In the context of the current multi-dimensional distribution of Internet information, the traditional paradigm of single-document information extraction no longer meets the demand for comprehensive event understanding. To address issues of information redundancy and fragmented viewpoints in multi-source text data, a multi-dimensional text summary generation model (FM-MDSG) based on a filtering mechanism is proposed. The proposed framework enables a structured fusion of cross-source information through a three-stage architecture. First, a hierarchical semantic representation is constructed by fine-tuning the RoBERTa model to capture contextual dependencies of the input text. Second, a two-layer filtering mechanism is designed to simultaneously perform saliency detection based on attention weights and domain-adaptive redundancy suppression, thus extracting semantic units with optimized information density. Finally, a knowledge-enhanced ERNIE decoder is employed, utilizing a dynamic gating strategy to enable the collaborative generation of multi-level semantic features. Experiments on the CSL dataset demonstrate that the proposed model achieves ROUGE-1/2/L F-scores of 55.37%, 47.28%, and 49.56%, respectively, ROUGE-L representing an improvement of 6.8 percentage points over classic baseline models. Further ablation studies confirm that the filtering mechanism contributes to a 9.22% performance gain in ROUGE-1 through noise suppression. The proposed model enables the systematic integration of heterogeneous source evidence and supports the reconstruction of complete event paradigms from multi-perspective observations in open-domain scenarios.