Ti-SIMS: Public Text and Annotation Release

Dataset summary

This public release contains 2,952 independently authored Tibetan-Chinese dialogue-text records with three-class sentiment annotations. The records are grouped into 1,152 dialogues. The fixed split contains 2,362 training samples, 295 validation samples, and 295 test samples.

The original research corpus was constructed for image-text sentiment analysis. Its television-frame images are not included in this repository because they are copyright-restricted third-party material. The image_id field is retained only as an opaque alignment identifier; it does not provide or link to an image.

Files

splits/train.csv
splits/validation.csv
splits/test.csv
splits/split_summary.json
release_manifest.json

Fields

• dialog_id: anonymized dialogue identifier.
• turn_index: zero-based turn index within a dialogue.
• image_id: opaque alignment identifier; no image is distributed.
• chinese: independently authored Chinese text.
• Tibetan: independently authored Tibetan text.
• sentiment: negative, neutral, or positive.

Split policy

Splitting is performed by dialogue rather than by individual utterance, so a dialogue cannot overlap between training, validation, and test sets. The split seed is 42. split_summary.json records the label counts and dataset fingerprint used by the TCAF experiments.

Copyright and access scope

Only the independently authored text, sentiment annotations, identifiers, and split metadata are distributed under CC BY-NC 4.0. No television frame, episode, subtitle, video, image path, or third-party media file is distributed or licensed by this repository.

Consequently, this public release supports verification of the textual data, labels, and split protocol, but it does not by itself enable exact reproduction of the image-based portions of the original multimodal experiments.

Intended use

The release is intended for non-commercial academic research on Tibetan text sentiment analysis, low-resource NLP, data splitting, and annotation analysis. Users should report results on the provided fixed splits and clearly state that the public release excludes the copyright-restricted visual modality.

Limitations

The dataset is small and domain-specific. It may contain linguistic, translation, cultural, and class-distribution biases. Results should not be interpreted as representative of all Tibetan speakers, domains, or dialects.

Citation

Please cite the TCAF paper and the accompanying code repository:

https://github.com/ChengGit001/TCAF