The quality of machine translation has long been a core issue in machine translation, and categorization of translation error types for quality assessment has attracted much attention. Under the guidance of Multidimensional Quality Metrics (MQM), this paper categorizes translation errors of specialized texts by DeepL, a neural network-based machine translation tool. The findings indicate that, despite optimal equivalence of machine translation between the source and target texts, typical errors are related to term inappropriateness, information ambiguity, text disfluency and nonstandard grammar. The research findings can provide professional translation training with practical pre-editing or post-editing suggestions.