Modified Caseframe Parsing for Speech Understanding Systems

Massimo Poesio, Claudio Rullent

Introduction

Traditional natural language processing (NLP) systems designed for written text often struggle when applied to the domain of continuous speech. In speech recognition, the output is rarely a clean string of words; instead, it is a "lattice" of word hypotheses, each associated with a confidence score.

Lexical hypotheses for the query "In quale regione si trova il Monte Rosa?" with associated confidence scores.

As shown in Figure 2, these hypotheses can vary significantly in certainty. A standard top-down parser, which typically begins with a "caseframe header" (such as a main verb), may fail if that specific word has a low recognition score or is missing from the top candidates.

The SUSY Architecture

To address these challenges, the authors developed the Speech Understanding SYstem (SUSY). SUSY moves away from rigid top-down parsing in favor of a hybrid strategy implemented within a blackboard architecture.

In this architecture, various "Knowledge Sources" (KSs) act as independent specialists that monitor a global memory (the blackboard). When a KS sees a hypothesis it can act upon, it posts a new "Deduction Instance." This allows for opportunistic processing: the system can start with high-confidence "islands"—words like "quale" (which) or "regione" (region)—and expand outward to predict missing or low-confidence components.

Hybrid Parsing and Unified Knowledge

A core innovation of SUSY is the unification of syntactic and semantic constraints. Instead of checking grammar and meaning in separate stages, the system compiles Dependency Grammars and Conceptual Graphs into unified Knowledge Sources.

A trace of the parsing process showing deduction instances (d1-d12) and the relationships between different syntactic/semantic categories.

As illustrated in Figure 1, the system builds a network of deductions. For example, identifying the adjective "QUAL" (d1) can trigger a search for a related noun like "REGIONE" (d4), which in turn helps anchor the verb "TROVA" (d12). This bidirectional triggering—where a KS is activated either by looking for a filler (top-down) or by the discovery of a filler it needs (bottom-up)—is essential for navigating the uncertainty of speech.

Evaluation and Insights

The system was evaluated using Italian geography queries, such as "In quale regione si trova il Monte Rosa?" (In which region is Mount Rosa located?). The results demonstrated that SUSY is significantly more robust than purely top-down systems. By starting with the most certain words, the system can effectively "fill in the blanks" for more ambiguous parts of the sentence.

Key insights from the research include:

  • The Island Strategy: Prioritizing high-certainty words regardless of their position in the sentence structure.
  • Efficiency through Pre-compilation: By compiling rules into KSs beforehand, the system avoids expensive runtime constraint checking.
  • Scalability: The automated compiler makes it relatively simple to extend the system to new domains by generating the necessary Knowledge Sources for new entities.

Conclusion

While the paper focuses on a specific geographic domain and lacks the large-scale statistical benchmarking common in modern NLP, it provides a foundational framework for robust speech understanding. The integration of blackboard architectures with hybrid, unified syntactic-semantic parsing remains a powerful concept for handling noisy, real-world linguistic data.

Summary updated

BibTeX
@inproceedings{poesio1987modified,
  author = {Massimo Poesio and Claudio Rullent},
  title = {{Modified Caseframe Parsing for Speech Understanding Systems}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1987},
}
Figure