The study introduces a novel framework for automated portfolio management, integrating Deep Reinforcement Learning (DRL) with sentiment analysis to optimize long-term investment strategies. A trading agent, trained via the Proximal Policy Optimization (PPO) algorithm in a Gymnasium environment, dynamically learns policies by interacting with a simulated financial market. Transformer-based feature extraction captures temporal patterns in historical stock data, while FinBERT-derived sentiment features from the FNSPID financial news dataset enhance decision-making with context-aware insights. The framework prioritizes long-term portfolio growth, incorporating transaction costs and risk-adjusted returns. Experiments on real-world datasets show that combining sentiment signals with DRL improves profitability and resilience compared to traditional strategies, advancing adaptive trading systems for volatile markets.