metadata
license: mit
task_categories:
- text-generation
- translation
language:
- en
- ar
- zh
- es
pretty_name: Multilingual Code-Switching Benchmark
tags:
- code-switching
- evaluation
- benchmark
- dialectal-nlp
Multilingual Code-Switching & Dialectal Evaluation Benchmark (Fatima Fellowship Application)
Question 1: Critical Blind Spot & Capability Gap
Standard NLP benchmarks (MMLU, GSM8K, HumanEval) evaluate language models on clean, standardized, monolingual inputs. However, for billions of global speakers, everyday digital communication occurs in low-resource code-switched vernaculars (e.g., Franco-Arabic/Arabizi, Singlish, Taglish, Hinglish, Naija Pidgin, Sheng, Spanglish, and Papiamento).
Current frontier small language models (0.6B–6B parameters) exhibit three systematic failure modes in these linguistic regimes:
- Semantic Hallucination & False Rules: When encountering regional idioms or transliterated vernaculars, models hallucinate fake explanations to justify literal translations (e.g., asserting that "pregnant" is a common Spanish slang term for "anxious").
- Subword Fragmentation & Transliteration Penalty: Tokenizers optimized for standard Western scripts fragment Romanized or creole dialects into excessive subwords, degrading downstream reasoning.
- Safety Guardrail & Tone Misalignment: Informal dialectal praise, regional exclamations, or colloquial slang are frequently misclassified or over-sanitized into generic boilerplate.
Question 2: Systematic Evaluation Narrative (google/gemma-2-2b-it)
Using a custom 10-item multilingual test suite covering 8 global dialect systems, we evaluated google/gemma-2-2b-it (2.6B parameters) in 4-bit precision.
Selected Empirical Results
| ID | Dialect / Language Pair | Prompt | Expected Intent | Model Output / Behavior | Failure Category |
|---|---|---|---|---|---|
| 1 | Arabizi + English | Ya zalameh, wallahi el mashroo3 dah 3al-alla, bas rabna hayastor. |
Project uncertainty / reliance on God | "I swear it's been a long time, but I'm back." | Hallucination: Completely missed project context; invented a return narrative. |
| 3 | Singlish | Don't keyi like that lah, the boss already kancheong, later he pok kai then how? |
Warning against reckless action causing bankruptcy/ruin. | Mapped kancheong to "cold shoulder" and pok kai to "scolding". |
Pragmatic Failure: Failed to recognize financial ruin vs scolding. |
| 5 | Taglish | Super nakakagigil yung deployment test... pero keribels lang kasi resolved na agad. |
Frustrating test, but manageable (keribels). |
Defined keribels as meaning "stressful". |
Polarity Inversion: Direct opposite meaning of slang term. |
| 9 | Spanglish | Estoy super embarazada por el error que cometí... |
Embarrassment over a meeting blunder. | Claimed "super pregnant" is Spanish slang for "extreme anxiety". | False Cognate Rationalization: Invented a false linguistic rule to defend literal error. |
Question 3: Path Forward & Proposed Solutions
To bridge the performance gap in sub-6B open-weights models without increasing parameters:
- Vocabulary Expansion & Byte-Level Subword Merging: Retrain base tokenizers on multi-dialectal social media corpora (X, Telegram, Reddit) to merge frequent Romanized and creole n-grams, reducing fragmentation penalties.
- Participatory Community Data Curation: Shift away from synthetic LLM-generated dialect corpora (which reproduce alignment biases). Recruit native speakers across dialect communities to curate authentic, pragmatically annotated conversational pairs.
- Direct Preference Optimization (DPO) on Pragmatic Alignment: Fine-tune models using LoRA adapters with preference pairs (y_win, y_lose), where y_lose represents literal translations or invented explanations, and y_win represents pragmatically accurate, context-aware responses.
Repository Artifacts
multilingual_evaluation_results.json: Raw evaluation logs and prompt outputs.evaluation_notebook.ipynb: Runnable Google Colab notebook used for benchmark execution.