Neural Symbolic Machines: Learning Semantic Parsers on Freebase with Weak Supervision

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 23–33 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1003 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 23–33 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1003 Neural Symbolic Machines: Learning Semantic Parsers on Freebase with Weak Supervision Chen Liang⇤, Jonathan Berant†, Quoc Le, Kenneth D. Forbus, Ni Lao Northwestern University, Evanston, IL Tel-Aviv University, Tel Aviv-Yafo, Israel Google Inc., Mountain View, CA {chenliang2013,forbus}@u.northwestern.edu, joberant@cs.tau.ac.il, {qvl,nlao}@google.com Abstract Harnessing the statistical power of neu- ral networks to perform language under- standing and symbolic reasoning is dif- ficult, when it requires executing effi- cient discrete operations against a large knowledge-base. In this work, we intro- duce a Neural Symbolic Machine (NSM), which contains (a) a neural “program- mer”, i.e., a sequence-to-sequence model that maps language utterances to programs and utilizes a key-variable memory to han- dle compositionality (b) a symbolic “com- puter”, i.e., a Lisp interpreter that performs program execution, and helps find good programs by pruning the search space. We apply REINFORCE to directly opti- mize the task reward of this structured prediction problem. To train with weak supervision and improve the stability of REINFORCE we augment it with an it- erative maximum-likelihood training pro- cess. NSM outperforms the state-of-the- art on the WEBQUESTIONSSP dataset when trained from question-answer pairs only, without requiring any feature engi- neering or domain-specific knowledge. 1 Introduction Deep neural networks have achieved impressive performance in supervised classification and struc- tured prediction tasks such as speech recognition (Hinton et al., 2012), machine translation (Bah- danau et al., 2014; Wu et al., 2016) and more. However, training neural networks for semantic parsing (Zelle and Mooney, 1996; Zettlemoyer and Collins, 2005; Liang et al., 2011) or program induction, where language is mapped to a sym- ⇤Work done while the author was interning at Google † Work done while the author was visiting Google x: Largest city in the US ⇒ y: NYC (USA) (Hop v0 CityIn) CityIn (Argmax v1 Population) Population Compositionality Large Search Space ( Argmax Hop v1 v0 Population Size Elevation ) v2 ← v1 ← v0 ← Figure 1: The main challenges of training a semantic parser from weak supervision: (a) compositionality: we use vari- ables (v0, v1, v2) to store execution results of intermediate generated programs. (b) search: we prune the search space and augment REINFORCE with pseudo-gold programs. bolic representation that is executed by an execu- tor, through weak supervision remains challeng- ing. This is because the model must interact with a symbolic executor through non-differentiable op- erations to search over a large program space. In semantic parsing, recent work handled this (Dong and Lapata, 2016; Jia and Liang, 2016) by training from manually annotated programs and avoiding program execution at training time. However, annotating programs is known to be ex- pensive and scales poorly. In program induc- tion, attempts to address this problem (Graves …

Content truncated to 500 words