This paper investigates statistical and neural approaches for automatic text summarization and keyword extraction in the Uzbek language. Traditional techniques like TF-IDF and Non-negative Matrix Factorization (NMF) are applied for identifying key terms and latent topics. Additionally, deep learning methods such as LSTM-based Seq2Seq models and Transformer architectures are used for abstractive summarization. Mathematical formulations describe term weighting, matrix decomposition, sequence modeling, and attention mechanisms. Experiments on Uzbek text datasets demonstrate the superiority of Transformer-based models in generating coherent and contextually appropriate summaries, validating the effectiveness of neural methods for Uzbek-language summarization.