Building a Location Dependent Dictionary for Speech Translation Systems

Mis-translation or dropping of proper nouns reduces the quality of machine translation output or speech recognition output as input of a dialog system. In this paper, we propose an automatic method of building a location dependent dictionary for speech recognition and speech translation systems. The method consists of two parts: location dependent word extraction and word classification. The first part extracts the word by using micro blog data based on Akaike's information criteria. The second part classifies the words by using the Convolutional Neural Net(CNN) trained on crawled data. According to the experimental results, the method extracted around 2,000 location dependent words in the Tokyo area with 75% accuracy.