A New Corpus and Imitation Learning Framework for Context-Dependent Semantic Parsing

A New Corpus and Imitation Learning Framework for Context-Dependent Semantic Parsing Andreas Vlachos Computer Science Department University College London a.vlachos@cs.ucl.ac.uk Stephen Clark Computer Laboratory University of Cambridge sc609@cam.ac.uk Abstract Semantic parsing is the task of translating natural language utterances into a machine- interpretable meaning representation. Most approaches to this task have been evaluated on a small number of existing corpora which assume that all utterances must be interpreted according to a database and typically ignore context. In this paper we present a new, pub- licly available corpus for context-dependent semantic parsing. The MRL used for the an- notation was designed to support a portable, interactive tourist information system. We develop a semantic parser for this corpus by adapting the imitation learning algorithm DAGGER without requiring alignment infor- mation during training. DAGGER improves upon independently trained classifiers by 9.0 and 4.8 points in F-score on the development and test sets respectively. 1 Introduction Semantic parsing is the task of translating natu- ral language utterances into a machine-interpretable meaning representation (MR). Progress in semantic parsing has been facilitated by the existence of cor- pora containing utterances annotated with MRs, the most commonly used being ATIS (Dahl et al., 1994) and GeoQuery (Zelle, 1995). As these corpora cover rather narrow application domains, recent work has developed corpora to support natural language in- terfaces to the Freebase database (Cai and Yates, 2013), as well as the development of MT systems (Banarescu et al., 2013). However, these existing corpora have some im- portant limitations. The MRs accompanying the utterances are typically restricted to some form of database query. Furthermore, in most cases each utterance is interpreted in isolation; thus utterances that use coreference or whose semantics are context- dependent are typically ignored. In this paper we present a new corpus for context-dependent seman- tic parsing to support the development of an interac- tive navigation and exploration system for tourism- related activities. The new corpus was annotated with MRs that can handle dialog context such as coreference and can accommodate utterances that are not interpretable according to a database, e.g. repetition requests. The utterances were collected in experiments with human subjects, and contain phe- nomena such as ellipsis and disfluency. We devel- oped guidelines and annotated 17 dialogs containing 2,374 utterances, with 82.9% exact match agreement between two annotators. pus.We also develop a semantic parser for this cor- As the output MRs are rather complex, in- stead of adopting an approach that searches the out- put space exhaustively, we use the imitation learning algorithm DAGGER (Ross et al., 2011) that converts learning a structured prediction model into learning a set of classification models. We take advantage of its ability to learn with non-decomposable loss func- tions and extend it to handle the absence of align- ment information during training by developing a randomized expert policy. Our approach improves upon independently trained classifiers by 9.0 and 4.8 F-score on the development and test sets. 2 Meaning Representation Language Our proposed MR language (MRL) was designed in the …

Content truncated to 500 words