Spanish-Language Solid-Waste Tweet Corpus (Colombia, 2019) — Manually Annotated

A two-class benchmark dataset of Spanish-language tweets manually annotated for the detection of urban solid-waste mentions. DOI (this record): 10.5281/zenodo.16040609 Version: v1 Release date: 2026-08-28 Language: Spanish (predominantly Colombian / Latin-American variety) Access model: the identifiers and labels are open; the tweet text is available on request under controlled access. Please write to research_NLP_ML_DL_CV[at]proton.me 1. Summary This record releases the manually annotated 2019 corpus used to fine-tune the transformer-based classifier described in the associated publication. Every item is an original Twitter/X post that trained annotators assigned to one of two classes: all_trash_related — the tweet refers to solid waste (garbage, littering, illegal dumping, or the shortcomings of solid-waste management), typically in an urban context; unrelated — the tweet is not about solid waste, even when a waste term appears in it (for example, when basura, “garbage”, is used figuratively or as an insult). The corpus was created to address a documented gap: at the time of writing, no publicly available, domain-specific, Spanish-language benchmark existed for detecting solid-waste mentions on social media. Releasing both classes with their labels makes the corpus directly usable for training and for reproducible evaluation of classifiers, rather than only for lexical inspection. 2. Provenance and collection Platform: Twitter/X Collection method: Streaming API v1.1 Spatial filter (bounding box): Longitude 79.83°W – 66.02°W, Latitude 4.68°S – 13.41°N (extent of Colombia) Temporal coverage: 2019 (streaming from week 5 to week 52) Raw sample (2019): 1,995,625 tweets Dominant language: Spanish (with a minority in Portuguese and English) The only sampling criteria applied at collection time were geographic location and time frame; no topical keyword filter was applied at the streaming stage. This deliberately unfiltered strategy means the raw pool reflects the natural (very low) prevalence of solid-waste discourse on the platform, rather than an artificially enriched sample. 3. From raw stream to annotated corpus Duplication 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 noise. Keyword pre-filtering. After the first annotation round, a TF-IDF was applied to the tweets labelled as all_trash_related and relevant keywords were extracted from this analysis. The keywords were used to filter a subsequent dataset that would be annotated. The same procedure was applied again, totalling 3 annotation rounds. Besides keywords extracted from the TF-IDF on tweets, other relevant keywords were manually extracted from newspaper articles from Colombia on the topic of solid waste. Manual annotation. Candidate tweets were labelled as all_trash_related or unrelated (see Section 4). For model training (reported in the paper, not included in this file), the positive class was subsequently expanded fourfold by data augmentation and combined with negatives into a balanced fine-tuning set. See Sections 5 and 6. 4. Annotation guidelines 4.1 Annotators and procedure Annotation was performed by two native Latin-American Spanish speakers. Before production labelling, the annotators completed a calibration phase: they independently labelled a common sample of tweets and compared results to establish a shared, operational understanding of the categories. Labelling was carried out over three successive annotation rounds. Splitting the work into rounds allowed the guidelines to be refined and difficult cases to be re-examined between rounds. The number of confirmed positive (all_trash_related) tweets accumulated across the three rounds to a final total of 2,573. The counts reported below and in the associated publication refer to the final, three-round total. 4.2 Label definitions all_trash_related (positive) — assign if the tweet focuses on garbage / solid-waste issues. This includes: complaints or reports about poor waste management or solid-waste-management (SWM) shortcomings; mentions of littering or illegal / open-air dumping in urban space; references to penalties, sanctions, or regulations associated with waste; calls to community action (e.g. organised clean-ups or collection events); indirect references to solid waste through closely related infrastructure (e.g. waste containers invading public space); posts by authorities or institutions reporting on waste-management actions or achievements. An ideal positive tweet contains three elements — object (the waste), place, and time — for example: “For several days, Avenue 55 in Medellín has been littered with garbage, with no intervention from local authorities.” Accompanying images are advantageous but were not required for a positive label. unrelated (negative) — assign if the tweet is not about the garbage / solid-waste topic, even when a waste word appears in it. In particular, assign this label when a waste term (e.g. basura) is used figuratively or as an insult. 