Control Structures and Theories of Interaction in Speech Understanding Systems

E.j. Briscoe, B.k. BoguraevACL 1984

TL;DR

This paper proposes a novel theory of "Interactive Determinism" for speech understanding systems (SUSs). It argues that efficient SUSs don't need complex, general-purpose control mechanisms like blackboards, but rather can rely on a simpler, domain-specific architecture driven by strictly defined, non-serial interactions between processing components. This approach promises more efficient and tractable SUSs by leveraging the inherent temporal organization of speech.

What Problem Does This Paper Address?

Building robust speech understanding systems (SUSs) is challenging because understanding speech involves integrating information from various sources – like recognizing words, understanding grammar (morphosyntax), and interpreting the rhythm and intonation (prosody). Traditional SUS architectures often struggle with how these different components should communicate and coordinate their efforts. Many existing systems use complex, general-purpose control mechanisms (like blackboard architectures) that are powerful but can be inefficient and difficult to manage. The authors argue that these general approaches are unnecessary and that a more domain-specific theory of interaction is needed to build efficient SUSs. They specifically focus on the interaction between morphosyntactic and prosodic analysis, highlighting the need for a theory that dictates when and how these components should communicate.

Methods

The paper introduces the concept of Interactive Determinism (ID). This theory posits that speech is temporally organized in a specific way, allowing for efficient, non-serial interactions (NSIs) between processing components. Instead of complex, general-purpose control, the ID model proposes a simpler architecture where components communicate through strictly defined, bidirectional channels.

Key aspects of their proposed method include:

  • Non-Serial Interactions (NSIs): Communication that goes beyond the standard, sequential flow of information between components. For example, prosodic information influencing morphosyntactic analysis.
  • Interactive Determinism Hypothesis: The core idea that speech is organized such that necessary information for resolving ambiguities is available precisely when it's needed by a processing component.
  • Modular Architecture: A system composed of independent components (e.g., word recognition, morphosyntax, prosody) that communicate via dedicated, bidirectional channels.
  • Restricted Interaction: NSIs are limited to simple, binary (yes/no) responses triggered by explicit requests from a component facing an ambiguity. This prevents unnecessary computation.
  • LEXICAT Parser: A specific morphosyntactic parser designed to operate deterministically, relying on these restricted NSIs to resolve ambiguities rather than complex look-ahead mechanisms.

Datasets

Not discussed in the paper. The paper focuses on theoretical modeling and experimental validation through existing linguistic observations and specific experimental studies (e.g., by Paul Warren).

Key Insights and Ideas

  • Speech is inherently organized for efficient processing: The paper strongly suggests that the temporal structure of speech isn't arbitrary but is organized to facilitate understanding, making complex, general-purpose control unnecessary.
  • Domain-specific interaction is key: Instead of universal problem-solving frameworks, a theory tailored to language processing (like the proposed Interactive Determinism) is more effective for building efficient SUSs.
  • Prosody and syntax are tightly coupled for ambiguity resolution: The interaction between prosodic cues (like pauses or pitch changes) and syntactic structure is crucial for resolving grammatical ambiguities in real-time.
  • Simplicity can be powerful: A system with a simple control structure and restricted, deterministic interactions can achieve high efficiency and accuracy, challenging the notion that complex architectures are always necessary.
  • "Garden Path" sentences are a consequence of this interaction: Sentences that typically trip up parsers are explained by the ID hypothesis as cases where the expected prosodic cues for disambiguation are absent or misleading.

Key Results

The paper presents experimental evidence supporting the Interactive Determinism hypothesis, primarily through studies by Paul Warren:

  • Prosodic cues correlate with syntactic ambiguity: Warren's experiments showed that prosodic breaks (lengthened vowels, pitch changes) were significantly more pronounced at points of syntactic ambiguity in speech, especially when a non-default interpretation was required.
  • Semantic context influences prosodic marking: When semantic context made a particular syntactic interpretation implausible, the prosodic cues for ambiguity resolution were less marked, supporting the idea that multiple information sources interact.
  • LEXICAT's deterministic parsing is viable: The paper argues that a parser like LEXICAT, which relies on these restricted interactions, can produce correct analyses even for highly ambiguous sentences (like noun compounds with numerous interpretations) by leveraging prosodic information.

Limitations and Future Work

  • Idealized Model: The authors acknowledge that their model is highly idealized. For instance, the assumption that word recognition can return a series of unique, correct lexical items is currently beyond our capabilities.
  • Scope of Interactions: While the paper focuses on prosodic-morphosyntactic interaction, it notes that other interactions (e.g., semantic anomaly detection, discourse context for referent resolution) will be required. The paper questions whether these other channels can be as restricted as the prosodic break channel.
  • Oversimplified Control Structure: The paper suggests that the proposed control structure might be oversimplified, particularly regarding the automatic generation of requests on all channels for every ambiguity.
  • Future Work: The authors advocate for continued research into discovering constraints on language processing that allow for efficient and tractable implementation. They emphasize developing computational theories of language derived from processing tasks, similar to Marr's approach to vision. Further investigation into the nature and restriction of various interactive channels is also implied.

Summary updated

BibTeX
@inproceedings{briscoe1984control,
  author = {E.J. Briscoe and B.K. Boguraev},
  title = {{Control Structures and Theories of Interaction in Speech Understanding Systems}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1984},
}
Figure