Author: chen_h
WeChat & QQ: 862251340
WeChat public account: coderpai
Starting today, I will do a short summary of some research papers, my personal focus is machine learning, reinforcement learning and natural language processing. I hope this short summary helps you. Of course, I also have other purposes. I hope my daily reading will help me improve my essay writing and analysis skills.
Today, let’s discuss the recent Facebook AI research (FAIR) paper on convolutional sequence learning. Here are the main points of reading this article:
major outcomes
For machine translation tasks, convolutional neural networks can also achieve better results.
Convolutional Neural Networks have a high degree of parallelism compared to Recurrent Neural Networks. Therefore, machine translation systems can do much faster. The authors improved the performance by a factor of 9.
Multi-hop mechanism: The network does not read sentence by sentence, but looks at multiple sentences together, so as to produce better translation.
Why choose CNN over RNN?
Speed and Scalability!
Experimental results
Achieved state-of-the-art results in WMT'16 English-Romanian translation, outperforming the previous best result by 1.8 BLEU.
On WMT'14 English-German translation, the results are even better than the powerful LSTM (Wu et al. (2016)).
In addition, the author says his translation speed has increased by a factor of 9.
code
Click here for the full code .
Source: Medium