MACHINE LEARNING OF MOTOR-VEHICLE ACCIDENT CATEGORIES FROM NARRATIVE DATA

Bayesian inferencing as a machine learning technique was evaluated for identifying pre-crash activity and crash type from accident narrative s describing 3,686 motor vehicle crashes. It was hypothesized that a B ayesian model could learn from a computer search for 63 keywords relat ed to accident categories. Learning was described in terms of the abil ity to accurately classify previously unclassifiable narratives not co ntaining the original keywords. When narratives contained keywords, th e results obtained using both the Bayesian model and keyword search co rresponded closely to expert ratings (P(detection) greater than or equ al to 0.9, and P(false positive)less than or equal to 0.05). For narra tives not containing keywords, when the threshold used by the Bayesian model was varied between p>0.5 and p>0.9, the overall probability of detecting a category assigned by the expert varied between 67% and 12% . False positives correspondingly varied between 32% and 3%. These lat ter results demonstrated that the Bayesian system learned from the res ults of the keyword searches.