Neural machine translation by jointly learning to align and translate
Download as PPTX, PDF0 likes486 views
This paper proposes a neural machine translation model that extends previous work by incorporating an attention mechanism, allowing the decoder to soft-search relevant parts of the input without explicitly segmenting it. The bidirectional RNN architecture with attention outperforms earlier encoder-decoder models on English-to-French translation tasks according to BLEU scores, handling variable length inputs more effectively. While promising, the model could be improved to better handle rare or unknown words.
1 of 8
Download to read offline
More Related Content
What's hot (20)
Ad
Similar to Neural machine translation by jointly learning to align and translate (20)
Ad
Recently uploaded (20)
Neural machine translation by jointly learning to align and translate
- 1. Neural Machine Translation By Jointly Learning To Align And Translation Published as a conference paper as ICLR 2015
- 2. What is this paper? In neural machine translation, it proposed the model that extended function which automatically (soft-)search for parts of an input sentence that are relevant to predicting a target word, without having to form these parts as a hard segment explicitly.
- 3. What’s better than previous paper? This model surpassed previous model significantly, regardless of the sentence length and is much stronger to the length of an input sentence. (Previous model’s performance becomes worse as increase the length of input sentence.)
- 4. What’re important parts of technics and method? Unlike previous, the new architecture consists of a bidirectional RNN. (Bidirectional means two opposite directions) By letting the decoder have an attention mechanism, this does not need to encode all information in the input sentence into a fixed-length vector.
- 5. How did they verify it? On the task of English-to-French, they compared RNN search (this model) with RNN Encoder- Decoder (previous model) in BLUE score using ACL WMT ‘14. (About BLUE score… https://to-in.com/blog/102282)
- 6. Is there a discussion? Considering that the proposed architecture, or the whole family of neural machine translation, has only been proposed as recently as this year. One if challenges left for the future is to better handle unknown, or rare words.
- 7. Is there a paper to read next? Sequence to sequence RNN Encoder-Decoder