badashin/lince-sa-refined

LINCE SA Refined — Cultural-Context Relabeling for Spanish-English Code-Switching Sentiment Analysis

This repository releases 763 sentiment label refinements for the Spanish-English code-switching subset (sa_spaeng) of the LINCE benchmark (Aguilar et al., 2020). Refinements were produced by a trilingual annotator (Spanish / English / Korean) drawing on Hispanic-American social media conventions, and validated through controlled mBERT experiments.

This work received an Honorable Mention at the Korea Software Congress (KSC) 2025.

What this is

A reproducible label-mapping layered on top of LINCE SA. The original LINCE data is not redistributed here — only the refinements and a script that applies them.

• label_mapping.json — 763 entries in {sample_id: {original, corrected}} format
• build_refined_dataset.py — loads the original LINCE from Hugging Face and applies the refinements to reconstruct the refined set

What this is NOT

• A replacement for LINCE — the original benchmark remains the canonical source
• A drop-in dataset — running the reproduction script is required to materialize the data

How to use

# 1. Install dependencies
pip install datasets

# 2. Clone this repo
git clone https://huggingface.co/datasets/badashin/lince-sa-refined
cd lince-sa-refined

# 3. Run the reproduction script
python build_refined_dataset.py
# Output: refined_dataset.json  (5,567 samples)

The script downloads LINCE SA sa_spaeng via datasets.load_dataset, filters for genuine Spanish-English code-switching (both languages with ≥ 2 tokens each), applies the 763 refinements where the recorded original label still matches, and drops samples whose sample_id appears in more than one split.

What's in the mapping

763 sentiment label refinements with the following transition distribution:

Original	→	Refined	Count	%
positive	→	neutral	194	25.4%
neutral	→	positive	156	20.4%
neutral	→	negative	133	17.4%
positive	→	negative	114	14.9%
negative	→	neutral	89	11.7%
negative	→	positive	77	10.1%

The dominant pattern is bidirectional positive ↔ neutral refinement (45.8%), reflecting cultural-pragmatic context — humor, formulaic greetings, sarcasm, and Hispanic-American social media conventions — where surface signals can mislead annotators less familiar with these registers.

Methodology (brief)

• Sample-level review of LINCE SA sa_spaeng train + validation splits
• Selection criteria: genuine Spanish-English code-switching (≥ 2 tokens in each language)
• Annotation by a trilingual researcher (ES / EN / KO) using Hispanic-American social media norms
• Conservative labeling for ambiguous cases (preferring neutral over false positives)
• Final refined dataset: 5,567 code-switched samples after deduplication

Full methodology, illustrative examples, and downstream results are in the paper.

Downstream impact

Controlled mBERT experiments using identical hyperparameters across original and refined data show +4.0pp improvement (56.6% → 60.6%) attributable to label refinement alone. A subsequent mBERT–XLM-R Late Fusion ensemble on the refined data reaches 67.15%.

Refined label distribution

After deduplication, the 5,567-sample refined set has the following label distribution:

Label	Count	Share
positive	3,021	54.3%
neutral	1,519	27.3%
negative	1,027	18.4%

Citation

If you use this dataset, please cite both the original LINCE benchmark and this work:

@inproceedings{aguilar2020lince,
  title     = {LinCE: A Centralized Benchmark for Linguistic Code-switching Evaluation},
  author    = {Aguilar, Gustavo and Kar, Sudipta and Solorio, Thamar},
  booktitle = {Proceedings of the 12th Language Resources and Evaluation Conference},
  year      = {2020}
}

@inproceedings{shin2025relabeling,
  title     = {Re-labeling Approach for Spanish-English Code-switching Sentiment Analysis: Impact of Data Quality Improvement},
  author    = {Shin, Bada and Kim, Sunoh},
  booktitle = {Korea Software Congress (KSC)},
  year      = {2025},
  note      = {Honorable Mention}
}

Acknowledgments

This work builds on the LINCE benchmark (Aguilar et al., 2020), which provides invaluable infrastructure for code-switching research. The Spanish-English sentiment analysis subset (sa_spaeng) used here was originally curated by the LINCE team.

This research was supported by the SW-Centered University Project (2024-0-00035) of the Ministry of Science and ICT and the Institute of Information & Communications Technology Planning & Evaluation (IITP), Republic of Korea.

License

• Label mapping (label_mapping.json) and reproduction script (build_refined_dataset.py) — CC-BY-4.0
• The original LINCE data is not redistributed and remains subject to its own terms (see https://ritual.uh.edu/lince/)

Links

• 📄 Paper: https://github.com/vamosbada/project-puente/blob/main/docs/paper.pdf
• 💻 Full project (code, notebooks, experiments): https://github.com/vamosbada/project-puente
• 🏛️ Original LINCE benchmark: https://huggingface.co/datasets/lince-benchmark/lince