Skip to content

T
Toy-MT-Introduction
Commit 99f5a117 by xiaotong
Options

parsing = string-based decoding

parent c36bc0ad
正在显示 包含 190 行增加 和 155 行删除
Section04-Phrasal-and-Syntactic-Models/section04-test.tex
......@@ -152,201 +152,118 @@
\subsection{规则匹配}
%%%------------------------------------------------------------------------------------------------------------
%%% 基于树的解码方法 - chart-based decoding
\begin{frame}{基于树的解码 - 基于chart的方法}
%%% 基于串的解码方法
\begin{frame}{基于串的解码}
\begin{itemize}
\item 基于chart这种结构，可以很容易的构建解码所用的超图。常用的方法是自底向上解码：
\item 不同于基于树的解码，\alert{基于串的解码}方法并不要求输入句法树，它直接对输入词串进行翻译，最终得到译文。
\begin{itemize}
\item 从源语言句法树的叶子节点开始，自下而上访问树的节点
\item 对于每个跨度，如果对应一个树节点，则匹配相应的规则
\item 从树的根节点可以得到翻译推导，最终选择最优推导所对应的译文输出
\item 这种方法适用于树到串、串到树、树到树等多种模型
\item 本质上，由于并不受固定输入的句法树约束，基于串的解码可以探索更多潜在的树结构，这也增大了搜索空间(相比基于串的解码)，因此该方法更有可能找到高质量翻译结果
\end{itemize}
\item<2-> 在基于串的方法中，句法结构被看做是翻译的隐含变量，而非线性的输入和输出。比如，层次短语翻译解码就是一种典型的基于串的解码方法，所有的翻译推导在翻译过程里动态生成，但是并不要输入或者输出这些推导所对应的层次结构
\end{itemize}
\visible<2->{
\begin{minipage}[b]{0.42\linewidth}
\begin{center}
\begin{tikzpicture}
\begin{scope}
\tikzstyle{chartnode}=[rectangle,minimum size=1.3em,draw]
\node [chartnode,anchor=north west] (cell11) at (0,0) {};
\node [chartnode,anchor=north west] (cell21) at ([yshift=-0em]cell11.south west) {};
\node [chartnode,anchor=west] (cell22) at (cell21.east) {};
\node [chartnode,anchor=north west] (cell31) at ([yshift=-0em]cell21.south west) {};
\node [chartnode,anchor=west] (cell32) at (cell31.east) {};
\node [chartnode,anchor=west] (cell33) at (cell32.east) {};
\node [chartnode,anchor=north west] (cell41) at ([yshift=-0em]cell31.south west) {};
\node [chartnode,anchor=west] (cell42) at (cell41.east) {};
\node [chartnode,anchor=west] (cell43) at (cell42.east) {};
\node [chartnode,anchor=west] (cell44) at (cell43.east) {};
\tikzstyle{chartnode2}=[rectangle,minimum size=1.3em,fill=orange!20]
\node<3-> [chartnode2,anchor=north west] (cell11) at (0,0) {};
\node<4-> [chartnode2,anchor=north west] (cell21) at ([yshift=-0em]cell11.south west) {};
\node<7-> [chartnode2,anchor=west] (cell22) at (cell21.east) {};
\node<5-> [chartnode2,anchor=north west] (cell31) at ([yshift=-0em]cell21.south west) {};
\node<8-> [chartnode2,anchor=west] (cell32) at (cell31.east) {};
\node<10-> [chartnode2,anchor=west] (cell33) at (cell32.east) {};
\node<6-> [chartnode2,anchor=north west] (cell41) at ([yshift=-0em]cell31.south west) {};
\node<9-> [chartnode2,anchor=west] (cell42) at (cell41.east) {};
\node<11-> [chartnode2,anchor=west] (cell43) at (cell42.east) {};
\node<12-> [chartnode2,anchor=west] (cell44) at (cell43.east) {};
\node [anchor=east] (s1) at (cell11.west) {\scriptsize{猫}};
\node [anchor=east] (s2) at (cell21.west) {\scriptsize{喜欢}};
\node [anchor=east] (s3) at (cell31.west) {\scriptsize{吃}};
\node [anchor=east] (s4) at (cell41.west) {\scriptsize{鱼}};
\node [anchor=north] (t5) at (cell41.south) {\tiny{$l$=1}};
\node [anchor=north] (t5) at (cell42.south) {\tiny{$l$=2}};
\node [anchor=north] (t5) at (cell43.south) {\tiny{$l$=3}};
