Integrating Back-Translation into BERT Model for Detecting Machine-Translated Text

• Item type: Symposium(1)
• 公開日: 2019-10-14
• タイトル: Integrating Back-Translation into BERT Model for Detecting Machine-Translated Text
• タイトル / 言語: en
• タイトル: Integrating Back-Translation into BERT Model for Detecting Machine-Translated Text
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: Machine-Translation Detection，Back-Translation，BERT Model，Paraphrasing，Adversarial Text
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_5794
• 資源タイプ: conference paper
• 著者所属: Indian Institute of Technology
• 著者所属: KDDI Research Inc.
• 著者所属: The University of Tokyo
• 著者所属: KDDI Research Inc.
• 著者所属: KDDI Research Inc.
• 著者所属(英): en
• 著者所属(英): Indian Institute of Technology
• 著者所属(英): en
• 著者所属(英): KDDI Research Inc.
• 著者所属(英): en
• 著者所属(英): The University of Tokyo
• 著者所属(英): en
• 著者所属(英): KDDI Research Inc.
• 著者所属(英): en
• 著者所属(英): KDDI Research Inc.
• 著者名: Gupta, Ishita; Hoang-Quoc, Nguyen-Son; Thao, Tran Phuong; Seira, Hidano; Shinsaku, Kiyomoto
• 著者名(英): Gupta, Ishita; Hoang-Quoc, Nguyen-Son; Thao, Tran Phuong; Seira, Hidano; Shinsaku, Kiyomoto
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Machine-generated text is being used by adversaries to support malicious purposes like spam mails and fake reviews etc. Recently a form of 'back translation' plagiarism has started where texts are paraphrased by translating into a different language and then back into the original language. Previous methods for detecting such machine-generated text are based only on the intrinsic content of the text. We propose a detector which exploits the external information obtained from back-translation; and integrates it into the BERT model. An evaluation of 90000 samples of original English sentences and translated French sentences shows that our detector can classify them with 83.8% accuracy. This is higher than previous methods whose best accuracy is 79.9%. Moreover, our detector can efficiently detect back-translated text with 87.1% accuracy when assessed on 20000 sentences. This is an improvement from 82.2% accuracy of the state of the art. We have also conducted experiments with low-resource language and reached similar results. This demonstrates the persistence of our detector on various tasks in both rich- and low-resource languages.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Machine-generated text is being used by adversaries to support malicious purposes like spam mails and fake reviews etc. Recently a form of 'back translation' plagiarism has started where texts are paraphrased by translating into a different language and then back into the original language. Previous methods for detecting such machine-generated text are based only on the intrinsic content of the text. We propose a detector which exploits the external information obtained from back-translation; and integrates it into the BERT model. An evaluation of 90000 samples of original English sentences and translated French sentences shows that our detector can classify them with 83.8% accuracy. This is higher than previous methods whose best accuracy is 79.9%. Moreover, our detector can efficiently detect back-translated text with 87.1% accuracy when assessed on 20000 sentences. This is an improvement from 82.2% accuracy of the state of the art. We have also conducted experiments with low-resource language and reached similar results. This demonstrates the persistence of our detector on various tasks in both rich- and low-resource languages.
• 書誌レコードID / 識別子タイプ: NCID
• 書誌レコードID / 関連識別子: ISSN 1882-0840
• 書誌情報: コンピュータセキュリティシンポジウム2019論文集; 巻 2019, p. 1349-1355, 発行日 2019-10-14
• 出版者 / 言語: ja
• 出版者: 情報処理学会