Introduction
In the mid-1970s, speech recognition faced a significant hurdle: acoustic-phonetic uncertainty. Even with advanced signal processing, spoken language often results in a "word lattice"—a complex graph of multiple possible word sequences that sound similar. Bertram Bruce’s 1975 paper from Bolt Beranek and Newman (BBN) addresses this by arguing that speech understanding requires more than just syntax and semantics; it requires "pragmatics." Pragmatics involves understanding the speaker's goals and the context of the conversation to resolve inherent ambiguities.
The Pragmatic Framework
The core of the SPEECHLIS system is a pragmatic framework that models dialogue as a series of "intents" and "interaction modes." By formalizing "context," the system can make intelligent choices about which words are most likely in a given situation, effectively pruning the search space of the word lattice.
Architectural Integration
The system integrates a dedicated Pragmatics Control component that draws upon various knowledge sources to interpret user utterances.
As shown in the architecture, the Pragmatics Control module interacts with several key data structures:
- Instances of Intents: Specific goal-directed actions like
ADD,EDIT, orQUERY. - Intent Preconditions and Outcome Conditions: Logical states that must exist before an intent is valid and the resulting changes after it is executed.
- Context Representation: Models of the user's knowledge and the specific task domain (e.g., travel budget management).
- Dialogue History: A record of previous interactions to resolve references like "that trip."
Modeling Interaction Modes
A key innovation is the use of "Modes of Interaction"—higher-level state-transition graphs that organize intents. These modes set expectations for what the user might say next. For example, if the system is in a mode for adding a new database item, it expects specific types of information or clarifications.
In Figure 2, we see how the system manages the flow of adding an item. If a user's reply is ambiguous or contradictory, the system enters a QUESTION/CLARIFY or CONFLICT state. This structured approach allows the system to deprioritize interpretations of the user's speech that do not align with the current interaction state.
Key Insights and Results
The paper emphasizes that language is a functional tool used to achieve goals. By adopting an expectation-driven processing model, the SPEECHLIS system can use pragmatics as a final filter. If the acoustic processor is uncertain between two words, the pragmatics component can "vote" for the one that fits the logical flow of the conversation.
Qualitatively, the system demonstrated success in handling complex clarification cycles. It could successfully resolve ambiguities by checking pragmatic constraints—such as verifying if a mentioned "trip" already exists in the database—which syntax alone could not achieve.
Limitations and Future Outlook
While foundational, the approach had limitations. The rules for intents and modes were largely hand-crafted for the specific travel budget domain, leading to a need for more general models of human motivation. Additionally, the computational cost of checking preconditions against large databases was significant. Future work identified handling anaphora (pronouns like "it") and ellipsis (omitted words) as ongoing challenges requiring more robust dialogue history components.
Summary updated