A Game-Theoretic Approach to Generating Spatial Descriptions

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 410–419, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics A Game-Theoretic Approach to Generating Spatial Descriptions Dave Golland UC Berkeley Berkeley, CA 94720 dsg@cs.berkeley.edu Percy Liang UC Berkeley Berkeley, CA 94720 pliang@cs.berkeley.edu Dan Klein UC Berkeley Berkeley, CA 94720 klein@cs.berkeley.edu Abstract Language is sensitive to both semantic and pragmatic effects. To capture both effects, we model language use as a cooperative game between two players: a speaker, who gener- ates an utterance, and a listener, who responds with an action. Specifically, we consider the task of generating spatial references to ob- jects, wherein the listener must accurately identify an object described by the speaker. We show that a speaker model that acts op- timally with respect to an explicit, embedded listener model substantially outperforms one that is trained to directly generate spatial de- scriptions. 1 Introduction Language is about successful communication be- tween a speaker and a listener. For example, if the goal is to reference the target object O1 in Figure 1, a speaker might choose one of the following two ut- terances: (a) right of O2 (b) on O3 Although both utterances are semantically correct, (a) is ambiguous between O1 and O3, whereas (b) unambiguously identifies O1 as the target object, and should therefore be preferred over (a). In this paper, we present a game-theoretic model that cap- tures this communication-oriented aspect of lan- guage interpretation and generation. Successful communication can be broken down into semantics and pragmatics. Most computational Figure 1: An example of a 3D model of a room. The speaker’s goal is to reference the target object O1 by de- scribing its spatial relationship to other object(s). The listener’s goal is to guess the object given the speaker’s description. work on interpreting language focuses on compo- sitional semantics (Zettlemoyer and Collins, 2005; Wong and Mooney, 2007; Piantadosi et al., 2008), which is concerned with verifying the truth of a sen- tence. However, what is missing from this truth- oriented view is the pragmatic aspect of language— that language is used to accomplish an end goal, as exemplified by speech acts (Austin, 1962). Indeed, although both utterances (a) and (b) are semantically valid, only (b) is pragmatically felicitous: (a) is am- biguous and therefore violates the Gricean maxim of manner (Grice, 1975). To capture this maxim, we develop a model of pragmatics based on game the- ory, in the spirit of J¨ager (2008) but extended to the stochastic setting. We show that Gricean maxims 410 fall out naturally as consequences of the model. An effective way to empirically explore the prag- matic aspects of language is to work in the grounded setting, where the basic idea is to map language to some representation of the non-linguistic world (Yu and Ballard, 2004; Feldman and Narayanan, 2004; Fleischman and Roy, 2007; Chen and Mooney, 2008; Frank et al., 2009; Liang et al., 2009). Along similar lines, past work has also focused on …

Content truncated to 500 words