Using Semantic Fluency Models Improves Network Reconstruction Accuracy of Tacit Engineering Knowledge

Abstract

Human- or expert-generated records that describe the behavior of engineered systems over a period of time can be useful for statistical learning techniques like pattern detection or output prediction. However, such data often assumes familiarity of a reader with the relationships between entities within the system — that is, knowledge of the system’s structure. This required, but unrecorded “tacit” knowledge makes it difficult to reliably learn patterns of system behavior using statistical modeling techniques on these written records. Part of this difficulty stems from a lack of good models for how engineers generate written records of a system, given their expertise, since they often create such records under time pressure using shorthand notation or internal jargon. In this paper, we model the process of maintenance work order creation as a modified semantic fluency task, to build a probabilistic generative model that can uncover underlying relationships between entities referenced within a complex system. Compared to more traditional similarity-metric-based methods for structure recovery, we directly model a possible cognitive process by which technicians may record work-orders. Mathematically, we represent this as a censored local random walk over a latent network structure representing tacit engineering knowledge. This allows us to recover implied engineering knowledge about system structure by processing written records. Additionally, we show that our model leads to improved generative capabilities for synthesizing plausible data.