语BERT发布时提供两种类型的预训练模型[1]:(1)BERTBASE模型,一个12层,768维,12个自注意头(self attention head),110M参数的神经网络结构;(2)BERTLARGE模型,一个24层,1024维,16个自注意头,340M参数的神经网络结构。两者的训练语料都是BooksCorpus[5]以及英語維基百科语料,单词量分别是8億以及25億。[6]
结构
BERT的核心部分是一个Transformer模型,其中编码层数和自注意力头数量可变。结构与Vaswani等人(2017)[7]的实现几乎“完全一致”。
BERT在两个任务上进行预训练: 语言模型(15%的token被掩盖,BERT需要从上下文中进行推断)和下一句预测(BERT需要预测给定的第二个句子是否是第一句的下一句)。训练完成后,BERT学习到单词的上下文嵌入。代价昂贵的预训练完成后,BERT可以使用较少的资源和较小的数据集在下游任务上进行微调,以改进在这些任务上的性能。[1][8]
性能及分析
BERT在以下自然语言理解任务上的性能表现得最为卓越:[1]
- GLUE(General Language Understanding Evaluation,通用语言理解评估)任务集(包括9个任务)。
- SQuAD(Stanford Question Answering Dataset,斯坦福问答数据集)v1.1和v2.0。
- SWAG(Situations With Adversarial Generation,对抗生成的情境)。
有關BERT在上述自然语言理解任务中為何可以達到先进水平,目前還未找到明確的原因[9][10]。目前BERT的可解释性研究主要集中在研究精心选择的输入序列对BERT的输出的影响关系,[11][12]通过探测分类器分析内部向量表示,[13][14]以及注意力权重表示的关系。[9][10]
历史
BERT起源于预训练的上下文表示学习,包括半监督序列学习(Semi-supervised Sequence Learning)[15],生成预训练(Generative Pre-Training),ELMo[16]和ULMFit[17]。与之前的模型不同,BERT是一种深度双向的、无监督的语言表示,且仅使用纯文本语料库进行预训练的模型。上下文无关模型(如word2vec或GloVe)为词汇表中的每个单词生成一个词向量表示,因此容易出现单词的歧义问题。BERT考虑到单词出现时的上下文。例如,词“水分”的word2vec词向量在“植物需要吸收水分”和“财务报表裡有水分”是相同的,但BERT根据上下文的不同提供不同的词向量,词向量与句子表达的句意有关。
2019年10月25日,Google搜索宣布他们已经开始在美国国内的英语搜索查询中应用BERT模型。[18]2019年12月9日,据报道,Google搜索已经在70多种语言的搜索采用了BERT。[19] 2020年10月,几乎每一个基于英语的查询都由BERT处理。[20]
获奖情况
在2019年计算语言学协会北美分会(NAACL)年会上,BERT获得了最佳长篇论文奖。[21]
参见
参考文献
- ^ 1.0 1.1 1.2 1.3 Devlin, Jacob; Chang, Ming-Wei; Lee, Kenton; Toutanova, Kristina. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. 2018-10-11. arXiv:1810.04805v2 [cs.CL].
- ^ Open Sourcing BERT: State-of-the-Art Pre-training for Natural Language Processing. Google AI Blog. [2019-11-27]. (原始内容存档于2021-01-13) (英语).
- ^ Understanding searches better than ever before. Google. 2019-10-25 [2019-11-27]. (原始内容存档于2021-01-27) (英语).
- ^ Rogers, Anna; Kovaleva, Olga; Rumshisky, Anna. A Primer in BERTology: What We Know About How BERT Works. Transactions of the Association for Computational Linguistics. 2020, 8: 842–866 [2021-11-24]. doi:10.1162/tacl_a_00349. (原始内容存档于2022-04-03).
- ^ Zhu, Yukun; Kiros, Ryan; Zemel, Rich; Salakhutdinov, Ruslan; Urtasun, Raquel; Torralba, Antonio; Fidler, Sanja. Aligning Books and Movies: Towards Story-Like Visual Explanations by Watching Movies and Reading Books: 19–27. 2015. arXiv:1506.06724 [cs.CV].
