Oregon Health & Science University
Syntactic analysis is important for many natural language processing (NLP) tasks, but constituency parsing is computationally expensive, often prohibitively so. Consumers who would be best served by constituency parsing are often forced by resource constraints to settle for less effective approaches. In this work, we examine the barriers to efficient context-free processing, and present several approaches to increasing throughput and reducing latency. We describe a matrix grammar representation and demonstrate that its memory- and cache-efficient properties yield considerable speedups in inference. We present an efficient and parallelizable refactoring of the standard dynamic programming algorithm, based on that matrix grammar, which further improves latency. We introduce several methods of targeting efficiency as an objective during model training, including some which are applicable outside of constituency parsing. We present several methods of incorporating efficiency concerns into the process of training latent-variable grammars, and experimental trials demonstrating the effects of each approach.
We present several methods of text normalization (prior to grammar induction) that reduce the grammar size considerably and improve parse efficiency with minimal accuracy degradation. We explore the characteristics of a grammar that impact efficient inference, and present a regression model predicting inference time from those characteristics. We incorporate the accuracy and efficiency models into latent-variable grammar training, allowing a controlled tradeoff between speed and accuracy, and optimizing the trained grammar for the desired operating point. Finally, we explore various optimization criteria for chart decoding, and efficient approximations thereof. We replicate prior results on max-rule decoding, and explore other options which have not been well explored in prior work. We present efficient approximations of these methods, capturing some of those gains without severe computational cost. In aggregate, our methods achieve a speedup of approximately 20 x for Viterbi decoding, and 3 x for alternate decoding methods.
Center for Spoken Language Understanding
School of Medicine
Dunlop, Aaron Joseph, "Efficient Latent-variable Grammars : Learning and Inference" (2014). Scholar Archive. 3487.