Designing a Message Handling Assistant Using the BDI Theory and Speech Act Theory

This thesis introduces a new approach to designing a Message Handling Assistant (MA). It presents a model of an MA and an intention extraction function for text messages, such as emails and Newsgroups articles. Based on a speech act theory and the belief-desire-intention (BDI) theory of rational agency, we define a generic MA. By interpreting intuitive descriptions of the desired behaviours of an MA using the BDI theory and speech act theory, we conjecture that intentions of messages alone provide enough information needed to capture user models and to reason how messages should be processed. To identify intentions of messages written in natural language, we develop a model of an intention extraction function that maps messages to intentions. This function is modelled in two steps. First, each sentence in a message is converted into a tuple (performative, proposition) using a dialogue act classifier. Second, the sender’s intentions are formulated from the tuples using constraints for felicitous human communication. As an investigation of the use of machine learning technologies for designing the intention extraction function, four dialog act classifiers are implemented and evaluated on Newsgroups articles. The thesis also proposes a semantic communication framework, which integrates the agent and Internet technologies for automatic message composing and ontology exchange services. Specifically, the major contributions of the thesis are: • A new approach to designing a personal assistant (PA) and a message handling assistant (MA). It demonstrates the use of the BDI theory and speech act theory in specifying smart personal assistants in general. • A new algorithm in formulating intentions from dialog acts of on-line messages based on speech act theory. • The application of the dialog act classifiers on Newsgroups articles. To our best knowledge, this is the first study on the application of the dialog act classifiers on on-line messages. • To our best knowledge, for the first time, a comparison of four dialog act classifiers is made for on-line messages.