Data extraction from machine-translated versus original language randomized trial reports: a comparative study
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
BACKGROUND: Google Translate offers free Web-based translation, but it is unknown whether its translation accuracy is sufficient to use in systematic reviews to mitigate concerns about language bias. METHODS: We compared data extraction from non-English language studies with extraction from translations by Google Translate of 10 studies in each of five languages (Chinese, French, German, Japanese and Spanish). Fluent speakers double-extracted original-language articles. Researchers who did not speak the given language double-extracted translated articles along with 10 additional English language trials. Using the original language extractions as a gold standard, we estimated the probability and odds ratio of correctly extracting items from translated articles compared with English, adjusting for reviewer and language. RESULTS: Translation required about 30 minutes per article and extraction of translated articles required additional extraction time. The likelihood of correct extractions was greater for study design and intervention domain items than for outcome descriptions and, particularly, study results. Translated Spanish articles yielded the highest percentage of items (93%) that were correctly extracted more than half the time (followed by German and Japanese 89%, French 85%, and Chinese 78%) but Chinese articles yielded the highest percentage of items (41%) that were correctly extracted >98% of the time (followed by Spanish 30%, French 26%, German 22%, and Japanese 19%). In general, extractors' confidence in translations was not associated with their accuracy. CONCLUSIONS: Translation by Google Translate generally required few resources. Based on our analysis of translations from five languages, using machine translation has the potential to reduce language bias in systematic reviews; however, pending additional empirical data, reviewers should be cautious about using translated data. There remains a trade-off between completeness of systematic reviews (including all available studies) and risk of error (due to poor translation).
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.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.312 | 0.117 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.023 | 0.003 |
| Bibliometrics | 0.000 | 0.002 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.002 | 0.001 |
| Open science | 0.003 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.027 | 0.026 |
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 it