A Massively Parallel Architecture for Natural Language Parsing – Some Results

Abstract

This paper explores a new architecture for massively parallel natural language parsing implemented in Squeak Smalltalk. The FastTalk architecture is based in relatively simple operations done by actors associated with each possible syntactic category that each word might have in an input text. The FastTalk architecture differs from previous approaches by enabling the analysis of word strings in any order, and not only from left to right. Thus, this approach provides much more parallelism and additionally does not require time complex operations such as backtracking. The growth observed for the number of actors necessary to analyze a single sentence by a naive algorithm is drastically reduced by applying some heuristics derived from special characteristics of natural language. We claim that with this reduction it is possible to efficiently parse real texts and obtain every possible parse of an input.