Fuzzy Logic-Based Natural Language Processing and Its Application to Speech Recognition JIPING SUN, FAKHRI KARRAY, OTMAN BASIR & MOHAMED KAMEL Department of Electrical and Computer Engineering University of Waterloo, Waterloo Ontario, N2L 3G1, Canada Abstract: In this paper we describe a fuzzy logic-based language processing method, which is applied to speech recognition. Our purpose is to create a system that can learn from a linguistic corpus the fuzzy semantic relations between the concepts represented by words and use such relations to process the word sequences generated by speech recognition systems. In particular, the system will be able to predict the words failed to be recognized by a speech recognition system. This will help to increase the accuracy of a speech recognition system. This will also serve as the first stage of deep semantic processing of speech recognition results by providing “semantic relatedness” between the recognized words. We report the fuzzy inference rule learning system, which we have developed and also report the experimental results based on the system. Key-words: Fuzzy logic, Natural Language Analysis, Speech Recognition, Corpus Linguistics. 1 Introduction The complexity of natural language has made people apply various kinds of “soft” computing techniques for its analysis. Besides statistical, connectionist and other approaches, the fuzzy logic-based approach provides another alternative for effective natural language analysis. It is commonly recognized that many phenomena in natural language lend themselves to descriptions by fuzzy mathematics, including fuzzy sets, fuzzy relations and fuzzy logic. By defining a fuzzy logic system and acquiring proper rules, we hope that difficulties in analysis of speech can be alleviated. As far as reference is concerned, words and their meanings (the referred objects or their measurements in the world) are often in a fuzzy relationship. This is important for grounded systems such as controlling robots. On the other hand, for the mainstream of NLP research, words themselves are the objects of description. It is natural to think that the language external fuzziness could be interpolated into language and fuzzy mathematical approaches are appropriate tools in solving problems. Fuzzy logic has been successfully applied to the description of words’ meanings as related to language external phenomena. Fuzzy linguistic descriptors have been used in control systems, in which mappings can be established between fuzzy linguistic terms and physical quantities. “Hot”, “cold”, for example, can serve as labels for fuzzy sets to which temperature readings can be mapped into membership degrees. Fuzzy logic rules for control systems can accept fuzzy descriptors in both the premises and the consequents to simulate human-like inferencing. Another case of fuzzy application is natural language-driven database search. Here the semantics of words can be expressed as fuzzy membership functions for certain database search keys [Medina, Vila]. A language internal fuzzy treatment is found in [Subasic], in which affect types of certain words in documents are dealt with as fuzzy sets. Words representing emotions are mapped to these fuzzy sets. The difference between this case and the previous two is that the latter dealt with language internal fuzzy phenomena.