BERT

基于变换器的双向编码器表示技术（英语：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）BERT_BASE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）BERT_LARGE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]以及英语维基百科语料，单词量分别是8亿以及25亿。^[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）。

有关BERT在上述自然语言理解任务中为何可以达到先进水平，目前还未找到明确的原因^[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（英语：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（英语：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“财务报表里有水分”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（英语：North American Chapter of the Association for Computational Linguistics））年会上，BERT获得了最佳长篇论文奖。^[21]

参见

Transformer模型
Word2vec
自编码器
文献-检索词矩阵（英语：Document-term matrix）
特征提取
特征学习
神经网络语言模型（英语：Neural network language model）
向量空间模型
概念向量（英语：Thought vector）
fastText（英语：fastText）
GloVe（英语：GloVe）
TensorFlow

参考文献

[1]
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].
[2]
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）.
[3]
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）.
[4]
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）.
[5]
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]. 缺少或|url=为空 (帮助)
[6]
Annamoradnejad, Issa. ColBERT: Using BERT Sentence Embedding for Humor Detection. 2020-04-27. arXiv:2004.12765  [cs.CL].
[7]
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].
[8]
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]
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]
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.
[11]
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.
[12]
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.
[13]
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.
[14]
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.
[15]
Dai, Andrew; Le, Quoc. Semi-supervised Sequence Learning. 2015-11-04. arXiv:1511.01432  [cs.LG].
[16]
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].
[17]
Howard, Jeremy; Ruder, Sebastian. Universal Language Model Fine-tuning for Text Classification. 2018-01-18. arXiv:1801.06146v5  [cs.CL].
[18]
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）.
[19]
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）.
[20]
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）.
[21]
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）

[:0-1] [1]
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].

[2] [2]
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）.

[3] [3]
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）.

[4] [4]
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）.

[5] [5]
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]. 缺少或|url=为空 (帮助)

[6] [6]
Annamoradnejad, Issa. ColBERT: Using BERT Sentence Embedding for Humor Detection. 2020-04-27. arXiv:2004.12765  [cs.CL].

[vaswani-7] [7]
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].

[8] [8]
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）.

[:1-9] [9]
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）.

[:2-10] [10]
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.

[11] [11]
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.

[12] [12]
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.

[13] [13]
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.

[14] [14]
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.

[15] [15]
Dai, Andrew; Le, Quoc. Semi-supervised Sequence Learning. 2015-11-04. arXiv:1511.01432  [cs.LG].

[16] [16]
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].

[17] [17]
Howard, Jeremy; Ruder, Sebastian. Universal Language Model Fine-tuning for Text Classification. 2018-01-18. arXiv:1801.06146v5  [cs.CL].

[18] [18]
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）.

[19] [19]
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）.

[20] [20]
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）.

[21] [21]
Best Paper Awards. NAACL. 2019 [2020-03-28]. （原始内容存档于2020-10-19）.

[1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]