Vector‐based language generation for associatively evoking environmental images

Abstract

A novel language generation method was proposed for spoken dialogue systems. The method selected verbal expressions for sensory inputs using a vector‐based method. A weather‐reporter system was implemented with seven input categories of weather parameters including temperature, humidity, wind velocity, and cloud cover. The system selected the most appropriate short phrase for representing the measured weather parameters to let the listener associatively imagine the reporter’s weather condition descriptions. The system included 149 short phrases for weather expressions, each represented by a 62‐element vector called the template vector. Given the 62‐element input vector for a set of weather parameters, one or two phrases were selected based on the similarity measured between the input and template vectors. The system was able to produce an expression by combining a couple of phrases. A dissimilarity measuring algorithm was proposed to assign the conjunction ‘‘and’’ or adversative ‘‘but’’ automatically to conjoin two phrases. A bilingual capability was also added for enabling the system to output English and Japanese phrases. For English outputs, sentence‐initial phrases were added to the selected phrases depending on the similarity value, while sentence‐finals were added for Japanese outputs. Evaluation experiments showed that automatic conjunction assignment can function with high reliability.