Annotating Preferences in Negotiation Dialogues

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 105–113, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Annotating Preferences in Negotiation Dialogues Ana¨ıs Cadilhac, Nicholas Asher and Farah Benamara IRIT, CNRS and University of Toulouse 118, route de Narbonne 31062 Toulouse, France {cadilhac, asher, benamara}@irit.fr Abstract Modeling user preferences is crucial in many real-life problems, ranging from individual and collective decision-making to strategic in- teractions between agents and game theory. Since agents do not come with their prefer- ences transparently given in advance, we have only two means to determine what they are if we wish to exploit them in reasoning: we can infer them from what an agent says or from his nonlinguistic actions. In this paper, we an- alyze how to infer preferences from dialogue moves in actual conversations that involve bar- gaining or negotiation. To this end, we pro- pose a new annotation scheme to study how preferences are linguistically expressed in two different corpus genres. This paper describes the annotation methodology and details the inter-annotator agreement study on each cor- pus genre. Our results show that preferences can be easily annotated by humans. 1 Introduction Modeling user preferences is crucial in many real- life problems, ranging from individual and collec- tive decision-making (Arora and Allenby, 1999) to strategic interactions between agents (Brainov, 2000) and game theory (Hausman, 2000). A web- based recommender system can, for example, help a user to identify (among an optimal ranking) the product item that best fits his preferences (Burke, 2000). Modeling preferences can also help to find some compromise or consensus between two or more agents having different goals during a nego- tiation (Meyer and Foo, 2004). Working with preferences involves three subtasks (Brafman and Domshlak, 2009): preference acquisi- tion, which extracts preferences from users, prefer- ence modeling where a model of users’ preferences is built using a preference representation language and preference reasoning which aims at computing the set of optimal outcomes. We focus in this paper on the first task. Handling preferences is not easy. First, specifying an ordering over acceptable outcomes is not trivial especially when multiple aspects of an outcome mat- ter. For instance, choosing a new camera to buy may depend on several criteria (e.g. battery life, weight, etc.), hence, ordering even two outcomes (cameras) can be cognitively difficult because of the need to consider trade-offs and dependencies between the criteria. Second, users often lack complete infor- mation about preferences initially. They build a partial description of agents’ preferences that typi- cally changes over time. Indeed, users often learn about the domain, each others’ preferences and even their own preferences during a decision-making pro- cess. Since agents don’t come with their preferences transparently given in advance, we have only two means to determine what they are if we wish to ex- ploit them in reasoning: we can infer them from what an agent says or from his nonlinguistic actions. In this paper, we analyze how to infer preferences from dialogue moves in …

Content truncated to 500 words