Task-Oriented Dialogue as Dataflow Synthesis

Task-Oriented Dialogue as Dataflow Synthesis Jacob Andreas John Bufe David Burkett Charles Chen Josh Clausman Jean Crawford Kate Crim Jordan DeLoach Leah Dorner Jason Eisner Hao Fang Alan Guo David Hall Kristin Hayes Kellie Hill Diana Ho Wendy Iwaszuk Smriti Jha Dan Klein Jayant Krishnamurthy Theo Lanman Percy Liang Christopher H. Lin Ilya Lintsbakh Andy McGovern Aleksandr Nisnevich Adam Pauls Dmitrij Petters Brent Read Dan Roth Subhro Roy Jesse Rusak Beth Short Div Slomin Ben Snyder Stephon Striplin Yu Su Zachary Tellman Sam Thomson Andrei Vorobev Izabela Witoszko Jason Wolfe Abby Wray Yuchen Zhang Alexander Zotov Microsoft Semantic Machines <sminfo@microsoft.com> Abstract We describe an approach to task-oriented dialogue in which dialogue state is rep- resented as a dataflow graph. A dialogue agent maps each user utterance to a program that extends this graph. Programs include metacomputation operators for reference and revision that reuse dataflow fragments from previous turns. Our graph-based state enables the expression and manipulation of complex user intents, and explicit meta- computation makes these intents easier for learned models to predict. We introduce a new dataset, SMCalFlow, featuring complex dialogues about events, weather, places, and people. Experiments show that dataflow graphs and metacomputation substantially improve representability and predictability in these natural dialogues. Additional experiments on the MultiWOZ dataset show that our dataflow representation enables an otherwise off-the-shelf sequence-to- sequence model to match the best existing task-specific state tracking model. The SM- CalFlow dataset, code for replicating exper- iments, and a public leaderboard are avail- able at https://www.microsoft.com/en- us/research/project/dataflow- based-dialogue-semantic-machines. 1 Introduction Two central design decisions in modern conversa- tional AI systems are the choices of state and ac- tion representations, which determine the scope of possible user requests and agent behaviors. Di- alogue systems with fixed symbolic state repre- sentations (like slot filling systems) are easy to train but hard to extend (Pieraccini et al., 1992). Deep continuous state representations are flexible enough to represent arbitrary properties of the dia- logue history, but so unconstrained that training a User: Where is my meeting at 2 this afternoon? User: Can you create a meeting with Megan right before that starts? User: Megan Bowen. place(findEvent(EventSpec(start=pm(2)))) createEvent(EventSpec( end=start(refer(Constraint[Event]())), attendee=PersonSpec(name='Megan') )) revise( new=PersonSpec(name='Megan Bowen'), oldLoc=Constraint[PersonSpec](), rootLoc=RoleConstraint(output) ) pm 2 findEvent Event(name=“kickoff”, place=…) Agent: It’s in Conference Room D. start 'Megan' Agent: Which person named Megan did you mean? UnderconstrainedException!(invite) Agent: Ok, I’ll book a meeting with Megan Bowen from 1:30 to 2:00 pm today. (1) (2) (3) createEvent name end att. Event(start=1:30, end=2:00, attendee=Person(…)) “Conference Room D” place EventSpec findEvent place start pm 2 EventSpec start EventSpec PersonSpec start 'Megan' createEvent name end att. findEvent place pm 2 EventSpec start EventSpec PersonSpec 'Megan Bowen' createEvent name att. EventSpec PersonSpec oldLoc New rootLoc Figure 1: A dialogue and its dataflow graph. Turn (1) is an ordinary case of semantic parsing: the agent predicts a compositional query that encodes the user’s question. Evaluating this program produces an initial graph frag- ment. In turn (2), that is used to refer to a salient Event; the …

Content truncated to 500 words