Efficient Representation of Linguistic Knowledge for Continuous Speech Understanding

TL;DR

This paper tackles the challenge of parsing continuous speech, which involves processing ambiguous word lattices. The authors propose a novel "rule fusion" technique to create more compact linguistic knowledge representations, significantly reducing parsing time and the number of intermediate phrase hypotheses generated. This leads to a more efficient and effective speech understanding system.

What Problem Does This Paper Address?

Understanding continuous speech is a complex task for computers. Unlike written text, where words are clearly separated, continuous speech recognition systems often produce a "lattice" of possible word sequences, each with an associated probability. This means a single utterance can be represented by hundreds of overlapping word hypotheses. Parsing this lattice to find the most likely and linguistically correct sentence requires a tremendous amount of search. Traditional approaches generate a high number of partial phrase hypotheses, leading to significant memory and processing demands. This paper aims to address this by developing a more efficient way to represent and use linguistic knowledge (syntax and semantics) to make the parsing process faster and less memory-intensive.

Methods

The core of the paper's contribution lies in its "rule fusion" technique for representing linguistic knowledge. Instead of having separate rules for every possible syntactic or semantic combination, the authors propose fusing similar rules into a single, more general "Knowledge Source" (KS).

  • High-Level Representation: The system uses Dependency Grammar for syntax and Caseframes for semantics.
  • Knowledge Source (KS) Compilation: These high-level representations are compiled into KSs, which are then used by the parser. Each KS encapsulates a set of linguistic constraints.
  • Rule Fusion:
    • Restricted Rule Fusion: This technique compacts rules that have the same constituent number and order but differ in their specific syntactic or semantic constraints. It creates a single KS that generalizes these constraints.
  • Generalized Rule Fusion: This is a more powerful approach that can compact rules with different constituent orders or even different numbers of constituents. It aims to create a single KS that covers a broader range of possibilities.
  • Phrase Hypothesis (PH) Representation: The system represents intermediate parsing results as "phrase hypotheses" (PHs). The fusion technique aims to generate "compact" PHs that account for multiple syntactic rules simultaneously, reducing redundancy.
  • Parsing Control Strategy: The parser uses a score-guided approach, prioritizing hypotheses with higher acoustic likelihood.

Datasets

Not discussed in the paper. The paper mentions using "real data on a non-toy application" and refers to a "787-word electronic mail management task" with utterances recorded from a PABX.

Key Insights and Ideas

  • The Power of Generalization: By fusing similar linguistic rules, the system can represent complex grammatical structures more compactly. This is crucial for handling the ambiguity inherent in speech lattices.
  • "Compact" Phrase Hypotheses: The goal is to create phrase hypotheses that represent multiple syntactic possibilities at once, rather than generating a separate hypothesis for each rule. This drastically reduces the search space.
  • Leveraging Ambiguity: The method is designed to cope with and even take advantage of the fact that short words are often missing in speech recognition lattices.
  • Efficiency Gains: The fusion technique leads to a significant reduction in the number of intermediate phrase hypotheses generated, which directly translates to faster parsing times and lower memory usage.

Key Results

The paper presents experimental results comparing three approaches: no fusion, restricted rule fusion, and generalized rule fusion.

  • Reduction in Phrase Hypotheses: Generalized rule fusion reduced the average number of generated phrase hypotheses from 806 (no fusion) to 91.
  • Speedup: Parsing time was reduced from 1.56 seconds (no fusion) to 0.38 seconds (generalized rule fusion) per sentence.
  • Improved Accuracy: The verification phase, applied after initial parsing, increased the percentage of correctly understood sentences from 66.0% to 72.2%.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the fusion techniques greatly reduce intermediate items, each structure might require more work to build. However, the overall reduction in structures compensates for this.
  • The example parsing scenario highlights that even with an "almost optimal" parser, the first solution might not always be correct, necessitating a verification phase.
  • Future Work: The paper doesn't explicitly detail future work, but the success of the generalized rule fusion suggests further exploration and refinement of this technique for even more complex linguistic phenomena and larger vocabularies. The mention of a verification phase also implies ongoing work in ensuring the robustness of the final solution.

Summary updated

BibTeX
@inproceedings{1991efficient,
  title = {{Efficient Representation of Linguistic Knowledge for Continuous Speech Understanding}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1991},
}
Figure