GOOGLE REVIEW SENTIMENT: PAIRED TEXT-FIELD REANALYSIS ReviewsBoost Editorial Team | Version 1.0.0 | 2026-11-07 Article: https://reviewsboost.ca/google-review-sentiment-preprocessing-study.html Licence for our aggregate release, charts and reproduction scripts: CC BY 4.0 https://creativecommons.org/licenses/by/4.0/ SCOPE Original secondary analysis, not newly collected customer reviews or a Canadian field survey. We rescore two stored text fields from the same 7,015 review records, representing 86 academic-library profiles worldwide. The result is model sensitivity, not authenticity, accuracy, policy eligibility or removal probability. No personal review data or library identifiers are exported. SOURCE AND ATTRIBUTION Ioannis C. Drivas, Academic Libraries User Perception: A VADER-Enhanced Corpus of 7,605 Google Maps Reviews, version v1. DOI: 10.5281/zenodo.18594870. https://zenodo.org/records/18594870 CC BY 4.0 (verified from repository API metadata). Declared publication 2025-09-10; repository record created 2026-02-10. Source describes worldwide libraries, collection March-July 2025, machine translation and duplicate/rating-only removal. Its title count is not the denominator observed in its downloadable file. Workbook MD5: 07a6f254b04b1bb821cc3ec304067fb1 Workbook SHA256: 62337cb2b304f7fbdbdd5ce646fa817eea3bb830ab37a872a29c1b45cde8bc8f NLTK lexicon TXT SHA256: 0aabac74cd95def2dbedc04ce55e39698d91113ffadb6d86b87dd337c4d2184d Additional source URLs, model hashes and CSV checksums are in the JSON release. METHOD Python 3.12.0; NLTK 3.9.4, sentiment.vader.SentimentIntensityAnalyzer. Input A: source "Google Translation". Input B: source "Processed Text". Only surrounding whitespace is stripped. No new translation, custom cleaning or fitting is applied. The supplied "Sentiment Score Compound" is NOT reused. The author’s original scoring/preprocessing pipeline is not reproduced here. The individual-star column is "review_rating", not the business "rating". Unique (google_id, review_id) pairs must have integer 1-5 stars and non-empty values for both text fields. Unexpected exclusions stop the builder. 41 score-only rows have no review IDs/profile IDs/stars and are excluded. There are 7,056 XML worksheet rows after the header and 7,015 eligible pairs. The 590 difference from the 7,605 title count remains unreconciled; do not interpret that difference as missing, removed or fake reviews. Positive: compound >= +0.05. Negative: compound <= -0.05. Neutral: between. The same thresholds and fixed lexicon apply to both fields. "Changed": different categorical labels for the paired inputs. "Direct opposite": one positive and one negative, in either direction. No human-labelled benchmark or Google moderation outcomes are available. DATA DICTIONARY (all CSV percentages are on a 0-100 scale) 1. google-sentiment-polarity-transitions.csv | 9 rows english_label, processed_label: model category for each input. review_count: number of paired records in that cell. english_label_denominator: sum across the English-label row. row_percent: cell count / English-label denominator * 100. all_review_percent: cell count / 7,015 * 100. 2. google-sentiment-rating-strata.csv | 5 rows stars: individual review rating, 1-5. reviews: eligible records within the star group. changed_labels, changed_percent: paired differences within that group. direct_opposite_labels: direct positive/negative changes within the group. english_negative/neutral/positive and processed_negative/neutral/positive: model-label counts for each input within the star group. 3. google-sentiment-threshold-sensitivity.csv | 50 rows absolute_cutoff: symmetric threshold t, 0.01 to 0.50 in 0.01 increments. Positive >= +t, negative <= -t, neutral strictly between. reviews: the SAME 7,015 paired observations at every cutoff. changed_labels, changed_percent, direct_opposite_labels, direct_opposite_percent: as above, using cutoff t. These are sensitivity scenarios, not independent samples or accuracy tests. 4. google-sentiment-library-weights.csv | 8 rows scope: all, each star group, combined 1_2_stars or combined 4_5_stars. reviews, changed_labels: pooled counts within the scope. eligible_libraries: profiles with at least one review in that scope. review_weighted_percent: sum(changes) / sum(reviews) * 100. equal_library_mean_percent: arithmetic mean of eligible within-profile percentages. Each eligible profile has equal weight. library_median_percent: median of eligible within-profile percentages. Groups are not independent. Small denominators can produce unstable rates. No profiles or institution identifiers are exported. 5. google-sentiment-data-audit.csv | 11 rows metric, value, unit, definition: source-title count, physical worksheet rows, score-only exclusions, eligible pairs, unreconciled gap, libraries, model evaluations, first/last review years, smallest/largest profile groups. Units differ. Do not sum these measures as a sample size. 6. google-review-sentiment-study.json Complete aggregate release: provenance, methods, summary, five data tables and SHA256 checksums of the CSV files. No reviewer text or identifiers. RESULT CHECKS 7,015 reviews; 86 libraries; 14,030 scorer evaluations. 1,403 changed labels (20.0%); 5,612 unchanged. 356 direct positive/negative changes (5.074839629...%). 1,047 changes to or from neutral. Review-weighted change rate 20.0%; equal-library mean 18.212939%. Percentages in CSV files are rounded to 6 decimal places. Displayed article/figure percentages are generally rounded to 1 decimal place. REPRODUCE Use a fresh folder containing the two scripts and the dependency file. Python 3.12.0 was used. The dependencies are research-only; no site runtime or production server changes are needed. python -m pip install -r google-sentiment-requirements.txt python build-google-sentiment-study.py --download python plot-google-sentiment-study.py The --download flag retrieves checksum-pinned public workbook/lexicon sources into .seo-private/google-star-text-study-2026-11-07/ next to the scripts. Source hashes are checked before use. Later runs can omit --download. The workbook contains public review records. Keep this cache private and out of your website/repository output; our release does not republish it. Only aggregate CSV/JSON tables and SVG/PNG figures are written to the output. No accounts, passwords, API keys, customer systems or network scraping are used. Do not upgrade the model or replace the source file and still call it v1.0.0. This script uses the NLTK implementation, not the separate vaderSentiment pip package. Outputs from another implementation need not be identical. Matplotlib 3.10.7 is used for figures, with system DejaVu Sans fonts. Plots can also be rebuilt directly from the supplied JSON without the workbook. LIMITATIONS Convenience/clustered corpus; one sector; source-provided translations; no independent human labels; original preprocessing steps not fully known; no fraud or removal outcomes; no Canadian prevalence estimate. No causal claim, population confidence interval or optimal-classification claim is made. NLTK/VADER retain their own licences. Their source code is not redistributed. CITE ReviewsBoost Editorial Team. (2026, November 7). Google review sentiment: a paired text-field reanalysis (Version 1.0.0). https://reviewsboost.ca/google-review-sentiment-preprocessing-study.html Based on Drivas, I. C., doi:10.5281/zenodo.18594870. When quoting 20.0%, specify 1,403/7,015 paired records in a worldwide academic-library corpus, comparing two stored inputs with NLTK VADER at the +/-0.05 cutoff. It is not a fake-review, accuracy or removal rate. Corrections: https://reviewsboost.ca/editorial-policy.html