Low-Resource Domain Adaptation for Compositional Task-Oriented Semantic Parsing

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 5090–5100, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 5090 Low-Resource Domain Adaptation for Compositional Task-Oriented Semantic Parsing Xilun Chen Asish Ghoshal Yashar Mehdad Luke Zettlemoyer Sonal Gupta Facebook Inc. {xilun,aghoshal,mehdad,lsz,sonalgupta}@fb.com Abstract Task-oriented semantic parsing is a critical component of virtual assistants, which is re- sponsible for understanding the user’s intents (set reminder, play music, etc.). Recent ad- vances in deep learning have enabled several approaches to successfully parse more com- plex queries (Gupta et al., 2018; Rongali et al., 2020), but these models require a large amount of annotated training data to parse queries on new domains (e.g. reminder, music). In this paper, we focus on adapting task- oriented semantic parsers to low-resource do- mains, and propose a novel method that outper- forms a supervised neural model at a 10-fold data reduction. In particular, we identify two fundamental factors for low-resource domain adaptation: better representation learning and better training techniques. Our representa- tion learning uses BART (Lewis et al., 2020) to initialize our model which outperforms encoder-only pre-trained representations used in previous work. Furthermore, we train with optimization-based meta-learning (Finn et al., 2017) to improve generalization to low- resource domains. This approach significantly outperforms all baseline methods in the experi- ments on a newly collected multi-domain task- oriented semantic parsing dataset (TOPv21). 1 Introduction Virtual Assistants now play an ever increasingly important role in our daily life, and can help users perform a wide spectrum of tasks ranging from set- ting personal reminders, checking local weather, to controlling smart home devices and online shop- ping. A critical step in any virtual assistant is to understand the user’s intent (e.g. set reminder, get weather info, etc.) given the user utterance. In recent years, a number of successful models have 1The dataset can be downloaded at https://fb.me/ TOPv2Dataset Driving directions to the Eagles game Utterance: Driving directions to the Eagles game Semantic Parse: [IN:GET_DIRECTIONS Driving directions to [SL:DESTINATION [IN:GET_EVENT the [SL:NAME_EVENT Eagles ] [SL:CAT_EVENT game ] ] ] ] Tree Representation: Figure 1: An compositional query from TOP dataset. emerged to tackle such task-oriented semantic pars- ing task, for both simple and more complex queries. Traditionally, task-oriented semantic parsers treat the problem as a joint intent classification and slot filling task (Liu and Lane, 2016), where the model first predicts the intent of the input ut- terance from a set of pre-defined intent labels, and then identify all the necessary slots for that intent. For instance, for the query “How’s the weather in San Francisco?”, the model would predict the GET_WEATHER intent, and tag San Francisco as a LOCATION slot. With the elevated expectation of virtual assistants, however, techniques for han- dling the more complex compositional queries have been proposed recently using neural parsers (Gupta et al., 2018) or Seq2Seq models (Rongali et al., 2020). In particular, these approaches can han- dle complicated queries with multiple intents or nested slots. For example, the following query from the TOP …

Content truncated to 500 words