The Effectiveness of Corpus-Induced Dependency Grammars for Post-processing Speech

Mary P. Harper, Christopher M. White, Wen Wang, Michael T. Johnson, Randall A. HelzermanANLP 2000 · OpenReview

Abstract

M. P. Harper, C. M. White, W. Wang, M. T. Johnson, R. A. Helzerman. 1st Meeting of the North American Chapter of the Association for Computational Linguistics.

  1. This paper demonstrates that learning complex syntactic constraints directly from a corpus significantly improves the accuracy of speech recognition systems. By using Constraint Dependency Grammars (CDG) as a post-processor, the researchers achieved a jump in sentence accuracy from 81.2% to nearly 92% on the Naval Resource Management task.

Introduction

Standard speech recognition systems often rely on "N-gram" models—simple statistical guesses about which word comes next based on the previous one or two words. While computationally efficient, these models are syntactically "shallow" and fail to capture long-distance relationships between words, such as those between a subject and a distant verb.

To improve accuracy, more powerful grammars are required. However, tightly integrating complex grammars like Context-Free Grammars (CFG) directly into acoustic search is computationally prohibitive. A more viable approach is "loose coupling," where a speech engine generates a word graph of possibilities, and a sophisticated parser acts as a "judge" to select the most grammatically sound hypothesis. The primary bottleneck in this approach is the difficulty of hand-writing these complex grammars. This study investigates whether such grammars can be automatically induced from annotated corpora.

Methods and System Architecture

The researchers utilize Constraint Dependency Grammar (CDG), which focuses on dependencies between words rather than building phrase-structure trees. CDG operates by filtering out impossible interpretations using unary and binary constraints.

The core innovation is the transition from a Conventional CDG (hand-written over a year) to Corpus-Induced CDGs. The process involves:

  1. Abstract Role Values (ARVs/ARVPs): Representing grammatical rules as templates describing spatial and syntactic relationships between word categories.
  2. Grammar Induction: Automatically extracting these templates from the Resource Management (RM) corpus.
  3. Template Expansion: Using templates for common entities (dates, times, coordinates) to improve generalization to unseen data.
The CDG-based speech recognition pipeline.

As shown in the system architecture, the pipeline begins with a speech signal processed by an HMM recognizer, which outputs a word graph. This graph is then processed by the CDG-based NL processor using induced constraints, followed by a hypothesis selection stage to determine the top sentence.

Experimental Setup

The study utilized two mid-size speech corpora:

  • Resource Management (RM): 2,845 distinct sentences used for training and extracting the induced grammars.
  • Extended Resource Management (RM2): 7,396 sentences from the same domain used to test the generalization capabilities of the learned grammars.

Results and Discussion

The experimental results demonstrate a significant advantage for the data-driven approach. The induced "Full Mod" grammar achieved a sentence accuracy of 91.94%, a substantial improvement over the 81.20% achieved by the hand-crafted conventional grammar.

Performance comparison of induced grammar variations versus the conventional grammar.

Word accuracy also saw an increase from 96.10% to 98.19%. Beyond accuracy, efficiency was a key finding: the "Feature Mod" variation was nearly eight times faster than the conventional grammar while maintaining superior accuracy. This efficiency is attributed to the use of fast table lookups for ARVs.

However, a trade-off exists between precision and coverage. Induced grammars are "tight," meaning they effectively filter out acoustic "garbage" but may fail on sentences slightly outside the training distribution. Template expansion helped mitigate this, improving coverage on the RM2 dataset by approximately 10%.

Conclusion

Corpus-induced dependency grammars provide a powerful and efficient alternative to hand-crafted syntactic models for speech recognition post-processing. By learning constraints directly from data, these systems achieve higher accuracy and better filtering of incorrect hypotheses. Future work will focus on expanding semantic templates and developing stochastic CDGs to improve robustness and coverage.

Summary updated

Abstract

M. P. Harper, C. M. White, W. Wang, M. T. Johnson, R. A. Helzerman. 1st Meeting of the North American Chapter of the Association for Computational Linguistics. 2000.
BibTeX
@inproceedings{harper2000effectiveness,
  author = {Mary P. Harper and Christopher M. White and Wen Wang and Michael T. Johnson and Randall A. Helzerman},
  title = {{The Effectiveness of Corpus-Induced Dependency Grammars for Post-processing Speech}},
  booktitle = {ANLP},
  year = {2000},
  url = {https://openreview.net/forum?id=sE3lv7B7D-k},
}
Figure