Spanish tweets classified as "solid waste related" or "unrelated" for the years 2020 until 2023

Spanish-Language Solid-Waste Tweet Classification Results (Colombia, 2020–2023) — Model-Generated Labels Machine-generated topic-classification results for a country-scale corpus of Spanish-language tweets, produced by a transformer pipeline (fine-tuned XLM-RoBERTa followed by zero-shot mDeBERTa) for the detection of urban solid-waste mentions. DOI (this record): 10.5281/zenodo.16098993 Version: v1 Release date: 2026-08-27 Licence (labels & documentation): [e.g. CC BY 4.0 — set in the record’s Licence field; see §11] Language: Spanish (predominantly Colombian / Latin-American variety) Access model: the identifiers and predicted labels are open; the tweet text is available on request under controlled access (see §7 and §10). Please write to research_NLP_ML_DL_CV[at]proton.me or the authors. Companion record (2019 manually annotated corpus): 10.5281/zenodo.16040609 1. Summary This record releases the model-generated classification results for the 2020–2023 country-scale Twitter/X corpus analysed in the associated publication. Each row is one original tweet, represented only by its identifier and the zero-shot topic ranking the pipeline assigned to it. These labels are predictions, not human annotations: the human-annotated gold data are released separately in the companion 2019 record and were used to fine-tune the classifier. The classification was produced by a sequential two-stage transformer pipeline: a fine-tuned XLM-RoBERTa (XLM-R) model separated waste-related from unrelated tweets across all 2020–2023 samples; the high-confidence waste subset (probability >0.9) was then re-classified by a zero-shot mDeBERTa model into five topical categories — waste garbage debris desechos desperdicios, sports, politics, economics, and other — each with a confidence score. Releasing these results as open identifiers plus predicted labels makes the country-scale inference reproducible and reusable — for confidence-based filtering, temporal and spatial analysis of waste discourse, topic modelling, and method comparison — while keeping the underlying tweet text under controlled access (§7, §10). 2. Provenance and collection Property Value Platform Twitter/X Collection method Streaming API v1.1, v2 Spatial filter (bounding box) Longitude 79.83°W – 66.02°W, Latitude 4.68°S – 13.41°N (extent of Colombia) Temporal coverage 2020–2023 Classified sample (2020–2023) 7,242,982 tweets Dominant language Spanish (with a minority in Portuguese and English) As with the 2019 record, the only sampling criteria applied at collection time were geographic location and time frame; no topical keyword filter was applied at the streaming stage. Unlike the 2019 annotated corpus, the 2020–2023 pipeline was also run without a keyword pre-filter — the classifier was applied to the full, unfiltered stream — so the results reflect the natural (very low) prevalence of solid-waste discourse on the platform. 3. From raw stream to classification results The classification results were derived from the 2020–2023 raw stream through the following steps: Duplicates elimination. Duplicate tweets and retweets were removed; only original (non-retweeted) posts were retained, to avoid inflating count-based results. Length filter. Tweets shorter than three words were removed to reduce classification noise. Supervised text classification (XLM-R). Every remaining tweet was classified as waste-related or unrelated by the fine-tuned XLM-RoBERTa model. 13.9 % of samples (6,843,795 tweets) received a probability ≥ 90 % of being waste-related; the remainder were classified as very likely unrelated (see §4.1, §5). Zero-shot re-classification (mDeBERTa). The high-confidence waste subset was re-classified by the zero-shot mDeBERTa model into five topical categories, each with a confidence score. This zero-shot ranking is what this record releases (see §4.2, §7). Confidence-based narrowing (analysis only). For the paper's downstream analyses, the subset whose top zero-shot label was waste… with probability ≥ 90 % was retained — 70,896 tweets, i.e. 0.14 % of all 2020–2023 samples — of which solid-waste tweets represented 6.9 %–7.6 % of the filtered set per year. This narrowing is a use of the released scores, not a separate file. The XLM-R classifier was fine-tuned on the balanced 25,723-tweet set derived from the 2019 corpus (positives augmented fourfold); the augmented and gold training data are not part of this record (they belong to the companion 2019 record and the paper). 