Introduction
Traditional natural language systems were primarily designed as passive information providers, responding directly to queries with neutral facts. However, human conversation is rarely neutral; speakers often have specific agendas and seek to influence the listener's opinion of the subject matter. In this paper, Anthony Jameson introduces IMP, a prototype dialog system designed for "evaluation-oriented dialog."
IMP moves beyond simple question-answering by strategically managing the listener’s impressions. It addresses three core human communicative tendencies: volunteering unsolicited information, maintaining a strategic bias, and performing anticipatory monitoring—recognizing that what is not said can be as informative as what is said.
The IMP Framework: Evaluation Forms and Impression Monitoring
The foundation of the IMP system is a structured knowledge representation called an evaluation form. This is essentially a multi-attribute utility model where each attribute of an object (e.g., a room for rent) is associated with a set of possibilities. Each possibility includes a numerical value (representing its desirability) and a frequency (representing its probability in the given domain).
To simulate social dynamics, IMP maintains a model of the listener’s current "expected value" of the object. Before generating a response, the system calculates the "anticipated impact" of a potential comment. It evaluates how a specific statement will shift the listener’s numerical rating of the object, allowing the system to select information that aligns with its internal goals.
Strategic Information Selection
IMP is assigned a "projected bias"—which can be positive, negative, or objective. The system then searches for comments that nudge the listener's impression in the desired direction while remaining strictly truthful. This process involves a sophisticated loop of anticipating how a listener will pragmatically interpret a statement.
The Informative Power of Silence
A key insight of the IMP system is the role of silence. If a speaker is known to have a positive bias but fails to mention a specific positive feature, a savvy listener may infer that the feature is absent. IMP models this by realizing that staying silent about a negative attribute might actually be more damaging to the desired impression than admitting it, especially if the silence makes the system's bias too obvious and undermines its credibility.
Connectives as Signposts
Beyond selecting facts, IMP strategically chooses sentential adverbs and connectives (e.g., "In fact," "By the way," "Unfortunately"). These serve as signals to the listener about how the new information relates to previous expectations. For instance, "but" is used when the upcoming information shifts the evaluation in the opposite direction of the preceding statement.
Qualitative Results and System Performance
The system was implemented in LISP/FUZZY and tested using a hand-coded knowledge base representing a "room for rent" scenario. In sample dialogs, IMP demonstrated sophisticated behavior:
- Volunteering Information: When asked about facilities, IMP might add, "And by the way, the room is very large," if it calculates that this unsolicited fact provides a necessary positive boost to the overall evaluation.
- Managing Credibility: If asked if a room is on a quiet street, a positively biased IMP might answer "No" and immediately add "In fact, unfortunately, there's a lot of noise from the street." It does this to prevent the listener from assuming the noise is even worse than it actually is, thereby protecting the system's "neutral image."
Limitations and Ethical Considerations
While pioneering, the IMP system has several limitations:
- Simplified Psychology: It assumes humans evaluate objects using a purely additive linear model, which may oversimplify complex human judgment.
- Scalability: The reliance on hand-coded evaluation forms makes it difficult to scale to broader or more complex domains.
- Ethics of Manipulation: The author acknowledges that a system designed to monitor and manipulate impressions raises ethical concerns regarding deception in AI, suggesting this as a critical area for future investigation.
Summary updated