Strategy for effective training of Korean-English idioms in NMT

• Korean Title : 한-영 관용구 기계번역을 위한 NMT 학습 방법
  (HCLT 2020 (Annual Conference on Human and Cognitive Language Technology))

• Download Paper (Korean)

  HCLT2020-한-영 관용구 기계번역을 위한 NMT 학습 방법_최민주.pdf
  

• HCLT 2020 Slide PPT (Korean)

  (구두발표)HCLT2020-한-영 관용구 기계번역을 위한 NMT 학습 방법_최민주.pdf
  

1. Abstract

• Training method using special token(tag) and KISS dataset
• After research of 'A Dataset and Evaluating Method for Korean-English Idiom Machine Translation’

2. Related Research

A Dataset and Evaluating Method for Korean-English Idiom Machine Translation

3. Dataset for experiment

• Special token(tag) to each 4 dataset
  • 1. Non-idiom sentence dataset
  • 2. Idiom sentence dataset
  • 3. idiom sentence dataset
    • Attach ‘’ tag in front of idiom word
  • 4. idiom sentence dataset
    • Attach ‘’ tag in front of idiom word and ‘’ tag back of idiom word

4. Experiment (Training)

OpenNMT-py, model hyper-parameters are:

• 2 Layer LSTM
• 4 Layer LSTM
• Transformer

Best : type 3 (‘’ tag & 4 layer transformer)

5. Conclusion

• In normal, idiom sentence BLEU score is lower than non-idiom sentence
• For efficiently train Korean idiom to NMT, attach special token only in front of idiom.

Reference Project 🌳🦜

https://github.com/itisused/2021_NLP_Project