A Pointer Network Architecture for Context-Dependent Semantic Parsing

A Pointer Network Architecture for Context-Dependent Semantic Parsing Xuanli He♣ Quan Hung Tran♥ Gholamreza Haffari♣ ♣Monash University, Australia ♥Adobe Research, San Jose, CA {xuanli.he1, gholamreza.haffari}@monash.edu qtran@adobe.com Abstract Semantic parsing targets at mapping human utterances into structured meaning represen- tations, such as logical forms, programming snippets, SQL queries etc. In this work, we focus on logical form generation, which is ex- tracted from an automated email assistant sys- tem. Since this task is dialogue-oriented, in- formation across utterances must be well han- dled. Furthermore, certain inputs from users are used as arguments for the logical form, which requires a parser to distinguish the func- tional words and content words. Hence, an intelligent parser should be able to switch be- tween generation mode and copy mode. In or- der to address the aforementioned issues, we equip the vanilla seq2seq model with a pointer network and a context-dependent architecture to generate more accurate logical forms. Our model achieves state-of-the-art performance on the email assistant task. 1 Introduction Recently, due to the breakthrough of the deep learning, numerous and various tasks within the filed of natural language processing (NLP) have made impressive achievements (Vaswani et al., 2017; Devlin et al., 2018; Edunov et al., 2018). However, most these achievements are assessed by automatic metrics, which are relatively superficial and brittle, and can be easily tricked (Paulus et al., 2017; Jia and Liang, 2017; L¨aubli et al., 2018). Hence, understanding the underlying meaning of natural language sentences is crucial to NLP tasks. As an appealing direction in natural language understanding, semantic parsing has been widely studied in the NLP community (Ling et al., 2016; Dong and Lapata, 2016; Jia and Liang, 2017). Semantic parsing aims at converting human ut- terances to machine executable representations. Most existing work focuses on parsing individual utterances independently, even they have an access to the contextual information. In spite of several pioneering efforts (Zettlemoyer and Collins, 2009; Srivastava et al., 2017), these pre-neural models suffer from complicated hand-crafted feature en- gineering, compared to their neural counterparts (Dong and Lapata, 2018; Rabinovich et al., 2017). One notable exception is the work of Suhr et al. (2018), who incorporate context into ATIS data with a neural approach. In this work, we propose a neural semantic parser for email assistant task which incorporates the conversation context as well as a copy mech- anism to fill-in the arguments of the logical forms from the input sentence. Our model achieves state- of-the-art (SOTA) performance. We further pro- vide details analysis about where these improve- ments come from. 2 Models To build our models, we follow a process of error- driven design. We first start with a simple seq2seq model, then we closely examine the errors, group them, and then propose a solution to each of these error groups. From our examination, we identify two main sources of errors of a seq2seq model: i) the overly strong influence of the language model component, and ii) the lack of contextual informa- tion. Thus we design our model …

Content truncated to 500 words