Conversational Semantic Parsing for Dialog State Tracking

Abstract

We consider a new perspective on dialog state tracking (DST), the task of estimating a user's goal through the course of a dialog. By formulating DST as a semantic parsing task over hierarchical representations, we can incorporate semantic compositionality, crossdomain knowledge sharing and co-reference. We present TreeDST, a dataset of 27k conversations annotated with tree-structured dialog states and system acts. 1 We describe an encoder-decoder framework for DST with hierarchical representations, which leads to 20% improvement over state-of-the-art DST approaches that operate on a flat meaning space of slot-value pairs. TurnUtterance and Annotation 1Hi can you book me a flight to Paris please. user.flight.book.object.equals .destination.equals.location.equals.Paris Sure, when and where will you depart? system.prompt.flight.book.object.equals .source .departureDateTime

Value-Agnostic Conversational Semantic Parsing

“…Perhaps closest to our work is that of Dong and Lapata (2018), which is also about decoupling decisions, but uses a dataset-specific notion of an abstracted program sketch along with different independence assumptions, and underperforms our model in comparable settings ( §3.2). Also close are the models of Cheng et al (2020) and . Our method differs in that our beam search uses larger steps that predict functions together with their arguments, rather than predicting the argument values serially in separate dependent steps.…”

“…We first report results on SMCALFLOW (Semantic Machines et al, 2020) and TREEDST (Cheng et al, 2020), two recently released large-scale conversational semantic parsing datasets. Our model makes use of type information in the programs, so we manually constructed a set of type declarations for each dataset and then used a variant of the Hindley-Milner type inference algorithm (Damas and Milner, 1982) to annotate programs with types.…”

“…To compare with prior work for SMCALFLOW (Semantic Machines et al, 2020) and TREEDST (Cheng et al, 2020), we replicated their setups. For SMCALFLOW, we predict plans always conditioning on the gold dialogue history for each utterance, but we consider any predicted plan wrong if the refer are correct flag is set to false.…”

“…We also canonicalize plans by sorting keyword arguments and normalizing numbers (so that 30.0 and 30 are considered equivalent, for example). For TREEDST, our model predicts programs that use the refer and revise operators, and we execute them against the dialogue history that consists of predicted programs and gold (oracle) system actions (following Cheng et al (2020)) when converting back to the original tree representation. We canonicalize the resulting trees by lexicographically sorting the children of each node.…”

Semantic Representation for Dialogue Modeling

“…They solve intent classification and slot-filling task via semantic parsing. Cheng et al (2020) design a rooted semantic graph that integrates domains, verbs, operators and slots in order to perform dialogue state tracking. All these structures are designed for specified task only.…”