Topics as contextual indicators for word choice in SMS conversations

SMS dictation by voice is becoming a viable alternative providing a convenient method for texting in a variety of environments. Contextual knowledge should be used to improve performance. We propose to add topic knowledge as part of the contextual awareness of both texting partners during SMS conversations. Topics can be used for speech applications, if the relation between the conversed topics and the choice of words in SMS dialogs is measurable. In this study, we collected an SMS corpus, developed a topic annotation scheme, and built a topic hierarchy in a tree structure. We validated our topic assignments and tree structure by the Agglomerative Information Bottleneck method, which also proved the measurability of the interrelation between topics and wording. To quantify this relation we propose a naïve classification method based on the calculation of topic distinctive word lists and compare the classifiers' topic recognition capabilities for SMS dialogs with unigram language models. The results demonstrate that the relation between topic and wording is significant and can be integrated into SMS dictation.