Computers getting ready for smalltalk
While fictional robots like Star Wars' R2D2 are depicted interacting with humans who seem to speak quite similar languages, real-life devices like mobile smart-phones or speech-controlled PC software must adapt to a wide range of users. Some may speak different languages and dialects, have different skill levels or ask their appliance to do quite different tasks for them. One example from the project is AthosMail, a multilingual speech-based e-mail application that can adapt to such different users, including Finnish, English and Swedish speakers. Its speech recognition interface can be used for most e-mail reading tasks. In addition, AthosMail, which is based on the open-source Mailman application, offers components for text-processing, dialogue management, semantic template construction, user modelling, integrated tutoring, and random indexing. Not only do Athos-based applications understand what a user wants to do, but they also analyse how a user responds and draws conclusions on his or her competence level. The system can change its behaviour accordingly in order to maintain the 'natural' feel of interaction. The 'cooperativity model' component decides on the level of language to use in dialogue with the user, based on observation of their experience levels. The application, which uses modern computing methods like symbolic computing and neural networks, is available in online and offline versions, which have slight differences. While the offline version stores a user's profile locally and keeps on learning about him whenever he uses the application, the online version learns from scratch in every new session.