\parinterval 图\ref{fig:6-5}是一个应用编码器-解码器结构来解决机器翻译问题的简单实例。给定一个中文句子``我\ 对\ 你\ 感到\ 满意'',编码器会将这句话编码成一个实数向量(0.2,-1,6,5,0.7,-2),这个向量就是源语言句子的``表示''结果。虽然有些不可思议,但是神经机器翻译模型把这个向量等同于输入序列。向量中的数字并没有实际的意义,然而解码器却能从中提取到源语句子中所包含的信息。也有研究者把向量的每一个维度看作是一个``特征'',这样源语言句子就被表示成多个``特征''的联合,而且这些特征可以被自动学习。有了这样的源语言句子的``表示'',解码器可以把这个实数向量作为输入,然后逐词生成目标语句子``I am satisfied with you''。
\parinterval 公式\ref{eq:6-30}展示了最基本的优化策略,也被称为标准的SGD优化器。实际上,训练神经机器翻译模型时,还有非常多的优化器可以选择,在第五章也有详细介绍,这里考虑Adam优化器。 Adam 通过对梯度的{\small\bfnew{一阶矩估计}}\index{一阶矩估计}(First Moment Estimation)\index{First Moment Estimation}和{\small\bfnew{二阶矩估计}}\index{二阶矩估计}(Second Moment Estimation)\index{Second Moment Estimation}进行综合考虑,计算出更新步长。
\parbox{36em}{During Soviet times, if a city’s population topped one million, it would become eligible for its own metro. Planners wanted to brighten the lives of everyday Soviet citizens, and saw the metros, with their tens of thousands of daily passengers, as a singular opportunity to do so. In 1977, Tashkent, the capital of Uzbekistan, became the seventh Soviet city to have a metro built. Grand themes celebrating the history of Uzbekistan and the Soviet Union were brought to life, as art was commissioned and designers set to work. The stations reflected different themes, some with domed ceilings and painted tiles reminiscent of Uzbekistan’s Silk Road mosques, while others ...}
\end{tabular}
};
%译文1--------------mt1
\node[font=\small] (mt1) at ([xshift=0em,yshift=-9.1em]original0.south) {译文1:};
\node[font=\small] (ts1) at ([xshift=0em,yshift=-4em]original1.south) {