Mining human interactions to construct a virtual guide for a virtual fair

Proceedings of the of the EACL 2014 Workshop on Dialogue in Motion (DM), pages 38–42, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Mining human interactions to construct a virtual guide for a virtual fair Andr´es Luna LIIS Group, FaMAF Universidad Nacional de C´ordoba C´ordoba, Argentina andres.ignacio.luna@gmail.com Luciana Benotti LIIS Group, FaMAF Universidad Nacional de C´ordoba C´ordoba, Argentina luciana.benotti@gmail.com Abstract In this paper we describe how we mine in- teractions between a human guide and a human visitor to build a virtual guide. A virtual guide is an agent capable of fulfill- ing the role of a human guide. Its goal is to guide visitors to each booth of a virtual fair and to provide information about the company or organization through interac- tive objects located at the fair. The guide decides what to say, using a graph search algorithm, and decides how to say using generation by selection based on contextual features. The guide decides where to speak at the virtual fair by creat- ing clusters using a data classification al- gorithm to learn in what positions the hu- man guide decided to talk. 1 Introduction and previous work Fairs are spaces where companies that offer simi- lar products and services meet to promote them. A virtual fair emulates a real fair and can be available before the real fair happens in order to promote it to its potential visitors. The virtual fair used in this work is a tourism fair that took place in Mexico, where visitors could find in each company’s booth interactive video and links to tourist companies’ websites promoting particular products. The goal of the virtual guide is to walk the user through the virtual fair, provid- ing information about the companies’ booths and inviting them to click on interactive objects to ob- tain more information. In (Jan et al., 2009) the authors describe a vir- tual guide used to promote an island in the online game Second Life whose goal was to provide in- formation to US army veterans. Our approach dif- fers to that of (Jan et al., 2009) in that the virtual guide learns where to speak and how to realize its contributions from an automatically annotated corpus, rather than by using manually designed rules. However, our guide is not able to interpret utterances from the visitor, its decisions are only based on the visitor behavior. Natural language generation is achieved by adapting the generation by selection method described in (Benotti and De- nis, 2011a; Benotti and Denis, 2011b). The generation by selection method affords the use of complex and human-like sentences, and it does not need rule writing by a dialogue ex- pert or manual annotations, among other of their many advantages. The disadvantage of corpus based generation is that the resulting dialogue may not be fully coherent. Shawar and Atwell (2003; 2005) present a method for learning pattern match- ing rules from corpora in order to obtain the dialogue manager for a chatbot. Gandhe and Traum (2007a; 2007b) …

Content truncated to 500 words