Identification of key sentences in a text

Human evaluation of students' summative assessments requires significant time and cognitive effort. Existing automated evaluation models struggle to provide accurate evaluations for various reasons. The complexity increases with lengthy student responses, which may contain both relevant and irrelevant information. Therefore, there is a need for an intermediate mechanism to assist human evaluation. To address this gap, this study presents, as a proof of concept, a semantic-aware model called the Key Sentence Identifier (KSI), which extracts sentences that are relevant to the topic and coherent with the meaning conveyed in the student's response. Unlike traditional approaches that rely primarily on keyword matching, KSI employs a dual-embedding framework that integrates transformer-based contextual embeddings (BERT) with semantic similarity-based sentence representations (SBERT). Furthermore, a dedicated dataset has been curated for this task to enable effective training and evaluation of the model. Ablation analysis indicates that the SFT BERT + SFT SBERT configuration achieves the best performance, improving the F1 score from 50.32% (BERT + SBERT) to 86.01%. Furthermore, comparisons with LoRA-based fine-tuning and standard baseline methods show that the proposed KSI model consistently outperforms alternative approaches in identifying contextually relevant sentences. The novelty of the proposed KSI lies in its ability to efficiently evaluate descriptive answers by identifying contextually relevant content for human evaluation. In addition to reducing evaluation time, KSI minimizes the cognitive effort required by human evaluators by eliminating the need to manually identify irrelevant content. Thus, the use of KSI has the potential to enhance the quality of evaluation by enabling evaluators to focus on contextually relevant content while reducing cognitive effort. As improvements in evaluation quality can positively influence learning outcomes, KSI can, more broadly, contribute to the realization of United Nations Sustainable Development Goal 4. A distinguishing feature of the KSI model is that it assists human evaluation and supports integration with future automated evaluation systems.