- ^ Annamoradnejad, Issa. ColBERT: Using BERT Sentence Embedding for Humor Detection. 2020-04-27. arXiv:2004.12765 [cs.CL].
- ^ Polosukhin, Illia; Kaiser, Lukasz; Gomez, Aidan N.; Jones, Llion; Uszkoreit, Jakob; Parmar, Niki; Shazeer, Noam; Vaswani, Ashish. Attention Is All You Need. 2017-06-12. arXiv:1706.03762 [cs.CL].
- ^ Horev, Rani. BERT Explained: State of the art language model for NLP. Towards Data Science. 2018 [27 September 2021]. (原始内容存档于2022-10-17).
- ^ 9.0 9.1 Kovaleva, Olga; Romanov, Alexey; Rogers, Anna; Rumshisky, Anna. Revealing the Dark Secrets of BERT. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). November 2019: 4364–4373 [2020-10-19]. doi:10.18653/v1/D19-1445. (原始内容存档于2020-10-20) (美国英语).
- ^ 10.0 10.1 Clark, Kevin; Khandelwal, Urvashi; Levy, Omer; Manning, Christopher D. What Does BERT Look at? An Analysis of BERT's Attention. Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP (Stroudsburg, PA, USA: Association for Computational Linguistics). 2019: 276–286.
- ^ Khandelwal, Urvashi; He, He; Qi, Peng; Jurafsky, Dan. Sharp Nearby, Fuzzy Far Away: How Neural Language Models Use Context. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (Stroudsburg, PA, USA: Association for Computational Linguistics). 2018: 284–294. Bibcode:2018arXiv180504623K. arXiv:1805.04623 . doi:10.18653/v1/p18-1027.
- ^ Gulordava, Kristina; Bojanowski, Piotr; Grave, Edouard; Linzen, Tal; Baroni, Marco. Colorless Green Recurrent Networks Dream Hierarchically. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) (Stroudsburg, PA, USA: Association for Computational Linguistics). 2018: 1195–1205. Bibcode:2018arXiv180311138G. arXiv:1803.11138 . doi:10.18653/v1/n18-1108.
- ^ Giulianelli, Mario; Harding, Jack; Mohnert, Florian; Hupkes, Dieuwke; Zuidema, Willem. Under the Hood: Using Diagnostic Classifiers to Investigate and Improve how Language Models Track Agreement Information. Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP (Stroudsburg, PA, USA: Association for Computational Linguistics). 2018: 240–248. Bibcode:2018arXiv180808079G. arXiv:1808.08079 . doi:10.18653/v1/w18-5426.
- ^ Zhang, Kelly; Bowman, Samuel. Language Modeling Teaches You More than Translation Does: Lessons Learned Through Auxiliary Syntactic Task Analysis. Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP (Stroudsburg, PA, USA: Association for Computational Linguistics). 2018: 359–361. doi:10.18653/v1/w18-5448.
- ^ Dai, Andrew; Le, Quoc. Semi-supervised Sequence Learning. 2015-11-04. arXiv:1511.01432 [cs.LG].
- ^ Peters, Matthew; Neumann, Mark; Iyyer, Mohit; Gardner, Matt; Clark, Christopher; Lee, Kenton; Luke, Zettlemoyer. Deep contextualized word representations. 2018-02-15. arXiv:1802.05365v2 [cs.CL].
- ^ Howard, Jeremy; Ruder, Sebastian. Universal Language Model Fine-tuning for Text Classification. 2018-01-18. arXiv:1801.06146v5 [cs.CL].
- ^ Nayak, Pandu. Understanding searches better than ever before. Google Blog. 2019-10-25 [2019-12-10]. (原始内容存档于2019-12-05).
- ^ Montti, Roger. Google's BERT Rolls Out Worldwide. Search Engine Journal. Search Engine Journal. 2019-12-10 [2019-12-10]. (原始内容存档于2020-11-29).
- ^ Google: BERT now used on almost every English query. Search Engine Land. 2020-10-15 [2020-11-24]. (原始内容存档于2022-05-06).
- ^ Best Paper Awards. NAACL. 2019 [2020-03-28]. (原始内容存档于2020-10-19).
外部链接
- 官方GitHub仓库 (页面存档备份,存于互联网档案馆)