Controlling Bidirectional Parsing

Traditional models of parsing as used in interfaces have shown to be weak and ine ective in complex tasks such as processing of naturally-occurring texts. Broad coverage parsers go mad when confronted with extended inputs without su cient information to control the interpretation process. The use of e ective control strategies is necessary to overcome these shortcomings. Extra-linguistic criteria (e.g. statistics, semantic or goal-driven heuristics) can be employed to reduce the combinatorics of parsing and to avoid misdirected e orts, focusing the analysis on the most promising solutions. In this paper we present an approach to text parsing where an agendabased bidirectional chart parser is coupled with a set of extra-linguistic control strategies. Those strategies a ect the agenda management favoring some tasks, delaying or pruning others. A preprocessing phase elaborate the text in order to collect information for the parsing control: text segmentation produces hints on the text supercial structure, and statistical classi cation provides the templates to ll. Afterwards the linguistic analyzer processes the text in two steps: segment parsing and segment combination. The control strategies are mainly applied during the latter phase. Currently implemented strategies integrate both general and domain-speci c criteria such as grammar rule scores and goal-driven (i.e. template-based) information. The two-step linguistic analyzer is able to gracefully cope with the linguistic complexity of the input, pursuing a complete analysis only when it is feasible; the feature of bidirectionality allows to maximize the coverage on chunks of input. Some preliminary results seem to show an improvement of the parsing e ciency.