Skip to content
Awsaf Alam
GitHubLinkedIn

Research paper, FSE 2026 Main Research Track

ToxiShield

A browser extension for GitHub pull requests. It catches a toxic code review comment before it's posted, explains what's wrong with it, and suggests a polite rewrite that keeps the technical point.

Md Awsaf Alam Anindya (first author), Showvik Biswas, Anindya Iqbal, Jaydeb Sarker, Amiangshu Bosu. ToxiShield: Promoting Inclusive Developer Communication through Real-Time Toxicity Filtering. Proceedings of the ACM on Software Engineering, Vol. 3, Issue FSE, Article FSE123, 2026. Presented at FSE 2026 in Montreal.

The problem

Code review is where developers talk to each other the most, and it's where harsh comments do real damage. A reviewer who's tired or annoyed can make a valid point in a way that hurts the person on the other end and puts them off contributing.

ToxiShield steps in at the moment of writing. Before a review comment goes up, it checks the text, and if the comment is toxic it tells the reviewer what kind of toxicity it is, why, and offers a rewrite. The reviewer stays in control: they can use the suggestion, edit it, or post their own words.

How it works

Three models run in a row, each answering one question. A clean comment stops after the first.

Review commenttyped in a GitHub PRToxicity FilterBERT classifierPostedif not toxicif toxicCommunication CoachClaude 3.5 Sonnettype + whyReframerfine-tuned Llama 3.2Suggestion in GitHubexplanation and rewrite
Illustrative. Simplified view of the pipeline described in the paper. Dashed line: the reviewer sees the suggestion in GitHub and decides what to post.

Results

Toxicity Filter

Is this comment toxic? BERT, trained on 38,761 code review comments.

98%
accuracy
97%
F1 score

Communication Coach

What kind of toxicity, and why? Prompt-tuned Claude 3.5 Sonnet.

39%
MCC, multiclass
42%
F1 score, multiclass

Reframer

How could it be said politely? Fine-tuned Llama 3.2, on 10,120 comments.

95.27%
style-transfer accuracy
97.03%
fluency
67.07%
content preservation
84%
J-score (accuracy, fluency and meaning combined)

In a user study with 10 participants based on the Technology Acceptance Model, people found ToxiShield useful and easy to adopt.

Try the flow

Click through what a reviewer sees.

Illustrative mockup

Pick a sample comment, or type your own.

Review comment on src/cart.ts, line 42
Illustrative. Canned answers written for this page; no model runs. The real extension runs the three models on any comment.

Read and cite

@article{anindya2026toxishield,
  title     = {ToxiShield: Promoting Inclusive Developer Communication through Real-Time Toxicity Filtering},
  author    = {Anindya, Md Awsaf Alam and Biswas, Showvik and Iqbal, Anindya and Sarker, Jaydeb and Bosu, Amiangshu},
  journal   = {Proceedings of the ACM on Software Engineering},
  volume    = {3},
  number    = {FSE},
  articleno = {FSE123},
  year      = {2026},
  doi       = {10.1145/3808130}
}

Back to the home page

© 2026 Awsaf Alam