<p>To solve the difficult data extraction and analysis in Japanese language education and assessment in universities, an innovative Chameleon algorithm based on an improved natural nearest neighbor graph to generate sub-clusters is proposed to achieve Japanese language education data clustering. Meanwhile, parallel networks and attention mechanisms are proposed for Japanese language education data classification, which are jointly used for university Japanese language education and assessment. The results showed that on the three Japanese datasets of CSJ, JNAS, and Laborotvspeech, the Chameleon mixed model based on the improved natural neighbor graph to generate sub-clusters and parallel network and attention mechanism achieved accuracy of 89.67%, 83.53%, and 85.36%, respectively, which were 11.41%, 13.21%, and 16.77% higher than that of the MemNet model. The measured F1 value reached 86.58%, 82.76%, and 83.24%, respectively, which increased by 14.63%, 13.53%, and 17.7% compared to the RAM model. The study demonstrates that the improved model has shown excellent performance in the processing and analysis of university Japanese language education and assessment data. The integration of these two data mining algorithms provides reliable algorithmic support for university Japanese language education and assessment models.</p>