Comparing Word Embeddings through Visualisation

Abstract

Asset management is a branch of facilities management that is responsible for the operation and maintenance of assets. The most common means of managing assets and their life-cycle is through requests and work orders. A request is used to report an occurrence that is detected either by a sensory device, a technician, or non-technical personnel; they are used to pointing out that something is wrong in a given asset, and needs appropriate attention. Depending on the problem, a request can give rise to a work order if the solution is not trivial. Work orders consist in technical reports that specify the asset that needs intervention and has the details about the work to be done or, in the case that the work is unknown from the start, the characteristics of the malfunctioning. Work orders contain a set of words, free text, that are not restricted from a fixed set of vocabulary, making it difficult to automatically analyse them. In this paper, we discuss the application of modern Natural Language Processing techniques to process the work order's description, while presenting a comparison between two Word Embedding models - Word2Vec and Fasttext- through semantic similarity tests between the encoded words, and a visualisation of the vector space through dimensionality reduction of the encoded vectors. The results show a better performance of the Fasttext approach, considering the semantics of the results.