Adapting Discriminative Reranking to Grounded Language Learning

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 218–227, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Adapting Discriminative Reranking to Grounded Language Learning Joohyun Kim Department of Computer Science The University of Texas at Austin Austin, TX 78701, USA scimitar@cs.utexas.edu Raymond J. Mooney Department of Computer Science The University of Texas at Austin Austin, TX 78701, USA mooney@cs.utexas.edu Abstract We adapt discriminative reranking to im- prove the performance of grounded lan- guage acquisition, specifically the task of learning to follow navigation instructions from observation. Unlike conventional reranking used in syntactic and semantic parsing, gold-standard reference trees are not naturally available in a grounded set- ting. Instead, we show how the weak su- pervision of response feedback (e.g. suc- cessful task completion) can be used as an alternative, experimentally demonstrat- ing that its performance is comparable to training on gold-standard parse trees. 1 Introduction Grounded language acquisition involves learn- ing to comprehend and/or generate language by simply observing its use in a naturally occur- ring context in which the meaning of a sentence is grounded in perception and/or action (Roy, 2002; Yu and Ballard, 2004; Gold and Scassel- lati, 2007; Chen et al., 2010). B¨orschinger et al. (2011) introduced an approach that reduces grounded language learning to unsupervised prob- abilistic context-free grammar (PCFG) induction and demonstrated its effectiveness on the task of sportscasting simulated robot soccer games. Sub- sequently, Kim and Mooney (2012) extended their approach to make it tractable for the more complex problem of learning to follow natural-language navigation instructions from observations of hu- mans following such instructions in a virtual envi- ronment (Chen and Mooney, 2011). The observed sequence of actions provides very weak, ambigu- ous supervision for learning instructional language since there are many possible ways to describe the same execution path. Although their approach im- proved accuracy on the navigation task compared to the original work of Chen and Mooney (2011), it was still far from human performance. Since their system employs a generative model, discriminative reranking (Collins, 2000) could po- tentially improve its performance. By training a discriminative classifier that uses global features of complete parses to identify correct interpreta- tions, a reranker can significantly improve the ac- curacy of a generative model. Reranking has been successfully employed to improve syntactic pars- ing (Collins, 2002b), semantic parsing (Lu et al., 2008; Ge and Mooney, 2006), semantic role la- beling (Toutanova et al., 2005), and named entity recognition (Collins, 2002c). Standard reranking requires gold-standard interpretations (e.g. parse trees) to train the discriminative classifier. How- ever, grounded language learning does not provide gold-standard interpretations for the training ex- amples. Only the ambiguous perceptual context of the utterance is provided as supervision. For the navigation task, this supervision consists of the observed sequence of actions taken by a hu- man when following an instruction. Therefore, it is impossible to directly apply conventional dis- criminative reranking to such problems. We show how to adapt reranking to work with such weak supervision. …

Content truncated to 500 words