A Natural Language Instructor for pedestrian navigation based in generation by selection

Proceedings of the of the EACL 2014 Workshop on Dialogue in Motion (DM), pages 33–37, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics A Natural Language Instructor for pedestrian navigation based in generation by selection Santiago Avalos LIIS Group, FaMAF Universidad Nacional de C´ordoba C´ordoba, Argentina santiagoe.avalos@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 a method for developing a virtual instructor for pedes- trian navigation based on real interactions between a human instructor and a human pedestrian. A virtual instructor is an agent capable of fulfilling the role of a human instructor, and its goal is to assist a pedes- trian in the accomplishment of different tasks within the context of a real city. The instructor decides what to say using a generation by selection algorithm, based on a corpus of real interactions generated within the world of interest. The instructor is able to react to different requests by the pedestrian. It is also aware of the pedes- trian position with a certain degree of un- certainty, and it can use different city land- marks to guide him. 1 Introduction and previous work Virtual instructors are conversational agents that help a user perform a task. These agents can be useful for many purposes, such as language learn- ing (Nunan, 2004), training in simulated envi- ronments (Kim et al., 2009) and entertainment (Dignum, 2012; Jan et al., 2009). Navigation agents generate verbal route direc- tions for users to go from point A to point B in a given world. The wide variety of techniques to accomplish this task, range from giving complete route directions (all route information in a single instruction), to full interactive dialogue systems which give incremental instructions based on the position of the pedestrian. Although it can recog- nize pre-established written requests, the instruc- tor presented in this work is not able to interpret utterances from the pedestrian, leaving it unable to generate a full dialogue. The instructor’s decisions are based on the pedestrian actual task, his posi- tion in the world, and the previous behavior from different human instructors. In order to guide a user while performing a task, an effective instruc- tor must know how to describe what needs to be done in a way that accounts for the nuances of the virtual world and that is enough to engage the trainee or gamer in the activity. There are two main approaches toward automat- ically producing instructions. One is the selection approach, in which the task is to pick the appropri- ate output from a corpus of possible outputs. The other is the composition approach, in which the output is dynamically assembled using some com- position procedure, e.g. grammar rules. The natural language generation algorithm used in this work is a modified version of the generation by selection method described in (Benotti and De- nis, 2011). The advantages of generation by selection are many: it affords the use of complex and human- like …

Content truncated to 500 words