Motivation-driven Cognitive Methodology For Natural Language Dialogue

Emulation of the human capability for natural language dialogue has been central to the enterprise of Artificial Intelligence since the very beginning of the field,going back to Alan Turing’s identification of fully human-like dialogue as a litmus test for successful AI(i.e.what is now called the "Turing Test").In recent years,natural language dialogue has also become important in the commercial AI sphere,with high profile offerings such as Apple’s Siri,and Microsoft’s Cortana and Xiao-ice.There have been dramatic recent advances in many related,supporting areas of computational linguistics and AI,such as speech recognition and generation,machine translation,information retrieval,parsing,semantic analysis,sentiment and opinion mining,and many others.In spite of this related progress,however,there have been no truly fundamental breakthroughs in the area of dialogue systems proper.Even the best existing systems operate mainly based on surface-level linguistic cues and simple heuristics,rather than based on richly mapping and responding to the semantic and pragmatic intentions implicit in the human portions of the dialogue.The result of these internal deficits in semantics and pragmatics are externally manifested as dialogue performance that generally becomes tiresome to the human user after a brief period.Creating truly human-level dialogue systems is a huge project,as witnessed by the large teams and large sums of money that technology firms have recently devoted to creating dialogue systems with rather severely limited functionality.Our goal in this thesis is not to fully solve the problem,but to lay a new conceptual and practical foundation for ongoing work.The key hypothesis underlying our work is that,to achieve a broadly useful level of functionality(and as well to ultimately achieve a truly human-level functionality),a dialogue system must be constructed according to the following four principles:deep representation:dialogue should be fundamentally founded on mapping natural language utterances into a rich,and flexibly manipulable,semantic representation,in which an utterance is given a "deep representation" that is largely separate from the properties of the surface-level language involved in an utteranceuncertain logical inference:dialogue should involve inference that is both logical(involving inferential steps such as deduction,induction and abduction)and uncertain(encompassing e.g.probabilistic and fuzzy aspects of uncertainty)foundation in speech act theory:linguistic interactions should be understood as embedded in a larger matrix of embodied interactions--,so that each utterance is understood as a "speech act" with some uniquely linguistic properties,and some pragmatic properties that are shared by speech acts and other sorts of actionsmotivated dialogue control:dialogue control(the choice of what to say when)should be founded on a model of intelligent agent motivation:the system should choose what to say based on its underlying motivations,rather than based on simple cue-response patternsWith the above four principles in mind,our research has been primarily explored them and have achieved to design and implement and a "cognitive dialogue model" embodying the four principles serving as the foundation of our investigation.The major works and contributions of this thesis can be summarized as follows:given a novel formulation of dialogue control in terms of Speech Act Theory and the Psi model of action selectiondemonstrated the utilization of uncertain logical reasoning(via the Probabilistic Logic Networks logic framework)to perform commonsense logical inference from statements uttered in natural languagecreated specific mechanisms for question-answering based on deep representations of natural language knowledge and queries,and demonstrated the utilization of these mechanisms for queries on a large common sense knowledge base(Simple English Wikipedia)designed and implemented a novel framework using hypergraph pattern matching to translate the output of a dependency parser into an abstract logic formalismdesigned and implemented a novel algorithm for surface realization,utilizing hypergraph pattern matching to generate natural language sentences corresponding to semantic hypergraphs,via leveraging knowledge obtained from sentences the system has previously interpretedVia these specific accomplishments,we have demonstrated the theoretical and practical viability of creating natural language dialogue systems based upon the four principles of deep representation,uncertain inference,speech act theory,and motivated dialogue control.