Skip to content

T
Toy-MT-Introduction
Commit d5ac0ef2 by xiaotong
Options

new updates

parent 90ed014e
正在显示 包含 80 行增加 和 407 行删除
Section04-Phrasal-and-Syntactic-Models/section04-test.tex
......@@ -149,373 +149,24 @@
\subsection{引入双语句法信息}
%%%------------------------------------------------------------------------------------------------------------
%%% 树到树规则抽取
\begin{frame}{引入双语句法信息}
%%% 翻译特征
\begin{frame}{特征}
\begin{itemize}
\item 对于树到树模型，源语和目标语端都有句法树，需要使用树片段到树片段的映射来描述翻译过程，这种映射关系被描述为树到树翻译规则。这里，把\\
\vspace{-1.3em}
\begin{eqnarray}
\langle\ \textrm{VP}, \textrm{VP}\ \rangle & \to & \langle\ \textrm{VP(PP}_{1}\ \textrm{VP(VV(表示) NN}_{2})), \nonumber \\
& & \ \ \textrm{VP(VBZ(was) VP(VBZ}_{2}\ \textrm{PP}_{1}))\ \rangle \nonumber
\end{eqnarray}
表示为\alert{树片段到树片段}的映射形式\\
\vspace{-1.3em}
\begin{eqnarray}
& & \textrm{VP(PP}_{1}\ \textrm{VP(VV(表示) NN}_{2})) \nonumber \\
& \to & \textrm{VP(VBZ(was) VP(VBZ}_{2}\ \textrm{PP}_{1})) \nonumber
\end{eqnarray}
\item<2-> 可以通过扩展GHKM方法进行树到树规则抽取
\item 与短语和层次短语模型一样，句法模型也使用判别式模型进行建模 - $\textrm{P}(d,\textbf{t}|\textbf{s}) = \frac{\exp(\sum_{i=1}^{M} \lambda_i \cdot h_i(d,\textbf{s},\textbf{t}))}{\sum_{d',t'}\exp(\sum_{i=1}^{M} \lambda_i \cdot h_i(d',\textbf{s},\textbf{t}'))}$。其中特征权重$\{\lambda_i\}$可以使用最小错误率训练进行调优，特征函数$\{h_i\}$需要用户定义。
\item<2-> 这里，所有规则满足$\langle\ \alpha_h, \beta_h\ \rangle \to \langle\ \alpha_r, \beta_r, \sim\ \rangle$的形式
\begin{itemize}
\item 双语端进行可信节点的识别，之后找到节点之间的对应
\item 基于对应的节点获得树片段的对应，即抽取树到树规则
\item 规则组合、SPMT等方法同样适用
\end{itemize}
\end{itemize}
\end{frame}
%%%------------------------------------------------------------------------------------------------------------
%%% 方法1：利用词对齐归纳句法映射
\begin{frame}{方法1：利用词对齐归纳树到树规则}
\begin{itemize}
\item 简单直接的方法是把GHKM方法扩展到双语的情况，利用词对齐归纳树到树映射
\begin{itemize}
\item<3-> 但是词对齐的错误往往会导致很多规则无法抽取
\end{itemize}
\end{itemize}
\begin{minipage}[c][5cm][t]{0.47\textwidth}
\begin{center}
\begin{tikzpicture}
\begin{scope}
\begin{scope}[scale=0.65, level distance=27pt]
\Tree[.S
[.NP
[.DT \node(ew1){the}; ]
[.NNS \node(ew2){imports}; ]
]
[.VP
[.VBZ \node(ew3){have}; ]
[.ADVP
[.RB \node(ew4){drastically}; ]
[.VBN \node(ew5){fallen}; ]
]
]
]
\end{scope}
\begin{scope}[scale=0.65, level distance=27pt, grow'=up, xshift=-13pt, yshift=-3.5in, sibling distance=22pt]
\Tree[.IP
[.NN \node(cw1){进口}; ]
[.VP
[.AD \node(cw2){大幅度}; ]
[.VP
