Societal Grounding Is Essential to Meaningful Language Use

David DeVault, Iris Oved, Matthew StoneAAAI 2006

Introduction: The Limits of Perceptual Grounding

The "Symbol Grounding Problem" has long haunted AI research, suggesting that internal symbols are meaningless unless linked to the physical world via sensors. However, this paper argues that perceptual grounding alone is insufficient for true linguistic meaning. Humans frequently use words meaningfully without perfect perceptual knowledge—for instance, asking about an "elm tree" without being able to distinguish it from a beech. True meaning requires societal grounding: a commitment to align internal symbols with the linguistic standards of a community.

Defining Societal Grounding

The authors propose a framework where meaning is a social contract, drawing on Kripke and Putnam’s theories of "meaning borrowing" and the "division of linguistic labor." They define societal grounding through two core conditions:

  1. Connection (D1): A symbol must be either perceptually grounded or borrowed from a community member who understands its reference.
  2. Commitment (D2): An agent must be "engaged," actively reconciling its private sensor data with the linguistic feedback provided by others.

Case Study: The COREF Interaction

To illustrate the necessity of this approach, the authors analyze the COREF agent, a dialogue system designed to identify objects in a shared visual space.

Visual objects used in the COREF agent interaction trace, including a red rectangle (a), green circle (b), blue diamond (c), and a gem (d).

In traditional setups, if a sensor error causes the system to misidentify a shape (e.g., calling a rectangle a "square"), it often fails to incorporate user corrections effectively. It treats its internal classifiers as absolute. A societally grounded agent, however, recognizes the user as a linguistic authority. When a disagreement occurs, the agent should pause its primary task to engage in a NegotiateFacts subtask, adjusting its internal definitions to match the community standard.

Architectural Implications

The paper proposes extending the Information-State approach to dialogue management. Instead of treating corrections as errors, they are viewed as collaborative opportunities. By adjusting internal thresholds (such as a "squareness" parameter) based on user feedback, the system demonstrates a "will" to align with the community. This commitment, rather than raw perceptual accuracy, is what justifies the attribution of genuine meaning to the agent's utterances.

Conclusion

Meaning is not a static property held within a system's code; it is an external target that must be constantly pursued through social interaction. By prioritizing societal alignment over internal certainty, AI researchers can develop systems that truly "mean" what they say, even when their sensors are limited. Future work remains in scaling these negotiation models to handle conflicting users and abstract concepts like "justice" or "inflation."

Summary updated

BibTeX
@inproceedings{devault2006societal,
  author = {David DeVault and Iris Oved and Matthew Stone},
  title = {{Societal Grounding Is Essential to Meaningful Language Use}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure