There has been an increasing demand for data-driven and machine learning-based bridge deterioration prediction approaches for supporting enhanced bridge maintenance decision making. Bridge inspection reports, which contain a wealth of information about bridge conditions, open opportunities for data analytics to better understand and predict bridge deterioration. However, learning from the reports is challenging, because they usually contain multiple - even ambiguous, uncertain, and conflicting - information about the same bridge element, its deficiencies, and its deficiency measurements. Learning from such data negatively affects the generalizability and the separability of machine learning models, which compromises the performance of data-driven prediction. To address this challenge, this paper proposes a hybrid information fusion method. The method includes two main components: named entity normalization for fusing concepts in ambiguous surface forms into a canonical form, and data fusion for fusing numerical deficiency measurements containing uncertainties and conflicts into a unified and consistent representation. This paper focuses on analyzing the data fusion requirements, and presenting the proposed data fusion method and its evaluation results. The results indicate that the proposed method can adequately address the fusion requirements.