[.VV \node(cw3){下降}; ]
[.AS \node(cw4){了}; ]
]
]
]
\end{scope}
\visible<2->{
\draw[-, dashed] (cw1) -- (ew2);
\draw[-, dashed] (cw2) -- (ew4);
\draw[-, dashed] (cw3) -- (ew5);
\draw[-, dashed] (cw4) .. controls +(north:1.0) and +(south:1.6) .. (ew1);
}
\visible<3->{
\draw[-, red, dashed,thick] (cw4) .. controls +(north:1.0) and +(south:1.6) .. (ew1);
}
\end{scope}
\end{tikzpicture}
\end{center}
\end{minipage}
\begin{minipage}[c][5cm][t]{0.50\textwidth}
\visible<2->{
\begin{tabular}{l l}
\multicolumn{2}{l}{\textbf{\scriptsize{抽取得到的规则}}} \\
\hline
\scriptsize{$r_1$} & \scriptsize{AS(了) $\rightarrow$ DT(the)} \\
\scriptsize{$r_2$} & \scriptsize{NN(进口) $\rightarrow$ NNS(imports)} \\
\scriptsize{$r_3$} & \scriptsize{AD(大幅度) $\rightarrow$ RB(drastically)} \\
\scriptsize{$r_4$} & \scriptsize{VV(下降) $\rightarrow$ VBN(fallen)} \\
\scriptsize{$r_5$} & \scriptsize{IP(NN$_1$ VP(AD$_2$ VP(VV$_3$ AS$_4$)) $\rightarrow$} \\
\multicolumn{2}{l}{\tiny{S(NP(DT$_4$ NNS$_1$) VP(VBZ(have) ADVP(RB$_2$ VBN$_3$))}} \\
\end{tabular}
}
\visible<3->{
\vspace{0.5em}
\begin{tabular}{l l}
\multicolumn{2}{l}{\textbf{\scriptsize{无法得到的规则}}} \\
\hline
\scriptsize{$r_{?}$} & \scriptsize{AS(了) $\rightarrow$ VBZ(have)} \\
\scriptsize{$r_{?}$} & \scriptsize{NN(进口) $\rightarrow$} \\
& \scriptsize{NP(DT(the) NNS(imports))} \\
\scriptsize{$r_{?}$} & \scriptsize{IP(NN$_1$ VP$_2$) $\rightarrow$ S(NP$_1$ VP$_2$)} \\
\end{tabular}
}
\end{minipage}
\end{frame}
%%%------------------------------------------------------------------------------------------------------------
%%% 方法2：直接进行节点对齐然后归纳句法映射
\begin{frame}{方法2：利用节点对齐抽取树到树规则}
\begin{itemize}
\item 另一种思路是直接获取源语言树节点到目标语树节点的对应关系，然后直接抽取规则，这样可避免词对齐错误
\begin{itemize}
\item 节点对其可以更准确的捕捉双语结构的对应
\end{itemize}
\end{itemize}
\begin{minipage}[c][5cm][t]{0.47\textwidth}
\begin{center}
\begin{tikzpicture}
\only<1>{
\begin{scope}
\begin{scope}[scale=0.65, level distance=27pt]
\Tree[.S
[.NP
[.DT \node(ew1){the}; ]
[.NNS \node(ew2){imports}; ]
]
[.VP
[.VBZ \node(ew3){have}; ]
[.ADVP
[.RB \node(ew4){drastically}; ]
[.VBN \node(ew5){fallen}; ]
]
]
]
\end{scope}
\begin{scope}[scale=0.65, level distance=27pt, grow'=up, xshift=-13pt, yshift=-3.5in, sibling distance=22pt]
\Tree[.IP
[.NN \node(cw1){进口}; ]
[.VP
[.AD \node(cw2){大幅度}; ]
[.VP
[.VV \node(cw3){下降}; ]
[.AS \node(cw4){了}; ]
]
]
]
\end{scope}