4. Classification pipeline and models 4.1 Supervised stage — XLM-RoBERTa (waste vs unrelated) The first stage used a multilingual masked language model, XLM-RoBERTa Base (Conneau et al., 2019; Liu et al., 2019), fine-tuned on the balanced 2019 training set (25,723 unique tweets; mini-batch size 4, gradient accumulation 16, token length 512, Adam optimiser, initial learning rate 1 × 10⁻⁵, 25 epochs with a 100-step warm-up). It distinguishes two categories: waste-related — tweets focused on garbage / solid-waste issues (complaints or reports on poor waste management and SWM shortcomings, littering or illegal dumping, associated penalties or regulations, and related infrastructure); unrelated — tweets not about the solid-waste topic, including figurative or insulting uses of waste terms (e.g. basura). On the balanced hold-out test set the supervised model reached an F1 of 0.931 (macro/weighted balanced accuracy ≈ 0.953), and F1 ≥ 0.90 across all years, confirming reliable in-domain performance. The evaluation-loss curve plateaus (≈ 0.18) while the test F1 stays ≈ 0.93, indicating overconfidence on the out-of-domain 2020–2023 data rather than loss of accuracy — a known trait of encoder-based transformers (Desai & Durrett, 2020). 4.2 Zero-shot stage — mDeBERTa (five topical categories) To reduce residual noise and separate the primary interest (solid waste) from competing topics, the high-confidence waste subset from §4.1 was re-classified by a zero-shot natural-language-inference model, mDeBERTa-v3-base-mnli-xnli (He et al., 2023), into five categories, chosen following prior work on general topic classification (Mast et al., 2025): waste garbage debris desechos desperdicios — the solid-waste category; Spanish and English synonyms of litter are included in the label to aid the multilingual model; sports; politics; economics; other. For each tweet the model returns a confidence (softmax) score for each of the five candidate labels; the five scores sum to ≈ 1. In the released file the five labels are sorted in descending order of confidence, so Predicted Class 1 is the model's top label and Confidence Score 1 its probability. 4.3 Nature of the labels The values in this record are model outputs, not adjudicated ground truth.  5. Distribution of results Stage Criterion Count Share of all 2020–2023 samples Raw (post de-duplication / length filter) all retained tweets 51,982,707 100 % Supervised XLM-R P(waste) ≥ 90 % 6,843,795 13.9 % Zero-shot top label waste… ≥ 90 % double-filtered solid-waste set 70,896 0.14 % 6. Performance and quality assessment Because the released labels are predictions, their reliability is characterised by the paper's automatic and manual quality assessment (Manual QA on stratified samples): Method Precision (waste class) Recall (waste class) F1 (waste class) F1 micro F1 macro Supervised XLM-R (auto) 0.931 0.931 0.931 0.931 0.931 Supervised XLM-R (manual QA, population-weighted) 0.483 0.963 0.643 0.698 0.691 Supervised + zero-shot combined (manual QA weighted) 0.506 0.896 0.646 0.559 0.531 7. Files, access model, and schema To respect the Twitter/X Terms of Service and applicable data-protection and copyright law (see §10), this record uses the same tiered-access model as the companion 2019 record. The predicted labels — the authors' own model output — are open; the tweet text is provided on request under controlled access. Openly available classification_2020_2023_ids_labels.csv — one row per tweet (plus header), UTF-8, comma-separated. Columns: Column Type Description id integer (int64) Tweet identifier assigned by Twitter/X. The authoritative reference for each record; text may be re-retrieved (“hydrated”) via the platform API, subject to current availability. year integer Collection year of the tweet (2020–2023). Predicted Class 1 string Highest-confidence zero-shot label (one of the five categories in §4.2). Confidence Score 1 float Softmax probability for Predicted Class 1 (the maximum of the five). Predicted Class 2–5 string Remaining zero-shot labels, in descending confidence order. Confidence Score 2–5 float Softmax probabilities for classes 2–5; the five scores sum to ≈ 1. The tweet text is excluded by design from the open file; the id is the authoritative reference and can be used to re-retrieve text or to request the controlled-access