The powerful computing capabilities of supercomputers play an important role in today’s scientific computing. A large number of high-performance computing jobs are submitted and executed concurrently in the system. Job failure can cause a waste of system resources and impact the efficiency of the system and user jobs. Job failure prediction can support fault-tolerant technology to alleviate this phenomenon in supercomputers. At present, the related work mainly predicts job failure by collecting the real-time performance attributes of jobs, but it is difficult to apply in the real environment because of the high cost of collecting job attributes. In addition to analyzing the time and resource attributes in the job logs, this study also explores the semantic information of jobs. We mine job application semantic information from job names and job paths, where job path is collected by additional monitoring of the job submitting process. A prediction method based on job application semantic enhancement is proposed, and the prediction results of the non-ensemble learning algorithm and the ensemble learning algorithm are compared under each evaluation indicator. This prediction method requires more miniature feature collection and computation overhead and is easy to apply. The experimental results showed that the prediction effect was promisingly improved with job application semantic enhancement, and the final evaluation indicator S_score was improved by 5%–6%, of which was 88.16% accuracy with 95.23% specificity and 88.24% sensitivity.