DRC students flagged curriculum gaps—but the AI analysis cannot represent every campus
A voluntary survey used multilingual BERT and topic modelling on feedback from 378 computer-science students at 18 institutions. Participation was uneven and sentiment accuracy ranged from 66.9% to 79.4% across the three topics.
The short version
- What happened
- Researchers surveyed 378 computer-science students at 18 public and private institutions in five DRC locations; the estimated eligible population was 1,866 and participation was 20.26%.
- Who it affects
- Students may gain a clearer channel for free-text feedback, while institutions can see recurring practical and structural concerns. Uneven participation and imperfect classification mean decisions still require direct consultation.
- What remains unproven
- Peer reviewed
Keep in mind
Peer reviewed
378 voluntary respondents
18 purposively selected institutions
20.26% estimated participation
Uneven campus participation
Computer-science students only
66.9%–79.4% sentiment accuracy
No longitudinal outcome
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The survey brought student comments into a system-level debate
A peer-reviewed study analysed how computer-science students in the Democratic Republic of the Congo described infrastructure, teaching, faculty and curriculum. The 2024 voluntary survey reached 378 students across 18 public and private universities and higher-education institutions in Kinshasa, Kongo-Central, Lubumbashi, Goma and Ilebo. Institutional records put the eligible computer-science population at about 1,866, giving a reported participation proportion of 20.26%.
Sixteen open and closed questions were asked in French. After cleaning, 376 usable composite responses remained for each of three aspects: infrastructure and teaching quality, faculty, and curriculum. A multilingual BERT model classified sentiment and latent Dirichlet allocation grouped recurring themes. Positive sentiment was the largest category for faculty competence and curriculum, while 21.8% of curriculum responses were classified negative and neutral responses ranged from 29.3% to 44.4%. Those proportions describe model-assigned categories, not independently verified levels of satisfaction.[1]
Human checking exposed uneven model accuracy
The researchers manually labelled a random 30% validation sample for each aspect. Two annotators agreed strongly, with Cohen's kappa between 0.94 and 0.97, and an NLP expert adjudicated the small number of disagreements. Against those labels, model accuracy was 79.4% for infrastructure and teaching quality, 66.9% for faculty and 70.5% for curriculum. Those figures are useful evidence that automated labels were imperfect, especially for complex or neutral feedback. They also show why aggregate sentiment should not be treated as an exact vote count.
Topic modelling identified two broad clusters: learning process and pedagogical conditions, and academic system and structural reform. Students called for more practical learning and better infrastructure, while comments about the recent licence–master–doctorate transition pointed to uncertainty about implementation. These topics summarize patterns in a survey corpus; they do not show how many campuses share each condition or prove that a particular reform caused student dissatisfaction.[1]
The distinction that matters
From voluntary feedback to cautious themes
Human validation matters because automated sentiment accuracy varied by topic.
Collect
378 students at 18 institutions answer a voluntary survey
376 usable composite responses
Analyse
Multilingual BERT and topic modelling classify text
Sentiment and two broad themes
Check
Two researchers label a 30% sample
66.9%–79.4% model accuracy
A baseline can amplify voices without replacing participation
For students, the practical issue is whether feedback changes curricula, laboratories, teaching support and the connection between study and employment. NLP can help institutions review many free-text responses without discarding nuance entirely. It should not become a substitute for representative consultation: institutions were purposively selected, recruitment depended partly on faculty promotion and participation varied across campuses. Students with the strongest views or better internet access may also have been more likely to respond. Computer-science students cannot represent every discipline or every higher-education learner in the country.
The authors report no funding or competing interests and say anonymised data are available on request rather than publicly released. A stronger follow-up would use balanced sampling, publish institution-level response rates, validate culturally specific French expressions, include other disciplines and repeat the survey after reforms. Linking specific concerns to documented institutional responses would show whether listening changed anything. Until then, the study provides a valuable DRC baseline and demonstrates a bounded AI-assisted listening method—not a national satisfaction score or a mandate for automated policy decisions.[1]