Pearl: A Probabilistic Chart Parser
David M. Magerman, Mitchell P. Marcus · 1991
This i)al)cr descrihcs a natural language i)ars - ing algorith,n for unrestricted text whicll uses a i)rol)ability-based scoring fimctiou to select the "}mst" I)arse of a sentence. The parser, earl, is a I. ime-a.synchronous bottom-ul) chart I)arscr with Earicy-type top-down prediction which pursues the highest-scoring theory i} the chart, where the score of a theory represents the cxteut I,o which t. he context of the sentmice predicts that iuterpretation. This parser differs h'om previous attempts at stochastic parsers in thai. it uses a richer form of conditional probalfilitics based on context to l)rcdiet likelihood. Pearl also provides a fralnework for incorporating l.he results of previous work iu Imrt-olLsl)cech assignnmnt, mlknown word models, and ol.her Irol)al)ilistic models of linguistic features iuto one parslug tool, interleaving these techniques instead of using the traditional pipeline archiLecture. In preliminary tests, 'Pearl has been successl'ul aL resolving parL-o[-speech and word (in speech processing) ambiguiLy, de[ermiuing categories [or unknown words, and selecLing cotreeL parses first. using a very loosely fiLing covering grammar. 1