Analyzing English syntax with a pattern-learning parser

Abstract

Abstract : A dependency analysis system based on pattern recognition and learning logic was developed to infer word-classes and rules of syntactic combination from experience with text which had been analyzed. The characteristics used to form word-classes are the depth in the dependency tree of each word, the direction of its governor, and the same features for each of its immediate neighbors. Syntactic rules of combination show the relation of a word to its governor in the depth pattern of the sentence. The system was tested on 400 elementary Basic English sentences including 300 used earlier by Knowlton in a different learning parser. The Pattern Learning Parser was able to derive a grammar which in a second pass allowed completely correct parsing of all 400 sentences. After experience with 300 sentences it was able to generalize with 77% accuracy to the next 100. In accumulative learning trials after the first 200 sentences it averaged a probability of .9 for accurately parsing each new sentence it encountered. It was concluded that the system was adequate for learning to parse the bulk of Basic English but that further development is required before conclusions about its application to ordinary English can be stated. The system is operational and available on the ARPA/SDC time-shared Q-32 computing system. (Author)