Automatic Understanding of Construction Schedules: Part-of-Activity Tagging

Fouad Amer, Mani Golparvar-Fard
University of Illinois at Urbana Champaign, United States of America

DOI: 10.35490/EC3.2019.196
Abstract: Nowadays, construction planning practices, whether conducted by human planners or artificial intelligence (AI) systems, rely heavily on manually elaborated descriptions of construction means and methods. As part of envisioning a new planning system that automatically learns construction knowledge from previous projects’ schedules, this paper introduces Part-of-Activity (POA) Tagging: a construction-specific word-category disambiguation method for decoding the constructional functionalities encoded in activity names. These functionalities represent the roles each token, i.e. word, in an activity name plays in understanding that activity from a construction point of view. The model is built using Bidirectional Long Short Term Memory Recurrent Neural Networks (BI-LSTM RNN). After training on a manually annotated dataset of more than 7000 activities, the model achieved a token accuracy of ~92%. The significance of this method lies in its ability to allow an AI System to decipher construction schedules. This schedule understanding opens the door for further applications such as the automated elaboration of weekly work plans and alignment of master schedules to weekly work plans.
Keywords: Data Mining, Natural Language Processing, Machine Learning, Automation
Pages: 190 - 197
Paper:
http://ec-3.org/conf2019/contribution_196_final/