Analysing Ill-formed Inputs with Parallel Chart-based Techniques

In this paper, we describe a parallel implementation of chart-based parser which performs robustness to analyse grammatical ill-formed inputs in loosely-coupled environment. The parser is composed of parallel bottom-up process which parses the inputs under the grammar rules and parallel top-down process which tries to recover the existing ill-formedness when the bottom-up process fails to nd a complete parse. The top-down process resembles a resolution of tree searching problem, utilizing the intermediate parsing information generated in the bottom-up process. We propose a method to use a dynamic task distribution as a core mechanism to control distributing tasks during the parsing process. Our parser is implemented on a parallel inference machine, named PIM. Using 256 processors, it ran 60-170 times faster than a serial version in the case of long ill-formed inputs with multiple errors.