Enhancing Knowledge Quality in Crowd-Sourced Developer Q&A Platforms through AI-driven Software Solutions
Bibliographic record
Abstract
Programming question-and-answer (Q&A) platforms have transformed how developers seek and share programming knowledge, addressing limitations of traditional documentation and tutorial resources. Among them, Stack Overflow (SO) stands as the most prominent community-driven repository where developers exchange solutions and insights. Despite its impact, SO faces persistent challenges that threaten the clarity, reliability, and longevity of its content. Many code snippets shared in answers lack inline comments, accepted answers can become outdated as technologies evolve, and subtle forms of technical debt (TD) often appear in optimization questions without explicit recognition. This thesis presents three interrelated studies that leverage Large Language Models (LLMs) to enhance the comprehensibility, adaptability, and quality awareness of knowledge shared on SO. In our first study, we focus on improving the interpretability of code snippets in SO answers. We observe that a large number of shared code examples lack meaningful inline comments, reducing their reusability, particularly for novice programmers. To address this, we develop AUTOGENICS, a context-aware tool powered by LLMs that automatically generates inline comments aligned with both the code and its corresponding question context. Through manual evaluation and practitioner feedback, AUTOGENICS demonstrates significant improvements in accuracy, adequacy, conciseness, and usefulness, offering developers clearer, noise-free code explanations that enhance comprehension. While addressing code comprehensibility, we notice that even well-explained answers may lose relevance over time as technologies change. The valuable discussions embedded in user comments often remain underutilized, despite containing suggestions and corrections that could improve answer quality. Motivated by these observations, our second study introduces AUTOCOMBAT, a tool that automatically synthesizes improvement-oriented comments to produce enhanced answers. Using the ReSOlve benchmark and multiple state-of-the-art LLMs, AUTOCOMBAT integrates comment-based feedback into semantically faithful answer revisions. Evaluations using syntactic and semantic similarity metrics, along with a user survey, confirm that AUTOCOMBAT effectively transforms community feedback into coherent and updated answers while preserving technical intent, thereby promoting the continuous evolution of SO knowledge. During the investigation of answer evolution, we discover that many optimization-related questions contain signs of unacknowledged quality concerns in the form of TD, such as inefficient design or premature optimization decisions. Recognizing this gap, our third study introduces DebtNetX, a multimodal late-fusion transformer that combines textual and code representations to detect TD in SO questions. By integrating DeBERTa for text encoding and CodeBERT for code encoding, DebtNetX outperforms text-only baselines, revealing that nearly one-third of optimization-related questions on SO show latent TD. A lightweight browser plugin operationalizes this approach by enabling developers to identify the presence and type of TD, as well as reflect on potential debt indicators while engaging with community content. Collectively, these studies present complementary AI-assisted approaches for improving the clarity, currency, and quality awareness of developer knowledge on SO. Together, they contribute toward a more comprehensible, adaptive, and sustainable ecosystem of community-driven software knowledge.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.034 | 0.140 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.006 | 0.002 |
| Science and technology studies | 0.002 | 0.002 |
| Scholarly communication | 0.008 | 0.009 |
| Open science | 0.004 | 0.012 |
| Research integrity | 0.003 | 0.003 |
| Insufficient payload (model declined to judge) | 0.006 | 0.003 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".