\node [anchor=north] (t5) at (cell44.south) {\tiny{$l$=4}};
\node [anchor=north] (chartlabel) at ([yshift=-1em]cell42.south east) {\footnotesize{\textbf{chart}}};
\node [anchor=north west] (w1) at ([yshift=-2.5em,xshift=-2.0em]cell41.south west) {猫};
\node [anchor=west] (w2) at ([xshift=0.3em]w1.east) {喜欢};
\node [anchor=west] (w3) at ([xshift=0.3em]w2.east) {吃};
\node [anchor=west] (w4) at ([xshift=0.3em]w3.east) {鱼};
\node [anchor=north east] (p0) at ([xshift=0.3em]w1.south west) {\blue{0}};
\node [anchor=north east] (p1) at ([xshift=0.3em]w2.south west) {\blue{1}};
\node [anchor=north east] (p2) at ([xshift=0.3em]w3.south west) {\blue{2}};
\node [anchor=north east] (p3) at ([xshift=0.3em]w4.south west) {\blue{3}};
\node [anchor=north west] (p4) at ([xshift=-0.4em]w4.south east) {\blue{4}};
\node [anchor=north] (slabel) at (p2.south) {\scriptsize{\textbf{源语言句子}}};
\begin{scope}[scale=0.9,level distance=15pt,sibling distance=0pt]
\end{scope}
{\scriptsize
\Tree[.\node(bsn0){IP};
[.\node(bsn1){NP};
[.\node(bsn2){NN}; \node(bsw1){猫}; ]
]
[.\node(bsn3){VP};
[.\node(bsn4){VV}; \node(bsw2){喜欢}; ]
[.\node(bsn5){VP}; \edge[roof]; \node(bsw3){吃 \ 鱼}; ]
]
]
\end{tikzpicture}
\end{center}
\node [anchor=west] (target) at ([xshift=1em]bsw3.east) {Cats like eating fish};
\node [anchor=north,inner sep=3pt] (cap1) at (target.south west) {(a) 基于树的解码};
\draw [->,thick] (bsw3.east) -- (target.west);
\node [anchor=west] (sourcelabel) at ([xshift=2em]bsn0.east) {显式输入的结构};
\node [anchor=west] (source2) at ([xshift=2em]target.east) {猫 喜欢 吃 鱼};
\node [anchor=west] (target2) at ([xshift=1em]source2.east) {Cats like eating fish};
\node [anchor=north,inner sep=3pt] (cap2) at (target2.south west) {(b) 基于串的解码};
\draw [->,thick] (source2.east) -- (target2.west);
\begin{scope}[xshift=2.45in,sibling distance=3pt]
\Tree[.\node(bsn0){IP};
[.\node(bsn1){NP};
[.\node(bsn2){NN}; ]
]
[.\node(bsn3){VP};
[.\node(bsn4){VV}; ]
[.\node(bsn5){VP}; ]
]
]
\begin{pgfonlayer}{background}
\node [draw,dashed,inner sep=2pt] (box) [fit = (bsn0) (bsn1) (bsn2) (bsn3) (bsn4) (bsn5)] {};
\node [anchor=north west] (boxlabel) at (box.north east) {隐含结构};
\end{pgfonlayer}
\end{minipage}
\begin{minipage}[b]{0.55\linewidth}
\end{scope}
{\footnotesize
\visible<3->{
\begin{tabular}{l l l l}
\visible<3->{序号} & \visible<3->{跨度} & \visible<3->{标记} & \visible<3->{源语句子片段} \\ \hline
\visible<3->{1} & \visible<3->{[{\blue 0},{\blue 1}]} & \visible<3->{NN \& NP} & \visible<3->{猫} \\
\visible<4->{2} & \visible<4->{[{\blue 1},{\blue 2}]} & \visible<4->{VV} & \visible<4->{喜欢} \\
\visible<5->{3} & \visible<5->{[{\blue 2},{\blue 5}]} & \visible<5->{VV} & \visible<5->{吃} \\
\visible<6->{4} & \visible<6->{[{\blue 3},{\blue 6}]} & \visible<6->{NN \& NP} & \visible<6->{鱼} \\
\visible<7->{5} & \visible<7->{[{\blue 0},{\blue 2}]} & \visible<7->{N/A} & \visible<7->{猫 喜欢} \\
\visible<8->{6} & \visible<8->{[{\blue 1},{\blue 3}]} & \visible<8->{N/A} & \visible<8->{喜欢 吃} \\
