AI Framework Compresses Thousands of MOOC Reviews Into Concise, Accurate Summaries

Massive open online courses have transformed global education, but they have also produced an avalanche of feedback that no human team can realistically read. Every day, platforms such as Coursera collect thousands of learner reviews spanning everything from instructor performance to technical glitches, and buried inside that text are the insights educators need to improve their courses. A new study published in Discover Artificial Intelligence proposes a way to tame this flood: a framework called ClusSumm, developed by Mohammed Abdul Al Arafat Tanzin and Md. Jisan Mashrafi of BRAC University in Dhaka, Bangladesh, which combines hierarchical clustering with adaptive transformer models to distill mountains of learner feedback into short, information-dense summaries.

The problem the researchers set out to solve is a familiar one in modern natural language processing. Transformer-based summarization models such as BART, PEGASUS, and T5 are remarkably fluent, but they face a persistent trade-off between brevity and completeness. Longer summaries tend to score better on automatic benchmarks like ROUGE because they overlap more with reference texts, yet they offer limited value to instructors and administrators who need rapid, actionable insights. Shorter summaries are more practical but often omit critical details or misrepresent the breadth of learner experiences. Compounding the issue, most existing systems apply a single pretrained model uniformly across all input content, treating highly informative sentences and semantically redundant ones in exactly the same way.

ClusSumm’s central innovation is to separate content selection from content generation. Rather than feeding an entire corpus of reviews into a summarizer and hoping for the best, the framework first organizes the material. Review texts are cleaned of URLs, punctuation noise, digits, and non-English tokens, then segmented into sentences and grouped by K-Means clustering so that semantically similar sentences end up together. The optimal number of clusters is determined automatically using the silhouette coefficient, a statistical measure that rewards clusters that are both internally compact and well separated from one another. The researchers searched cluster counts from 2 to 18 and selected the configuration yielding the highest silhouette score.

Within large clusters, where redundancy is most likely to accumulate, the framework applies TF-IDF sentence ranking. This classic information retrieval technique scores each sentence by combining how frequently its terms appear within the sentence with how rare those terms are across the wider corpus. Sentences with high scores are retained; low-ranked ones are filtered out before any generation takes place. This pre-selection step is what allows the downstream models to concentrate their generative capacity on genuinely informative content instead of diluting it across repetitive praise and boilerplate recommendations.

The framework then deploys what the authors call a model-adaptive assignment strategy. Large clusters, which serve as a practical proxy for input complexity, are routed to PEGASUS, a model known for strong abstractive capabilities on complicated inputs. Smaller, simpler clusters are handled by DistilBART, a computationally efficient distilled variant. Finally, the individual cluster summaries are merged into one coherent output by a hierarchical BART-based fusion module. According to the authors, this coupling of cluster-size-adaptive model assignment with hierarchical fusion within a single pipeline is what distinguishes ClusSumm from earlier hierarchical approaches that applied one model uniformly across all clusters.

To test the framework, the team first had to build the data. Because no publicly available dataset offered sufficient course-level review information across multiple MOOC domains, they constructed one through systematic web scraping of Coursera using Selenium WebDriver with Chrome automation. The collection covered eight domains: Artificial Intelligence, Machine Learning, Deep Learning, Natural Language Processing, Python Programming, Java Programming, Web Development, and Generative AI. Roughly 1,000 course links were gathered per domain, producing approximately 8,000 course records, each containing quantitative metadata such as ratings, enrollment figures, and like ratios alongside up to three learner reviews. The full collection effort took around eighty hours, and the curated dataset has been released publicly on Kaggle to support reproducibility.

The researchers benchmarked eight state-of-the-art transformer models against ClusSumm, including BART, PEGASUS, T5, FLAN-T5, DistilBART, LED, LongT5, and BigBird PEGASUS. All models were run in a zero-shot configuration using their standard Hugging Face settings, with no domain-specific fine-tuning, a choice the authors say reflects realistic deployment scenarios where labeled training data is scarce. The results revealed striking differences in compression behavior. LED produced the shortest summaries at a summarization ratio of just 3.50 percent, but paid a devastating price in quality, recording the lowest ROUGE-1 F-measure at 6.47 percent. At the other extreme, PEGASUS generated the least compressed summaries at 57.49 percent, achieving the highest F-measures largely by retaining vast portions of its input.

ClusSumm carved out a distinctive position. It achieved a summarization ratio of 12.65 percent, the second most concise of all nine systems evaluated, while maintaining ROUGE-1 precision of 98.31 percent and ROUGE-2 precision of 90.12 percent. In other words, nearly everything the framework chose to include was relevant, even though it included far less than its competitors. The authors are careful to note that ClusSumm’s recall and F-measure scores were lower than several baselines, and they characterize the framework as offering high precision under strong compression rather than overall superiority across every dimension. Still, the findings suggest that high ROUGE recall can sometimes reflect extensive text retention rather than effective abstraction, and that extreme compression without structured content selection, as LED demonstrated, severely degrades semantic coverage.

Perhaps the most consequential methodological choice in the study is its evaluation framework. Conventional assessments rely almost exclusively on ROUGE metrics, which measure overlap with reference summaries but say nothing about what actual users value. The researchers instead derived a composite Overall Score by surveying 30 participants with backgrounds in computer science, education technology, and instructional design, who ranked precision, recall, F-measure, and summarization ratio by perceived importance. Respondents assigned the highest weight to the summarization ratio, reflecting a consistent preference for concise outputs. Under this user-calibrated scoring, PEGASUS ranked first with 87.97 points, followed by BART-large-cnn and DistilBART, while ClusSumm ranked sixth overall with 74.10 points but earned the highest contribution from the compression component at 34.94 points. LongT5 and BigBird PEGASUS fared worst, suggesting that long-sequence architectures alone do not guarantee strong performance on heterogeneous, semantically sparse learner feedback.

The study’s limitations are acknowledged candidly by its authors. All models were evaluated without domain-specific fine-tuning, and no component-wise ablation was performed to isolate the individual contributions of clustering, sentence filtering, and hierarchical fusion. The reliance on ROUGE introduces lexical bias, and semantic similarity measures, human evaluation, and statistical significance testing were left for future work. Even so, the findings offer a practical blueprint for educational platforms drowning in feedback. By clustering semantically similar reviews, filtering redundancy before generation, and matching model capacity to input complexity, ClusSumm demonstrates that structured preprocessing and hierarchical summarization can turn unstructured learner commentary into decision-ready summaries, supporting instructors, course developers, and researchers in improving course quality at a scale that manual analysis could never match.

Subject of Research: Automated summarization of MOOC learner reviews using hierarchical clustering and transformer models

Article Title: ClusSumm a user centered hierarchical clustering and adaptive transformer framework for efficient summarization of MOOC learner reviews

Article References: Tanzin, M. A. A. A., & Mashrafi, M. J. (2026). ClusSumm a user centered hierarchical clustering and adaptive transformer framework for efficient summarization of MOOC learner reviews. Discover Artificial Intelligence, 6(1), Article 1461. https://doi.org/10.1007/s44163-026-02491-7

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02491-7

Keywords: MOOC reviews, text summarization, ClusSumm, hierarchical clustering, transformer models, K-Means, TF-IDF, PEGASUS, BART, ROUGE metrics, learning analytics, natural language processing