Automatically Learning Construction Injury Precursors from Text

1. Introduction

• Digitally available, unstructured data are growing, and learning from construction incidents is important for safety in an industry with elevated occupational fatality rates.The passage reports rates around three times the national average for main industries in the USA and Europe.Source: §1 · p. 2
• The study compares CNN, HAN, and TF-IDF + SVM, then identifies predictive text patterns for each safety outcome and verifies that valid precursors occur among them.The methods can also visualize and help interpret model predictions.Source: §1 · pp. 2–3
• Predictive skill is high for all models, while TF-IDF + SVM is on par with or outperforms deep learning most of the time.Source: §1 · p. 3

1.1. Background: deep learning in NLP

• Deep NLP models build document representations through successive layers, while dense word embeddings capture semantic and syntactic similarities that sparse BoW vectors do not.BoW also ignores word order and word similarity.Source: §1.1 · pp. 4–5

1.2. Background: NLP in Construction

• Construction NLP had mostly used BoW/TF-IDF, prior rule-based precursor extraction required substantial human input, and surveyed construction NLP had not tested deep learning.The earlier rule-based work extracted 81 attributes using a hand-written lexicon and rules.Source: §1.2 · p. 6

2. Models

• The section compares feedforward CNN and recurrent HAN architectures with a linear SVM using TF-IDF features.The CNN and HAN are structurally different deep learning models.Source: §2 · p. 7, §2.4 · p. 12

2.1. Convolutional Neural Network (CNN)

• Global 1-max pooling keeps the largest response from each CNN feature map, while filters of sizes 2, 3, and 4 capture complementary local features.The CNN uses 100 filters per branch for most outcomes and 300 for incident_type.Source: §2.1 · pp. 7–8

2.2. Hierarchical Attention Network (HAN)

• HAN encodes words within sentences and sentences within documents, using self-attention at both levels to weight their importance.Its sentence and document encoders are bidirectional GRUs, and the bidirectional RNN uses context from both directions.Source: §2.2 · pp. 8–10

2.3. Softmax layer

• A softmax layer converts each deep model’s document vector into class probabilities; separate models are trained for each outcome and architecture.The study constructs eight models: four outcomes multiplied by two architectures.Source: §2.3 · pp. 11–12

2.4. TF-IDF + SVM

• TF-IDF weights document terms by their within-document frequency and inverse corpus frequency, and a linear SVM classifies the resulting vectors using a separating hyperplane.For multiple categories, the SVM uses one-versus-rest classification; its C parameter controls tolerance for misclassification and regularization.Source: §2.4 · pp. 12–13

3.1. Text and outcomes

• More than 90,000 reports from work on five continents and spanning the early 2000s to 2018 form the dataset.At least one of four fields was available per report: title, description, details, or root_cause; four outcomes were available for some reports.Source: §3.1 · pp. 13–14

3.2. Text preprocessing

• Text preprocessing included sentence and word tokenization, lowercasing, removal of non-ASCII characters and HTML, and correction of word-segmentation errors.For example, “fellfrom” was corrected to “fell from.”Source: §3.2 · p. 14

3.3. Train/test splits and class imbalance

• Separate 90%/10% train/test splits were made for each outcome, with multi-label assignments applied afterward and inverse-frequency training weights used to address class imbalance.Injury_type and bodypart included 301 and 519 reports, respectively, associated with more than one level.Source: §3.3 · p. 14

3.3.1. Vocabulary construction

• The vocabulary was built from more than 218K reports outside all outcome test sets, retaining tokens appearing at least five times to produce about 24K tokens.Zero-padding and out-of-vocabulary words were assigned reserved indices.Source: §3.3.1 · p. 15

3.3.2. Pre-training word vectors

• The models used 64-dimensional skip-gram word2vec vectors trained on the study’s incident reports.The authors chose in-domain pretraining because incident reports are specific and the dataset was large enough for effective pretraining.Source: §3.3.2 · p. 15