file. Available on request (controlled access) classification_2020_2023_text_restricted.csv — the same rows with the tweet text added; columns id, text, year, Predicted Class 1…Confidence Score 5. Please write to research_NLP_ML_DL_CV[at]proton.me or the authors. Column Type Description text string Tweet text, lightly normalised to a single line (whitespace collapsed); articles, prepositions, and special characters are retained. Truncated as delivered by the Twitter/X API and typically ending in a shortened t.co link (now largely inactive). Not the complete verbatim post. This file is deposited under Zenodo's Restricted access; access is granted to researchers for non-commercial academic use who accept the terms in §10. 8. Recommended uses and limitations Recommended uses Confidence-based extraction of high-precision solid-waste tweet subsets for downstream analysis. Temporal and (via the companion metadata / geolocation work) spatial analysis of urban solid-waste discourse in Colombia. Seeding topic modelling and discourse analysis of waste narratives in Colombian Spanish. Benchmarking and error analysis of supervised vs zero-shot pipelines for low-resource, highly imbalanced environmental NLP. Limitations to be aware of Predicted, not gold, labels. All labels are model outputs; use the human-annotated companion 2019 record for training or for gold evaluation. Zero-shot over-prediction. At low confidence the waste… label has low precision; results are only reliable after confidence thresholding (§6, §9). Overconfidence out-of-domain. The supervised model's high scores on 2020–2023 data reflect overconfidence; calibrated interpretation (e.g. temperature scaling) is advisable for probabilistic use. Geographic and platform bias. Coverage is skewed toward large Colombian urban centres; rural and peri-urban areas, and users without internet access, are under-represented. Tweet availability. Re-hydrating text from id may no longer be possible for deleted or protected tweets, and the Twitter/X API is now access-restricted. Requesting the controlled-access text file avoids this. 9. Ethics and privacy The tweets originate from public posts collected under the Twitter/X Terms of Service. The open file contains only tweet id, year, and predicted topic labels with scores — no text, usernames, profile metadata, or user coordinates. The tweet text is held in a controlled-access file (§7) and released only to researchers who accept the terms below. Even so, tweet text may contain @-mentions or self-identifying content; users intending to publish derivatives should consider further anonymisation where appropriate. Because the labels are automated and imperfect, users should avoid drawing conclusions about identifiable individuals from single predicted labels. Individual posts can be removed on request to the maintainer. 10. Licence, terms of use and compliance What is licensed. The predicted labels and confidence scores and the accompanying documentation are the authors' contribution and are released under the licence stated in this record's Licence field [e.g. CC BY 4.0]. The underlying tweet content remains the property of its respective authors and of the platform, and is provided (under controlled access only) solely to support the labels. 11. Citation If you use this dataset, please cite both the dataset and the associated publication. Dataset XXXXX. ([year]). Spanish-language solid-waste tweet classification results (Colombia, 2020–2023) [Data set]. Zenodo. doi.org Associated publication XXXXX ([year]). Bridging NLP and Urban Waste Monitoring: Transformer-Based Detection and Mapping of Spanish-Language Solid Waste Mentions on Social Media. Online Social Networks and Media. [DOI when available] (Author list and order shown as in the manuscript; adjust for the double-blind review stage if the record must remain anonymous.) BibTeX @dataset{xxxxxxxxx_2020_2023_swtweets, author = {XXXXXXXXXXX}, title = {{Spanish-language solid-waste tweet classification results (Colombia, 2020--2023)}}, year = {[year]}, publisher = {Zenodo}, doi = {10.5281/zenodo.16098993}, url = {doi.org } References: Conneau, A., et al. (2019/2020). Unsupervised Cross-lingual Representation Learning at Scale (XLM-R). — He, P., Gao, J., & Chen, W. (2023). DeBERTaV3. — Desai, S., & Durrett, G. (2020). Calibration of pre-trained transformers. — Liu, Y., et al. (2019). RoBERTa. (Full references as cited in the associated publication.)