In this article, we present a novel approach for representing text as fixed-length semantic vectors. Building upon previous research, we propose a modified method that combines key phrase extraction techniques with clustering techniques to generate concept-based, low-dimensional directed representations of documents. To evaluate the effectiveness of the resulting vectors, we conducted experiments using them in an aggregation task against a set of robust baselines following a similar approach. Our experimental findings demonstrate that the proposed modification significantly enhances the quality of the vector representation of documents. Specifically, the assembly accuracy improved by at least 3% on the V1 scale, validating the efficacy of our approach. This research contributes to the field of natural language processing by offering an efficient method for representing text as semantic vectors, thereby enabling more accurate and comprehensive analysis of textual data.