Applying Natural Language Processing to Identify Critical Textual Phrases for Classification of Wrong-Way Driving Crashes

Advances in natural language processing models have opened a nearly unlimited opportunity to analyze human-written and generate humanlike texts. Researchers in various disciplines have utilized these models to perform different tasks. However, few studies have applied natural language processing algorithms in highway safety and transportation in general. In this research, we applied natural language processing models to crash narratives written by law enforcement officers to classify and extract relevant textual phrases to facilitate the detection of wrong-way driving crashes in a database. The proposed natural language processing model was implemented using a deep learning algorithm where various tokenization approaches or sequences of adjoining n words (n-grams) in sentences were assessed. The proposed algorithm showed promising results in classifying wrong-way driving and can identify the most important keywords that produce high discriminating accuracy. The prediction accuracy of the classifiers ranged between 84% and 92%, with unigrams and bigrams having relatively higher accuracy than trigrams and tetragrams. Moreover, unigrams and bigrams revealed generic keywords, which are usually less useful in wrong-way driving identification. However, trigrams and tetragrams provided clear explanations of the critical phrases for classifying wrong-way driving crashes compared with unigrams and bigrams. The findings are consistent with prior knowledge, and the fitted model can suitably be used for inference. Furthermore, the trigrams and tetragrams can be powerful tools for identifying critical phrases associated with wrong-way driving crashes for traffic safety analysis.