Abstarct: Traditional clustering requires processing the entire dataset from scratch, and with the arrival of new data, the whole dataset must be reprocessed. This approach is inefficient and costly for continuous data, such as social networks. Incremental clustering processes new data without the need to reprocess the entire dataset, reducing resource consumption and enabling real-time processing. This method is suitable for applications with continuous or large-scale data. However, it has limitations such as sensitivity to the order of data arrival, reduced accuracy if inappropriate representatives are chosen, and the need for precise parameter tuning, making its successful implementation dependent on careful design and configuration. Given these challenges, the proposed method combines word embedding models and topic modeling to assess word importance. This combination not only considers the meaning of words in the text but also their position and significance within a segment of the data stream. The result of this process is a high-quality representative for the text, eliminating the need for past data. The proposed method has been evaluated on Persian and English datasets. Evaluation results show that, compared to existing methods, the proposed method achieves significant improvements in evaluation metrics. For the Persian dataset Tasnim, improvements in homogeneity, completeness, and NMI metrics were 17%, 21%, and 18%, respectively, while for the Fars News dataset, the improvements were 15%, 26%, and 12%, respectively. For the English dataset, in terms of homogeneity, the proposed method outperformed EStream in the SO-T, News-Trends, and Trends-T datasets by 1.3%, 2.4%, and 3.4%, respectively. Compared to MStream, the improvements were 61.3%, 6.4%, and 17.9%, respectively. In completeness, the proposed method showed improvements over EStream in the SO-T, News-Trends, and Trends-T datasets by 12.4%, 12.9%, and 35.7%, respectively, and compared to MStream, by 12.6%, 1.8%, and 38.8%, respectively. For NMI, the proposed method outperformed EStream in the SO-T, News-Trends, and Trends-T datasets by 6.4%, 5.9%, and 20.6%, respectively. Compared to MStream, the improvements were 60.6%, 3.8%, and 11.8%, respectively.