KcELECTRA-base-v2022
| Property | Value |
|---|---|
| Model Size | 475M |
| License | MIT |
| Language Support | Korean, English |
| Author | beomi |
What is KcELECTRA-base-v2022?
KcELECTRA-base-v2022 is a specialized ELECTRA model designed specifically for Korean user-generated content and noisy text analysis. Unlike traditional Korean language models trained on formal texts, this model excels at processing informal language, including social media comments, colloquialisms, and internet vernacular.
Implementation Details
The model was trained on approximately 17GB of data collected from Korean news comments and replies between 2019-2021, comprising over 180 million sentences. It implements the ELECTRA architecture with significant improvements over its predecessor KcBERT, showing approximately 1%p performance enhancement across most downstream tasks.
- Specialized tokenizer trained with BertWordPieceTokenizer (vocab size: 30,000)
- Trained on TPU v3-8 for 848k steps
- Preserves emojis and special characters in preprocessing
Core Capabilities
- NSMC Classification: 91.97% accuracy
- Naver NER: 87.35% F1 score
- KorSTS: 83.67 Spearman correlation
- Question Pair Analysis: 95.12% accuracy
Frequently Asked Questions
Q: What makes this model unique?
The model's unique strength lies in its specialized training on user-generated content, making it particularly effective for analyzing informal Korean text, social media content, and comments with non-standard language patterns.
Q: What are the recommended use cases?
KcELECTRA-base-v2022 is ideal for sentiment analysis, named entity recognition, and text classification tasks involving informal Korean text, particularly user-generated content like comments, reviews, and social media posts.
