kobart2ksl-translation

KoBART-based Korean to KSL Translation Model

This is a Seq2Seq translation model based on KoBART, fine-tuned to convert Korean sentences into grammatically aligned Korean Sign Language (KSL) expressions.

---

Model Details

• Base model: gogamza/kobart-base-v2
• Architecture: Encoder-decoder (BART)
• Task: Korean → Korean Sign Language (KSL) grammatical transformation
• Tokenizer: KoBARTTokenizer with special tokens <s>, </s>, <pad>

---

Training Data

The model was trained on the Korean-KSL parallel corpus provided by the National Institute of Korean Language (국립국어원).
This dataset contains Korean sentences and their grammatical KSL counterparts, prepared for machine translation.

📢 Additional sentence-level variations were generated using GPT-based augmentation to improve generalization.
The final training dataset consists of approximately 35,000 sentence pairs.

---

Training Objective

This model aims to transform standard Korean sentences into KSL-style grammatical structures, including:

• Subject-Object-Verb (SOV) word order
• Omission of particles and formal endings
• Simplified expressions suitable for signing

---

Example Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("your-username/kobart2ksl_translation")
model = AutoModelForSeq2SeqLM.from_pretrained("your-username/kobart2ksl_translation")

input_text = "오늘 날씨 어때?"
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=64)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))