Lavie, A., "GLR* : a Robust Grammar-Focused Parser For
Total Page:16
File Type:pdf, Size:1020Kb
GLR* : A Robust GrammarFocused Parser for Spontaneously Spoken Language Alon Lavie May 1996 CMUCS96126 School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213 Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy Thesis Committee: Masaru Tomita, Chair Jaime Carbonell Alex Waibel Edward Gibson, MIT Copyright c 1996 Alon Lavie Keywords: Natural Language Processing, Speech Understanding, Machine Translation, Parsing, Generalized LR Parsing, JANUS Abstract The analysis of spoken language is widely considered to be a more challenging task than the analysis of written text. All of the difficulties of written language can generally be found in spoken language as well. Parsing spontaneous speech must, however, also deal with problems such as speech disfluencies, the looser notion of grammaticality, and the lack of clearly marked sentence boundaries. The contamination of the input with errors of a speech recognizer can further exacerbate these problems. Most natural language parsing algorithms are designed to analyze “clean” grammatical input. Because they reject any input which is found to be ungrammatical in even the slightest way, such parsers are unsuitable for parsing spontaneous speech, where completely grammatical input is the exception more than the rule. This thesis describes GLR*, a parsing system based on Tomita's Generalized LR parsing algorithm, that was designed to be robust to two particular types of extragrammaticality: noise in the input, and limited grammar coverage. GLR* attempts to overcome these forms of extragrammaticality by ignoring the unparsable words and fragments and conducting a search for the maximal subset of the original input that is covered by the grammar. The parser is coupled with a beam search heuristic, that limits the combinations of skipped words considered by the parser, and ensures that the parser will operate within feasible time and space bounds. The developed parsing system includes several tools designed to address the difficulties of parsing spontaneous speech. To cope with high levels of ambiguity, we developed a statistical disambigua tion module, in which probabilities are attached directly to the actions in the LR parsing table. The parser must also determine the “best” parse from among the different parsable subsets of an input. We thus designed a general framework for combining a collection of parse evaluation measures into an integrated heuristic for evaluating and ranking the parses produced by the GLR* parser. This framework was applied to a set of four parse scoring measures developed for the JANUS scheduling domain and the ATIS domain. We added a parse quality heuristic, that allows the parser to selfjudge the quality of the parse chosen as best, and to detect cases in which important information is likely to have been skipped. To demonstrate its suitability to parsing spontaneous speech, the GLR* parser was integrated into the JANUS speech translation system. Our evaluations on both transcribed and speech recognized input have indicated that the version of the system that uses GLR* produces between 15% and 30% more acceptable translations, than a corresponding version that uses the original nonrobust GLR parser. We also developed a version of GLR* that is suitable to parsing word lattices produced by the speech recognizer, and investigated how lattice parsing can potentially overcome errors of the speech recognizer and further improve endtoend performance of the speech translation system. Acknowledgements There are many who have helped me along the long road that has culminated in this thesis, some in direct and obvious ways, others in small and supposedly unrelated ways, but important nonetheless. I would like to thank my advisor, Masaru Tomita, for inspiring and supporting my research interests in parsing algorithms, for introducing me to the problems of spoken language parsing and understanding, for suggesting the topic of this thesis, and for guiding me in my thesis work. I also wish to thank the other members of my thesis committee: Jaime Carbonell, Alex Waibel and Ted Gibson. Alex brought me into the JANUS speechtospeech translation project, which provided a natural and practical setting for applying and testing my work. I am particularly grateful for his guidance on system performance issues and evaluation methods. Ted Gibson provided me with objective insight about my work, and with careful and well thought comments on my thesis draft. Special thanks are due to Jaime Carbonell, for sharing his experienced perspective on Machine Translation and AI, for stepping into the “advisor shoes” in Tommy's absence, and for his very helpful comments and suggestions on the preliminary draft of this thesis. I would like to thank all my friends and colleagues in the JANUS project, for providing a fun and exciting environment for conducting research on speech understanding and translation. Particular