4.3 Disagreement patterns and hard cases Because waste vocabulary is heavily overloaded in everyday Spanish, several categories of tweet were systematically difficult and were the main sources of annotator disagreement and false positives. Documenting them here is intended to help future annotators and to characterise the decision boundary: Ambiguous or context-poor tweets, where relevance to solid waste cannot be established. For example, “some people’s junk is other people’s treasure …” or “… one more season thrown in the trash …” Figurative uses labelled unrelated. Adjacent-but-excluded topics, such as recycling (“… donating two recycling containers for plastic bottles …”) or pollution / e-waste risk (“electronic cigarettes and tobacco … are toxic and contaminate the Earth …”). These resemble solid-waste content but did not meet the strict inclusion criteria and were labelled unrelated. Waste terms used as insults, e.g. “human garbage” (basura humana) or “do I need more reasons to say that the State is a piece of rubbish?” — labelled unrelated. 4.4 Inter-annotator agreement Agreement over the calibration sample was measured with Fleiss’ Kappa κ = 0.712, indicating substantial agreement (Landis & Koch, 1977). This value is comparable to the downstream model’s own performance ceiling, suggesting that annotation reached close to the practical upper bound of reliable human agreement for this task. Reference: Landis, J. R., & Koch, G. G. (1977). The measurement of observer agreement for categorical data. Biometrics, 33(1), 159–174. 5. Class distribution all_trash_related — solid-waste-related (positive): 2,573 (16.7 %) unrelated — not solid-waste-related (negative): 12,856 (83.3 %) Total annotated: 15,429 The positive class is the minority — roughly a 1 : 5 positive-to-negative ratio — which realistically reflects the scarcity of solid-waste discourse on the platform, consistent with the paper’s finding that high-confidence solid-waste tweets constitute only about 0.14 % of the country-scale corpus. Users should account for this imbalance in training and evaluation (e.g. stratified splits, class weighting, or resampling). 6. File contents and schema To respect the Twitter/X Terms of Service and applicable data-protection and copyright law (see §7, §8), this record uses a tiered-access model. The annotations — the authors’ own contribution — are open; the tweet text is provided on request under controlled access. Openly available corpus_annotated_2019_repository_ids_only.csv — 15,429 rows (plus header), UTF-8, comma-separated, all fields quoted. 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. label string all_trash_related or unrelated. Available on request (controlled access) corpus_2019_text_restricted.csv — the same 15,429 rows with the tweet text added; columns id, text, label. 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 https://t.co/… link (now largely inactive). Not the complete verbatim post. This file has Restricted access. Access is granted to researchers for non-commercial academic use who accept the terms of the license. Please write to research_NLP_ML_DL_CV[at]proton.me or the authors. 7. Recommended uses and limitations Recommended uses: Fine-tuning and benchmarking of Spanish-language text classifiers for solid-waste / sanitation detection. Transfer-learning and domain-adaptation studies for low-resource environmental NLP. Lexical, topic, and discourse analysis of urban solid-waste narratives in Colombian Spanish. Limitations to be aware of: 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. Temporal scope. Collection is limited to 2019. Candidate pool. Both classes were drawn from a keyword-pre-filtered pool, so the unrelated examples are “hard negatives” lexically close to the positive class rather than a random sample of all Twitter/X traffic. This makes the benchmark demanding but not representative of the platform at large. Ambiguity ceiling. The decision boundary is intrinsically fuzzy (κ = 0.712); some residual label noise is expected. Text form of a few negatives. For the vast majority of tweets the text field is the lightly-processed, single-line form. Tweet availability. Because only IDs and truncated text