<p>Due to the inherent complexity of trading markets and the variability of influencing factors, forecasting agricultural futures prices remains a significant challenge. Numerous studies have demonstrated that incorporating sentiment indices extracted from news data into predictive models, as opposed to relying solely on historical price data, can effectively enhance forecasting performance. However, the discontinuity of news data poses a major challenge to the efficient and accurate extraction of sentiment indices. To address these issues, this paper proposes a text-based framework for predicting agricultural futures prices, integrating Multi-Scale Adaptive Graph Neural Network (MAGNN) with Dual-Clustering Sentiment Forecasting (DCSF). Specifically, the proposed framework first employs MAGNN to dynamically construct a sentiment network from news data, thereby capturing the intricate inter-commodity relationships and enabling robust estimation of missing sentiment values. Next, the DCSF is introduced to address data instability and to learn correlations among features. Within DCSF, a Temporal Clustering Module (TCM) is utilized to conduct distributed clustering on price and sentiment sequences, which helps to capture non-stationary temporal patterns and facilitates multi-level feature extraction. Subsequently, a Sentiment Decay Fusion Module (SDFM) is applied to model the dynamic and multi-scale effects of sentiment on prices, generating fused features for the final prediction. Empirical results using corn and soybean futures prices demonstrate that: (a) the proposed sentiment imputation approach retains more valuable information and significantly outperforms traditional methods, yielding average reductions of 5.55% in MAE, 4.66% in RMSE, and 4.51% in MAPE; (b) the DCSF module achieves higher predictive accuracy than other baseline models, highlighting the effectiveness and robustness of the proposed framework.</p>