Select-then-generate paradigm for feedback generation on overlength documents¶

Theory · All theories
Evidence · no claims cited

Description¶

The select-then-generate paradigm decomposes feedback generation for long documents into two sequential sub-problems. As the article states, it resolves "the issue of overlength input documents: 1) an unsupervised sentence-level extractive summarization task, and 2) a supervised PLM-based text-to-text generation task". An extractive summarizer (cross-entropy extraction optimizing vocabulary diversity plus a length term) produces a summarized report that becomes the input to the generation function implemented by a pre-trained language model (BART).

Design Implications¶

Context¶

Requirements¶

• An extractive summarizer that retains diverse salient content within the language model's input length limit
• A pre-trained language model fine-tuned on summarized reports paired with reference feedback

Constraints¶

• Summarization may drop sentences that expert suggestions focus on, which the authors speculate contributes to the system's weaker Suggestions performance

Target Learners¶

• students submitting long-form project reports

Target Learning Objectives¶

• receiving instant textual feedback on complex, open-ended student work

Claims¶

Related Theories¶

Examples¶

• Insta-Reviewer: a select-then-generate system for instant feedback on student project reports

Key Sources¶

• Qinjin Jia, Mitchell Young, Yunkai Xiao, Jialin Cui, Chengyuan Liu, Parvez Rashid, Edward Gehringer. (2022). Automated Feedback Generation for Student Project Reports: A Data-Driven Approach. Journal of Educational Data Mining, Volume 14, No 3. https://github.com/qinjinjia/JEDM22_Automated_Feedback_Generation