4.1. Deep learning models

• Deep-learning models minimized categorical cross-entropy using stochastic gradient descent with cyclical learning-rate and momentum schedules.Learning-rate bounds were estimated separately for each outcome using validation data; dropout and early stopping were also used.Source: §4.1 · pp. 16–17

4.2. TF-IDF + SVM

• TF-IDF + SVM used bag-of-words dimensions up to trigrams, with the linear SVM’s C parameter tuned on a validation set.Terms occurring in more than 90% of reports were ignored.Source: §4.2 · pp. 17–18

4.3. Performance metrics

• Because class imbalance made accuracy inadequate, performance was assessed with class-level precision, recall, and F1-score derived from confusion matrices.Precision and recall use the predicted-category column and true-category row, respectively; F1 is their harmonic mean.Source: §4.3 · p. 18

4.4. Configuration

• The experiments used Python 3.6, TensorFlow-GPU 1.5.0, Keras 2.2.0, and scikit-learn for TF-IDF vectorization and SVM classification.The TF-IDF + SVM approach used the same preprocessing and train/test splits as the deep-learning models.Source: §4.4 · p. 18

Overall performance

• Performance was always much better than random, with CNN reaching 95.49 F1 for injury_type’s FOB class and HAN reaching 97.04 F1 for bodypart’s eye category.For the harder Rules incident_type and Pain injury_type categories, TF-IDF + SVM attained F1 scores of 61.75 and 66.17, respectively.Source: §5 · p. 19, Table 10 · p. 20, Table 11 · p. 20

Model comparisons

• HAN outperformed CNN almost everywhere except injury_type’s FOB and bodypart’s upper extremity, while taking three times longer per epoch on average.The authors suggest attention or the ability to capture word order beyond small local windows may explain HAN’s advantage.Source: §5 · p. 19, Table 9 · p. 19
• TF-IDF + SVM outperformed deep learning on all outcomes except bodypart, where it did not; the authors hypothesize that deep learning may need larger datasets to stand out.The authors do not conclude that deep learning should be abandoned, noting that pattern detection can suffice for these classification tasks while other applications require deeper input understanding.Source: §5 · pp. 19–20, Table 12 · p. 20

Training checks

• The CNN severity model diverged after a few epochs, so its results were not shown.The other models and outcomes trained successfully according to the authors’ checks.Source: §5 · p. 21, Table 13 · p. 21

Embedding analysis

• After training, HAN bodypart report embeddings were almost perfectly grouped by category and nearly linearly separable, unlike their random initial organization.The authors note that high-quality report embeddings can also support semantic report retrieval.Source: §5 · p. 21, Figure 5 · p. 22

6. Automatically learning incident precursors from text

• The methods identify, for each model, report text that is most predictive of safety outcomes on average, and valid injury precursors appear among those patterns.The visualizations also let users inspect report sections that matter to model predictions.Source: §6 · pp. 22–23

6.1. CNN: Predictive regions and word saliency

• CNN predictive regions are selected from high-norm receptive-field embeddings before pooling; corpus-level regions are collected by recording each report’s most predictive regions.Table 14 presents examples for incident_type and bodypart categories.Source: §6.1 · p. 23, Table 14 · p. 23
• CNN predictive-region examples include “trip hazard,” “blocked by,” and “slippery floor,” which describe observable conditions that can precede incidents.The table also includes outcome mentions such as “access” or “his eye,” which models use when reports explicitly state the category being predicted.Source: §6.1 · pp. 23–24, Table 14 · p. 23
• CNN word saliency uses derivatives of the model output with respect to input words, and highlights “without wearing” and “face shield” in an incident_type example.The visualization also shows a probabilistic forecast from the softmax output.Source: §6.1 · p. 24, Figure 6 · p. 24
• CNN predictive regions and word saliency find some common precursors, such as “trip hazard,” but mostly extract complementary patterns, so the authors advise using both.For Access, predictive regions include “blocked by” and “obstructed by,” while salient word sequences include “slippery floor” and “slipping hazard.”Source: §6.1 · p. 25, Table 16 · p. 25

