Jeju Translation

? Jeju language, standard language two -way voice translation model creation project

Model use

 import torch
from transformers import AutoTokenizer , AutoModelForSeq2SeqLM
  
## Set up the device (GPU or CPU)
device = torch . device ( "cuda" if torch . cuda . is_available () else "cpu" )

## Load the tokenizer and model
tokenizer = AutoTokenizer . from_pretrained ( "Junhoee/Kobart-Jeju-translation" )
model = AutoModelForSeq2SeqLM . from_pretrained ( "Junhoee/Kobart-Jeju-translation" ). to ( device )

## Set up the input text
## 문장 입력 전에 방향에 맞게 [제주] or [표준] 토큰을 입력 후 문장 입력
input_text = "[표준] 안녕하세요"

## Tokenize the input text
input_ids = tokenizer ( input_text , return_tensors = "pt" , padding = True , truncation = True ). input_ids . to ( device )

## Generate the translation
outputs = model . generate ( input_ids , max_length = 64 )

## Decode and print the output
decoded_output = tokenizer . decode ( outputs [ 0 ], skip_special_tokens = True )
print ( "Model Output:" , decoded_output )

 Model Output : 안녕하수꽈 

? My role

• Dataset creation
  • Jeju-Standard Dataset Collection and Preparation for New Dataset (Dataset)
    • Data collection such as AI-Hub, Github, etc.
• Translation model logic design
  • Fine tuning as a KOBART model
  • Among the text2Text models in Korean in Korean, Kobart is the best and fastest model.
  • In the process of designing two -way logic , entering [Jeju] and [Standard] tokens in front of the sentence to make the model easy to understand (BLEU Score 0.5-> 0.7, up to 1 standard)
  • Due to the lack of RAM, only 700,000 data were learned , but the dataset format method was changed from Float16 to Unit16 to solve memory shortages (GPU memory, resource saving)

1. Introduction of project

?‍?‍? Team member

• Vitamin 12: Leader, Lee Seo -hyun, Lee Yerin
• Vitamin 13: Kim Yun -young, Kim Jae -gyeom, Lee Hyung -seok

? period

• 2024 first semester

? ️ theme

• Create Jeju dialect and standard language bidirectional translation model

target

• We would like to promote the understanding of Jeju dialects and contribute to the preservation of Jeju culture.
• We promote smooth communication with citizens in Jeju.
• We develop a two -way translation model that connects the Jeju dialect and the Korean standard language.
• Implementing voice recognition and user interface.

2. Data collection

• Data collected by AI-HUB

  • Korean dialect ignition data
  • Korean and older Korean dialect data

• Data collected by Github

  • Kakao JIT Jeju Tongue Data

• Other data

  • Living Province Data (Jeju Preliminary web page crawling)
  • Well, Lang Harman Data (YouTuber Collection of Data by referring to the lyrics translation video among the Langhaman videos)
  • Jeju dialect that taste and stylish data (data collected from the book 'Jeju Tongue Taste and Prize')
  • Data even if it passes off, even if it passes, it collects data from the book 'Even if it is gone')
  • 2018 Jeju Language Oral Materials Collection (Collected for Evaluation)

3. Model learning

3-1. Model related

• I learned in a way to bring in the pre-learning model and Fine-Tuning .

• Pre -learning model used to develop translation models:

  • GOGAMZA/KOBART-Base-V2

• Pre -learning model selection criteria

  • Is it the right model for translation?
  • Is it learned in Korean?
  • Is the model capacity so big and the learning speed is fast?

• Models that have been considered but not selected:

  • T5 (There is a problem with too long learning time)
  • Jebert (performance was not satisfactory)

3-2. Learning method

• Learning methodology

  • Source-> Learning in Target format
  • Prior to entering the sentence, adding [Jeju] or [Standard] tokens to specify the direction of translation and learning together
  • Using the Dataset of the DataSets package, converting it to an optimized form for language model learning

• Main parameter settings

  • MAX_LENGTH: 64
  • batch_size: 32
  • Transing_rate: At first, starting from 2E-5 and learning progresses gradually reduced
  • EPOCHS: 3

? 4. Main achievements

• Final BLEU Score -Jeju Language Oral Data Book Data Standards
  • Jeju language-> standard language: 0.76
  • Standard language-> Jeju language: 0.5
• BLEU Score Performance Table

date	04-13	05-03	05-06	05-13	05-21	05-24	05-26	05-30
Jeju language-> standard language bleu score	0.56	0.59	0.42	0.64	0.70	0.74	0.76	0.74
Standard language-> Jeju Bleu Score	0.35	0.37	0.26	0.37	0.39	0.46	0.50	0.49

• Overall, we recorded the BLEU Score .

• Interface implementation

• Voice recognition function

  • Stt
    • Receive Whisper models from Hugging Face and proceed with Fine-Tuning
    • Jeju language conversion to text and convert to text
  • TTS
    • Receive the Glos TTS, HIFIGAN model from Hugging Face and proceed with Fine-Tuning
    • I tried to express voice in Jeju, but failed ...
    • Expression instead of standard language voice (using GTTS)

? 5. Future plans

• Preliminary processing through additional data collection and grammatic micro -adjustment to secure quality data
• Improvement of the ability to recognize the accent of the voice recognition model
• Web implementation and mobile app development plan

? 6. Reference

• Data source
  • Korean dialect ignition data (provided by AI-HUB): https://www.aihub.or.kr/aihubdata/data/View.do?curmenu=115&topmenu
  • Middle and older Korean dialect data (AI-HUB): https://www.aihub.or.kr/aihubdata/data/VIEW.DO?Curmenu=115&topmenu
  • Kakao JIT Jeju Tongue Data (see Kakaobrane GitHub): https://github.com/kakaobrain/jejuo
  • Living Living Side Data (see Jeju Language Preliminary): https://www.jeju.go.kr/culture/dialect/lifedialect.htm
• Model source
  • Kobart Hugging Face: https://huggingFace.co/gogamza/kobart-base-v2
  • Whisper Hugging Face: https://huggingFace.co/openai/whisper-large-v2
  • Kobart github: https://github.com/skt-ai/kobart