a b s t r a c tThe needs to ground construction safety-related decisions under uncertainty on knowledge extracted from objective, empirical data are pressing. Although construction research has considered machine learning (ML) for more than two decades, it had yet to be applied to safety concerns. We applied two state-of-the-art ML models, Random Forest (RF) and Stochastic Gradient Tree Boosting (SGTB), to a data set of carefully featured attributes and categorical safety outcomes, extracted from a large pool of textual construction injury reports via a highly accurate Natural Language Processing (NLP) tool developed by past research. The models can predict injury type, energy type, and body part with high skill (0.236 b RPSS b 0.436), outperforming the parametric models found in the literature. The high predictive skill reached suggests that injuries do not occur at random, and that therefore construction safety should be studied empirically and quantitatively rather than strictly being approached through the analysis of subjective data, expert opinion, and with a regulatory and managerial perspective. This opens the gate to a new research field, where construction safety is considered an empirically grounded quantitative science. Finally, the absence of predictive skill for the output variable injury severity suggests that unlike other safety outcomes, injury severity is mainly random, or that extra layers of predictive information should be used in making predictions, like the energy level in the environment. In the context of construction safety analysis, this study makes important strides in that the results provide reliable probabilistic forecasts of likely outcomes should an accident occur, and show great potential for integration with building information modeling and work packaging due to the binary and physical nature of the input variables. Such data-driven predictions had been absent from the field since its inception.

In the United States like in many other countries throughout the globe, construction workers are more likely to be injured on the job than workers in any other industry. This poor safety performance is responsible for huge human and financial losses and has motivated extensive research. Unfortunately, safety improvement in construction has decelerated in the last decade and traditional safety programs have reached saturation. Yet, major construction companies and federal agencies possess a wealth of empirical knowledge in the form of huge databases of digital construction injury reports. This knowledge could be used to better understand, predict, and prevent the occurrence of construction accidents. Unfortunately, due to the lack of a clear methodology and the high costs of manual large-scale content analysis, these valuable data have yet to be extracted and leveraged. Recently, researchers have proposed a framework allowing meaningful empirical data to be extracted from accident reports. However, the resource limitations inherent to manual content analysis still remain. The present study tested the proposition that manual content analysis of injury reports can be eliminated using natural language processing (NLP). This paper describes (1) the overall strategy and methodology used in developing the system, and specifically how key challenges with decoding unstructured reports were overcome; (2) how the system was built through an iterative process of coding and testing against manual content analysis results from a team of seven independent analysts; and (3) the implications and potential uses of the data extracted. The results indicate that the NLP system is capable of quickly and automatically scanning unstructured injury reports for 101 attributes and outcomes with over 95% accuracy. The main contribution of this research is to empower any organization to quickly obtain a large and highly reliable structured attribute and outcome data set from their databases of unstructured accident reports. Such structured data are a necessary prerequisite to the application of statistical modeling techniques allowing the extraction of new safety knowledge and finally the amelioration of safety management. 1. MOTIVATION Construction is constantly ranked as one of the most dangerous industries worldwide (Sacks et al. 2009). In the United States, despite the improvements that followed the Occupational Health and Safety Act of 1970, construction still accounts for 17% of all work-related deaths while only employing 7% of the national workforce (CPWR 2013). In fact, according to the Bureau of Labor Statistics (2014), approximately 700 workers die each year. Construction fatalities and injuries result in immense societal costs, totaling approximately $15 billion in

We introduce a novel graph-based framework for abstractive meeting speech summarization that is fully unsupervised and does not rely on any annotations. Our work combines the strengths of multiple recent approaches while addressing their weaknesses. Moreover, we leverage recent advances in word embeddings and graph degeneracy applied to NLP to take exterior semantic knowledge into account, and to design custom diversity and informativeness measures. Experiments on the AMI and ICSI corpus show that our system improves on the state-of-the-art. Code and data are publicly available 1 , and our system can be interactively tested 2 .

We operate a change of paradigm and hypothesize that keywords are more likely to be found among influential nodes of a graph-ofwords rather than among its nodes high on eigenvector-related centrality measures. To test this hypothesis, we introduce unsupervised techniques that capitalize on graph degeneracy. Our methods strongly and significantly outperform all baselines on two datasets (short and medium size documents), and reach best performance on the third one (long documents).

No abstract