摘要：

针对传统评估教学未充分利用教学过程信息, 导致泛化能力差和预测精度低等问题, 提出一种基于SA-BPNN多模态融合的教学质量评价模型。该模型包括模态特征提取模块和多模态融合预测模块。首先构建全过程多维度的教学质量评价体系, 包括人工评分、在线教育平台和教学视频3种模态数据, 通过自注意力机制改进反向传播神经网络模块(SA-BPNN)提取各模态特征信息。其次鉴于评教过程存在模态数据缺失的可能, 在预测模块的早期融合阶段引入多模态混合融合策略以改进反向传播神经网络(MF-BPNN), 该策略按照特定规则将不同模态的特征信息混合再融合, 以减轻模型对某一模态数据的依赖。随后将融合后的特征信息输入到MF-BPNN模块中, 生成最终的评教结果。该模型融合多模态信息进行评价, 避免了传统评教的主观性, 且混合融合策略提升了模态缺失时的预测精度。在某高校真实数据集上的实验结果表明, 相较于BPNN和GA-BPNN方法, 该模型的均方误差分别提高了2.4~3.9个百分点, 能够高效准确地评估教学质量, 为检验教师教学效果和优化教学管理, 最终全面提升教学质量提供理论支持。

关键词: 质量评价, 自注意力机制, 多模态融合, 特征提取, 神经网络

Abstract:

A Simulated Annealing Backpropagation Neural Network (SA-BPNN)-based multi-modal fusion model for evaluating teaching quality is proposed to address the issues of poor generalization ability and low prediction accuracy caused by the insufficient utilization of teaching process information in traditional teaching evaluations. The model comprises two parts: a modal feature extraction module and a multi-modal fusion prediction module. First, a multidimensional teaching quality evaluation system with three modal data types: manual scoring, online education platforms, and teaching videos, is constructed. The SA-BPNN module is improved using a self-attention mechanism to extract feature information from each modality. Second, considering the possibility of missing modal data in the evaluation process, a multi-modal hybrid fusion strategy is introduced in the early fusion stage of the prediction module to improve the MF-BPNN. This strategy combines and fuses the feature information of different modalities according to specific rules to reduce the reliance of the model on a certain modality. The fused feature information is then input into the MF-BPNN module to generate the final evaluation results. The model that fuses multi-modal information for evaluation avoids the subjectivity of traditional evaluations, and the hybrid fusion strategy improves the prediction accuracy when modal data are missing. Experimental results based on the real dataset of a university show that, compared with the BPNN and GA-BPNN methods, the Mean Square Error (MSE) of this model increases by 2.4 and 3.9 percentage point, respectively, which can effectively and accurately evaluate the teaching quality and provide theoretical support for testing the teaching effect of teachers, optimizing teaching management, and comprehensively improving the teaching quality.

Key words: quality evaluation, self-attention mechanism, multi-modal fusion, feature extraction, neural network