Knowledge tracing involves predicting learners' understanding of concepts over time, and fusion networks have emerged as effective tools for integrating diverse data sources into the learning models. This study focuses on understanding the necessity and influence of various components within the fusion network. The traditional model for KT has limitations in capturing the nuanced and evolving nature of a student's understanding. These models primarily focus on tracking learners' conceptual understanding, neglecting factors such as guess and slip, which can significantly impact learning. Additionally, for educational content that requires open-ended answer questions, KT only considers the binary label of correctness that excludes the student's text information. The model was compared with existing knowledge tracing models using two datasets, ASSISTments. The results demonstrated that AEKT outperformed all other models, achieving a 10% improvement in the area under the receiver operating characteristic curve (AUC). Additionally, the study introduced the method of integrated gradients to interpret correctness in open-ended answers and understand the influence of students' responses. In this paper, We proposed novel model designed to enhance the prediction of student knowledge states by incorporating answer text information. Building upon the DKVMN model, our proposed model integrates the BERT language model to achieve state-of-the-art performance in knowledge tracing, and understanding the necessity and influence of answert text. Keywords: deep learning, knowledge tracing, explainable AI, education, personalized learning
