Organizing Tagged Knowledge: Similarity Measures and Semantic Fluency in Structure Mining
Thurston B. Sexton, Mark Fuge
Recovering a system's underlying structure from its historical records (also called structure mining) is essential to making valid inferences about that system's behavior. For example, making reliable predictions about system failures based on maintenance work-order data requires determining how concepts described within the work order are related. Obtaining such structural information is challenging, requiring system understanding, synthesis, and representation design. This is often either too difficult or too time-consuming to produce. Consequently, a common approach to quickly eliciting tacit structural knowledge from experts is to gather uncontrolled keywords as record labels-i.e., "tags." One can then map those tags to concepts within the structure and quantitatively infer relationships between them. Existing models of tag similarity tend to either depend on correlation strength (e.g. overall co-occurrence frequencies), or on conditional strength (e.g. tag sequence probabilities). A key difficulty in applying either model is understanding under what conditions one is better than the other for overall structure recovery. In this paper, we investigate the core assumptions and implications of these two classes of similarity measures on structure recovery tasks. Then, using lessons from this characterization, we borrow from recent psychology literature on semantic fluency tasks to construct a tag similarity measure that emulates how humans recall tags from memory. We show through empirical testing that this method combines strengths of both common modeling paradigms. We also demonstrate its potential as a pre-processor for structure mining tasks via a case study in semi-supervised learning on real excavator maintenance work-orders.