DKPrompt: Developer-side Code Review Prompts

• DKPrompt is a knowledge-guided prompt learning framework that enhances public code review by focusing on the developer’s request quality.
• It reformulates review tasks into masked language model problems using hard prompts and PDG-based code encoding to capture both linguistic and structural signals.
• Empirical results indicate that DKPrompt outperforms baselines, improving request necessity prediction by up to 8.4% and tag recommendation by up to 6.9%.

Knowledge-guided prompt learning for public code review is a prompt-based methodology for developer-based code review request quality assurance in Public Code Review (PCR). Rather than centering the reviewer—an emphasis that has dominated prior PCR work on reviewer finding, comment-quality prediction, and review comment recommendation or generation—the method targets the quality of the request posted by the developer, on the premise that request visibility and necessity signaling are prerequisites for better review responses. The framework, named Knowledge-guided Prompt learning for Public Code Review (KP-PCR), addresses two subtasks: request necessity prediction and tag recommendation (Li et al., 2024).

1. Research setting and problem focus

PCR is described as a direction within the Software Question Answering (SQA) community that assists developers in obtaining high-quality and efficient review services. In this setting, the paper identifies a structural asymmetry in the literature: most existing methods operate from the reviewer’s perspective, whereas the quality assurance of the developer’s own review request is comparatively under-studied. KP-PCR is presented as a response to that gap, explicitly treating the developer’s request as a first-order object of modeling rather than as a passive input to downstream review systems (Li et al., 2024).

The central problem formulation is narrow and operational. The framework does not attempt to replace the full review pipeline; instead, it concentrates on two request-side decisions. The first is whether a posted request expresses sufficient necessity. The second is which tags should be recommended for that request. This focus is significant because it repositions request formulation as an upstream control point for review quality. A common misconception is that PCR quality is determined chiefly by reviewer selection or comment generation. The motivation of KP-PCR suggests a different causal ordering: the request itself influences visibility and, by extension, the conditions under which useful review can occur (Li et al., 2024).

2. Prompt-based task reformulation

The method is organized around a prompt-learning reformulation of both subtasks. Specifically, KP-PCR uses text prompt tuning with hard prompt templates to convert request necessity prediction and tag recommendation into Masked LLM (MLM) problems. In this design, the original PCR subtasks are no longer treated as standalone classifiers in the usual sense; instead, they are cast into masked-token prediction through prompt construction (Li et al., 2024).

This reformulation is paired with an answer engineering module, through which both subtasks output their predicted results. The available description identifies the existence of this module but does not provide the underlying verbalizer design, label mapping, or decoding procedure. Published descriptions also indicate that the study evaluates four prompt templates, implying that prompt-template choice is treated as a substantive modeling variable rather than a cosmetic wrapper (Li et al., 2024).

Conceptually, this places KP-PCR within the family of prompt methods that align downstream tasks with pretrained objectives instead of introducing wholly separate heads or objective forms. In the present case, the relevant pretrained objective is MLM, and the prompt functions as the bridge between developer-request semantics and masked-token inference.

3. Knowledge guidance and code-aware prefix tuning

KP-PCR supplements hard-prompt reformulation with knowledge and code prefix tuning based on soft prompts. The method is described as introducing knowledge guidance from fine-tuned LLMs and as using the program dependence graph (PDG) to characterize code snippets. In the available account, these two sources of structure—external knowledge guidance and program-structure encoding—constitute the principal additions beyond ordinary prompt templating (Li et al., 2024).

Descriptions of the method further indicate that the soft-prompt component uses Wikipedia-derived knowledge to generate prefix vectors. This suggests a hybrid architecture in which request text, code structure, and external knowledge are combined at the prompt level rather than through late feature fusion. The PDG contribution is especially notable because it implies that code is not treated merely as raw text; it is represented through dependency structure, which plausibly provides a more faithful characterization of snippet semantics than lexical surface form alone (Li et al., 2024).

