Approaching Natural Language Understanding:Study Of Taxonomy Of Intentions Of Speech

Making machine understand human’s language has always been a dream to many people since the start even long before computer was invented. However, until today it is still one of the obscurest problems of artificial intelligence. One of the reasons is that it involves too many complicated issues, like philosophy, linguistics, mathematics, psychology, cognitive science, etc., nevertheless among all of which the most central problems are still meaning and understanding.The meaning of a sentence is not only about its propositional content, but also about the behavior of speaking out this sentence, and the purpose of speaking out, i.e. speak, goal. Accordingly, understanding is a comprehensive process for it not only includes parsing the propositional content of a sentence, it also includes reasoning about the relationship between "speak" and "goal", and constructing the details of the "goal". However, former NLU studies only concentrate on the literal meanings, but overlooked the multiple layers of meanings. The studies of intentions of speech open a new window for the NLU studies. This dissertation is a study of intentions of speech. It considered both theoretical andpractical issues about how to make a machine understand humans’intentions through natural languages, which includes philosophy of intentions, linguistic features, and imitation of interpreting intentions of speech through computational programming, etc. In this study, I totally rejected Searle’s classification and his principles of "illocutionaryforce","illocutionary point" and so on; instead, I see intentions of speech from the angle of "world". One of the merits of classifying intentions of speech in this way is that it furthers our steps to seeing how understanding works in our head. Another merit is that it assigns each proposition with a world label. This paved the way for explaining semantics of the logic representation of intention with Possible Worlds Theory. The research of linguistic features of intentions of speech is intended to bridge the gapbetween theoretical thinking and practical operating. In order to have a deep understanding of the relationship between predicate verbs and its arguments, I classified verbs into types. During classifying the verbs, I mainly considered the image schemata of verb types, and I found that verbs with similar schemata have similar syntactic performances. I also investigated English and Chinese particles. I found that they indeed have the functions of indicating "illocutionary force", just as what Searle had hypothesized; besides, I found they can indicate semantic roles of arguments. Finally, I summed up all of these primary jobs into two sets of syntactic features of intentions of speech. They are the basis of parsing technologies in the following step.Understanding is no way a one-step matter. If it were, then it would be fair enough for us to explain indirect meanings, context, and image, all of them together. I suppose that it is a comprehensive process that includes analyzing, reasoning, and constructing.In this dissertation, I presented two methodologies to analyze intentions of speech. One is based on clustering; the other is based on structure parsing. Clustering is more flexible but it cannot handle the structures of intentions of speech. Structure parsing is very smart in handling structures; however it is too rigid to be fitted into the opening system of natural languages. It would be so great, if there were a method that could melt them into one. But until today, this problem is still so obscure in knowledge representation that nobody has ever found a solution.Interpreting intentions of speech would not be achieved unless it is through reasoning. So, in this study, I also proposed a logic representation of intention and showed how it works in reasoning processes. I argued that the logic representation of intentions of speech has no difference from the one of intentions of ordinary acts."Illocutionary act" can be considered as the means and "perlocutionary act" can be seen as the goal. They are like two ends of one relation. However, I proved that reasoning based on knowledge is unsound no matter what forms it is represented as; because, knowledge is incomplete. Therefore, the reasoning about intentions is intrinsically unsound. But, anyhow, as Reiter said,"we need to reason". I proved that fuzzy knowledge can make up the deficiency of incompleteness and hence fuzzy rules might be better for a natural language understanding system.In constructing process, the function of speech is to provide us with information to evoke knowledge; conversely, the function of knowledge is to provide us with details to "prove" speech, and to make up omitted information. Without these details, we wouldn’t be capable of planning our acts or answering questions.