Deep learning and network analysis: Classifying and visualizing accident narratives in construction

Highlights

• We propose a deep learning-based approach for analyzing the text with accident reports.

• A CNN model is developed to classify accident narratives without using manual features.

• An LDA-based network analysis method is used to provide a visualization of factors contributing to accidents.

• The proposed approach can provide managers with much-needed information and knowledge to improve safety on-site.

Abstract

If headway is to be made to improve safety performance in construction, then there is a need to learn from past accidents. Accident reports provide a useful source of information to make sense as to why and how events occurred. Analyzing such reports, however, can be a lengthy and challenging process as there is a tendency for data to be presented in an unstructured or semi-structured free-text format. Thus, being able to classify and analyze the narrative that surrounds accidents and to better understand their causal nature is a challenge. Text classification using shallow machine learning with sophisticated manual lexical, syntactic, and semantic features engineering has been typically used to mine accident data. However, this approach requires highly skilled experts with domain knowledge to undertake this task. A limited number of studies have employed deep learning models to examine the text of safety reports in construction. In consideration of this limitation, word embedding is used to model the semantic narratives of accidents. Then, a Convolution Neural Network (CNN) model is trained to automatically extract text features and classify accident narratives without manual feature processing. The Latent Dirichlet Allocation (LDA) model is used to examine the interdependency that exists between causal variables to visualize the accident narratives. The proposed automated classification model and LDA-based network analysis method provide a useful approach to enable machine-assisted interpretation of texts-based accident narratives. Moreover, the proposed approach can provide managers with much-needed information and knowledge to improve safety on-site.

Introduction

Safety analysis in construction can broadly be classified into two different categories, namely predictive and retrospective methods [1]. Retrospective methods rely on past experiences and accident records (e.g., lessons learned and safety checklists) to avoid them in the future. Retrospective management requires a mechanism to prevent occurrence of similar accidents and promote workplace safety [2]. The crucial function of such a mechanism is the ability to analyze accident narratives collected over a period of time and derive knowledge about what previously went wrong [3]. More often than not managers are not provided with timely and fact-based information about accident causation as it is typically in an unstructured or semi-structure format [4]. Having to manually analyze accident text is a time-consuming and inefficient task [5].

Text mining has been identified as a potential technique that can be used to analyze and classify data contained within safety reports [6]. Existing text classification approaches that have been used to examine safety reports have tended to combine lexical, syntactic, and semantic features manually [[7], [8], [9]]. Such approaches are referred as shallow machine learning and include Support Vector Machine (SVM), Naive Bayes (NB) and K-Nearest Neighbor (KNN) with Natural Language Processing (NLP). This manual feature extraction process is limited by a person's domain knowledge and only can learn using the human-specified shallow feature. Contrastingly, deep learning algorithms (e.g., Convolution Neural Networks (CNN) and Recurrent Neural Networks (RNN)) can automatically identify features and use multiple single functions to learn complex tasks with a nonlinear combination of parameters from training data [10].

While there have been a number of studies that have used deep learning methods in construction [11], there is a paucity of research that has focused its use with text classification to examine accident reports. Managers often glimpse over reports as they can be rich in content and in some cases lengthy. As consequence valuable information that describes circumstances and conditions may be overlooked. Being able to analyze reports so that we can better understand the conditions and circumstances of an accident as well as their relationships others that have occurred can help managers make more informed-decisions about how manage safety. Thus, there is a need to automate the process of analyzing accident reports so that managers can learn and put in place processes to mitigate their future occurrence in a timely manner.

Against this contextual backdrop, we develop an Artificial Intelligence (AI) solution to automate the process of analyzing accident reports. We integrate therefore NLP with deep learning to automatically extract features and effectively classify accident narratives. The developed deep learning model incorporates topic mining and visual network to analyze and interpret texts-based accident narratives. The effectiveness of the deep learning model is verified using an experiment and compared with SVM, NB and KNN shallow machine learning methods. Each narrative category based on the CNN's classification is examined using the Latent Dirichlet Allocation (LDA) to garner further insights into the causes of the accident. The keywords derived from the LDA are analyzed using network analysis to identify and visualize the causal nature of an accident.

The aim of this paper is not to provide new insights into the causes of accidents per se, but demonstrate that deep learning can be used to extract unstructured safety data from accident text narratives automatically. As a result, managers will be better positioned to make timely and better-informed decisions about how to ensure the safety of their workforce on-site. The paper's contributions are twofold: (1) a novel deep learning approach is developed to analyze accident reports automatically; and (2) the narrative in form of text can be extracted and the causal variables of accidents visualized.