\draw[-, dashed] (cw1) -- (ew2);
\draw[-, dashed] (cw2) -- (ew4);
\draw[-, dashed] (cw3) -- (ew5);
\draw[-, dashed] (cw4) .. controls +(north:1.0) and +(south:1.6) .. (ew1);
\end{scope}
}
\begin{scope}
\visible<2->{
\begin{scope}[scale=0.65, level distance=27pt]
\Tree[.\node[draw](en1){S};
[.\node[draw](en2){NP};
[.DT the ]
[.NNS imports ]
]
[.\node[draw](en3){VP};
[.\node[draw](en4){VBZ}; have ]
[.ADVP
[.\node[draw](en5){RB}; drastically ]
[.\node[draw](en6){VBN}; fallen ]
]
]
]
\end{scope}
\begin{scope}[scale=0.65, level distance=27pt, grow'=up, xshift=-13pt, yshift=-3.5in, sibling distance=22pt]
\Tree[.\node[draw](cn1){\ \ IP\ \ };
[.\node[draw](cn2){NN}; 进口 ]
[.\node[draw](cn3){VP};
[.\node[draw](cn4){AD}; 大幅度 ]
[.VP
[.\node[draw](cn5){VV}; 下降 ]
[.\node[draw](cn6){AS}; 了 ]
]
]
]
\end{scope}
}
\visible<3->{
\draw[latex-latex, dotted, thick, red] (cn4.east) .. controls +(east:0.5) and +(west:0.5) .. (en5.west);
\draw[latex-latex, dotted, thick, red] (cn5.east) .. controls +(east:0.5) and +(south:0.5) .. (en6.south west);
\draw[latex-latex, dotted, thick, red] (cn6.north west) .. controls +(north:1.5) and +(south:2.5) .. (en4.south west);
\draw[latex-latex, dotted, thick, red] (cn3.north west) -- (en3.south west);
\draw[latex-latex, dotted, thick, red] (cn2.west) .. controls +(west:0.6) and +(west:0.6) .. (en2.west);
\draw[latex-latex, dotted, thick, red] (cn1.north west) .. controls +(north:4) and +(south:5.5) .. (en1.south west);
}
\end{scope}
\end{tikzpicture}
\end{center}
\end{minipage}
\begin{minipage}[c][5cm][t]{0.50\textwidth}
\only<1>{
\begin{tabular}{l l}
\multicolumn{2}{l}{\textbf{\scriptsize{抽取得到的规则(词对齐)}}} \\
\hline
\scriptsize{$r_1$} & \scriptsize{AS(了) $\rightarrow$ DT(the)} \\
\scriptsize{$r_2$} & \scriptsize{NN(进口) $\rightarrow$ NNS(imports)} \\
\scriptsize{$r_3$} & \scriptsize{AD(大幅度) $\rightarrow$ RB(drastically)} \\
\scriptsize{$r_4$} & \scriptsize{VV(下降) $\rightarrow$ VBN(fallen)} \\
\scriptsize{$r_5$} & \scriptsize{IP(NN$_1$ VP(AD$_2$ VP(VV$_3$ AS$_4$)) $\rightarrow$} \\
\multicolumn{2}{l}{\tiny{S(NP(DT$_4$ NNS$_1$) VP(VBZ(have) ADVP(RB$_2$ VBN$_3$))}} \\
\end{tabular}
}
\visible<4->{
\begin{tabular}{l l}
\multicolumn{2}{l}{\textbf{\scriptsize{抽取得到的规则(子树对齐)}}} \\
\hline
{\color{gray!70} \scriptsize{$r_1$}} & {\color{gray!70} \scriptsize{AS(了) $\rightarrow$ DT(the)}} \\