6.2. HAN: word and sentence attention

• HAN combines word- and sentence-level attention to highlight important text; in the example, “without wearing proper face shield” and “piece of metal” are identified as valid precursors.Corpus-level word scores multiply each word’s attention coefficient by its sentence’s coefficient, then aggregate across sentences and reports.Source: §6.2 · pp. 25–26, Figure 7 · p. 26, Table 17 · p. 26

6.3. TF-IDF+SVM

• TF-IDF+SVM uses signed token weights to indicate each token’s contribution to a one-vs-rest classification decision.Large positive coefficients indicate tokens associated with the category; large negative coefficients indicate tokens associated with not belonging to it.Source: §6.3 · pp. 26–27
• TF-IDF+SVM predictive-token lists tend to start with mostly unigrams, then bigrams and rarely trigrams; valid precursors require going further down, since the top 10 almost entirely concern outcomes.Examples were manually selected by inspecting the top 50 elements of each list.Source: §6.3 · p. 27

7.1. Recommendations for improving precursor extraction

• Models tend to rely on narrative passages that discuss outcomes rather than incident circumstances and environmental conditions, reducing their ability to extract meaningful precursors.Suggested remedies include omitting fields rich in outcome information, filtering highly relied-on text and retraining, or removing sentences around outcome-indicating keywords.Source: §7.1 · pp. 27–28
• Reducing access to outcome information may yield more valid precursors but could lower prediction performance; this setting is closer to prediction using only information available before an accident.Filtering text may also remove legitimate precursors, so the precursor-to-solution ratio should be monitored.Source: §7.1 · p. 28
• The methods can use longer, more detailed documents, such as root-cause analyses or work-condition descriptions, which may improve the quality and relevance of extracted precursors.The authors note that such documents are commonly available for medium- and higher-severity reports.Source: §7.1 · p. 28

7.2. Recommendations related to prediction

• Prediction was not the study’s focus; field use would require workers to describe work conditions and access to an internet-connected computer, smartphone, or tablet.The authors note that these resources are not always available, though practical constraints may change as literacy, devices, and connectivity increase.Source: §7.2 · p. 29

8. Contributions to practice

• The methods can support work planning, job hazard analyses, pre-job meetings, and audits, as well as post-incident analysis and project statistics.The extracted information can complement expert knowledge when professionals focus safety efforts and budgets on incident types tied to specific activities.Source: §8 · p. 29
• Processing a full organizational database can provide a standardized view of safety issues, helping share less-obvious precursors when knowledge is scattered across projects, divisions, and geographic areas.Some extracted precursors are already known to safety professionals and serve as a sanity check; the approach is most useful for less frequent, less obvious ones.Source: §8 · p. 29
• The authors suggest applying the methods to continuously updated reports across the industry and to other text, including contracts, lessons learned, and quality records.They state that the observed quantitative and qualitative results across a diverse dataset suggest robustness to varied report styles, including machine-translated reports.Source: §8 · pp. 29–30

9. Conclusion

• The paper proposes automatically extracting valid accident precursors from raw construction injury reports.The methods identify textual patterns that are most predictive of each safety outcome on average.Source: §9 · p. 30
• For three supervised models, including two deep-learning models, the paper provides post-training methods to identify predictive text patterns and visualize model predictions.The authors report verifying that learned precursors are valid and suggest ways to improve results.Source: §9 · p. 30
• Predictive skill is high across the models, while TF-IDF+SVM is on par with or outperforms deep learning most of the time.Source: §9 · p. 30

• Antoine Tixier designed the study, curated the dataset, implemented models, ran experiments, interpreted and visualized results, and wrote the paper’s core.Source: §11 · p. 30
• Henrietta Baker wrote parts of the introduction, literature review, contributions to practice section, and conclusion, reviewed the paper, and led submission.Source: §11 · p. 30
• Matthew Hallowell reviewed the paper.Source: §11 · p. 30