\visible<9->{7} & \visible<9->{[{\blue 2},{\blue 4}]} & \visible<9->{VP} & \visible<9->{吃 鱼} \\
\visible<10->{8} & \visible<10->{[{\blue 0},{\blue 3}]} & \visible<10->{N/A} & \visible<10->{猫 喜欢 吃} \\
\visible<11->{9} & \visible<11->{[{\blue 1},{\blue 4}]} & \visible<11->{VP} & \visible<11->{喜欢 吃 鱼} \\
\visible<12->{10} & \visible<12->{[{\blue 0},{\blue 4}]} & \visible<12->{IP (\alert{root})} & \visible<12->{猫 喜欢 吃 鱼} \\
\\
\end{tabular}
}
}
\end{minipage}
\end{scope}
\end{tikzpicture}
\end{center}
}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%% 规则使用 - 基于树的匹配
\begin{frame}{使用树结构匹配树到串规则}
%%%------------------------------------------------------------------------------------------------------------
%%% 基于串的解码方法本质上和句法分析一样
\begin{frame}{基于串的解码 $\approx$ 句法分析}
\begin{itemize}
\item 对于规则的源语言部分，可以使用树片段的匹配找到可以使用这条规则位置
\begin{itemize}
\item 匹配的规则会被存入相应的表格单元中
\end{itemize}
\item 基于串的翻译和传统\alert{句法分析}的任务很像：对于一个输入的词串，找到生成这个词串的最佳推导。唯一不同的地方，在于机器翻译需要考虑译文的生成(语言模型的引入会使问题稍微复杂一些)，但是源语言部分的处理和句法分析一模一样
\item<2-> 这个过程仍然可以用基于chart的方法实现，即对于每一个源语言片段，都匹配可能的翻译规则，之后填入相应的表格单元，这也构成了一个超图，最佳推导可以从这个超图得到
\end{itemize}
\vspace{-2em}
\visible<2->{
\begin{center}
\begin{tikzpicture}
\begin{scope}
{\scriptsize
\begin{scope}[sibling distance=2pt,level distance=20pt,grow'=up]
\Tree[.\node(treeroot){IP};
[.NP [.NR 阿都拉$_1$ ]]
[.\node(tn1){VP};
[.\node(tn2){PP};
[.\node(tn3){P}; \node(cw1){对$_2$}; ]
[.\node(tn4){NP}; \edge[roof]; {自己$_3$ 四$_4$\ 个$_5$\ 多$_6$\ 月$_7$\ 以来$_8$\ 的$_9$\ 施政$_{10}$\ 表现$_{11}$} ]
]
[.\node(tn5){VP};
[.VV 感到$_{12}$ ]
[.NN 满意$_{13}$ ]
]
]
]
\end{scope}
\visible<2->{
\node [anchor=west,fill=green!20!white] (rulepart1) at ([yshift=2.0in,xshift=-1.3in]treeroot.east) {VP(PP(P(对) NP$_1$) VP$_2$)};
\node [anchor=north west] (rulepart2) at (rulepart1.south west) {$\to$ VP$_2$ with NP$_1$};
}
\node [anchor=west] (sw1) at (0,0) {阿都拉$_1$};
\node [anchor=west] (sw2) at ([xshift=0.1em]sw1.east) {对$_2$};
\node [anchor=west] (sw3) at ([xshift=0.1em]sw2.east) {自己$_3$ 四$_4$\ 个$_5$\ 多$_6$\ 月$_7$\ 以来$_8$\ 的$_9$\ 施政$_{10}$\ 表现$_{11}$};
\node [anchor=west] (sw4) at ([xshift=0.2em]sw3.east) {感到$_{12}$ 满意$_{13}$};
\begin{pgfonlayer}{background}
\visible<2->{
\node[rectangle,draw,inner sep=2pt] [fit = (rulepart1) (rulepart2)] (rulemark) {};
}
\end{pgfonlayer}
\visible<2->{
\node [anchor=south west] at (rulemark.north west) {\tiny{\textbf{树到串翻译规则}}};
}
}
\begin{scope}[xshift=1.5in,yshift=1.6in]
\visible<3->{
\node[anchor=center, minimum size=10pt,draw] (cell1and1) at (0,0) {};
\node[anchor=center, minimum size=10pt,draw] (cell2and1) at ([xshift=10pt]cell1and1.center) {};
\node[anchor=center, minimum size=10pt,draw] (cell3and1) at ([xshift=10pt]cell2and1.center) {};
\node[anchor=center, minimum size=10pt] (cell4and1) at ([xshift=15pt]cell3and1.center) {\tiny{$\dots$}};
\node[anchor=center, minimum size=10pt,draw] (cellnand1) at ([xshift=13pt]cell4and1.center) {};
\node[anchor=center, minimum size=10pt,draw] (cell1and2) at ([yshift=10pt]cell1and1.center) {};
\node[anchor=center, minimum size=10pt,draw] (cell2and2) at ([xshift=10pt]cell1and2.center) {};
