Print Email Facebook Twitter Virtual Conversation: Automated Free Speech Dialogue with a Computer Character Title Virtual Conversation: Automated Free Speech Dialogue with a Computer Character Author Ter Heijden, N. Contributor Brinkman, W.P. (mentor) Faculty Electrical Engineering, Mathematics and Computer Science Department Department of Mediamatics Programme Man-Machine Interaction Date 2011-04-20 Abstract Virtual reality exposure therapy (VRET) is already used for treatment of phobias like fear of flying, heights and small places. It is shown to be effective in these areas prompting research into new phobia treatment possibilities. Current papers concerned with social phobia and the use of VRET do not seem to focus on person to person interaction. One of the problems that might have caused this is the difficulty of automating the interaction between a human and a computer character. In this thesis the focus lies on evaluating a possible technique for automating the interaction between a human and a computer for use in a therapeutic setting. The chosen method used avatars which behaviour was simulated using various artificial intelligence techniques. Some of the techniques used in this thesis were borrowed from research related to chatbots and embodied conversational agents. Because these type of conversational agents are not used in a therapeutic setting, the techniques had to be altered to fit this extra requirement. This produced a semi-scripted dialogue structure that would be executed. The user was asked a question and in return the computer would respond on the given answer. A speech recognizer was used, to figure out what the user had said in his free speech response, to choose a predefined response for the computer to utter. These predefined sentences encapsulated human intelligence in response to the computer to improve the realism of the conversation. The dialogues were developed in an iterative process where the dialogues were used in a chatbot like interaction with a test group. The semi-scripted dialogue was tested with four conditions; (1) using only a speech detector, (2) using a speech recognizer to detect a limited set of words, (3) some keywords added to this list and (4) using a human as the controller. The human controller was used as a benchmark point for the automatic techniques to strive for. The experiment put 24 participants in a situation where they had to give a presentation in front of four avatars after which they asked questions about a specific topic. This was repeated four times with four different dialogues resulting in a within-subject comparison of the four possible conditions. The experiment showed that the realism of the dialogue was influenced by two important factors: (1) the speed of the avatar responses and (2) the correctness of these responses. The automated techniques could not yet compete with the human control but were sufficient enough to make them hard to distinguish from the human control condition. To verify the potential usefulness of the program in a therapeutic setting a case study was conducted with one therapist and two clients. The human control condition has a higher workload than the automated conditions. The therapist therefore preferred the automatic conditions because she could actually spend time observing the clients and work on the therapy. The therapist remarked that the dialogues should be focused more on therapy purposes than correctness of the responses. Subject social phobiaVRETcase studyconversational agentsdialogue developmentfree speech interaction To reference this document use: http://resolver.tudelft.nl/uuid:f139cd09-1db6-439c-941a-f9a66dd9df6b Part of collection Student theses Document type master thesis Rights (c) 2011 Ter Heijden, N. Files PDF Virtual_Conversation_Auto ... racter.pdf 3.59 MB Close viewer /islandora/object/uuid:f139cd09-1db6-439c-941a-f9a66dd9df6b/datastream/OBJ/view