Unsupervised PCFG Induction for Grounded Language Learning with Highly Ambiguous Supervision

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 433–444, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Unsupervised PCFG Induction for Grounded Language Learning with Highly Ambiguous Supervision Joohyun Kim Raymond J. Mooney Department of Computer Science The University of Texas at Austin 1616 Guadalupe, Suite 2.408 Austin, TX 78701, USA {scimitar,mooney}@cs.utexas.edu Abstract “Grounded” language learning employs train- ing data in the form of sentences paired with relevant but ambiguous perceptual contexts. B¨orschinger et al. (2011) introduced an ap- proach to grounded language learning based on unsupervised PCFG induction. Their ap- proach works well when each sentence po- tentially refers to one of a small set of pos- sible meanings, such as in the sportscasting task. However, it does not scale to prob- lems with a large set of potential meanings for each sentence, such as the navigation in- struction following task studied by Chen and Mooney (2011). This paper presents an en- hancement of the PCFG approach that scales to such problems with highly-ambiguous su- pervision. Experimental results on the naviga- tion task demonstrates the effectiveness of our approach. 1 Introduction The ultimate goal of “grounded” language learning is to develop computational systems that can acquire language more like a human child. Given only su- pervision in the form of sentences paired with rel- evant but ambiguous perceptual contexts, a system should learn to interpret and/or generate language describing situations and events in the world. For example, systems have learned to commentate sim- ulated robot soccer games by learning from sample sportscasts (Chen and Mooney, 2008; Liang et al., 2009; B¨orschinger et al., 2011), or understand nav- igation instructions by learning from action traces produced when following the directions (Chen and Mooney, 2011; Tellex et al., 2011). B¨orschinger et al. (2011) recently introduced an approach to grounded language learning using un- supervised induction of probabilistic context free grammars (PCFGs) to learn from ambiguous con- textual supervision. Their approach first constructs a large set of production rules from sentences paired with descriptions of their ambiguous context, and then trains the parameters of this grammar using EM. Parsing a novel sentence with this grammar gives a parse tree which contains the formal mean- ing representation (MR) for this sentence. This ap- proach works quite well on the sportscasting task originally introduced by Chen and Mooney (2008). In this task, each sentence in a natural-language commentary describing activity in a simulated robot soccer game is paired with the small set of actions observed within the past 5 seconds, one of which is usually described by the sentence. Even with this low level of ambiguity in a constrained domain, their method constructs a PCFG with about 33,000 pro- ductions. More fundamentally, their approach is re- stricted to a finite set of potential meaning represen- tations, and the grammar size grows at least linearly with the number of possible MRs, which in turn is inevitably exponential in the number of …

Content truncated to 500 words