NLP analysis of student feedback reveals teaching quality concerns in the DRC

A survey of computer science students across the Democratic Republic of Congo has been analyzed with natural language processing, offering one of the first systematic pictures of how students in the country perceive the quality of their own education. The study, published in Discover Artificial Intelligence, collected open-ended responses from 378 students at 18 universities and higher education institutions and processed them using sentiment analysis and topic modeling. The results show a broadly positive outlook, tempered by substantial ambivalence and by pointed criticism of the country’s recently introduced LMD degree system.

Education in the Democratic Republic of Congo, as in most countries, rests on three primary stakeholders: the government, teachers, and students. Yet student voices are frequently underrepresented in decisions about teaching standards, a pattern the study’s authors describe as particularly evident in the Congolese academic ecosystem. Until now, the country has lacked a comprehensive report reflecting computer science students’ feedback on key aspects of teaching quality. The new research set out to fill that gap by applying computational text analysis to free-text survey answers focused on infrastructure and teaching quality, faculty, and curriculum.

The survey was designed in French, the official language of the DRC, and distributed online through Google Forms. Students participated voluntarily during the 2023–2024 academic year, with data collected between May and August 2024. Institutions were purposively selected according to two criteria: geographic location, spanning Kinshasa, Kongo-Central, Lubumbashi, Goma, and Ilebo, and institutional type, covering both public and private establishments. Participants came from a range of universities including the Université de Kinshasa, Université de Lubumbashi, Université Pédagogique Nationale, and Institut Facultaire des Sciences de l’Information et de la Communication, among others. According to institutional enrollment records, the total estimated computer science student population across the 18 institutions was approximately 1,866, meaning the 378 respondents represented a participation proportion of 20.26 percent.

The questionnaire contained sixteen questions targeting three aspects of education: Infrastructures and Teaching Quality, Faculty, and Curriculum. Only the open-ended responses were used for the core analysis, since free text carries richer perceptual information than numerical ratings. Before any modeling could begin, the researchers undertook extensive preprocessing tailored to French text, a language for which many standard NLP techniques were not originally designed. Punctuation, numbers, emojis, non-alphabetic characters, French diacritics, and stop words were removed; text was lowercased, tokenized, and lemmatized. Institution names were manually normalized to standard abbreviations so that responses could be aggregated and stratified consistently, consolidating variations in spelling, spacing, capitalization, and abbreviation.

For sentiment classification, the team applied a pre-trained BERT-based model from Hugging Face’s Transformers library, specifically the multilingual “nlptown/bert-base-multilingual-uncased-sentiment” model, which outputs ratings on a one-to-five star scale. These were mapped to three categories: positive for four or five stars, negative for one or two stars, and neutral for three stars. The model was applied to three composite text corpora, one for each educational aspect, built by merging the relevant open-ended answers. Two observations were removed before analysis because their entire rows contained many missing values and unidentifiable institution names, leaving 376 observations per aspect.

The sentiment results reveal a nuanced picture rather than simple satisfaction. Infrastructure received the most positive sentiment at 52.7 percent, while a plurality of students expressed positive views toward faculty competency, at 44.9 percent, and curriculum content, at 48.4 percent. At the same time, 21.8 percent of responses about the curriculum were negative, and the proportion of neutral responses ranged from 29.3 to 44.4 percent across the three aspects. The authors interpret this high neutral share as widespread ambivalence: students see potential in their education but recognize substantial room for improvement. Manual inspection of model outputs showed that the classifier tended to label all non-negative responses, including those suggesting improvements, as neutral, which partly explains the pattern.

To verify the model’s reliability, the researchers randomly sampled 30 percent of responses across each aspect as a validation set, which two independent researchers, an NLP expert and a French-speaking researcher, annotated manually. Inter-annotator agreement measured by Cohen’s Kappa reached 0.94, 0.94, and 0.97 for the three aspects respectively, indicating almost perfect agreement. Against these human labels, the BERT model achieved overall accuracies of 79.4 percent for Infrastructures and Teaching Quality, 66.9 percent for Faculty, and 70.5 percent for Curriculum. It consistently outperformed a lexicon-based TextBlob baseline, with improvements ranging from 10.7 percentage points for curriculum to 31.2 percentage points for infrastructure, suggesting that infrastructure-related feedback contains complex, context-dependent sentiment that simple word-polarity methods struggle to capture.

The second analytical layer used Latent Dirichlet Allocation, an unsupervised topic modeling algorithm, on 324 usable improvement-oriented responses drawn from a question asking what students would change. The researchers trained models with between two and ten topics and evaluated them using coherence scores and perplexity, both of which peaked at two topics, indicating that larger models overfit the small corpus. The two resulting themes were labeled Learning Process and Pedagogical Conditions, and Academic System and Structural Reform. The first theme, marked by words such as “student,” “course,” “lecturer,” and “condition,” reflects demands for more applied, hands-on learning and better study conditions, a gap the authors consider particularly critical in computer science, where technical skills are paramount. The second theme, marked by words such as “change,” “system,” and “LMD,” points to student perceptions of challenges related to the LMD system, a Licence-Master-Doctorate structure introduced in the DRC in 2023 as part of harmonization with international education standards. The authors suggest students may perceive its implementation as rushed and insufficiently prepared in terms of infrastructure, trained faculty, and clear guidelines.

Based on these findings, the authors propose several policy considerations. They recommend a phased, well-resourced implementation strategy for the LMD reform and a participatory review involving students, faculty, and administrators before full-scale implementation proceeds. They also suggest institutionalizing student voice mechanisms, such as regular feedback surveys, student representation on curriculum committees, and public reporting of feedback outcomes, alongside investment in laboratory equipment, internet access, and updated software, potentially through public-private partnerships with technology companies. The study is grounded in the “student voice movement” in educational research and in a system dynamics perspective, and the authors interpret the two topics as interconnected micro-level and macro-level concerns that may form a reinforcing cycle: systemic challenges may constrain classroom-level improvements, which in turn may perpetuate systemic weaknesses.

The authors are careful to note the study’s limitations. The sample of 378 students is relatively small and participation was imbalanced across institutions, with higher representation where faculty collaborators actively promoted the survey. The annotation procedure relied on adjudication by a single NLP expert in cases of disagreement, and the use of off-the-shelf models without domain-specific fine-tuning limited the ability to handle culturally specific expressions, sarcasm, irony, and other nuanced patterns in African French educational text. The findings also speak only to computer science students, not to higher education in the DRC as a whole. Even so, the study demonstrates the feasibility of applying NLP techniques to student feedback in a French-speaking, resource-constrained setting where such methods had not previously been assessed, and offers the Congolese government and academic institutions an evidence-based baseline for understanding how their students perceive the education they receive.

Subject of Research: Natural language processing analysis of computer science students' perceptions of teaching quality in the Democratic Republic of Congo

Article Title: An AI-driven empirical analysis of educational perceptions from the Congolese academic ecosystem using NLP techniques

Article References: Mukala, P., Muka, G., & Yamba, L. (2026). An AI-driven empirical analysis of educational perceptions from the Congolese academic ecosystem using NLP techniques. Discover Artificial Intelligence, 6(1), Article 1705. https://doi.org/10.1007/s44163-026-02675-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02675-1

Keywords: natural language processing, sentiment analysis, topic modeling, Democratic Republic of Congo, higher education, student feedback, computer science education, LMD system, BERT, Latent Dirichlet Allocation, survey research, educational quality