LINCE SA — 763 Refined Sentiment Labels Available
Hi @gaguilar ,
I conducted a systematic quality analysis of the LINCE SA (sa_spaeng) sentiment labels and found a ~17% labeling error
rate, primarily Positive-Neutral confusion stemming from missing Hispanic-American cultural context.
Leveraging my fluency in Spanish, English, and Korean, I manually reviewed and corrected 763 labels across 5,567
samples. With no model changes, mBERT accuracy improved from 56.6% to 60.6% (+4.0pp).
This work received an Honorable Mention at KSC 2025 (Korea Software Congress).
The corrections are released as a reproducible label-mapping file (no redistribution of the original data):
🔗 GitHub: https://github.com/vamosbada/project-puente
📄 Label mapping (763 entries, {sample_id: {original, corrected}}):
https://github.com/vamosbada/project-puente/blob/main/data/label_mapping.json
📄 Paper: https://github.com/vamosbada/project-puente/blob/main/docs/paper.pdf
A reproduction script (build_refined_dataset.py) rebuilds the refined split from the original LINCE source.
Happy to contribute these corrections back to the dataset in whatever format would be most useful. Thank you for
creating such a valuable benchmark for code-switching research.
Best,
Bada Shin