Character-Level Question Answering with Attention读书笔记

针对simple question

char-level LSTM得到句子表征序列

char-level CNN建模entity和predicate

利用LSTM decoder生成entity和predicate:

    每次输入context向量和input向量,t=0是input为0,t=1 train时 input为true entity embedding,test时input为上一步生成的entity embedding

    context为hidden和question表征的attention加权平均

    输出h与entity/predicate计算余弦相似度,归一化作为预测概率

 

entity candidate生成:entity是question的substring且不是另一entity的子序列

Character-Level Question Answering with Attention读书笔记