Understanding of Navy Technical Language via Statistical Parsing

摘要

A key problem in indexing technical information is the interpretation of technical words and word senses, expressions not used in everyday language. This is important for captions on technical images, whose often pithy descriptions can be valuable to decipher. We describe the natural- language processing for MARIE-2, a natural-language information retrieval system for multimedia captions. Our approach is to provide general tools for lexicon enhancement with the specialized words and word senses, and to learn word usage information (both on word senses and word-sense pairs) from a training corpus with a statistical parser. Innovations of our approach are in statistical inheritance of binary co-occurrence probabilities and in weighting of sentence subsequences. MARIE-2 was trained and tested on 616 captions (with 1009 distinct sentences) from the photograph library of a Navy laboratory. The captions had extensive nominal compounds, code phrases, abbreviations, and acronyms, but few verbs, abstract nouns, conjunctions, and pronouns. Experimental results fit a processing time in seconds of 876.2 0858.0 n and a number of tries before finding the best interpretation of 668 .1809.1 n where n is the number of words in the sentence. Use of statistics from previous parses definitely helped in reparsing the same sentences, helped accuracy in parsing of new sentences, and did not hurt time to parse new sentences. Word-sense statistics helped dramatically; statistics on word-sense pairs generally helped but not always.