Span-based Hierarchical Semantic Parsing for Task-Oriented Dialog

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1520–1526, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1520 Span-based Hierarchical Semantic Parsing for Task-Oriented Dialog Panupong Pasupat∗ Stanford University ppasupat@cs.stanford.edu Sonal Gupta Facebook Assistant sonalgupta@fb.com Karishma Mandyam∗ University of Washington krm28@cs.washington.edu Rushin Shah∗ Google rushinshah@google.com Mike Lewis Facebook AI Research mikelewis@fb.com Luke Zettlemoyer Facebook AI Research lsz@fb.com Abstract We propose a semantic parser for parsing com- positional utterances into Task Oriented Parse (TOP), a tree representation that has intents and slots as labels of nesting tree nodes. Our parser is span-based: it scores labels of the tree nodes covering each token span indepen- dently, but then decodes a valid tree globally. In contrast to previous sequence decoding ap- proaches and other span-based parsers, we (1) improve the training speed by removing the need to run the decoder at training time; and (2) introduce edge scores, which model rela- tions between parent and child labels, to mit- igate the independence assumption between node labels and improve accuracy. Our best parser outperforms previous methods on the TOP dataset of mixed-domain task-oriented utterances in both accuracy and training speed. 1 Introduction Most commercial conversational AI systems parse task-oriented utterances using intent classification and slot-filling models (He and Young, 2003; Ray- mond and Riccardi, 2007; Mesnil et al., 2015), where the intent is the task of the utterance (e.g., IN:GET DIRECTION) and the slots are the parameters needed to complete the task (e.g., SL:DESTINATION). This limited representation typically allows only a single intent per utterance and at most one slot label per token. Dialog sys- tems using such a flat representation would strug- gle to handle compositional tasks that involve in- voking multiple backend services (e.g., “direction to John’s party”: find John’s address, and then find the direction to that address). To support compositional utterances, the hierar- chical Task Oriented Parsing (TOP) representation has recently been introduced (Gupta et al., 2018). As illustrated in Figure 1, the TOP representation ∗work done while at Facebook Assistant. IN:GET DIRECTION SL:DESTINATION IN:FIND EVENT SL:CATEGORY party4 ’s3 SL:ORGANIZER John2 directions0 to1 T = {(0, 5) : (IN:GET DIRECTION), (2, 5) : (SL:DESTINATION, IN:FIND EVENT), (2, 3) : (SL:ORGANIZER), (4, 5) : (SL:CATEGORY), (0, 1) : ∅, (0, 2) : ∅, (0, 3) : ∅, . . . } Figure 1: An example TOP tree and its mapping repre- sentation T. (IN: = intent; SL: = slot) is a tree where intents and slots are nested alter- natively to model composition. The values inside intent and slot subtrees can then be used by down- stream dialog modules to invoke appropriate ser- vices in a hierarchical fashion. We propose a span-based semantic parser for parsing utterances into the TOP representation. In its most basic form, the parser embeds each to- ken span (e.g., x2:5 = “John ’s party” in Fig- ure 1) as a vector, and then uses it to predict the …

Content truncated to 500 words