Evaluating the strength of text classification categories for supporting construction field inspection

•We proposed two measures for assessing the strength of candidate TC categories.•We proposed a methodology for applying the category evaluation measures.•We used construction field inspection as the domain for research illustration.•We showed how our measures are able to filter out weak TC categories•Our research offers a possible standard for classifying fall protection issues

Field inspection is a common approach to the prevention of on-site accidents in the construction industry, which aims to identify and correct violations before they result in accidents. While conducting a field inspection, quite often safety professionals need to consult applicable construction safety standards. By doing so, they can make informed judgments on the violations and reference applicable standards. Text classification (TC) can be used to classify safety standards based on the types and causes of violations. Safety professionals can therefore use violation types and causes as indices to quickly locate applicable standards. Defining TC categories (or labels) is the first important step in performing TC, because satisfactory results cannot be achieved without appropriate TC categories. Researchers often determine applicable TC categories based on the important topics within a knowledge domain. However, not all TC categories can yield satisfactory TC results because some of them are not associated with strong and specific keywords that can be identified by text classifiers. This paper proposes a methodology with two strength measures for evaluating the appropriateness of candidate TC categories. The measures were tested on two alternative sets of candidate categories that were drafted for supporting construction field inspections. The results showed that the measures could accurately predict the relative TC performance and the satisfaction levels (satisfactory or unsatisfactory) of TC categories. Beyond the construction domain, this research provides a generalized procedure for evaluating the strength of candidate TC categories.