CS224N: Natural Language Processing with Deep Learning
Lecture 6: Dependency Parsing
Lecture 7: Tensorflow Tutorial
Lecture 8: RNN
vanishing gradient:initialization + Relu
exploding gradient:clipping trick
beam search
bidirectional RNN:双向rnn
Lecture 9: Fancy Recurrent Neural Networks for Machine Translation
MT:machine translation
机器翻译原始模型:
改进:
- Encoder和Decoder使用不同的W,不再共享
- 解码过程,最后一个时刻的c,要传送给decoder的每个时刻的ht和yt;
同时,当前时刻的输出y
t
也要传给下一时刻的h
t+1
和y
t
+1
因此,encoder的h
t
取决于c、h
t-1、
y
t-1
- deep RNN with multiple layers
- bidirectional encoder
- 把输入序列反转
Lecture 10: Midterm recap
SGD:
Lecture 11: machine translation and models with attention
相比于传统机器翻译,NMT的优点:
Attention
Global VS Local:
Beam Search
– 只用在test阶段,train和valid不用;
– 用在seq2seq生成模型中 decode阶段,要使生成序列的概率最大,取每个时刻上概率最大不可行;
– 贪心的思想;
在每个时刻,对每个已有序列,对词典中所有的词计算概率,取概率最大的前k个序列;
在每个时刻,对每个已有序列,对词典中所有的词计算概率,取概率最大的前k个序列;
– 结果不一定是全局最优;
greedy search:
贪婪地选取当前最可能的那个单词
贪婪地选取当前最可能的那个单词
解决Large-vocab问题
(下一讲最后)
vocab太大,每次预测概率时求softmax太耗时
解决办法:
1.segment data:
2.Select candidate words
Lecture 12: Further topics in Neural Machine Translation and Recurrent Models
LSTM
Lecture 13: End-to-end models for Speech Processing
Lecture 14:Convolutional Neural Networks (for NLP)
Lecture 15:Tree Recursive Neural Networks and Constituency Parsing
treeRNN
Lecture 16:Coreference Resolution
指代消解
Lecture 17:Dynamic Neural Networks for Question Answering