Flood susceptibility assessment for identifying flood-prone areas plays a significant role in flood hazard mitigation. Machine learning is an optional assessment method because of its high objectivity and computational efficiency, but how to get enough and accurate information of historical flood locations to train the machine learning models has been a key problem. In recent years, news media data from both news websites and social media accounts has emerged as a promising source for natural science studies. However, the application of news media data in urban flood susceptibility assessment is still inadequate. This study proposed an approach to fill this gap. Firstly, flood locations were extracted from news media data based on a named entity recognition (NER) model. Then, a frequency or distance-based data quality control method was employed to improve the representativeness of the extracted flooded locations. Finally, flood conditioning factors with information of historical flood locations were input into a Support Vector Machine (SVM) model for flood susceptibility assessment. We took the central city of Dalian, China as a case study. The T-test results show that there was no significant difference between the distributions of most flood conditioning factors at the flood locations from the news media data and the official planning report. In the obtained flood susceptibility map, the high flood susceptibility areas got a recall of 90% compared with the high flood hazard areas in the planning report. Performing data quality control in the frequency-based method can improve the precision of the flood susceptibility map by up to 5%, while the distance-based method is ineffective. This study provides an example and offers the value of applying new data sources and modern deep learning techniques for urban flood management.