Introduction
Continuous speech recognition presents a significant computational challenge compared to processing written text. In speech, words lack clear delimiters, and acoustic noise often results in a "word lattice"—a dense graph of overlapping word hypotheses with varying probability scores. To extract meaning, a linguistic processor must navigate this lattice to find the most probable path that satisfies both grammatical and semantic constraints. In languages with relatively free word order, such as Italian, the number of potential partial phrases can explode, leading to massive memory requirements and redundant calculations.
This paper introduces "Rule Fusion," a technique designed to streamline this process. By merging multiple grammatical and semantic rules into compact structures, the authors achieved an 8x reduction in phrase hypotheses and a 4x speedup in processing time without sacrificing accuracy.
The Challenge of Linguistic Parsing
Standard parsers typically create a separate memory object for every possible rule application. In a speech recognition context, where the input is inherently uncertain, this leads to an exponential increase in redundant partial phrases. The authors argue that while individual rules are human-readable, they are computationally inefficient when applied to noisy word lattices.
To address this, the proposed system utilizes a two-step approach to make linguistic knowledge more "parsimonious":
- Knowledge Compilation: High-level rules (Dependency Grammar for syntax and Caseframes for semantics) are compiled into Knowledge Sources (KSs) that define all possible connections between word classes.
- Rule Fusion: This core innovation groups rules with similar structures into a single "compact" KS. This allows the system to account for many syntactic variations simultaneously using a single template.
Methods and Implementation
The system employs Generalized Rule Fusion to handle constituent orders and varying numbers of elements within a single structure. To maintain high performance, bit-vector operations and bit-masks are used to track which specific rules within a fused KS remain "active" as new words are added to a phrase.
Furthermore, the parser is designed to handle the common problem of missing words. Short prepositions (e.g., "of," "to," or the Italian "di," "da") are frequently missed by acoustic recognizers. The system compensates by creating virtual placeholders, allowing the search to continue even when the input lattice is incomplete.
Experimental Results
The effectiveness of Generalized Rule Fusion was evaluated using 600 word lattices generated from an electronic mail management task (787-word vocabulary) across 10 different speakers. The results demonstrate a dramatic improvement in efficiency:
| Metric | No Fusion | Restricted Fusion | Generalized Fusion |
|---|---|---|---|
| Phrase Hypotheses (PHs) | 806 | 383 | 91 |
| CPU Time (seconds) | 1.56s | 0.86s | 0.38s |
Generalized fusion reduced the number of generated hypotheses by nearly 90% compared to the baseline. Additionally, the authors implemented an "Acoustical Verification" safety net. By collecting the top linguistic candidates and re-scoring them with detailed acoustic data, the system improved the percentage of correctly understood sentences from 66.0% to 72.2%.
Conclusion
The "Rule Fusion" approach demonstrates that while a compact Knowledge Source is more complex to process individually, it prevents the exponential explosion of redundant data structures. By prioritizing "super-structures" over thousands of nearly identical partial hypotheses, the system achieves significant gains in both speed and memory efficiency, making it highly suitable for the complexities of continuous speech in languages with flexible syntax.
Summary updated