Natural Language Processing Based On Scenarized Knowledge Representation And Its Application In Automatic Text Correction

Knowledge representation is one of the key problem in NLP, but there still has no KR system that is capable to embody the features of natural language expression. Based on analyzing the natural process by which human beings express their knowledge, this paper proposed a major model of the natural language expression of knowledge and the corresponding Scenarized Knowledge Representation system (SKR), which was applied in a semantic text-correction software and achieved a good result.One of the key step of natural language expression of knowledge is to extract the key inforrmation in human's naturally-formed knowledge, which is the base of human's logical reasoning too. based on this point of view, Scenarized knowledge representation (SKR) is proposed by analyzing which information that is needed in the natural language expression of knowledge and how this information should be organized. The proccess to exact the key information in naturally-formed knowledge is called scenarization of knowledge, which involves three major steps: compositional recognition, feature recognition and scenical recognition. Through these process, the compositional information of knowledge, feature and category of elements in knowledge and their roles in special scene are extracted and organized in a granular form, which is the scene struct of knowledge.Besides proposing SKR, the formal reasoning methods in SKR were studied too, which includes inherited, instantial and negative reasoning. Additionally, a group of transformation rules are defined to convert the scenical structure of knowledge into the calculus in first-order predicate logic and makes it possbile to reason by the approaches of auto-reasoning.Based on SKR, a major model of natural language expression of knowledge was proposed. In this model, there are five major methods to describe knowledge in natrual language, ie. naming description, referent description, super-class description, internally structural description and relevant description. The previous three methods are also called Lexicalizing Description because they directly express knowledge with some special words. The last two arecalled Structural Description. The approach of internal structural description converts the scene structure of knowledge into a sequence in language by using the linguistic transformation templates, which primarily generate the sentence structures in natural language. The approach of relevant description express knowledge with its relevant knowledge, and it primarily generate the structures of phrase, attributive clause and parentheses in natural language.Based on this model, we analyzed various general or special clause and phrase structures in Chinese, and proposed a group of corresponding syntactic scene models and linguistic transformation templates. By using these models and templates, it can be implemented to dual-directional convertion between the scene structure of knowledge and its natrual language expression, which makes it possible to use the SKR as the fundation of semantical expression that is needed in various applications of NLP.According to the requirement of natural langauge understanding, this paper proposed the basic means of analyzing sentence with linguistic tranformation templates, and simultaneously proposed three improvements to the algorithms of tranditional grammar parsing, ie. the Chart parser with grammatical type expection, the Chart paser based on prohibited regions and the semantical assistance for grammar parsing based on the knowledge's weak consistancy. In additional, this paper designed a group of tree tranformation rules to convert grammatical structures generated in the grammar parsing into the scenical structures of knowledge.Eventually, the theories and methods proposed in this paper were applied in the development of a software that can semantically find out the improper expressions in documents in special fields.