Cloze: A Building Metadata Model Generation System based on Information Extraction

Recently, we have seen a flourish of data-driven building applications. It has also been noted that the main effort in application development today is on data preprocessing. More specifically, buildings have entities, e.g., a chiller. To extract their data values or to control the entities, applications need to refer to the metadata of a building, i.e., the data describe the entities in a building. Data preprocessing organizes the raw metadata of a building into a form that can be easily recognized by applications. Different buildings have different metadata conventions. Data preprocessing today is largely an ad hoc process and is manually done in a building-by-building manner. How to automate data preprocessing is challenging. In this paper, we first formulate a problem on converting building raw metadata with ad hoc conventions into a building metadata model that follows a standard convention, e.g., the Brick metadata schema. Depending on application scenarios, we present three variants of the problem. This problem is intrinsically a text analysis problem. We thus propose to leverage the information extraction paradigm, a type of document processing to extract structured information from unstructured documents/texts. We analyze real-world building metadata and present a set of challenges on corpus denoise, coreference resolution, disambiguity, etc. We develop a system, Cloze with corresponding solutions. We evaluate Cloze with six real-world buildings. Our results show that Cloze can learn and automatically recognize raw metadata and their relations with an accuracy of 96.3%.