KNOWLEDGE BASED NEURAL NETWORKS FOR NATURAL LANGUAGE PARSING

Abstract

The work presents a connectionist approach to parsing of English sentences using Rule based connectionism. To implement the scheme, a general connectionist architecture has been developed which is capable of accepting variable number of input and output nodes. The basic grammar rules representing a fragment of English language have been considered to determine the initial topology of the neural network, which includes deciding the number of layers, input, output and hidden nodes. The preterminals associated with a sentence are mapped onto the input nodes and the start symbol is mapped onto the single output node. The training instances are fixed maximum length sentences. The connectionist neural network program checks for grammatical validity of the sentence input. For valid sentences, parse trees are obtained as a by-product of the grammaticality determination procedure. To determine the validity of complex and compound sentences, these are first broken up into_a main clause and a.. subordinate -clause by the clause analyser. These clauses are independently tested for validity and if both are found valid then the complex/compound sentence is said to be valid. The program has been tested for a large number of English sentences using two grammars. The program is written in `C' language and runs on TATA ELXSI under UNIX environment. It can also be run on DOS environment. The program code is about 1300 lines in length.