{\color{gray!70} \scriptsize{$r_2$}} & {\color{gray!70}\scriptsize{NN(进口) $\rightarrow$ NNS(imports)}} \\
\scriptsize{$r_3$} & \scriptsize{AD(大幅度) $\rightarrow$ RB(drastically)} \\
\scriptsize{$r_4$} & \scriptsize{VV(下降) $\rightarrow$ VBN(fallen)} \\
{\color{gray!70} \scriptsize{$r_5$}} & {\color{gray!70} \scriptsize{IP(NN$_1$ VP(AD$_2$ VP(VV$_3$ AS$_4$)) $\rightarrow$}} \\
\multicolumn{2}{l}{{\color{gray!70} \tiny{S(NP(DT$_4$ NNS$_1$) VP(VBZ(have) ADVP(RB$_2$ VBN$_3$))}}} \\
\alert{\scriptsize{$r_6$}} & \alert{\scriptsize{AS(了) $\rightarrow$ VBZ(have)}} \\
\alert{\scriptsize{$r_7$}} & \alert{\scriptsize{NN(进口) $\rightarrow$ }} \\
& \alert{\scriptsize{NP(DT(the) NNS(imports))}}\\
\alert{\scriptsize{$r_8$}} & \alert{\scriptsize{VP(AD$_1$ VP(VV$_2$ AS$_3$)) $\rightarrow$}} \\
& \alert{\scriptsize{VP(VBZ$_3$ ADVP(RB$_1$ VBN$_2$)}} \\
\alert{\scriptsize{$r_9$}} & \alert{\scriptsize{IP(NN$_1$ VP$_2$) $\rightarrow$ S(NP$_1$ VP$_2$)}} \\
\end{tabular}
}
\end{minipage}
\end{frame}
%%%------------------------------------------------------------------------------------------------------------
%%% 抽取更多的规则：节点对齐矩阵
\begin{frame}{节点对齐矩阵}
\begin{itemize}
\item 节点对齐的自动获取：1）基于分类模型的方法；2）无指导节点对齐的方法
\item 使用节点对齐的另一个好处是，我们可以直接用节点对齐矩阵进行规则抽取，而不是用单一的对齐结果
\begin{itemize}
\item 对齐矩阵可以帮助抽取更多样的规则
\item $\alpha_h$和$\beta_h$是规则左部的源语和目标语部分，对应树结构的根节点
\item $\alpha_r$和$\beta_r$是规则右部的源语和目标语部分，对应树结构
\item $\sim$表示$\alpha_r$和$\beta_r$中叶子非终结符的对应
\item 此外，定义$r(\alpha_r)$和$r(\beta_r)$为源语和目标语树结构的叶子节点序列。例如，对于规则$\langle\ \textrm{VP}, \textrm{VP}\ \rangle \to \langle\ \textrm{VP(PP}_{1}\ \textrm{VP(VV(表示) NN}_{2})), \textrm{VP(VBZ(was) VP(VBZ}_{2}\ \textrm{PP}_{1}))$，有 \\
\vspace{-1.5em}
\begin{eqnarray}
r(\alpha_r) & = & \textrm{PP}_1\ \textrm{表示 NN}_2 \nonumber \\
r(\beta_r) & = & \textrm{was}\ \textrm{VBZ}_2\ \textrm{PP}_1\nonumber
\end{eqnarray}
\end{itemize}
\end{itemize}
\vspace{-0.2em}
\centering
\begin{tikzpicture}
\begin{scope}[scale=0.7]
\begin{scope}[sibling distance=17pt, level distance=25pt]
\Tree[.\node(en1){VP$^{[1]}$};
[.\node(en2){VBZ$^{[2]}$}; have ]
[.\node(en3){ADVP$^{[3]}$};
[.\node(en4){RB$^{[4]}$}; drastically ]
[.\node(en5){VBN$^{[5]}$}; fallen ]
]
]
\end{scope}
\begin{scope}[grow'=up, yshift=-2.7in, sibling distance=32pt, level distance=25pt]
\Tree[.\node(cn1){VP$^{[1]}$};
[.\node(cn2){AD$^{[2]}$}; 大幅度 ]
[.\node(cn3){VP$^{[3]}$};
