A new study published in Frontiers in Built Environment by John Ofori Datsa and William H. Collinge of the School of Civil Engineering at the University of Leeds set out to test whether automated sentiment analysis of Twitter discourse could reliably detect emotional patterns linked to occupational stress in the construction industry. The work is explicitly framed as a proof of concept rather than a snapshot of the sector today. The researchers mined a historical dataset spanning May 2008 to March 2016, a deliberate choice that allowed them to establish the technical feasibility of the approach without claiming contemporary generalizability. Their central finding is that social media platforms do contain measurable emotional signals about work experiences, and that these signals can be extracted and classified with reasonable accuracy using established natural language processing techniques.

The methodology began with careful data collection. Using Twitter’s X API V2 accessed through the Tweepy Python library, the team deployed six sequential search queries built around keywords and hashtags such as “construction,” “#constructionindustry,” “construction worker,” “construction site,” and “construction management.” Geographical filtering for UK-origin tweets was implemented through two complementary approaches: explicit GPS-derived coordinates falling within UK boundaries, and inferred location based on user profile fields matched against a predefined list of UK cities, postal code prefixes, and regional descriptors. Validation against tweets with explicit geocoding showed 94.2 percent agreement between the two methods, suggesting that profile-based inference is reasonably reliable. Because many users disable location services, relying exclusively on explicit geolocation would have excluded a substantial portion of UK discourse.

From an initial retrieval of 3,847 tweets, the researchers applied progressive filtering to arrive at a final analytical dataset of 2,322 tweets. Retweets were excluded to capture primary authorial intent, tweets shorter than 20 characters were removed as truncated content, non-English posts were dropped, and tweets discussing suicide or self-harm by third parties were excluded on ethical grounds to avoid amplifying harmful narratives. Preprocessing removed URLs, special characters, and stop words, generated tokens using the Natural Language Toolkit, and excluded tweets containing 50 percent or more non-dictionary tokens as likely noise or spam. The final corpus contained 42,344 tokens after stop-word removal, with 8,456 unique tokens and an average of 18.2 tokens per tweet. Notably, the researchers deliberately avoided lemmatization and stemming, preserving words in their original grammatical forms to retain semantic nuance for more accurate emotion detection.

Emotion labels were assigned using NRCLex, a lexicon-based tool derived from crowdsourced annotations by the National Research Council of Canada that maps 14,182 English words to ten emotion categories: anger, anticipation, disgust, fear, joy, negative, positive, sadness, surprise, and trust. Each tweet could carry multiple labels, with a mean of 2.3 emotions per tweet. To validate the automated labels, two independent human raters blind to classifier results assigned emotion categories to a random sample of 200 tweets. Inter-rater agreement between the humans was substantial at Cohen’s kappa of 0.82, and agreement between human ratings and NRCLex labels was 0.79, comfortably above the 0.70 threshold for acceptable agreement. This validation step is critical, as it establishes that the lexicon-based emotional classifications reasonably reflect human interpretation of the tweets despite relying on automated word matching rather than context-sensitive reading.

The machine learning phase was ambitious in scale. The team trained ten classifiers spanning diverse algorithmic families, including logistic regression, support vector classifier, random forest, decision tree, gradient boosting, XGBoost, CatBoost, AdaBoost, multilayer perceptron, and naive Bayes. Features were generated at three n-gram levels and weighted using term frequency-inverse document frequency vectorization. The dataset was partitioned using stratified random sampling into 80 percent training, 10 percent validation, and 10 percent test sets, with stratification preserving emotion class distribution across partitions. In total, 300 model instances were trained across all combinations of ten classifiers, ten emotion categories, and three feature types. Hyperparameters were deliberately left at published scikit-learn defaults rather than tuned, enabling reproducible baseline performance.

The results revealed a clear pattern: uni-gram features, meaning individual words, consistently outperformed bi-gram and tri-gram feature sets across all classifiers. This suggests that individual word choices carry sufficient discriminatory power for emotion detection in occupational discourse, without needing longer contextual phrases. For anticipation detection, logistic regression and CatBoost achieved performance comparable to the multilayer perceptron neural network, with AUC values of 0.93 versus 0.90. For fear detection, the multilayer perceptron achieved the superior AUC of 0.90. In the anger detection task, CatBoost, the multilayer perceptron, and the support vector classifier produced similar performance at approximately 0.76. Overall classification performance ranged from 0.76 to 0.93 AUC, with macro-averaged accuracies between 62 and 72 percent, substantially exceeding the naive baseline of 30.7 percent that would result from always predicting the most frequent emotion, trust.

Beyond classification, topic modelling using Latent Dirichlet Allocation identified ten substantive themes in construction-related discourse, with coherence scores peaking at ten topics. The three dominant themes, project scheduling and deadlines, employment opportunities and career advancement, and workplace safety and risk management, together accounted for 49.2 percent of the corpus. Employment security concerns were found to be embedded throughout technical discussions, suggesting that anxieties about job stability permeate even conversations ostensibly about equipment, scheduling, or management. By contrast, proactive discussions of occupational stress and organizational innovation received minimal attention, at less than 3.2 percent combined. This asymmetry is telling: workers discuss the operational pressures of their jobs far more readily than they discuss the psychological toll those pressures take.

The implications of this work extend beyond the construction sector. Traditional studies on worker wellbeing often rely on surveys, interviews, or organizational reports, which can be limited by response bias, restricted sample sizes, and the inherent delay between data collection and analysis. Social media analysis offers a complementary stream of naturally occurring text data that could enable more immediate, and potentially real-time, monitoring of workforce sentiment. By identifying patterns of negative sentiment, frustration, or stress within online discourse, organizations, policymakers, and industry stakeholders may be able to detect emerging workplace challenges earlier and develop proactive intervention strategies aimed at improving worker wellbeing and organizational productivity. The finding that default-parameter traditional classifiers match neural network approaches also has practical significance, as it suggests that meaningful emotion detection can be implemented without extensive model optimization or sophisticated architectures.

Important limitations temper these findings. Social media users may not fully represent the broader construction workforce, as participation in online discussions varies across demographic groups and job roles. Sentiment interpretation can be influenced by linguistic context, sarcasm, or ambiguous phrasing, which may affect classification accuracy. Most significantly, the historical timeframe of the dataset predates major industry transformations, including pandemic-era workforce disruptions and accelerated technology adoption. The researchers recommend contemporary replication using 2020 to 2026 data to assess whether emotional patterns have shifted with industry evolution and post-pandemic work adaptations, alongside expansion to multiple platforms such as LinkedIn and industry forums, longitudinal tracking of emotion signals, and integration with qualitative methods to contextualize the quantitative findings. As a methodological proof of concept, the study establishes that the emotional pulse of an industry can, in principle, be taken from the digital traces its workers leave behind.

Subject of Research: Social media sentiment analysis as a tool for monitoring mental wellbeing in the UK construction industry

Article Title: Investigating mental wellbeing in the construction industry using social media sentiment analysis: a UK case study

Article References: Datsa, J. O., & Collinge, W. H. (2026). Investigating mental wellbeing in the construction industry using social media sentiment analysis: a UK case study. Frontiers in Built Environment, 12, Article 1981301. https://doi.org/10.3389/fbuil.2026.1981301

Image Credits: AI Generated

DOI: 10.3389/fbuil.2026.1981301

Keywords: mental health, construction industry, sentiment analysis, machine learning, social media, occupational wellbeing, UK, Investigating, mental, wellbeing, construction, industry