Thurstone models have been widely applied in wisdom-ofthe-crowd applications to aggregate individual rankings due to their ability to represent individual knowledge and achieve high accuracy. However, they lack the ability to generalize even across highly similar items and cannot leverage external knowledge bases or learned machine representations. In this work, we extend Thurstone models for partial ranking data by introducing a latent construct that maps pretrained vector representations to latent truths. These representations are finetuned through a single neural network layer, enhancing the model’s ability to capture meaningful ranking structures. We evaluate our neural Thurstone model across objective ranking tasks, including animal speeds, material hardness, and the longitudinal positioning of U.S. states from west to east. Our results demonstrate that the extended model improves aggregation accuracy in sparse data settings and generalizes to novel items with moderate predictive accuracy, highlighting its potential to enhance collective intelligence in ranking-based inference.