\node[anchor=center, minimum size=10pt,draw] (cell1and3) at ([yshift=10pt]cell1and2.center) {};
\node[anchor=center, minimum size=10pt] (cell1and4) at ([yshift=20pt]cell1and3.center) {\tiny{$\vdots$}};
\node[anchor=center, minimum size=10pt] (cell2and3) at ([yshift=20pt]cell2and2.center) {\tiny{$\vdots$}};
\node[anchor=center, minimum size=10pt] (cell3and2) at ([yshift=20pt]cell3and1.center) {\tiny{$\vdots$}};
\node[anchor=center, minimum size=10pt,draw] (cell4andn) at ([yshift=20pt,xshift=10pt]cell3and1.center) {};
%\node[anchor=center, minimum size=10pt] (cell4and2) at ([yshift=20pt]cell4and1.center) {\tiny{$\ddots$}};
\node[anchor=center, minimum size=10pt,draw] (cell1andn) at ([yshift=10pt]cell1and4.center) {};
\node[anchor=center, minimum size=10pt,draw] (cell2andn) at ([xshift=10pt,yshift=-10pt]cell1andn.center) {};
%\node[anchor=center, minimum size=10pt,draw] (cell3andn) at ([xshift=10pt,yshift=-10pt]cell2andn.center) {};
\node[anchor=west] (chartlabel) at ([xshift=-0em,yshift=-0.8em]cell1and1.south east) {\scriptsize{Chart}};
\node [fill=red!20,inner sep=0pt] (box1) [fit = (sw1)] {};
\node [fill=green!20,inner sep=0pt] (box2) [fit = (sw3)] {};
\node [fill=orange!20,inner sep=0pt] (box3) [fit = (sw4)] {};
\node [anchor=south,align=center] (box1label) at (box1.north) {[{\blue 0},{\blue 1}]\\VP};
\node [anchor=south,align=center] (box2label) at (box2.north) {[{\blue 2},{\blue 11}]\\NP};
\node [anchor=south,align=center] (box3label) at (box3.north) {[{\blue 11},{\blue 13}]\\VP};
}
\end{scope}
\end{pgfonlayer}
\visible<3->{
\path [draw,thick,blue,->,dashed] (rulemark.north east) .. controls +(60:2.0) and +(north east:1.5) .. ([yshift=0.05em,xshift=0.05em]cell4andn.north east) node[pos=0.25, below,yshift=-0.2em] (spanlabel) {};
}
\draw[decorate,decoration={brace,mirror,,amplitude=3mm}] (sw1.south west) -- (sw4.south east);
\node [anchor=north] (label) at ([yshift=-1em]sw3.south) {在跨度[{\blue 0},{\blue 13}]上进行规则匹配};
\node [anchor=north] (rule) at ([yshift=-0.3em]label.south) {{\footnotesize 比如：IP({\color{red} NP$_1$} VP(PP(P(对) {\color{ugreen} NP$_2$}) {\color{orange} VP$_3$}))}};
\node [anchor=north west] (rule2) at ([yshift=0.2em]rule.south west) {{\footnotesize \hspace{2.8em} $\to$ NP$_1$ VP$_3$ with NP$_2$}};
\begin{pgfonlayer}{background}
\visible<2->{
\path [fill=green!20] (tn1.south west) -- ([yshift=0.2em]tn2.south west) -- ([yshift=0.2em,xshift=-0.30em]tn3.south west) -- (cw1.north west) -- (cw1.north east) -- ([xshift=0.30em]tn3.north east) -- (tn4.north east) -- (tn4.south east) -- ([xshift=0.5em]tn2.north east) -- (tn5.north east)-- ([yshift=0.2em]tn5.south east) -- (tn1.south east) -- (tn1.south west);
}
\end{pgfonlayer}
\visible<2->{
\path [draw,thick,blue,<->] ([xshift=-2em]rulepart1.south east) .. controls +(south:1.7) and +(north:1.3) .. ([xshift=0em,yshift=0.10em]cw1.north) node[pos=0.5, below,xshift=0.3in] {\scriptsize{树片段的匹配}};
}
\end{scope}
\end{tikzpicture}
\end{center}
}
\end{frame}
......
