Abstract
Teaching style is a critical factor in student engagement and academic success. However, quantifying these styles has traditionally been a subjective and manual process. This paper presents a novel approach to understanding teaching styles by mapping them into a Teaching Style Semantic Space (TSSS) based on the pleasure-arousal dimensional theory. Using a newly developed dataset of 4,541 annotated cross-modality teaching segments, we propose the Attention-based Multi-path Multi-task Deep Neural Network (AMMDNN). This model integrates audio, visual, and textual features to predict teaching styles with high accuracy. Our results show that AMMDNN outperforms baseline methods, achieving a significant improvement in Concordance Correlation Coefficient (CCC). Furthermore, we demonstrate the practical utility of this model through case studies on teacher fatigue, subject-specific style variations, and the relationship between teaching style and student attention.
Introduction
While educators recognize that a teacher's "style"—their unique blend of delivery, enthusiasm, and interaction—is vital for learning, the field has lacked a standardized, quantitative way to measure it. Most existing evaluations rely on student surveys or manual observations, which are prone to bias and difficult to scale. Previous automated attempts often simplified the problem into discrete categories (e.g., "happy" vs. "sad") or relied on a single data source, such as audio.
This research addresses these gaps by proposing a continuous, multi-modal system. By analyzing how a teacher speaks, moves, and uses language, the system can objectively place their performance within a nuanced semantic space. This allows for a more granular understanding of teaching behaviors and provides a foundation for automated, real-time feedback in educational settings.
Quantifying Style: The Teaching Style Semantic Space (TSSS)
To move beyond rigid categories, the authors developed the Teaching Style Semantic Space (TSSS). This space is rooted in the psychological Pleasure-Arousal dimensional model. To ensure the model reflected real-world educational contexts, the researchers analyzed over 10,000 parent feedback questionnaires from an educational enterprise.
From this data, 41 descriptive adjectives (such as "humorous," "strict," "passionate," and "dull") were identified. Human annotators then mapped these words onto a 2D coordinate system:
- Pleasure (X-axis): Ranging from unpleasant/serious to pleasant/engaging.
- Arousal (Y-axis): Ranging from soporific/low-energy to energized/high-intensity.
This mapping allows any teaching segment to be represented as a coordinate, providing a mathematical way to compare different styles and track changes over time.
The AMMDNN Architecture
The core of the predictive system is the Attention-based Multi-path Multi-task Deep Neural Network (AMMDNN). The architecture is designed to handle the complexity of cross-media data:
- Multi-path Processing: To prevent the "curse of dimensionality" when combining massive feature sets (audio, visual, and text), the model processes each modality through separate local regressors before integration.
- Acoustic-Guided Attention: Recognizing that a teacher's voice often signals the importance of their gestures or words, the network uses high-level acoustic features to weight the visual and textual inputs. This attention mechanism ensures the model focuses on the most relevant cues at any given moment.
- Multi-task Learning: Since Pleasure and Arousal are psychologically linked, the model predicts both dimensions simultaneously. This shared learning approach allows the model to capture the internal correlations between the two axes, improving overall robustness.
The TEG18 Dataset
To train and validate the model, the authors introduced the TEG18 dataset, a comprehensive benchmark of real-world teaching. It consists of 4,541 10-second clips from primary school classes, fully annotated for cross-modality analysis:
- Acoustic: 1,582 features extracted via openSMILE.
- Visual: 14 features covering facial expressions, body orientation, and posture.
- Textual: 4,200 features derived from word2vec embeddings of transcribed speech.
Four independent annotators provided the ground-truth Pleasure and Arousal values, ensuring high inter-rater reliability for the dataset.
Experimental Results
The AMMDNN model demonstrated superior performance over traditional machine learning baselines, including Support Vector Machines (SVM) and Random Forests. In terms of the Concordance Correlation Coefficient (CCC):
- Pleasure: The model achieved a CCC of 0.739, a +0.067 improvement over the Random Forest baseline.
- Arousal: The model achieved a CCC of 0.808, a +0.048 improvement over the baseline.
Ablation studies confirmed that both the attention mechanism and the multi-task learning framework were essential to these performance gains.
Case Studies and Practical Insights
The researchers applied their model to several real-world scenarios, yielding insightful results:
- Subject-Specific Styles: The model revealed that STEM courses (like Physics) tend to occupy the lower-pleasure, high-arousal quadrant (serious/strict), while humanities (like Chinese) often skew toward higher pleasure.
- The Fatigue Effect: By analyzing a single class period, the model quantitatively tracked teacher fatigue. As the lesson progressed, "pleasure" scores often remained stable or rose slightly, but "arousal" (energy) significantly decreased.
- Student Engagement: A study tracking student gaze against the teacher's TSSS coordinates found that students were most attentive when teachers exhibited high arousal, even if the "pleasure" score was low (e.g., a "severe" or "passionate" style).
Conclusion and Future Work
This research provides a robust framework for the automated analysis of teaching styles. By combining a psychologically grounded semantic space with a sophisticated multi-modal neural network, the authors have created a tool that can provide objective feedback to educators. While the current reliance on some manually annotated visual features presents a bottleneck for full automation, future work in computer vision could bridge this gap, leading to real-time teaching quality analysis systems that help educators optimize their classroom presence for better student outcomes.
Summary updated