Robust reasoning: integrating rule-based and similarity-based reasoning

摘要

The paper attempts to account for common patterns in commonsense reasoning through integrating rule-based reasoning and similarity-based reasoning as embodied in connectionist models. Reasoning examples are analyzed and a diverse range of patterns is identified. A principled synthesis based on simple rules and similarities is performed, which unifies these patterns that were before difficult to be accounted for without specialized mechanisms individually. A two-level connectionist architecture with dual representations is proposed as a computational mechanism for carrying out the theory. It is shown in detail how the common patterns can be generated by this mechanism. Finally, it is argued that the brittleness problem of rule-based models can be remedied in a principled way, with the theory proposed here. This work demonstrates that combining rules and similarities can result in more robust reasoning models, and many seemingly disparate patterns of commonsense reasoning are actually different manifestations of the same underlying process and can be generated using the integrated architecture, which captures the underlying process to a large extent.