Controllable Semantic Parsing via Retrieval Augmentation

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7683–7698 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 7683 Controllable Semantic Parsing via Retrieval Augmentation Panupong Pasupat and Yuan Zhang and Kelvin Guu Google Research {ppasupat,zhangyua,kguu}@google.com Abstract In practical applications of semantic parsing, we often want to rapidly change the behav- ior of the parser, such as enabling it to han- dle queries in a new domain, or changing its predictions on certain targeted queries. While we can introduce new training examples ex- hibiting the target behavior, a mechanism for enacting such behavior changes without ex- pensive model re-training would be preferable. To this end, we propose ControllAble Seman- tic Parser via Exemplar Retrieval (CASPER). Given an input query, the parser retrieves re- lated exemplars from a retrieval index, aug- ments them to the query, and then applies a generative seq2seq model to produce an output parse. The exemplars act as a control mech- anism over the generic generative model: by manipulating the retrieval index or how the augmented query is constructed, we can ma- nipulate the behavior of the parser. On the MTOP dataset, in addition to achieving state- of-the-art on the standard setup, we show that CASPER can parse queries in a new domain, adapt the prediction toward the specified pat- terns, or adapt to new semantic schemas with- out having to further re-train the model. 1 Introduction Semantic parsing is the task of mapping input queries to their meaning representations. In practi- cal applications of semantic parsing, such as con- versational agents, we often want to rapidly control the behavior of the parser. We particularly focus on three scenarios: (1) Domain bootstrapping: mak- ing the parser handle queries in a new domain (Su and Yan, 2017; Hou et al., 2020a; Li et al., 2021b). This requires predicting new semantic labels (e.g., intents and slots) unseen during training, and as- signing correct values to such labels. (2) Parse guiding: influencing the prediction toward a spe- cific parse pattern. This can be used as an override for sensitive queries or queries that the model strug- gles on. (3) Schema refactoring: adapting the [IN create call = [SL contact = [IN get contact = [SL contact related = Jay Brown] [SL type relation = mom]]]] Retriever call Zoey’s wife [IN create call = [SL contact = [IN get contact = [SL contact related = Zoey] [SL type relation = wife]]]] Make a call to … [IN create call = … Get number of … [IN get number = … make a call to Jay Brown’s mom Generator make a call to Jay Brown’s mom @@ call Zoey’s wife ## [IN create call = [SL contact = [IN get contact = [SL contact related = Zoey] [SL type relation = wife]]]] @@ Make a call … ## [IN create call … @@ Get number … ## [IN get number … Retrieval Index Figure 1: The CASPER model. (1) Given a query x, we retrieve exemplars (x′ i, …

Content truncated to 500 words