FeasPar - a feature structure parser learning to parse spontaneous speech

Traditionally, automatic natural language parsing and translation have been performed with various symbolic approaches. Many of these have the advantage of a highly speciic output formalism, allowing ne-grained parse analyses and, therefore, very precise translations. Within the last decade, statistical and connectionist techniques have been proposed to learn the parsing task in order to avoid the tedious manual modeling of grammar and malformation. How to learn a detailed output representation and how to learn to parse robustly even ill-formed input, has until now remained an open question. This thesis provides an answer to this question by presenting a connectionist parser that needs a small corpus and a minimum of hand modeling, that learns, and that is robust towards spontaneous speech and speech recognizer eeects. The parser delivers feature structure parses, and has a performance comparable to a good hand modeled uniication based parser. The connectionist parser FeasPar consists of several neural networks and a Consistency Checking Search. The number of, architecture of, and other parameters of the neural networks are automatically derived from the training data. The search nds the combination of the neural net outputs that produces the most probable consistent analysis. To demonstrate learnability and robustness, FeasPar is trained with transcribed sentences from the English Spontaneous Scheduling Task and evaluated for network, overall parse, and translation performance, with transcribed and speech data. The latter contains speech recognition errors. FeasPar requires only minor human eeort and performs better or comparable to a good symbolic parser developed with a 2 year, human expert eeort. A key result is obtained by using speech data to evaluate the JANUS speech-to-speech translation system with diierent parsers. With FeasPar, acceptable translation performance is 60.5 %, versus 60.8 % with a GLR* parser. FeasPar requires two weeks of human labor to prepare the lexicon and 600 sentences of training data, whereas the GLR* parser required signiicant human expert grammar modeling. Presented in this thesis are the Chunk'n'Label Principle, showing how to divide the entire parsing tasks into several small tasks performed by neural networks , as well as the FeasPar architecture, and various methods for network performance improvement. Further, a knowledge analysis and two methods for improving the overall parsing performance are presented. Several evaluations and comparisons with a GLR* parser, producing exactly the same output formalism , illustrate FeasPar's advantages. Ausgabeformalismus produzieren, zeigen deutlich die Vorteile von FeasPar.