Current approaches to Chinese Named Entity Recognition (CNER) often struggle with the accurate identification of entities due to the inherent complexity of the Chinese language, including its script and phonetics. This study introduces three innovative solutions to address specific challenges in CNER. Firstly, the prevalent issue of irrelevant vocabulary inclusion is tackled by implementing a semantic similarity-based filtering method. This method employs cosine similarity calculations between matched words and the entire sentence, selecting only those words that exceed a predefined similarity threshold for integration into the character sequence. Secondly, the challenge of distinguishing relevant from irrelevant vocabulary is addressed through contrastive learning. By minimizing the representational distance between characters and matching words, and maximizing the distance from non-matching segmented words, the model’s discriminatory power is significantly enhanced. Lastly, the complexity of Chinese characters, both in terms of their structure and phonetics, is addressed by incorporating multimodal features. Phonetic features are extracted by converting characters into pinyin, followed by word2vec embeddings, while morphological features are obtained through character images processed by a Vision Transformer (VIT) model. These multimodal features provide a comprehensive understanding of Chinese characters, thereby improving the accuracy and robustness of the CNER model. Collectively, these solutions present a significant advancement in addressing the unique challenges of CNER, offering a more nuanced and effective approach to entity recognition in Chinese texts.