Section04-Phrasal-and-Syntactic-Models/section04.tex
......@@ -4485,7 +4485,7 @@ NP-BAR(NN$_1$ NP-BAR$_2$) $\to$ NN$_1$ NP-BAR$_2$
%%% 基于树的解码方法 - 超图
\begin{frame}{基于树的解码 - 超图}
\begin{itemize}
\item 如果源语言输入的是句法树，基于树的解码会找到一个推导覆盖整个句法树，之后输出所对应的目标语词串作为译文
\item 如果源语言输入的是句法树，\alert{基于树的解码}会找到一个推导覆盖整个句法树，之后输出所对应的目标语词串作为译文
\item 比如，可以从树的叶子结点开始，找到所有能匹配到这个节点的规则，当所有节点匹配完之后，本质上获得了一个超图
\begin{itemize}
\item<2-> 图的节点对应一个句法树句法节点
......@@ -4808,6 +4808,121 @@ NP-BAR(NN$_1$ NP-BAR$_2$) $\to$ NN$_1$ NP-BAR$_2$
\end{frame}
%%%------------------------------------------------------------------------------------------------------------
%%% 基于串的解码方法
\begin{frame}{基于串的解码}
\begin{itemize}
\item 不同于基于树的解码，\alert{基于串的解码}方法并不要求输入句法树，它直接对输入词串进行翻译，最终得到译文。
\begin{itemize}
\item 这种方法适用于树到串、串到树、树到树等多种模型
\item 本质上，由于并不受固定输入的句法树约束，基于串的解码可以探索更多潜在的树结构，这也增大了搜索空间(相比基于串的解码)，因此该方法更有可能找到高质量翻译结果
\end{itemize}
\item<2-> 在基于串的方法中，句法结构被看做是翻译的隐含变量，而非线性的输入和输出。比如，层次短语翻译解码就是一种典型的基于串的解码方法，所有的翻译推导在翻译过程里动态生成，但是并不要输入或者输出这些推导所对应的层次结构
\end{itemize}
\visible<2->{
\begin{center}
\begin{tikzpicture}
\begin{scope}[scale=0.9,level distance=15pt,sibling distance=0pt]
{\scriptsize
\Tree[.\node(bsn0){IP};
[.\node(bsn1){NP};
[.\node(bsn2){NN}; \node(bsw1){猫}; ]
]
[.\node(bsn3){VP};
[.\node(bsn4){VV}; \node(bsw2){喜欢}; ]
[.\node(bsn5){VP}; \edge[roof]; \node(bsw3){吃 \ 鱼}; ]
]
]
\node [anchor=west] (target) at ([xshift=1em]bsw3.east) {Cats like eating fish};
\node [anchor=north,inner sep=3pt] (cap1) at (target.south west) {(a) 基于树的解码};
\draw [->,thick] (bsw3.east) -- (target.west);
\node [anchor=west] (sourcelabel) at ([xshift=2em]bsn0.east) {显式输入的结构};
\node [anchor=west] (source2) at ([xshift=2em]target.east) {猫 喜欢 吃 鱼};
\node [anchor=west] (target2) at ([xshift=1em]source2.east) {Cats like eating fish};
\node [anchor=north,inner sep=3pt] (cap2) at (target2.south west) {(b) 基于串的解码};
\draw [->,thick] (source2.east) -- (target2.west);
\begin{scope}[xshift=2.45in,sibling distance=3pt]
\Tree[.\node(bsn0){IP};
[.\node(bsn1){NP};
[.\node(bsn2){NN}; ]
]
[.\node(bsn3){VP};
[.\node(bsn4){VV}; ]
[.\node(bsn5){VP}; ]
]
]
\begin{pgfonlayer}{background}
\node [draw,dashed,inner sep=2pt] (box) [fit = (bsn0) (bsn1) (bsn2) (bsn3) (bsn4) (bsn5)] {};
\node [anchor=north west] (boxlabel) at (box.north east) {隐含结构};
\end{pgfonlayer}
\end{scope}
}
\end{scope}
\end{tikzpicture}
\end{center}
}
\end{frame}
%%%------------------------------------------------------------------------------------------------------------
%%% 基于串的解码方法本质上和句法分析一样
\begin{frame}{基于串的解码 $\approx$ 句法分析}
\begin{itemize}
