In this paper, to address the challenges of semantic granularity and limited flexibility in effective portfolio investment caused by the inadequate perception of stock price fluctuations in portfolio return prediction models, this paper proposes a comprehensive system prediction model for stock returns by integrating the Sentiment Situation Evaluation Score Model (SESTM) and Graphical Lasso (GLASSO). Firstly, we introduce quantile regression to model stock price volatility, defining volatility width sequence and volatility mean sequence to identify vocabulary related to positive return fluctuations. Next, the SESTM model is employed to extract perception vocabulary related to stock price volatility from news announcements and generate news sentiment scores based on closely related themes and matching dictionaries associated with policies, valuations, and market sentiments. Finally, by combining the GLASSO method, we construct a network structure of interdependence among stock prices and develop individual stock portfolio strategies based on this network. Empirical experiments are conducted using stocks from the biotechnology vaccine sector during the epidemic period to compare network interdependence and sentiment situation evaluation models. The results show that firstly, investment strategies constructed based on perception vocabulary are more suitable for short-term predictions; secondly, incorporating information the reflected partial correlations in the interdependent network,the average daily logarithmic returns of the investment portfolio reached 1.6%, which is a 14.3% improvement by 14.3% compared to not considering partial correlations,and it is twice the average daily logarithmic returns of a random combination 0.7%.Moreover, the highest return increased from 3.117 for a random combination to 3.605, showing a significant improvement of 15.6%.These results indicate that the combination model of SESTM+GLASSO provides an efficient and superior system prediction model through a comprehensive approach that can analyze the network interdependence among stock prices and accurately predict stock returns for formulating corresponding investment strategies. It has positive implications for advancing statistical research in dynamic price perception and deepening generation-based cross-modal tasks within large language models.