thanks go to Lori Levin, Donna Gates, Oren Glickman, Noah Coccaro, Carolyn Rose,´ Marsal Gavalda,` Laura Mayfield, Keiko Horiguchi and Kaori Shima, who worked closely with me on the project, and assisted me in a variety of experiments and evaluations reported in this thesis. On the personal side, there is a whole bunch of friends, colleagues and family members, whom I would like to thank for their support and encouragement. I will not even attempt to list them all, for fear that surely someone will be forgotten. Yet, I feel a need to mention a handful of people to whom special thanks are due: To my friend and exofficemate Shai, who greeted and hosted me when I first arrived here in Pittsburgh, spent five years with me in a windowless Wean office, and helped me with so many “system hacking” questions... To my best friend and next door neighbor Dean, who is always there to listen and give advice, and is really good at putting things in perspective. To my good friend Orna, for being just that, but also for the numerous dinners and coffee breaks during the last two intense months of writing, that made them so much more bearable. To my family in Israel, for their support, and in particular to my father, who not only provided constant encouragement and advice, but has also been a true role model for me to follow. And most of all, to Bob, for taking care of “things” during the very busy months of final thesis writing, and for traveling with me along the longest and most difficult segments of the road to the PhD. I couldn't have done it without you. Contents 1 Introduction and Overview 7 1.1 Introduction ¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡ 7 1.1.1 Extragrammaticalities in Spontaneous Speech ¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡ 8 1.2 Research Goals ¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡ 10 1.3 Thesis Summary ¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡ 11 1.3.1 Foundation: GLR Parsing ¡¢¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡ 11 1.3.2 The GLR* Parsing Algorithm ¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡ 12 1.3.3 Statistical Disambiguation ¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡¢¡£¡¢¡£¡ 13 1.3.4 Parse Evaluation Heuristics ¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡ 16 1.3.5 Parsing Spontaneous Speech using GLR* ¡¢¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡ 18 1.3.6 Parsing Speech Lattices using GLR* ¡¢¡£¡¢¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡ 20 1.4 Thesis Contributions ¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡¢¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡¥¡¤¡¢¡£¡ 22 1.5 Previous Related Work ¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡¢¡£¡¢¡£¡ 22 1.5.1 Other Approaches to Robust Parsing ¡¢¡£¡¢¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡¢¡£¡¢¡£¡ 22 1.5.2 Other Work on Parsing Speech ¡£¡¢¡¤¡¥¡¤¡¢¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡ 24 1.5.3 Adaptive and Selflearning Approaches ¡¤¡¢¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡ 27 2 Generalized LR Parsing 28 2.1 Principles of LR Parsing ¡¢¡£¡¢¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡¢¡£¡¢¡£¡ 28 2.2 The GLR Parsing Algorithm ¡¢¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡¢¡£¡¢¡£¡ 29 2.2.1 The Graph Structured Stack ¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡ 29 2.2.2 Local Ambiguity Packing ¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡ 30 2.2.3 Shared Packed Forests ¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡ 31 2.3 Computational Complexity and Performance of the GLR Parser ¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡ 32 2.4 GLR and Unification Based Grammars ¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡ 33 3 The GLR* Parsing Algorithm 35 3.1 Introduction ¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡ 35 3.2 The Unrestricted GLR* Parsing Algorithm ¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡¢¡£¡¢¡£¡ 36 3.2.1 Design Considerations ¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡¢¡£¡¢¡£¡ 36 3.2.2 Outline of the Unrestricted GLR* Algorithm ¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡ 37 3.3 Enhanced Local Ambiguity Packing ¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡ 38 3.4 An Example ¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡¢¡£¡¢¡£¡ 40 3.5 Complexity and Performance of the Unrestricted GLR* Algorithm ¡¥¡¤¡¢¡¥¡¤¡¢¡£¡ 46 3.5.1 Time complexity of Unrestricted GLR* ¡£¡¢¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡ 46 3.5.2 Runtime Performance of Unrestricted GLR* ¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡ 49 3.6 Controlling Parser Search ¡¤¡¥¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡£¡ 50 ¡¢¡¤¡¥¡¤¡¢¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡ 3.6.1 The ¦ word Skip Limit Heuristic 52 1 3.6.2 The Beam Search Heuristic ¡£¡¢¡£¡¢¡£¡¢¡¤¡¥¡¢¡¤¡¥¡¤¡¢¡£¡¢¡£¡¢¡£¡¢¡¢¡£¡¢¡¤¡ 53 3.6.3 Empirical Evaluation