Multi-task and Multi-lingual Joint Learning of Neural Lexical Utterance Classification based on Partially-shared Modeling

Proceedings of the 27th International Conference on Computational Linguistics, pages 3586–3596 Santa Fe, New Mexico, USA, August 20-26, 2018. 3586 Multi-task and Multi-lingual Joint Learning of Neural Lexical Utterance Classification based on Partially-shared Modeling Ryo Masumura, Tomohiro Tanaka, Ryuichiro Higashinaka, Hirokazu Masataki and Yushi Aono NTT Media Intelligence Laboratories, NTT Corporation, Japan ryou.masumura.ba@hco.ntt.co.jp Abstract This paper is an initial study on multi-task and multi-lingual joint learning for lexical utter- ance classification. A major problem in constructing lexical utterance classification modules for spoken dialogue systems is that individual data resources are often limited or unbalanced among tasks and/or languages. Various studies have examined joint learning using neural-network-based shared modeling; however, previous joint learning studies focused on either cross-task or cross- lingual knowledge transfer. In order to simultaneously support both multi-task and multi-lingual joint learning, our idea is to explicitly divide state-of-the-art neural lexical utterance classification into language-specific components that can be shared between different tasks and task-specific components that can be shared between different languages. In addition, in order to effectively transfer knowledge between different task data sets and different language data sets, this pa- per proposes a partially-shared modeling method that possesses both shared components and components specific to individual data sets. We demonstrate the effectiveness of the proposed adversarial training using Japanese and English data sets with three different lexical utterance classification tasks. 1 Introduction Modern spoken dialogue systems use multiple lexical utterance classification modules that can detect dialogue act (Stolcke et al., 2000; Khanpour et al., 2016), intent (Tur et al., 2011), domain (Xu and Sarikaya, 2014), question type (Wu et al., 2005), etc. to properly understand natural languages. The modules are typically trained using the machine learning technologies developed for individual language- specific systems (Higashinaka et al., 2014). A common issue is that the data resources for such individual training are often limited or unbalanced among different tasks and/or different languages. For modeling lexical utterance classification, various modeling methods have been examined. Re- cently, neural lexical utterance classification, which is a fully neural-network-based modeling method, has demonstrated substantial performance without the use of manual feature engineering. The networks include long short-term memory recurrent neural networks (LSTM-RNNs) (Ravuri and Stolcke, 2015b; Ravuri and Stolcke, 2015a; Ravuri and Stolcke, 2016), convolution neural networks (Kim, 2014), and more advanced networks (Zhou et al., 2016a; Yang et al., 2016; Sawada et al., 2017). In addition, neural networks are suitable for performing joint learning; the paucity of data is tackled by transferring knowledge between different tasks or different languages. Various joint learning methods have been examined for leveraging different tasks or different language data sets in the natural language processing field. Multi-task joint learning can transfer knowledge between tasks by sharing task-invariant layers (Collobert and Weston, 2008; Liu et al., 2015; Liu et al., 2016c; Zhang and Weng, 2016). In lexical utterance classification, multi-task joint learning has been shown to effectively improve individual tasks (Liu et al., 2016b; Liu et al., 2016a; Liu et al., 2017). In addition, multi-lingual joint learning can …

Content truncated to 500 words