Dataset Creation and Preparation for the AISE Project
A multilingual dataset of 2,617 student essays was collected from schools participating in the AISE project across Greece, Serbia, Croatia, and Cyprus, each graded by teachers on three pedagogically meaningful dimensions: Content, Organization, and Language. A thorough data preparation pipeline was implemented, including text cleaning and normalization, removal of punctuation and special characters, stemming, lemmatization, and sentence tokenization, while preserving the natural writing characteristics relevant to human-style evaluation. A multilabel stratified train/test split strategy was applied to ensure that the joint distribution of all three grading dimensions was proportionally preserved across partitions, while strict student-level separation prevented any data leakage between training and test sets. The final dataset was divided into 2,356 training essays and 261 test essays. Using the training subset, three independent FAISS-based knowledge bases were constructed (one per grading dimension) using MiniLM (all-MiniLM-L6-v2) embeddings to enable efficient semantic retrieval of representative graded examples. These knowledge bases formed the core retrieval infrastructure of the Retrieval-Augmented Generation (RAG) approach which was used for automated essay assessment to ensure that the AI predictions are grounded in real, teacher-graded examples rather than generic model reasoning.