Abstract: This paper presents a method for maintaining logical consistency within hierarchical knowledge representations. By organizing concepts into tree partitions and applying structured propagation rules, the proposed Truth Value Tree (TVT) mechanism identifies and resolves contradictions between memory-based facts and plausible inferences.
Introduction
Early Knowledge Representation (KR) systems often struggled with conflicting information. An AI might "know" a specific fact from its memory while simultaneously "inferring" a contradictory property based on categorical defaults. John L. Goodson's 1980 paper addresses this by proposing a "local and focused" approach to consistency, leveraging the inherent structure of hierarchical concept trees.
The Truth Value Tree (TVT) Framework
The core innovation is the Truth Value Tree (TVT). In this model, knowledge is organized into tree partitions where categories at the same level are mutually exclusive. Each node in the tree is assigned a truth value: T (True), F (False), or Q (Question/Indeterminate).
To manage these values, every node maintains a "ledger" consisting of a true-list and a false-list. These lists track both support (the source of the inference) and requirements (logical constraints that forced the value). If a node ends up with entries in both its True and False lists, the system has detected an inconsistency.
Tree-Consistency Propagation Rules
Consistency is maintained through four primary propagation rules that dictate how truth values flow through the hierarchy:
- Upward Propagation: If a leaf node is determined to be True, all its ancestor nodes must also be True.
- Downward/Sideways Propagation: If a node is True, all its siblings and their descendants are automatically False, enforcing the mutual exclusivity of the partition.
- Exclusionary Propagation: If a node is False, all its descendants must also be False.
- Exhaustive Propagation: This rule relates the truth of a parent node to the collective truth values of its children. For instance, if all siblings of a node are False, the parent's truth state is affected accordingly to ensure the partition remains exhaustive.
Conflict Resolution and Consistent Subtrees
When contradictions arise—such as memory stating an individual is 40 years old while plausible reasoning suggests they are "Young"—the system employs a "consistent subtree" strategy. Rather than failing, it attempts to find the largest possible subset of beliefs rooted at the top of the hierarchy that remains logically sound. This involves a heuristic procedure to select nodes with "maximum support," effectively guessing the most likely state of the world when data sources clash.
Conclusion and Limitations
The TVT mechanism was implemented within the AIMDS (Artificial Intelligence Meta-Description System). While effective for strict tree partitions, the approach has limitations: it does not natively support "tangled hierarchies" (directed acyclic graphs) where categories overlap, and it focuses on static consistency rather than the temporal dynamics of belief revision. Nevertheless, it provides a robust foundation for localized truth maintenance in structured domains.
Summary updated