Continuous Learning for Large-scale Personalized Domain Classification

Proceedings of NAACL-HLT 2019, pages 3784–3794 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3784 Continuous Learning for Large-scale Personalized Domain Classification Han Li1, Jihwan Lee2, Sidharth Mudgal2, Ruhi Sarikaya2, and Young-Bum Kim2 1University of Wisconsin, Madison hanli@cs.wisc.edu 2Amazon Alexa AI {jihwl, sidmsk, rsarikay, youngbum}@amazon.edu Abstract Domain classification is the task of mapping spoken language utterances to one of the natu- ral language understanding domains in intelli- gent personal digital assistants (IPDAs). This is a major component in mainstream IPDAs in industry. Apart from official domains, thou- sands of third-party domains are also created by external developers to enhance the capa- bility of IPDAs. As more domains are de- veloped rapidly, the question of how to con- tinuously accommodate the new domains still remains challenging. Moreover, existing con- tinual learning approaches do not address the problem of incorporating personalized infor- mation dynamically for better domain classi- fication. In this paper, we propose CONDA, a neural network based approach for domain classification that supports incremental learn- ing of new classes. Empirical evaluation shows that CONDA achieves high accuracy and outperforms baselines by a large margin on both incrementally added new domains and existing domains. 1 Introduction Domain classification is the task of mapping spo- ken language utterances to one of the natural lan- guage understanding (NLU) domains in intelligent personal digital assistants (IPDAs), such as Ama- zon Alexa, Google Assistant, and Microsoft Cor- tana, etc. (Sarikaya, 2017). Here a domain is de- fined in terms of a specific application or function- ality such as weather, calendar or music, which narrows down the scope of NLU. For example, given an utterance “Ask Uber to get me a ride” from a user, the appropriate domain would be one that invokes the “Uber” app. Traditionally IPDAs have only supported dozens of well-separated domains, where each is defined in terms of a specific application or func- tionality such as calendar and weather (Sarikaya et al., 2016; Tur and De Mori, 2011; El-Kahky et al., 2014). In order to increase the domain cov- erage and extend the capabilities of the IPDAs, mainstream IPDAs released tools to allow third- party developers to build new domains. Ama- zons Alexa Skills Kit, Googles Actions and Mi- crosofts Cortana Skills Kit are examples of such tools. To handle the influx of new domains, large- scale domain classification methods like SHORT- LISTER (Kim et al., 2018b) have been proposed and have achieved good performance. As more new domains are developed rapidly, one of the major challenges in large-scale domain classification is how to quickly accommodate the new domains without losing the learned predic- tion power on the known ones. A straightforward solution is to simply retraining the whole model whenever new domains are available. However, this is not desirable since retraining is often time consuming. Another approach is to utilize con- tinual learning where we dynamically evolve the model whenever a new domain is available. There is extensive work on the topic of continual learn- ing, …

Content truncated to 500 words