TFLEX: Speeding Up Deep Parsing with Strategic Pruning

This paper presents a method for speeding up a deep parser through backbone extraction and pruning based on CFG ambiguity packing.1 The TRIPS grammar is a wide-coverage grammar for deep natural language understanding in dialogue, utilized in 6 different application domains, and with high coverage and sentence-level accuracy on human-human task-oriented dialogue corpora (Dzikovska, 2004). The TRIPS parser uses a best-first beam search algorithm and a chart size limit, both of which are a form of pruning focused on finding an n-best list of interpretations. However, for longer sentences limiting the chart size results in failed parses, while increasing the chart size limits significantly impacts the parsing speed. It is possible to speed up parsing by implementing faster unification algorithms, but this requires considerable implementation effort. Instead, we developed a new parser, TFLEX, which uses a simpler technique to address efficiency issues. TFLEX combines the TRIPS grammar with the fast parsing technologies implemented in the LCFLEX parser (Rosé and Lavie, 2001). LCFLEX is an all-paths parser which uses left-corner prediction and ambiguity packing, and which was shown to be efficient on other unification augmented context-free grammars. We describe a way to transfer the TRIPS grammar to LCFLEX, and a pruning method which achieves significant improvements in both speed and coverage compared to the original TRIPS parser.