Cascaded Partial Parsing (Análise sintáctica parcial em cascata)
Jan 1999
Natural language understanding systems need to be robust and efficient when we change from simple toy problems to text from real sources. Syntactic parsing is particularly sensible to real texts, due to the incomplete nature of the grammar and the lexicon, the large variety of syntactic structures and words that occur in the texts and the inevitable mistakes that occur either in the text being parsed or in the grammar and lexicon coding. In this paper, we assume the inevitable partiality of syntactic parses and we propose an efficient architecture for a Portuguese parsing system using a wide coverage grammar. The proposed method is based on the division of the process into three self-contained tasks arranged in a cascade, and pruning the search space between each of two of these tasks. We propose a mixed search strategy and show its superiority to strictly top-down or bottom-up strategies. We emphasize the advantage of submitting the text to a previous part-of-speech tagging process in order to eliminate lexical ambiguity and assigning categories to unknown words. To justify the decisions made, we show experimental results, comparing the performance of several tests with different search strategies and by using or not part-of-speech tagging.
Linguística Computacional: Investigação Fundamental e Aplicações
