Detecting Speech Lock Expressions in Caregiving Conversations Using NLP

Abuse in caregiving environments has emerged as a critical social issue in Japan. While abuse manifests in both physical and psychological forms, the latter presents unique detection challenges from external perspectives. Among psychological abuse variants, "speech lock"—verbal expressions that constrain care recipients' actions—is particularly problematic, as caregivers may unconsciously employ such language, complicating identification efforts. This study proposes an automated system for detecting speech lock expressions through speech-to-text conversion of caregiver-recipient conversations. In the preparatory phase, conversation samples were collected from caregiving-related academic literature and professional resources, encompassing both standard interactions and those exhibiting speech lock characteristics. Identification criteria were established based on prior research on caregiver awareness and professional practice guidelines. To enhance the dataset, supplementary sentences were generated via ChatGPT-4o under controlled prompt conditions. All textual data underwent morphological analysis before being vectorized using a Doc2Vec model, followed by cosine similarity score computation. Classification was performed by comparing these similarity values against a predefined threshold. Experimental results demonstrated the method's effectiveness in distinguishing between ordinary and speech lock conversations, suggesting potential applications for early detection and prevention of psychological abuse in caregiving environments.