Classification of the Requirement Sentences of the US DOT Standard Specification Using Deep Learning Algorithms

Abstract

This aim of this study is to classify requirement sentences from the specifications of US DOT using natural language processing (NLP) and a deep neural network. At the contract phase of the project, the requirements analysis of contract documents is a significant task to prevent claims or disputes caused by ambiguous or missing clauses, but it is highly human-intensive and difficult to identify requirements within a given short period. In this article, the requirement sentences identification model was proposed based on deep-learning algorithms. First, the critical terms that define what the requirement sentence is were identified, and then all sentences were labeled using the pre-defined critical terms. Second, three vectorizing methods were used, including two pre-trained methods—GloVe and Word2Vec—and a self-trained method to produce word embedding. Third, the automated classification of requirements sentences was experimented using three deep-learning models: the convolutional neural network (CNN), the long-short-term memory (LSTM), and the combination of CNN+LSTM. In the evaluation of nine total experiments, the results showed that the F1 scores of the CNN model were the highest at 92.9% and 92.4% for both the Word2Vec model and the Glove model. This study provided a way to achieve a high level of classification accuracy with simple deep-learning models and pre-trained embedding models.