Figure 2
The workflow diagram illustrates a step-by-step natural language processing pipeline that transforms an input accident report into contextualized key phrases. At the top left, a box labeled “Input Report” contains a paragraph describing an incident: an employee assisting with roofing activity during extreme heat, feeling unwell, and being hospitalized for heat exhaustion. An arrow labeled “Remove Stopwords” leads to a simplified “Cleaned Report” box on the right, which contains condensed text such as “assisting roofing activity performed complained feeling well hospitalized exhaustion”. From the cleaned report, a vertical arrow labeled “Tokenizing” leads below to a “Tokenized Words” box listing individual terms like “assisting”, “roofing”, “activity”, “performed”, “complained”, “feeling”, “well”, “hospitalized”, and “exhaustion”. A left pointing arrow from “Tokenized Words” leads to a large section titled “Embedded Words (with R o B E R T a)”, showing three groups of vector representations: “Unigram Vectors”, with ovals V 1 assisting, V 2 roofing, ellipsis, and V 9 exhaustion; “Bigram Vectors”, with ovals V 10 assisting roofing, V 11 roofing activity, ellipsis, and V 17 hospitalized exhaustion; and “Trigram Vectors”, with ovals V 18 assisting roofing activity, V 19 roofing activity performed, ellipsis, and V 24 well hospitalized exhaustion. Each vector is represented as an oval with labels. These vectors feed into a “Calculating Cosine Similarity” section at the bottom left, where arrows radiate from a central point labeled “V 0 (Cleaned Report)”. Each arrow corresponds to a vector (V 1, V 2, V 9, V 10, V 11, V 17, V 18, V 19, V 24), and the angle between vectors represents similarity using cosine theta. To the right, a “Ranking and Selecting Top Vectors” panel lists selected vectors such as V 24, V 10, V 11, and V 18 with a green check mark, while less relevant vectors (V 2, V 9, V 1) are crossed out in red. Finally, the selected vectors lead to a box labeled “Contextualized Key Phrases”, which outputs a summarized phrase such as “A 8: Tasks being conducted on roofs, an ellipsis”, representing the extracted contextual meaning from the original report.Phrase extraction process. Source: Authors’ own work