[.\node(cn4){VV$^{[4]}$}; 下降 ]
[.\node(cn5){AS$^{[5]}$}; 了 ]
]
]
\end{scope}
\begin{scope}[xshift=1.7in, yshift=-0.4in]
\node[anchor=west, rotate=60] at (0.8,-0.6) {VP$^{[1]}$};
\node[anchor=west, rotate=60] at (1.8,-0.6) {VBZ$^{[2]}$};
\node[anchor=west, rotate=60] at (2.8,-0.6) {ADVP$^{[3]}$};
\node[anchor=west, rotate=60] at (3.8,-0.6) {RB$^{[4]}$};
\node[anchor=west, rotate=60] at (4.8,-0.6) {VBN$^{[5]}$};
\node[] at (6.2,-1) {VP$^{[1]}$};
\node[] at (6.2,-2) {AD$^{[2]}$};
\node[] at (6.2,-3) {VP$^{[3]}$};
\node[] at (6.2,-4) {VV$^{[4]}$};
\node[] at (6.2,-5) {AS$^{[5]}$};
\foreach \i in {1,...,5}{
\foreach \j in {-5,...,-1}{
\node[fill=blue,scale=0.2] at (\i,\j) {};
}
}
\visible<2-3>{
\node[fill=blue, scale=1.2] at (1,-1) {};
\node[fill=blue, scale=1.2] at (4,-2) {};
\node[fill=blue, scale=1.2] at (2,-5) {};
}
\visible<2>{
\node[fill=blue, scale=1.2] at (5,-4) {};
}
\visible<3>{
\node[fill=red, scale=1.2] at (5,-4) {};
}
\visible<4-5>{
\node[fill=blue, scale=1.1] at (1,-1) {};
\node[fill=blue, scale=0.5] at (1,-3) {};
\node[fill=blue, scale=0.6] at (2,-2) {};
\node[fill=blue, scale=0.7] at (2,-3) {};
\node[fill=blue, scale=0.7] at (2,-5) {};
\node[fill=blue, scale=0.4] at (3,-1) {};
\node[fill=blue, scale=0.6] at (3,-2) {};
\node[fill=blue, scale=0.5] at (3,-3) {};
\node[fill=blue, scale=0.9] at (4,-2) {};
\node[fill=blue, scale=0.7] at (5,-3) {};
\node[fill=blue, scale=0.4] at (5,-5) {};
}
\visible<4>{
\node[fill=blue, scale=0.6] at (3,-4) {};
\node[fill=blue, scale=0.8] at (5,-4) {};
}
\visible<5>{
\node[fill=red, scale=0.6] at (3,-4) {};
\node[fill=red, scale=0.8] at (5,-4) {};
}
\visible<2-3>{
\node[] at (4,-5.8) {\footnotesize{{\color{blue} $\blacksquare$} = extractable node-pair}};
}
\visible<4-5>{
\node[] at (4,-5.8) {\footnotesize{{\color{blue} $\blacksquare$} = possible alignment}};
}
\end{scope}
\visible<3>{\draw[<->, red, thick] (cn4.east) .. controls +(east:0.9) and +(west:0.9) .. (en5.west);}
\visible<5>{\draw[<->, red, dotted, very thick] (cn4.east) .. controls +(east:0.9) and +(west:0.9) .. (en5.west);}
\visible<5>{\draw[<->, red, dotted, very thick] (cn4.west) .. controls +(west:1.0) and +(west:2) .. (en3.west);}
\end{scope}
\end{tikzpicture}
\end{frame}
%%%------------------------------------------------------------------------------------------------------------
......
