Automatically Learning Construction Injury Precursors from Text (1907.11769v3)

Abstract: In light of the increasing availability of digitally recorded safety reports in the construction industry, it is important to develop methods to exploit these data to improve our understanding of safety incidents and ability to learn from them. In this study, we compare several approaches to automatically learn injury precursors from raw construction accident reports. More precisely, we experiment with two state-of-the-art deep learning architectures for NLP, Convolutional Neural Networks (CNN) and Hierarchical Attention Networks (HAN), and with the established Term Frequency - Inverse Document Frequency representation (TF-IDF) + Support Vector Machine (SVM) approach. For each model, we provide a method to identify (after training) the textual patterns that are, on average, the most predictive of each safety outcome. We show that among those pieces of text, valid injury precursors can be found. The proposed methods can also be used by the user to visualize and understand the models' predictions.

• The paper introduces a pioneering methodology that leverages CNN and HAN to capture both high-level features and hierarchical context from accident reports.
• The paper compares deep learning models with a TF-IDF+SVM baseline, revealing that simpler token-based techniques can sometimes match advanced approaches in detecting key safety phrases.
• The paper demonstrates that automated precursor extraction enables targeted safety interventions by identifying predictive risk patterns from unstructured text.

Automatically Learning Construction Injury Precursors from Text

The paper "Automatically Learning Construction Injury Precursors from Text," authored by Henrietta Baker, Matthew R. Hallowell, and Antoine J.-P. Tixier, introduces a pioneering methodology for extracting injury precursors from construction accident reports using machine learning and NLP. The paper is situated within the context of increasing volumes of digital, unstructured construction data, which include incident reports that are critical for enhancing workplace safety strategies.

The authors deploy three computational models: Convolutional Neural Networks (CNNs), Hierarchical Attention Networks (HANs), and a traditional machine learning approach using TF-IDF combined with Support Vector Machines (SVMs). By training these models on diverse sets of injury reports, the authors aim to identify textual patterns that can predict safety outcomes and consequently, extract valid precursors that can inform preventive measures in the construction industry.

Key Aspects and Results

1. Model Selection and Methodology:
   • The paper experiments with CNN and HAN architectures to leverage their ability to capture different levels of semantic information. The CNN is utilized for its proficiency in identifying high-level features through local receptive fields, while HAN, with its hierarchical structure, integrates word and sentence-level attention to capture global contextual information.
   • The TF-IDF + SVM baseline serves as a comparison model and focuses on simpler token-based features, providing a stark contrast to what deep learning models can achieve with semantic understanding.
2. Data Handling and Preprocessing:
   • As a preliminary step, the raw reports are preprocessed and transformed into formats suitable for training the models. A tokenizer is applied, and non-essential characters are filtered out. Documents are transformed into structured forms, such as matrices for CNNs and hierarchical sequences for HANs.
   • A critical aspect is ensuring alignment between training and testing splits to maintain operational integrity and preventing data leakage.
3. Performance Evaluation:
   • Each model's performance is assessed using confusion matrices and derived metrics such as precision, recall, and F1-score, providing nuanced insights into their predictive capabilities.
   • The authors report that while both CNN and HAN achieve robust predictive performance, the TF-IDF + SVM approach often matches or exceeds deep learning models, particularly in cases where detecting precise key phrases correlates strongly with incident outcomes.
4. Precursor Extraction:
   • For CNN, predictive regions within the input text are identified by evaluating intermediate outputs before pooling—a mechanism that highlights the most impactful phrases.
   • HAN's attention mechanism inherently elucidates which words and sentences most heavily influence predictions, allowing for straightforward extraction of precursor information.
   • TF-IDF + SVM elucidates contributions of specific tokens to classification decisions, revealing which words or phrases are most indicative of safety outcomes.

Implications and Future Directions

The methodologies proposed for automatic precursor extraction present significant implications for constructing more efficient and preemptive safety measures. By systematically identifying conditions leading to incidents, this research offers a practical, data-driven tool for safety professionals, facilitating more targeted safety interventions and potentially reducing incidents through strategic focuses on previously identified hazards.

Moreover, the research invites further examination of the complex trade-offs between model transparency and predictive accuracy. As the authors suggest, removing explicit outcome-oriented information from input texts might further refine precursor identification. Future research could explore larger and more diverse datasets, integrate multi-language capabilities, and apply these models to other domains within construction safety or related sectors.

The paper contributes valuable insights into harmonizing advanced machine learning techniques with practical objectives within construction management, pushing forward the capacity to utilize massive, unstructured data for tangible safety improvements.