This study addresses the challenges of unstructured expressions, semantic ambiguity, and noise interference in named entity recognition tasks on social texts. A recognition method is proposed that integrates the BiLSTM-CRF model with multi-source social features. The method uses a bidirectional long short-term memory network to extract contextual semantic information and applies a conditional random field for globally optimal sequence labeling. On this basis, social semantic features such as user interaction relations and topic labels are incorporated through feature concatenation. This enhances the model's ability to distinguish entity boundaries and categories. Experiments are conducted on the Twitter NER dataset. A systematic comparison is performed across different word embedding strategies, multi-source feature fusion settings, and input sequence lengths. The results show that the proposed method outperforms the baseline models in accuracy, precision, recall, and F1 score. In particular, it demonstrates stronger robustness and recognition ability when dealing with non-standard social texts. The model framework and experimental analysis presented in this paper offer effective technical support and methodological reference for named entity recognition in social text environments.