--- pretty_name: Chilean Spanish Attitude Corpus language: - es language_details: es-CL license: other license_name: cc-by-4.0-annotations-only license_link: LICENSE task_categories: - token-classification annotations_creators: - expert-generated source_datasets: - original size_categories: - 1K- This corpus contains unmoderated Chilean Spanish text about political conflict, including hate speech and threats. The annotations and documentation are licensed under CC BY 4.0; the source texts are not. Access to source-text fields is provided only for non-commercial research. By requesting access you agree not to attempt re-identification, not to redistribute source texts outside your research group, and to honour takedown requests forwarded by the maintainers. extra_gated_fields: Name: text Affiliation: text Intended research use: text I will use source texts only for non-commercial research: checkbox I will not redistribute source texts outside my research group: checkbox I will not attempt re-identification: checkbox I accept the terms in LICENSE: checkbox configs: - config_name: gold default: true data_files: - split: train path: data/gold/train-*.parquet - split: validation path: data/gold/validation-*.parquet - split: test path: data/gold/test-*.parquet - config_name: annotations data_files: - split: full path: data/annotations/attitude-*.parquet --- # Chilean Spanish Attitude Corpus The Chilean Spanish Attitude Corpus contains 2,546 texts annotated for **Attitude** within Appraisal Theory in Systemic Functional Linguistics (SFL). It supports two span-level sequence-labeling tasks: **Attitude Types** (Affect, Judgment, Appreciation) and **Judgment Subtypes** (normality, capacity, tenacity, veracity, propriety). The default `gold` configuration provides consolidated, token-aligned labels and fixed train/validation/test splits. The `annotations` configuration provides privacy-redacted per-annotator character spans. Exact reproduction of the agreement results reported in the paper also requires the historic token-level BIO sidecar maintained with the companion analysis code. > **Content warning.** The corpus was collected around episodes of political conflict in Chile and contains insults, hate speech, threats and references to violence. Texts are reproduced as written, without moderation, because the annotation targets the evaluative language itself. ## Configurations | Config | Rows | What it is | |---|---|---| | `gold` (default) | 2,546 | Consolidated corpus: one row per text, three label layers, the splits used in the paper. | | `annotations` | 3,033 | One row per (text, annotator): the independent layers, for agreement and adjudication studies. | ```python from datasets import load_dataset gold = load_dataset("jorgeortizfuentes/chilean-spanish-attitude-corpus") layers = load_dataset("jorgeortizfuentes/chilean-spanish-attitude-corpus", "annotations", split="full") ``` Access is gated. Authenticate with `hf auth login` before loading the files. ## Tasks and label hierarchy The three arrays in `gold` represent successive levels of the same hierarchy: ```text Attitude ├── Affect ├── Appreciation └── Judgment ├── Social Esteem: normality, capacity, tenacity └── Social Sanction: veracity, propriety ``` Labels are stored as flat token classes without BIO prefixes. A predicted span is a maximal contiguous run of the same non-`O` label. The benchmark evaluates exact labeled spans: both the category and the two boundaries must match. ### `gold` | Split | Texts | |---|---| | train | 1782 | | validation | 382 | | test | 382 | Fields: `id`, `text`, `tokens`, `attitude_tags`, `types_jugdment_tags`, `subtypes_jugdment_tags`, `genre`, `domain`, `n_annotators`. The three label layers are token-aligned with `tokens`; the field name `types_jugdment_tags` carries a typo that belongs to the released schema and is kept so code written against earlier versions keeps working. See the [data dictionary](docs/data_dictionary.md). ### `annotations` Fields: `annotation_id`, `text_id`, `in_gold_corpus`, `campaign`, `annotator`, `text`, `tokens`, `spans`. Each span is `{label, start, end}` with character offsets into `text`. `text_id` joins to `gold.id`. It is null for the 187 rows whose text was annotated but did not enter the consolidated corpus. | Annotators per gold text | Texts | |---|---| | 1 | 2396 | | 3 | 150 | 150 gold texts carry three independent layers and support inter-annotator agreement; 0 gold texts have no row in this configuration. ### Agreement provenance The paper computes agreement over shared texts from all nine annotation subcampaigns, including pre-campaign texts that are not part of `gold`. In total, 210 texts have three independent layers: 150 are in `gold` and 60 are outside it. The privacy-redacted character spans in this configuration do not preserve every original token-level BIO decision. Exact reproduction therefore joins `annotations.annotation_id` to a