摘要
Neural machine translation : 神經(jīng)機(jī)器翻譯。
神經(jīng)機(jī)器翻譯模型經(jīng)常包含編碼器和解碼器:an encoder and a decoder.
編碼器: 從一個變長輸入序列中提取固定長度的表示。a fixed-length representation.
解碼器:從表示中生成一個正確的翻譯。generates a correct translation
本文使用模型:?RNN Encoder–Decoder、
? ? ?a newly proposed gated recursive convolutional neural network (門遞歸卷積神經(jīng)網(wǎng)絡(luò))
?a grammatical structure 語法結(jié)構(gòu)
介紹
問題:
it is crucial to understand the properties and behavior of this new neural machine translation approach in order to determine future research directions. Also, understanding the weaknesses and strengths of neural machine translation might lead to better ways of integrating SMT and neural machine translation systems
變長序列神經(jīng)網(wǎng)絡(luò):
循環(huán)神經(jīng)網(wǎng)絡(luò)、門遞歸卷積神經(jīng)網(wǎng)咯
RNN更新:
RNN序列可以有效的學(xué)習(xí)一個概率分布:
Gated Recursive Convolutional Neural Network?
?
Purely Neural Machine Translation
編碼器和解碼器方法
實(shí)驗
數(shù)據(jù)集:?English-to-French translation.
模型:
兩個模型的共同點(diǎn):an RNN with gated hidden units as a decoder
優(yōu)化算法:minibatch stochastic gradient descent with AdaDelta
使用beam-search去發(fā)現(xiàn)最大化條件概率分布。
結(jié)果和分析
*?the BLEU score 作為評價指標(biāo)。
the fixed-length vector representation does not have enough capacity to encode a long sentence with complicated structure and meaning
固定長度的向量表示并沒有足夠的能力去編碼一個包含復(fù)雜語義信息的長句子。文章來源:http://www.zghlxwxcb.cn/news/detail-533255.html
感悟
需要徹底理解以下條件概率分布以及RNN網(wǎng)絡(luò)的數(shù)學(xué)推導(dǎo)。研讀論文需要注意以下其算法原理。以及評價指標(biāo)。文章來源地址http://www.zghlxwxcb.cn/news/detail-533255.html
到了這里,關(guān)于On the Properties of Neural Machine Translation: Encoder–DecoderApproaches的文章就介紹完了。如果您還想了解更多內(nèi)容,請在右上角搜索TOY模板網(wǎng)以前的文章或繼續(xù)瀏覽下面的相關(guān)文章,希望大家以后多多支持TOY模板網(wǎng)!