Automated Planning for Situated Natural Language Generation

Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1573–1582, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Automated planning for situated natural language generation Konstantina Garoufi and Alexander Koller Cluster of Excellence “Multimodal Computing and Interaction” Saarland University, Saarbr¨ucken, Germany {garoufi,koller}@mmci.uni-saarland.de Abstract We present a natural language genera- tion approach which models, exploits, and manipulates the non-linguistic context in situated communication, using techniques from AI planning. We show how to gen- erate instructions which deliberately guide the hearer to a location that is convenient for the generation of simple referring ex- pressions, and how to generate referring expressions with context-dependent adjec- tives. We implement and evaluate our approach in the framework of the Chal- lenge on Generating Instructions in Vir- tual Environments, finding that it performs well even under the constraints of real- time generation. 1 Introduction The problem of situated natural language gen- eration (NLG)—i.e., of generating natural lan- guage in the context of a physical (or virtual) environment—has received increasing attention in the past few years. On the one hand, this is be- cause it is the foundation of various emerging ap- plications, including human-robot interaction and mobile navigation systems, and is the focus of a current evaluation effort, the Challenges on Gener- ating Instructions in Virtual Environments (GIVE; (Koller et al., 2010b)). On the other hand, situated generation comes with interesting theoretical chal- lenges: Compared to the generation of pure text, the interpretation of expressions in situated com- munication is sensitive to the non-linguistic con- text, and this context can change as easily as the user can move around in the environment. One interesting aspect of situated communica- tion from an NLG perspective is that this non- linguistic context can be manipulated by the speaker. Consider the following segment of dis- course between an instruction giver (IG) and an instruction follower (IF), which is adapted from the SCARE corpus (Stoia et al., 2008): (1) IG: Walk forward and then turn right. IF: (walks and turns) IG: OK. Now hit the button in the middle. In this example, the IG plans to refer to an ob- ject (here, a button); and in order to do so, gives a navigation instruction to guide the IF to a conve- nient location at which she can then use a simple referring expression (RE). That is, there is an inter- action between navigation instructions (intended to manipulate the non-linguistic context in a cer- tain way) and referring expressions (which exploit the non-linguistic context). Although such subdi- alogues are common in SCARE, we are not aware of any previous research that can generate them in a computationally feasible manner. This paper presents an approach to generation which is able to model the effect of an utter- ance on the non-linguistic context, and to inten- tionally generate utterances such as the above as part of a process of referring to objects. Our ap- proach builds upon the CRISP generation system (Koller and Stone, 2007), which translates gener- …

Content truncated to 500 words