With the continuous development of third-generation RNA-seq, obtaining accurate splice-aware alignments for the long RNA-seq reads to reference genomes has become one of the major challenges in transcriptomic analysis. To mitigate the erroneous alignments arising from existing long-read RNA-seq aligners, we propose GLASS, a novel splice-aware alignments filtering approach based on third-generation transcriptome data. GLASS introduces a newly designed \'Read-AS Map\' model and integrates graph machine learning techniques for detecting and removing falsely spliced aligned reads from alignment files. Experimental results demonstrate that GLASS significantly reduces the error rate of spliced alignmnt and enhance the accuracy of subsequent transcriptome reconstruction. Additionally, GLASS demonstrates strong generalization ability in data from species such as Mus musculus and Arabidopsis thaliana, providing a new approach for the efficient processing of third-generation transcriptome data and offering more reliable data support for subsequent gene expression analysis and functional studies.