Section04-Phrasal-and-Syntactic-Models/section04.tex
......@@ -4287,6 +4287,8 @@ NP-BAR(NN$_1$ NP-BAR$_2$) $\to$ NN$_1$ NP-BAR$_2$
\end{scope}
\begin{scope}[xshift=1.7in, yshift=-0.4in]
{\footnotesize
\node[anchor=west, rotate=60] at (0.8,-0.6) {VP$^{[1]}$};
\node[anchor=west, rotate=60] at (1.8,-0.6) {VBZ$^{[2]}$};
\node[anchor=west, rotate=60] at (2.8,-0.6) {ADVP$^{[3]}$};
......@@ -4298,6 +4300,7 @@ NP-BAR(NN$_1$ NP-BAR$_2$) $\to$ NN$_1$ NP-BAR$_2$
\node[] at (6.2,-3) {VP$^{[3]}$};
\node[] at (6.2,-4) {VV$^{[4]}$};
\node[] at (6.2,-5) {AS$^{[5]}$};
}
\foreach \i in {1,...,5}{
\foreach \j in {-5,...,-1}{
......@@ -4370,8 +4373,23 @@ NP-BAR(NN$_1$ NP-BAR$_2$) $\to$ NN$_1$ NP-BAR$_2$
%%%------------------------------------------------------------------------------------------------------------
%%% 翻译特征
\begin{frame}{翻译特征}
% NiuTrans Manual
\begin{frame}{特征}
\begin{itemize}
\item 与短语和层次短语模型一样，句法模型也使用判别式模型进行建模 - $\textrm{P}(d,\textbf{t}|\textbf{s}) = \frac{\exp(\sum_{i=1}^{M} \lambda_i \cdot h_i(d,\textbf{s},\textbf{t}))}{\sum_{d',t'}\exp(\sum_{i=1}^{M} \lambda_i \cdot h_i(d',\textbf{s},\textbf{t}'))}$。其中特征权重$\{\lambda_i\}$可以使用最小错误率训练进行调优，特征函数$\{h_i\}$需要用户定义。
\item<2-> 这里，所有规则满足$\langle\ \alpha_h, \beta_h\ \rangle \to \langle\ \alpha_r, \beta_r, \sim\ \rangle$的形式
\begin{itemize}
\item $\alpha_h$和$\beta_h$是规则左部的源语和目标语部分，对应树结构的根节点
\item $\alpha_r$和$\beta_r$是规则右部的源语和目标语部分，对应树结构
\item $\sim$表示$\alpha_r$和$\beta_r$中叶子非终结符的对应
\item 此外，定义$r(\alpha_r)$和$r(\beta_r)$为源语和目标语树结构的叶子节点序列。例如，对于规则$\langle\ \textrm{VP}, \textrm{VP}\ \rangle \to \langle\ \textrm{VP(PP}_{1}\ \textrm{VP(VV(表示) NN}_{2})), \textrm{VP(VBZ(was) VP(VBZ}_{2}\ \textrm{PP}_{1}))$，有 \\
\vspace{-1.5em}
\begin{eqnarray}
r(\alpha_r) & = & \textrm{PP}_1\ \textrm{表示 NN}_2 \nonumber \\
r(\beta_r) & = & \textrm{was}\ \textrm{VBZ}_2\ \textrm{PP}_1\nonumber
\end{eqnarray}
\end{itemize}
\end{itemize}
\end{frame}
%%%------------------------------------------------------------------------------------------------------------
......
Section06-Neural-Machine-Translation/section06.tex
......@@ -1148,10 +1148,10 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\hspace*{-0.6cm}
\begin{tikzpicture}
\setlength{\base}{0.9cm}
\tikzstyle{rnnnode} = [rounded corners=1pt,minimum height=0.5\base,minimum width=1\base,draw,inner sep=0pt,outer sep=0pt]
\tikzstyle{wordnode} = [font=\tiny]
% RNN translation model
\begin{scope}[local bounding box=RNNMT]
% RNN Encoder
......@@ -1165,11 +1165,11 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\node[rnnnode,fill=purple!30!white] (enclabel3) at (enc3) {\tiny{$\textbf{h}_{m}$}};
\node[wordnode,left=0.4\base of enc1] (init1) {$\cdots$};
\node[wordnode,left=0.4\base of eemb1] (init2) {$\cdots$};
\node[wordnode,below=0pt of eemb1] () {走};
\node[wordnode,below=0pt of eemb2] () {吗};
\node[wordnode,below=0pt of eemb3] () {$\langle$eos$\rangle$};
% RNN Decoder
\foreach \x in {1,2,...,3}
\node[rnnnode,minimum height=0.5\base,fill=green!30!white,anchor=south] (demb\x) at ([yshift=\base]enc\x.north) {\tiny{$e_y()$}};
......@@ -1180,7 +1180,7 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\node[wordnode,right=0.4\base of demb3] (end1) {$\cdots$};
