In recent years, translation software has faced linguistic complexity and contextual variability, so the article proposed an improved algorithm based on Bayesian networks. During the experimental phase, the performance of algorithm models such as Bayesian networks in Word Sense Disambiguation (WSD) tasks was evaluated. In the accuracy comparison experiment, the Bayesian network showed an accuracy of 85%. In the recall measurement experiment, the recall rate of the Bayesian network was 81.63%. In performance evaluations with different contextual complexities, the accuracy of Bayesian networks remains at 88% under high complexity. In the final robustness testing experiment, the accuracy of the Bayesian network can also reach 65% at a high noise level of 0.5. From the data conclusion, it can be seen that in the field of natural language processing (NLP), Bayesian networks have higher performance compared to some traditional algorithms, proving their applicability as optimal models.