Abstract
An implemented system called ATT-Meta is sketched. The system can perform an important type of metaphor-based reasoning. It is based on an emphasis on source-domain reasoning and a de-emphasis of the idea of creating new mappings between source domain and target domain. The metaphor-based reasoning is fully integrated into a general framework for uncertain reasoning. The system thereby copes with various different types of uncertainty involved in metaphor understanding. The view of metaphor on which the system is founded is highly understander-relative, and does not require prior assumptions about what literal discourse is.
Introduction
Traditional computational approaches to metaphor have long prioritized the mechanism of mapping: the process of finding structural analogies between a source domain (e.g., physical space) and a target domain (e.g., the human mind). While mapping is essential for interpreting entirely novel metaphors, it often struggles with "view-transcending" metaphors. These are cases where a general metaphorical view is already established (such as MIND AS PHYSICAL SPACE), but the speaker introduces a specific, novel detail (e.g., "ideas in different store-rooms").
The ATT-Meta system addresses this by shifting the focus from the discovery of new mappings to deep reasoning within the source domain. By treating metaphorical statements as a "pretence" and applying common-sense logic within that sandbox, the system can derive complex meanings from familiar metaphors without needing to invent new rules for every unique phrasing.
The ATT-Meta Architecture
ATT-Meta is an implemented reasoning system written in Prolog, designed to handle the logical complexities of metaphorical language. Its architecture is built upon several novel concepts that distinguish it from purely analogy-based systems:
The Source-Based Pretence Cocoon
When the system encounters a metaphor, it creates a "pretence cocoon." This is a logical sandbox where the metaphorical statement is treated as literally true, regardless of its objective absurdity. Inside this cocoon, the system applies common-sense rules about the source domain. For instance, if a metaphor describes ideas as being in different rooms, the system uses its knowledge of physical space to infer that those objects are not spatially together.
Map-Extension Minimization
A core heuristic of ATT-Meta is Map-Extension Minimization. Rather than attempting to find a direct target-domain equivalent for every specific word (e.g., searching for a mental equivalent of a "store-room"), the system continues to reason within the source domain until it reaches a concept that already possesses a known mapping. This prevents the computational explosion that occurs when a system tries to create new mappings for every idiosyncratic detail.
Conversion Rules and Uncertainty
Once the system reaches a conclusion inside the pretence cocoon, "Conversion Rules" act as bridges to translate those findings into the target domain. For example, a conclusion that "ideas are physically separated" might be converted to "the ideas cannot interact mentally."
Crucially, this reasoning is integrated into an uncertainty framework. The system utilizes four levels of certainty—certain, presumed, suggested, and possible—allowing it to handle defaults and potential conflicts between metaphorical information and literal facts about the target domain.
Key Insights: Reasoning Over Mapping
The fundamental "aha moment" of the ATT-Meta project is that a massive dictionary of metaphorical links is unnecessary if the system possesses a robust model of the physical world. By utilizing high-level "bridge" rules, the system can leverage its existing knowledge of physics or biology to understand complex metaphors about abstract concepts like time or cognition.
Furthermore, the authors argue for understander-relativity. Whether a statement is perceived as a metaphor depends entirely on the listener's knowledge base. If a listener already has a lexicalized entry for a term (e.g., "kill" meaning "defeat"), it is processed literally for them, even if it remains metaphorical for another listener. This perspective allows ATT-Meta to treat metaphor as a valid, albeit uncertain, source of information rather than merely decorative language.
Case Study: The "Store-Room" Example
The efficacy of the system is demonstrated through its qualitative trace of the sentence: "The two ideas were in different store-rooms in John’s mind."
- Input: The system receives the hand-coded logical form of the sentence.
- Source Reasoning: Using common-sense physics, the system infers that objects in different rooms are not "spatially together."
- Conversion: It applies a rule mapping "physical conjoint operation" to "mental conjoint operation."
- Conclusion: The system concludes (with a certainty of presumed) that John cannot mentally operate on those two ideas at the same time, such as comparing them or using them together for a complex inference.
Limitations and Future Work
While ATT-Meta provides a powerful framework for metaphorical reasoning, it currently faces several limitations:
- Lack of NLP Front-End: The system processes hand-coded logical forms rather than raw English text.
- Static Reasoning: It lacks the temporal logic required to handle metaphors describing processes that evolve over time.
- Mapping Discovery: While excellent at using existing mappings, the system cannot yet "discover" or invent an entirely new metaphorical view (e.g., creating a new mapping for a domain it has never associated with the target before).
- Contextual Sensitivity: Future iterations require more robust methods for using surrounding discourse to determine which source-domain connotations are relevant to the current context.
Summary updated