\item 基于串的翻译和传统\alert{句法分析}的任务很像：对于一个输入的词串，找到生成这个词串的最佳推导。唯一不同的地方，在于机器翻译需要考虑译文的生成(语言模型的引入会使问题稍微复杂一些)，但是源语言部分的处理和句法分析一模一样
\item<2-> 这个过程仍然可以用基于chart的方法实现，即对于每一个源语言片段，都匹配可能的翻译规则，之后填入相应的表格单元，这也构成了一个超图，最佳推导可以从这个超图得到
\end{itemize}
\visible<2->{
\begin{center}
\begin{tikzpicture}
\begin{scope}
{\scriptsize
\node [anchor=west] (sw1) at (0,0) {阿都拉$_1$};
\node [anchor=west] (sw2) at ([xshift=0.1em]sw1.east) {对$_2$};
\node [anchor=west] (sw3) at ([xshift=0.1em]sw2.east) {自己$_3$ 四$_4$\ 个$_5$\ 多$_6$\ 月$_7$\ 以来$_8$\ 的$_9$\ 施政$_{10}$\ 表现$_{11}$};
\node [anchor=west] (sw4) at ([xshift=0.2em]sw3.east) {感到$_{12}$ 满意$_{13}$};
\begin{pgfonlayer}{background}
\visible<3->{
\node [fill=red!20,inner sep=0pt] (box1) [fit = (sw1)] {};
\node [fill=green!20,inner sep=0pt] (box2) [fit = (sw3)] {};
\node [fill=orange!20,inner sep=0pt] (box3) [fit = (sw4)] {};
\node [anchor=south,align=center] (box1label) at (box1.north) {[{\blue 0},{\blue 1}]\\VP};
\node [anchor=south,align=center] (box2label) at (box2.north) {[{\blue 2},{\blue 11}]\\NP};
\node [anchor=south,align=center] (box3label) at (box3.north) {[{\blue 11},{\blue 13}]\\VP};
}
\end{pgfonlayer}
\draw[decorate,decoration={brace,mirror,,amplitude=3mm}] (sw1.south west) -- (sw4.south east);
\node [anchor=north] (label) at ([yshift=-1em]sw3.south) {在跨度[{\blue 0},{\blue 13}]上进行规则匹配};
\node [anchor=north] (rule) at ([yshift=-0.3em]label.south) {{\footnotesize 比如：IP({\color{red} NP$_1$} VP(PP(P(对) {\color{ugreen} NP$_2$}) {\color{orange} VP$_3$}))}};
\node [anchor=north west] (rule2) at ([yshift=0.2em]rule.south west) {{\footnotesize \hspace{2.8em} $\to$ NP$_1$ VP$_3$ with NP$_2$}};
}
\end{scope}
\end{tikzpicture}
\end{center}
}
\end{frame}
%%%------------------------------------------------------------------------------------------------------------
%%% 基于串的解码
......
Section06-Neural-Machine-Translation/section06.tex
......@@ -3997,12 +3997,13 @@ $\textrm{``you''} = \argmax_{y} \textrm{P}(y|\textbf{s}_1, \alert{\textbf{C}})$
\item 通过自注意机制能够直接获取全局信息，不像RNN需要逐步进行信息提取，也不像CNN只能获取局部信息，可以并行化操作，提高训练效率
\item Transformer不仅仅被用于神经机器翻译任务，还广泛用于其他NLP任务、甚至图像处理任务。目前最火的预训练模型Bert也基于Transformer
\item<2-> Transformer不仅仅被用于神经机器翻译任务，还广泛用于其他NLP任务、甚至图像处理任务。目前最火的预训练模型Bert也基于Transformer
\end{itemize}
\vspace{0em}
\visible<2->{
{
\footnotesize
\begin{center}
......@@ -4027,6 +4028,8 @@ $\textrm{``you''} = \argmax_{y} \textrm{P}(y|\textbf{s}_1, \alert{\textbf{C}})$
\end{tabular}
\end{center}
}
}
\end{frame}
%%%------------------------------------------------------------------------------------------------------------
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论