The paper also reports a time-complexity analysis and emphasizes that the introduction of knowledge guidance is implemented through a lightweight prefix-based operation. The significance of that claim is methodological: the framework is presented not as full model retraining, but as a comparatively parameter-efficient adaptation layer built around prompt/prefix mechanisms.

4. Position within prompt-learning research

KP-PCR belongs to a broader line of research in which prompt learning is used to recast downstream tasks into forms closer to pretrained objectives. In Declaration-based Prompt Tuning for VQA, downstream answer prediction is reformulated into MLM and ITM objectives through declarative prompting, explicitly to reduce the pretrain–fine-tune mismatch (Liu et al., 2022). In Adaptive Prompt Learning with Distilled Connective Knowledge for implicit discourse relation recognition, continuous prompt templates and virtual answers are combined with knowledge distillation, again using prompt structure as the primary adaptation interface (Wang et al., 2023).

Other prompt-learning work has pushed the adaptation interface in different directions. Dynamic Prompting treats prompt position, length, and representation as variables to be chosen dynamically rather than fixed globally (Yang et al., 2023). PromptKD uses prompt tuning to extract student-friendly knowledge for generative language-model distillation, showing that prompt parameters can be used not only for task conditioning but also for teacher adaptation (Kim et al., 2024). Against this background, KP-PCR can be understood as a prompt-learning instantiation specialized to public code review requests, with its distinctive elements being developer-side quality assurance, knowledge-guided prefix tuning, and PDG-based code characterization (Li et al., 2024).

This comparison suggests that KP-PCR is not simply “prompt engineering” in the narrow sense of hand-written instructions. It is better viewed as a structured prompt-learning system that combines hard prompts, soft prompts, external knowledge guidance, and code-aware structural signals.

5. Reported empirical evidence

The reported empirical evaluation uses a PCR dataset for the period 2011–2023. On this benchmark, the paper states that KP-PCR outperforms baselines by 2.3%–8.4% in request necessity prediction and by 1.4%–6.9% in tag recommendation (Li et al., 2024).

Published descriptions additionally state that the evaluation includes task accuracy on both subtasks, four prompt templates, comparisons against LLMs, a time-complexity analysis, and a case study. Those elements indicate that the study is intended to cover both effectiveness and deployment-oriented considerations, although the available high-level description does not enumerate the specific baselines, prompt-template contents, or case-study protocol (Li et al., 2024).

The paper also reports that the code implementation has been released. From an encyclopedia perspective, that matters because it situates KP-PCR not only as a conceptual proposal but also as a reproducible research artifact. The strongest explicit empirical claim, however, remains the performance improvement margins on the two developer-side subtasks.

6. Significance, interpretation, and unresolved technical granularity

The principal significance of KP-PCR lies in its reframing of PCR quality assurance as a developer-side problem. This shifts attention from the downstream review artifact to the upstream request that conditions whether high-quality review is likely to occur at all. The method’s combination of prompt reformulation, knowledge-guided soft prefixes, and PDG-based code characterization suggests a view of PCR requests as mixed natural-language–code objects that benefit from both linguistic priors and program-structural signals (Li et al., 2024).

A second important point is methodological. It would be inaccurate to interpret KP-PCR as a purely textual prompt method. The framework is described as combining hard prompt MLM reformulation with soft prompt/prefix tuning, fine-tuned LLM guidance, Wikipedia-derived prefix vectors, and PDG-based code encoding. This suggests a layered prompt-learning architecture in which prompting serves as the interface for integrating heterogeneous sources of information rather than merely as a template wrapper (Li et al., 2024).

At the same time, the published high-level description leaves several technical details unspecified. It does not provide the actual prompt templates, the exact MLM objective, the answer-engineering realization, the full algorithmic procedure, the dataset size, the baseline inventory, ablation numbers, or the derivation of the stated time complexity. Those omissions do not alter the main identity of KP-PCR, but they do limit fine-grained reconstruction of the method from summary material alone. Accordingly, the framework is best characterized as a prompt-based, knowledge-guided PCR request-quality system whose contribution is clearly defined at the architectural and task levels, while some lower-level implementation specifics remain outside the available description (Li et al., 2024).