Domainless Adaptation by Constrained Decoding on a Schema Lattice

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2051–2060, Osaka, Japan, December 11-17 2016. Domainless Adaptation by Constrained Decoding on a Schema Lattice Young-Bum Kim Microsoft Redmond, WA ybkim@microsoft.com Karl Stratos∗ Bloomberg L. P. New York, NY me@karlstratos.com Ruhi Sarikaya† Amazon Seattle, WA rsarikay@amazon.com Abstract In many applications such as personal digital assistants, there is a constant need for new domains to increase the system’s coverage of user queries. A conventional approach is to learn a separate model every time a new domain is introduced. This approach is slow, inefficient, and a bottleneck for scaling to a large number of domains. In this paper, we introduce a framework that allows us to have a single model that can handle all domains: including unknown domains that may be cre- ated in the future as long as they are covered in the master schema. The key idea is to remove the need for distinguishing domains by explicitly predicting the schema of queries. Given permitted schema of a query, we perform constrained decoding on a lattice of slot sequences allowed under the schema. The proposed model achieves competitive and often superior performance over the conventional model trained separately per domain. 1 Introduction Recently, there has been much investment on the personal digital assistant (PDA) technology in industry (Sarikaya, 2015). Apple’s Siri, Google Now, Microsoft’s Cortana, and Amazon’s Alexa are some exam- ples of personal digital assistants. Spoken language understanding is an important component of these examples that allows natural communication between the user and the agent (Tur, 2006; El-Kahky et al., 2014; Kim et al., 2015a; Kim et al., 2016b). PDAs support a number of scenarios including creating reminders, setting up alarms, note taking, scheduling meetings, finding and consuming entertainment (i.e. movie, music, games), finding places of interest and getting driving directions to them. The number of domains supported by these systems constantly increases, and whether there is a method that allows us to easily scale to a larger number of domains is an unsolved problem (Kim et al., 2015d; Kim et al., 2016a). The main reason behind the need for additional domains is that we require a new set of schema (i.e., query topics), composed of intents, and slots for processing user queries in a new category. For example, a query in the TAXI domain is processed according to domain-specific schema that is different from those in the HOTEL domain. This in turn requires collecting and annotating new data, which is time consuming and expensive. Once the data is prepared, we also need to build a new system (i.e., models) for this specific domain. In particular, slot modeling is one of the most demanding components of the system in terms of costs in annotation and computation. In this paper, we introduce a new approach that entirely removes the costs traditionally associated with enlarging the set of supported domains while significantly improving performance. This approach uses a single model to handle all domains: including unknown domains that …

Content truncated to 500 words