Image integrity and data authenticity report for 'Boosting Semi-Supervised Video Captioning via Learning Candidates Adjusters' (DOI: 10.1145/3652838)

Academic fraud report · Geng Detector

Summary

Verdict: No confirmed academic misconduct identified. The report concludes the paper is substantially free of substantive fraud, finding only editorial, typographical, and metadata inconsistencies. Key issues: (1) widespread spelling/OCR-like errors throughout the manuscript (e.g., 'deine', 'irst', 'efectiveness', 'luency', 'proit', 'pesudo') that impair readability without affecting scientific content; (2) an author affiliation contradiction listing Beijing Jiaotong University as 'Beijing, United States' for two authors; (3) a chronology inconsistency where the stated 'revised' date (10 July 2024) postdates the 'accepted' date (25 February 2024); (4) unusually monotonic improvement across all metrics (B@4, ROUGE, CIDEr) and all labeled-data ratios (30/50/80/100%) for the proposed LCABM model, e.g., CIDEr rising 83.13 → 95.05 → 106.1 → 111.76. Confidence is moderate: the issues flag warrant scrutiny but are not proof of fabrication. Limits include reliance on the public PDF, absence of raw experimental logs, and inability to verify peer-review records.

Verdict

No conclusive evidence of academic fraud (data fabrication, image manipulation, or systematic misconduct) was detected in the paper *Boosting Semi-Supervised Video Captioning via Learning Candidates Adjusters* (DOI: 10.1145/3652838, ACM TOMM, 2024). All anomalies identified are editorial, typographical, or metadata-related.

Key findings

Typographical errors (text quality, not science): The manuscript contains numerous spelling mistakes consistent with OCR or inadequate proofreading (examples: 'deine' for 'define', 'irst' for 'first', 'efectiveness' for 'effectiveness', 'luency' for 'fluency', 'proit' for 'profit', 'pesudo' for 'pseudo'). These recur across abstract, introduction, and methods sections.
Affiliation error: Authors ZHENJIANG MIAO and YIGANG CEN are listed under 'Beijing Jiaotong University Institute of Information Science, Beijing, United States' — a geographic impossibility, since the institution is in Beijing, China. Co-authors from the same affiliation are correctly listed as 'Beijing, China'.
Timeline anomaly: The copyright page shows: Received 5 August 2023; revised 10 July 2024; accepted 25 February 2024. The 'revised' timestamp (10 July 2024) is later than the 'accepted' timestamp (25 February 2024), violating the standard submission → revision → acceptance sequence.
Suspiciously smooth performance scaling: In Table 2 and Table 3, the proposed LCABM model shows strictly monotonic gains in B@4, ROUGE, and CIDEr across labeled-data ratios of 30%, 50%, 80%, and 100%, and outperforms all baselines at every ratio. Example CIDEr trajectory for LCABM: 83.13 → 95.05 → 106.1 → 111.76.

Evidence highlights

Typographical pattern is dense and uniform, suggesting automated extraction or insufficient copy-editing rather than authorial sloppiness in scientific reasoning.
The affiliation error is internally inconsistent within the same author block, pointing to a metadata/transcription mistake.
The chronology reversal is explicit in the ACM copyright block (DOI: 10.1145/3652838) and cannot be reconciled without publisher clarification.
The performance trajectory (CIDEr 83.13, 95.05, 106.1, 111.76 at 30/50/80/100%) shows no noise or non-monotonic deviation typical of empirical learning curves; this is suggestive but not probative of data manipulation.

Notes

Findings 1–3 are best characterized as editorial/production oversights and do not constitute misconduct.
Finding 4 ('perfect' trend) warrants mild skepticism but is consistent with a genuinely strong model; verification would require access to raw training logs, which are not provided in the PDF.
Recommended follow-up (optional): request raw experimental data from the corresponding author to verify Table 2/3 entries, and ask the publisher (ACM) to correct the affiliation and timeline metadata.
This automated check cannot detect issues beyond the text and figures supplied; image forensics on figures did not reveal manipulation.