AI의 112 상황실 적용 및 기대효과 - 신고 접수와 경찰 언어모델 개발을 중심으로 -

Abstract

In recent years, Artificial Intelligence(hereafter AI) has been recognized as a mainstream means of business innovation regardless of the field. This study was conducted to explore the scope and expected effects of applying AI to the 112 calls, which is the first point of contact with citizens during police work, and the 112 situation control center, which receives calls and maintains and oversees the police response. Until now, there have been studies on the application of AI to police work in general and investigative work, but there has not yet been a study to apply AI to the 112 situation control center. We conducted a literature review on the application of AI in police work, as well as interviews with eight police experts who are involved in the procedural handling of 112 calls or who have studied AI in policing. The results showed that the greatest need for AI was in the area of 112 call receiving. Currently, there is variation in the quality of calls received by police officers due to individual competence and emotions. In order to homogenize this, the need for AI to analyze the caller's voice and STT to display it with high accuracy was identified. To do this, it is necessary to have an STT with an AI language model that goes beyond the level of a simple engine. We also realized that the shortcomings of the 112 system, such as the difficulty in obaining the desired spot in the location search, could be offset by replacing it with a learned AI-based search. In addition, we could expect that when an incompetent caller misses an essential question, the AI can help the caller understand the situation accurately by asking the missing question. To achieve this, it is not enough to simply train a general large language model. This is because police data can only work in a closed network to maintain security, and it is difficult for a model that is not trained with police domain knowledge to be effective in actual 112 work. Therefore, it is necessary to build an AI 112 language model by pre-training police data on an open source large language model that has been trained with data and then fine-tuning it with actual 112 data. However, it is necessary to continuously support the huge investment of resources as specialized personnel must run a large number of GPUs simultaneously to progress the learning process. In order to prioritize rapid business application, AI tasks that reduce the learning process by injecting lightweight resources are required, so we also described suggestions centered on prompt engineering. Through this research, we hope that this study will serve as a foundation for fundamentally and dramatically increasing the police's ability to respond to emergency incidents in the 112 call handling process, which receives many requests for police dispatch from the public.