privacy-safe sidecar maintained with the companion analysis code. That sidecar contains public annotation IDs, opaque shared-text IDs, BIO arrays and token hashes. It contains no text, tokens, character offsets, source annotation IDs or annotator names. It is not part of this dataset release. Five `Positive` or `Negative` spans from a legacy polarity pass are omitted from the public character-span inventory because polarity is not a benchmark task. Their original BIO labels remain in the sidecar used for general-level agreement. ## Composition | Genre | Texts | |---|---| | tweet | 2420 | | other | 93 | | letter_to_editor | 14 | | consumer_complaint | 13 | | opinion_column | 6 | ## Span labels in `annotations` | Label | Spans | |---|---| | `NG` | 18327 | | `Appreciation` | 3108 | | `propriety (J3)` | 1715 | | `capacity (J3)` | 1095 | | `Affect` | 605 | | `normality (J3)` | 515 | | `tenacity (J3)` | 466 | | `veracity (J3)` | 199 | | `Judgment (J1)` | 19 | | `Social Esteem (J2)` | 18 | | `Social Sanction (J2)` | 7 | `NG` is the retained structural label for a candidate nominal or adjectival group. Annotators first delimited these groups and then assigned Attitude labels. ## Pseudonymization Every Twitter handle was replaced by a stable `@userN` pseudonym. The same real account always receives the same pseudonym, across both configurations, and the pseudonym numbering is the one used throughout the project. **What pseudonymization does not remove.** The texts discuss Chilean public life and name public figures in running prose — politicians, officials, journalists. Those names are part of what is being evaluated and were left in place. This is a pseudonymized corpus, not an anonymous one. ## Data sources and annotation Texts were sampled from a 4.7-million-text Chilean Spanish collection covering 2010--2022. Sources include Twitter, news, opinion columns and a consumer complaint website. A classifier trained for offensive-language detection was used to enrich the sample for negative Judgment; this was a sampling device, not a released target. Three linguists trained in SFL annotated the data in two stages: structural group delimitation followed by semantic labeling. The saved Attitude stage is organized into 9 batches. The Spanish [annotation guidelines](docs/annotation_guidelines_es.pdf) are included in the repository. ## Intended use The corpus is intended for non-commercial research on SFL Appraisal, span-level sequence labeling, annotation agreement and Chilean Spanish NLP. The fixed splits support comparison with the accompanying paper. The intermediate Judgment Types layer may also be used to study hierarchy-aware models. It is not intended for identifying authors, profiling individuals, moderating users, making decisions about people, or training a general-purpose sentiment system. The corpus should not be used as a representative sample of Chilean public discourse. ## Licensing and access terms The annotations and documentation are licensed under CC BY 4.0. The source texts in `text` and `tokens` are not covered by that license; copyright remains with their authors and, where applicable, their publishers. Access to those fields is gated and provided only for non-commercial research under the conditions in [LICENSE](LICENSE). This notice does not grant rights in the source texts. Users must not attempt re-identification or redistribute source texts outside their research group and are responsible for determining whether their use is permitted. A takedown process is provided. ## Related resources - [Dataset repository](https://huggingface.co/datasets/jorgeortizfuentes/chilean-spanish-attitude-corpus) - [Selected supervised checkpoints and TULIO](https://huggingface.co/collections/jorgeortizfuentes/chilean-spanish-attitude-emnlp-2026-6a8ce4d92cabb99418a1204b) - [TULIO base model](https://huggingface.co/dccuchile/tulio-chilean-spanish-bert) ## Repository contents and reproducibility - [`docs/data_dictionary.md`](docs/data_dictionary.md): field types and labels. - [`docs/annotation_guidelines_es.pdf`](docs/annotation_guidelines_es.pdf): Spanish annotation manual. - [`docs/build_manifest.json`](docs/build_manifest.json): source paths and data-derived build counts. - [`SHA256SUMS`](SHA256SUMS): checksums for every released file. The release is rebuilt from the canonical Parquet and annotation pickles. It does not use the legacy TSV exports. ## Citation ```bibtex @inproceedings{ortizfuentes-attitude-2026, title = {Attitude Analysis in Systemic Functional Linguistics: A New Corpus and Benchmark Task}, author = {Ortiz Fuentes, Jorge Luis and Bravo-Marquez, Felipe and Quiroz, Beatriz}, booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026}, year = {2026}, note = {To appear} } ``` ## Contact and takedown Open a discussion on this repository for questions or takedown requests. Requests from an author to remove a text are honoured without requiring justification.