Flexible semantic matching of rich knowledge structures

Many AI tasks require matching two knowledge representations to determine whether (and how) they match. For example, rule based classification requires matching rule antecedents with working memory; knowledge based information retrieval requires matching queries with an ontology; and discourse understanding requires matching the speaker's utterances with background knowledge to build a coherent model of what was said. Solving this matching problem is difficult because similar information can be expressed in very different ways. Existing solutions use either syntactic measures, such as maximum common subgraph, or shallow semantics, such as taxonomic knowledge. These solutions, however, overlook many mismatches between representations - adversely affecting performance. Our goal is to improve the performance of existing semantic matchers. Our solution is to augment these matchers with transformation rules that achieve broad coverage. To achieve this coverage, we built a library of transformations based on:

1. a domain independent upper ontology and 2) a recurring pattern called "Transfers Thru". We systematically enumerated all valid instantiations of this pattern, and the result was a comprehensive library of about 200 transformations. We evaluated our matcher by applying it to several different tasks including course-of-action critiquing, information retrieval, discourse understanding, and sense disambiguation and semantic role labeling. In each case, we found that our matcher significantly improved matching and outperformed the state-of-the-art systems for each specific task.