An Improved Approach to Generation and Detection of Out-of-Domain Texts

• Item type: SIG Technical Reports(1)
• 公開日: 2021-09-08
• タイトル: An Improved Approach to Generation and Detection of Out-of-Domain Texts
• タイトル / 言語: en
• タイトル: An Improved Approach to Generation and Detection of Out-of-Domain Texts
• 言語: eng
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Kyushu Universtiy
• 著者所属: Kyushu Universtiy
• 著者所属(英): en
• 著者所属(英): Kyushu Universtiy
• 著者所属(英): en
• 著者所属(英): Kyushu Universtiy
• 著者名: Bo, Wang; Tsunenori, Mine
• 著者名(英): Bo, Wang; Tsunenori, Mine
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Many approaches have been proposed to detect Out-of-Domain texts in user intent classification, but most of them are trained on In-Domain data which cannot utilize the huge potential of unlabeled data, or need many hard-to-obtain real Out-of-Domain data. Recently, an Out-of-Domain generation framework has been proposed, which overcame the drawbacks of previous work and got better results. However, the current implementation is far from practical because of its common but ineffective network implementation and unconsidered potential conflicts in the GAN training procedure. In this paper, we propose an improved approach to realize a better Out-of-Domain texts generation, where we modify the Autoencoder for faster learning of context data, and also regularize the output to decrease the difficulty of GAN imitation afterwards. For GAN, we utilize different activation scheme and a more moderate training signal to solve the training conflicts. Comprehensive experiments on three datasets and efficiency measurements show the practicality and efficiency of our new approach.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Many approaches have been proposed to detect Out-of-Domain texts in user intent classification, but most of them are trained on In-Domain data which cannot utilize the huge potential of unlabeled data, or need many hard-to-obtain real Out-of-Domain data. Recently, an Out-of-Domain generation framework has been proposed, which overcame the drawbacks of previous work and got better results. However, the current implementation is far from practical because of its common but ineffective network implementation and unconsidered potential conflicts in the GAN training procedure. In this paper, we propose an improved approach to realize a better Out-of-Domain texts generation, where we modify the Autoencoder for faster learning of context data, and also regularize the output to decrease the difficulty of GAN imitation afterwards. For GAN, we utilize different activation scheme and a more moderate training signal to solve the training conflicts. Comprehensive experiments on three datasets and efficiency measurements show the practicality and efficiency of our new approach.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AA11135936
• 書誌情報: 研究報告知能システム（ICS）; 巻 2021-ICS-204, 号 4, p. 1-8, 発行日 2021-09-08
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-885X
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会