Semantic insights into public trust: A textual data approach for automated vehicle interaction

Trust in automated vehicles (AVs) significantly impacts user experience and safety. This research employs social media data as a proxy for voice-to-text translation, proposing a novel approach to directly discern the public’s varied trust levels in AVs systems through semantic analysis of textual data. We provide a viable data source and model for indirect perception–based (non-contact) trust evaluations, traditionally done through voice recognition. The study accesses 55,187 pieces of public expression text from first-person demonstration videos on TikTok and Bilibili, capturing data akin to real driving contexts without experimental interference. By utilizing word embedding models, such as Word2Vec, to create digital representations of textual data, and capture semantic relationships between words, the research applies deep learning-based trust classification models to identify and classify instances of public overtrust and distrust. Further advancing the research, this study pioneers the application of the Bi-directional Encoded Representation for Transformers (BERT) in the field of trust recognition for AVs, significantly enhancing classification accuracy to an exemplary 92.23%. The approach stands out in its ability to unravel complex trust dynamics through text semantic analysis, effectively capturing the subtleties of real-life interactions between the public and AV systems. Such progress is essential for advancing non-contact trust assessments in AVs, demonstrating the significant benefits of integrating advanced deep learning, large language models, and social media texts to enhance the understanding and measurement of trust in interactions with AVs.