應用深度學習於教學問卷之探勘與反饋

Deep Learning Applied in Analysis of the Student Surveys for Feedback to Instructors

連鑫(Sin Lian)

指導教授 ： 周瑞仁

國立臺灣大學/生物資源暨農學院/生物產業機電工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201902868

摘要

本研究與國立宜蘭大學校務研究辦公室合作，針對教學問卷文字意見進行文字情緒分析，並將結果分別應用於教師教學之反饋與優良教師之遴選。教師教學可由學生教學問卷得到回饋，惟需花費大量人力及時間進行意見提取，因此本研究開發一套系統，進行學生教學問卷文字意見分析的自動化，及時提供教師教學之參考並進而檢討改進，以及提供優良教師遴選之參考。
本研究利用資料探勘方法來對教學問卷進行質性評估。接著對教學問卷文字意見進行資料前處理並給予正面、中性及負面情緒的標註，之後將其向量化與分析。本研究比較循環神經網路、長短期記憶、專注循環神經網路、專注長短期記憶及卷積神經網路五種分類模型之文字情緒性能，最後選出最適合的分類器。
經各模型之分析性能比較後，從少量詞組去進行特徵提取效果較佳。本研究推薦卷積神經網路來作為教師教學改善及優良教師遴選目標之分類器，其分別於非負面情緒辨識率達96.0%，以及負面情緒辨識率達94.2%；於非正面情緒辨識率達94.1%，以及正面情緒辨識率達95.9%。

關鍵字

教師教學反饋 ； 優良教師遴選 ； 教學問卷 ； 文字情緒分析 ； 卷積神經網路

並列摘要

The study was conducted in collaboration with the office of Institutional Research (IR) at National Ilan University (NIU) in Taiwan to analyze textual opinions in teaching evaluation questionnaires, then apply the analysis results to the teaching of teachers and the selection of outstanding teachers. Teachers’ teaching can be awarded by the teaching evaluation questionnaires, but it takes a lot of manpower and time to extract the students’ opinions. Therefore, the research develops a set of system to automate the analysis of textual opinions in teaching evaluation questionnaires, not only providing reference for the improvement of teachers' teaching, but also providing a reference for outstanding teacher selection.
The study used data mining methods to qualitatively evaluate teaching questionnaires. First, the textual of the teaching questionnaire is pre-processed and labeled with positive, neutral and negative sentiments, then vectorized and analyzed. The study compared the textual sentiment performance of five kinds of classification models: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Attention RNN, Attention LSTM and Convolutional Neural Network (CNN), and finally selected the most suitable classifier.
We found that classifiers with a small number of phrases from feature extraction is better. Among those models CNN model has best performance with non-negative sentiment recognition rate of 96.0% and negative sentiment recognition rate of 94.2%, as well as non-positive sentiment recognition rate of 94.1%, and positive sentiment recognition rate of 95.9%. Thus, the study chose CNN as a classifier for teacher teaching improvement and outstanding teacher selection goals.

並列關鍵字

Teacher teaching feedback ； Selection of outstanding teachers ； Teaching evaluation questionnaires ； Text sentiment analysis ； Convolution neural network

參考文獻

曾秋旺。2018。教學問卷文字意見探勘應用於優良教師之遴選。碩士論文。台北: 臺灣大學生物產業機電工程學研究所。

Bengio, Y., P. Simard, and P. Frasconi. 1994. Learning long-term dependencies with gradient descent is difficult. IEEE Transactions on Neural Networks 5(2): 157-166.

Collobert, R., and J. Weston. 2008. A Unified Architecture For Natural Language Processing: Deep Neural Networks With Multitask Learning. In "Proc. 25th International Conference on Machine Learning.", 160-167.

Dutt, A., M. A. Ismail, and T. Herawan. 2017. A Systematic Review on Educational Data Mining. IEEE Access 5:15991-16005.

Hochreiter, S., and J. Schmidhuber. 1997. Long Short-term Memory. Neural computation (9)8: 1735-1780.