Against the backdrop of the booming development of the Internet, the prevalence of Internet language has brought new challenges to intelligent translation systems. The problem is that Internet language has complex and diverse characteristics, such as non-standard grammar, new vocabulary, and rich cultural connotations, which makes it difficult for existing algorithms to translate accurately. This paper first analyzes the characteristics of Internet language, and then proposes a multi-step optimization algorithm, including data preprocessing to clean and standardize Internet language data, using an architecture-optimized deep learning-based neural machine translation model to better capture semantic features, and further refining the post-processing of the translation results. The experiments show that the importance of multimodal embedding technology has been verified in the evaluation of the model. After removing this module, the BLEU value of the model dropped by 1.8 points, which shows that multimodal embedding can effectively integrate non-text information such as emoticons, significantly improve the model's ability to understand network language sentences, and thus improve the translation quality.