Quality Estimation (QE) for machine translation aims to evaluate translation quality without reference translations. The WMT 2022 competition introduced a novel QE task focused on predicting MQM labels, i.e. human annotations for the locations, categories, and severity of errors in machine translations under the Multidimensional Quality Metrics framework. However, such annotation is costly and results in data scarcity. Inspired by named entity recognition tasks, we leverage a sequence labeling model to fully exploit the temporal dependencies and contextual information in MQM data. The model is trained to automatically simulate human annotations by identifying error locations and severity in machine translations. Experimental results demonstrate that the pseudo MQM data generated by this sequence labeling model can significantly enhance the performance of quality estimation models.