This study examines how prediction performance and sentence information utilization differ according to methods for aggregating sentence-level embeddings to the document level in Korean essay assessment using Automated Essay Scoring (AES). Non-trainable aggregation methods (such as Mean, Max, and Min Pooling) were compared with attention-based aggregation methods that learn sentence-level importance (such as Gated MIL Attention and TransMIL Attention). The results show that simple pooling methods achieved a certain level of performance, but they did not demonstrate consistent advantages across evaluation metrics compared with attention-based approaches. In contrast, Gated MIL Attention and TransMIL Attention performed relatively better on several metrics, indicating that sentence-level information can be more effectively utilized when sentences are differentially weighted or modeled within inter-sentence relational contexts. Result of attention-based aggregation further revealed that sentences assigned higher importance weights tended to perform core argumentative functions, such as presenting claims, providing evidence, and emphasizing key points. These findings suggest that attention-based sentence aggregation is not only beneficial for score prediction but also informative and explainable for interpreting essay structure. Despite limitations related to exploratory analysis and the absence of rater identification data, this study empirically demonstrates that attention-based aggregation constitutes a structurally meaningful approach in AES.