Automatic Review of Construction Specifications Using Natural Language Processing

Abstract

Since construction specifications are normally over 1000 pages and are complicated and often inconsistent, reviewing them is a labor-intensive and time-consuming activity. Thus, the aim of this study was to automate the review process by comparing construction specifications with standard specifications using natural language processing. Standard specifications for road construction projects were collected from 43 different states in the U.S. and used as experimental data. Doc2Vec, cosine similarity, and named entity recognition (NER) were used to recognize construction objects, standard values, and execution conditions, which can be used to find specification errors. As an early stage of the research, most of related sentences were found from standard specifications with high relevancy, and the average F1 score of NER was 0.256. The research findings will contribute to enhancing the efficiency of checking for specification errors by automatically detecting abnormalities and the absence of specific standards.