Download as PDF, PPTX
Uploaded byLifeng (Aaron) Han
Incorporating Chinese Radicals Into Neural Machine Translation: Deeper Than Character Level
The document presents a proposed model to apply Chinese radicals into neural machine translation. It discusses related work on machine translation and neural networks. The proposed model would combine radical-level machine translation with an attention-based neural model, incorporating radicals into the input data. Experiments would evaluate the model on various data settings and metrics, comparing performance with and without radicals. Future work could improve parameter tuning and include more diverse data.
Incorporating Chinese Radicals Into Neural Machine Translation: Deeper Than Character Level
- 1. Apply Chinese RadicalsInto Neural Machine Translation: Deeper Than Character Level AARON POET https://github.com/poethan DCU LPRC 2018, Limerick, Ireland, May 24
- 2. Agenda • Myself • Intro •Related work • Proposed idea/model • Experiments design • Evaluation results • Future work
- 3. Myself • PhD student,Dublin • Student Researcher, Amsterdam • Master degree of Sci., Macau • Bachelor of Maths, Shijiazhuang • Primary ~ high school, Handan • No kindergarten
- 4. Intro • Machine Translation •- what I m doing. Translate human language. • Natural Language Processing • - different processing of human languages • Artificial Intelligence • - teach machine to perform human intelligences
- 5.
- 6. Related work • MachineTranslation: Rule to Neural • - rule, example-based, statistical, phrase-based, hierarchical structure, tree-best, forest, neural • Neural MT, sequence to sequence, attention, coverage • - word embessing, sequence to sequence encoding-decoding, attention, coverage, document/discourse • Chinese NLP, radical applications • - Word Segmentation, Entity Rec., MT, Sentiment Analysis, text mining
- 7.
- 8.
- 9. Proposed Model • ApplyChinese Radical into Translation • - how to apply redical into MT • - how to split character into redical • Combine radical-level MT with Neural Model • - attention-based Neural MT • Redical combination into input data
- 10.
- 11. Experiments • Attention NeuralMT • Word+Character+Radical • Word+Character • Character+Radical • Word+Radical • Data prepration • Training: 1.25 million parallel Chinese-English sentences / 80.9 millions Chinese words and 86.4 millions English • Developing / testing: NIST06/NIST08 (National Institute for Standards and Technology, USA)
- 12.
- 13. Evaluation • Broader EvaluationMetrics • -hLEPOR, BEER, CharacTER -> BLEU, NIST • Evaluation Scores • - inside analysis LEPOR WIKI: https://en.wikipedia.org/wiki/LEPOR
- 14.
- 15.
- 16.
- 17.
- 18.
- 19.
- 20. Future work • Improveparameter optimisation/tuning models • Include more testing data • Include different domain data • This paper: https://arxiv.org/pdf/1805.01565.pdf
- 21. Follow me • Github:https://github.com/poethan • Poems: https://poethan.wordpress.com • Photography: https://poethan.wordpress.com/ photography/ • Cook: https://poethan.wordpress.com/activities/ • Instagram/smapchat: aaronlifeng