This paper systematically investigates a passenger comfort quantification model for automated vehicle by integrating attention mechanism-based bidirectional long short-term memory (AM-Bi-LSTM) network, semi-supervised learning strategy, and passenger heterogeneity. Firstly, a comfort quantification dataset is gathered from a physical vehicle. More than thirty volunteers are recruited. Passenger classification based on volunteer heterogeneity is introduced for label reconstruction. Secondly, this paper innovatively develops a comfort degree quantification model based on the combination of AM-Bi-LSTM and semi-supervised learning. The semi-supervised learning strategy is developed to generate pseudo-labels and reduce dataset noise caused by individual variances of volunteers. The final model precision increases by 22.7% compared with the model trained on raw dataset. Finally, the deployment of model on physical automated vehicle is accomplished. The outputs of onboard real-time comfort quantification system are consistent with the driving intensity of automated vehicle. The results illustrate that the proposed model can accurately identify different comfort degrees of lane change scenarios, which can enhance the evaluation capacity of automated vehicle and optimize the decision-making and path planning modules of automated driving framework with more personalized driving styles.