Location
https://www.kennesaw.edu/ccse/events/computing-showcase/sp24-cday-program.php
Event Website
https://sites.google.com/view/hkllm
Document Type
Event
Start Date
25-4-2024 4:00 PM
Description
Police officers spend many hours a week documenting their findings when reporting to a 911 call. There is so much detail in these reports that they remain an untapped resource for future data analytics by the police department. The reports are currently being analyzed by human experts and categorized into the following categories: “Substance Abuse”, “Mental Health”, “Domestic/Social”, “Nondomestic/Social”, and “Other”. To assist the experts and reduce the amount of time that is spent on reading and analyzing, we are proposing the use of large language models (LLMs) to tag police reports based on their content. Two models, Mistral-7B and TinyLlama, have been trained and fine-tuned to reduce the time needed to complete police report documentation. Both models output both the tag and the reason for the chosen tag, so one of the potential uses is for it to be used to train human analyzers in the future. For the finetuned Mistral-7B model we observed a 84% and 88% agreement with both human annotators and a 96% and 92% agreement with at least one human annotator on sample tagging and sample reasoning respectively.
Included in
UC-85 HELPR: Helping Extrapolate Labels for Police Reports using Large Language Models
https://www.kennesaw.edu/ccse/events/computing-showcase/sp24-cday-program.php
Police officers spend many hours a week documenting their findings when reporting to a 911 call. There is so much detail in these reports that they remain an untapped resource for future data analytics by the police department. The reports are currently being analyzed by human experts and categorized into the following categories: “Substance Abuse”, “Mental Health”, “Domestic/Social”, “Nondomestic/Social”, and “Other”. To assist the experts and reduce the amount of time that is spent on reading and analyzing, we are proposing the use of large language models (LLMs) to tag police reports based on their content. Two models, Mistral-7B and TinyLlama, have been trained and fine-tuned to reduce the time needed to complete police report documentation. Both models output both the tag and the reason for the chosen tag, so one of the potential uses is for it to be used to train human analyzers in the future. For the finetuned Mistral-7B model we observed a 84% and 88% agreement with both human annotators and a 96% and 92% agreement with at least one human annotator on sample tagging and sample reasoning respectively.
https://digitalcommons.kennesaw.edu/cday/Spring_2024/Undergraduate_Capstone/22