Modified Caseframe Parsing for Speech Understanding Systems

TL;DR

This paper introduces a novel caseframe parsing strategy for speech understanding systems that deviates from traditional approaches by integrating syntactic and semantic knowledge earlier in the process. This allows for more robust and efficient interpretation of spoken language, especially in the presence of speech recognition errors and incomplete utterances.

What Problem Does This Paper Address?

Parsing spoken language presents unique challenges compared to typed text. Speech recognizers produce lattices of lexical hypotheses, not definitive word sequences, leading to ambiguity and the possibility of multiple valid interpretations for a single utterance. Furthermore, speech understanding systems must cope with unrecognised words and overlapping hypotheses. Traditional caseframe parsing, while promising for anchoring interpretation to meaningful parts of an utterance, often struggles with the complexities of speech. Existing approaches either rely too heavily on top-down instantiation of caseframes or keep syntactic and semantic processing separate, leading to inefficiencies and difficulties in handling complex sentences. This paper aims to bridge this gap by proposing a more integrated and flexible caseframe parsing approach suitable for the blackboard architecture commonly used in speech understanding systems.

Methods

The authors propose a modified caseframe parsing strategy for their Speech Understanding System (SUSY). Key innovations include:

  • Alternating Top-Down and Bottom-Up Parsing: Instead of a purely top-down approach, SUSY alternates between predicting caseframes and predicting casefillers based on already recognized components. This allows for more dynamic hypothesis generation and validation.
  • Early Integration of Syntax and Semantics: Caseframes are not used in isolation. Instead, they are merged with syntactic knowledge before being used by Knowledge Sources (KSs). This means that both syntactic and semantic constraints are applied simultaneously, even at early stages of fragment aggregation.
  • Knowledge Source (KS) Compilation: KSs are automatically generated by compiling dependency rules, caseframes, and mapping rules. This declarative approach allows for flexibility and easier modification of the grammar.
  • Blackboard Architecture: The system utilizes a blackboard architecture where KSs cooperate to build and validate "Deduction Instances" (DIs) representing sentence fragments.
  • Representations: Dependency Grammars are used for syntactic rules, and Conceptual Graphs (CGs) are used to represent caseframes. CGs offer powerful unification operations that preserve selectional restrictions.

Datasets

Not discussed in the paper. The system is described as understanding and answering continuous speech queries about Italian geography.

Key Insights and Ideas

  • Syntax and Semantics are Inseparable for Robust Speech Parsing: The paper strongly argues that integrating syntactic and semantic constraints early and simultaneously is crucial for efficient and accurate speech understanding, especially when dealing with the inherent uncertainties of speech recognition.
  • Flexibility Through KS Compilation: Compiling syntactic and semantic knowledge into KSs offers a more flexible and maintainable approach than traditional, monolithic grammars. This allows for easier updates and additions to the system's knowledge.
  • Blackboard Architecture as a Natural Fit: The blackboard architecture proves to be well-suited for managing the complex, cooperative, and iterative nature of hypothesis generation and validation in speech understanding.
  • Bridging the Gap Between Typed and Spoken Language Parsing: The proposed modifications directly address the limitations of applying typed-language parsing techniques to the more challenging domain of speech.

Key Results

The paper describes the implementation and testing of the SUSY system. While specific quantitative results (e.g., accuracy percentages) are not detailed, the authors claim that their approach leads to:

  • More efficient parsing: By applying constraints simultaneously, the system reduces the search space and inferential activity.
  • Improved robustness: The integrated approach helps handle incomplete utterances and speech recognition errors more effectively.
  • Flexibility: The system can handle complex sentences and is adaptable to changes in syntactic rules and domain knowledge.
  • Demonstrated capability: The system successfully processes an example Italian geographical query, illustrating the integration of top-down and bottom-up processing and the final interpretation through instantiated caseframes.

Limitations and Future Work

  • Limitations: The authors acknowledge that while their approach reduces the number of KSs generated compared to some other methods, adding new entity types to the domain can still lead to the definition of a significant number of new KSs (around 15 new KSs per new entity type in their experience). They also note that representing complex restrictions within CAGRs can lead to unnecessary complexities.
  • Future Work: The authors state that they are currently working on improving the compiler and the overall blackboard architecture of the system.

Summary updated

BibTeX
@inproceedings{1987modified,
  title = {{Modified Caseframe Parsing for Speech Understanding Systems}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1987},
}
Figure