\node[wordnode,right=0.4\base of dec3] (end2) {$\cdots$};
\node[wordnode,right=0.4\base of softmax3] (end3) {$\cdots$};
% Decoder input words
\node[wordnode,below=0pt of demb1] (decwordin) {$\langle$sos$\rangle$};
\ExtractX{$(demb2.south)$}
......@@ -1189,7 +1189,7 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\ExtractX{$(demb3.south)$}
\ExtractY{$(decwordin.base)$}
\node[wordnode,anchor=base] () at (\XCoord,\YCoord) {you};
% Decoder output words
\node[wordnode,above=0pt of softmax1] (decwordout) {Do};
\ExtractX{$(softmax2.north)$}
......@@ -1198,7 +1198,7 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\ExtractX{$(softmax3.north)$}
\ExtractY{$(decwordout.base)$}
\node[wordnode,anchor=base] () at (\XCoord,\YCoord) {know};
% Connections
\draw[-latex'] (init1.east) to (enc1.west);
\draw[-latex'] (dec3.east) to (end2.west);
......@@ -1213,11 +1213,11 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\draw[-latex'] (enc\x.east) to (enc\y.west);
\draw[-latex'] (dec\x.east) to (dec\y.west);
}
\coordinate (bridge) at ([yshift=0.4\base]enc2.north west);
\draw[-latex'] (enc3.north) .. controls +(north:0.3\base) and +(east:\base) .. (bridge) .. controls +(west:2.7\base) and +(west:0.3\base) .. (dec1.west);
\end{scope}
\begin{scope}
\coordinate (start) at (5.8\base,0.3\base);
\visible<2->{
......@@ -1240,9 +1240,9 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
-2 & .3 & \cdots & .1
\end{bmatrix}
$}};
\draw [decorate,decoration={brace,mirror}] ([shift={(6pt,2pt)}]mat.south west) to node [auto,swap,font=\scriptsize] {词嵌入矩阵} ([shift={(-6pt,2pt)}]mat.south east);
\visible<3->{
\draw [-latex'] ([xshift=-2pt,yshift=-0.65cm]one.east) to ([yshift=-0.65cm]words.west);
}
......@@ -1251,10 +1251,10 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
}
\draw [-latex'] ([yshift=-0.4cm]w.south) to ([yshift=2pt]w.south);
\node [anchor=north] (wlabel) at ([yshift=-0.6em]w.south) {\scriptsize{输入的单词}};
\node [draw=ugreen,densely dashed,thick,rounded corners=3pt,fit=(one) (words) (mat) (w)] (input) {};
\end{scope}
\draw [->,thick,densely dashed,ugreen] ([yshift=-0.2em]demb3.east) to [out=0,in=180] ([yshift=-1cm]input.west);
\end{tikzpicture}
\end{center}
......@@ -1275,10 +1275,10 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\hspace*{-0.6cm}
\begin{tikzpicture}
\setlength{\base}{0.9cm}
\tikzstyle{rnnnode} = [rounded corners=1pt,minimum height=0.5\base,minimum width=1\base,draw,inner sep=0pt,outer sep=0pt]
\tikzstyle{wordnode} = [font=\tiny]
% RNN translation model
\begin{scope}[local bounding box=RNNMT]
% RNN Encoder
......@@ -1292,11 +1292,11 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\node[rnnnode,fill=purple!30!white] (enclabel3) at (enc3) {\tiny{$\textbf{h}_{m}$}};
\node[wordnode,left=0.4\base of enc1] (init1) {$\cdots$};
\node[wordnode,left=0.4\base of eemb1] (init2) {$\cdots$};
\node[wordnode,below=0pt of eemb1] () {走};
\node[wordnode,below=0pt of eemb2] () {吗};
\node[wordnode,below=0pt of eemb3] () {$\langle$eos$\rangle$};
% RNN Decoder
\foreach \x in {1,2,...,3}
\node[rnnnode,minimum height=0.5\base,fill=green!30!white,anchor=south] (demb\x) at ([yshift=\base]enc\x.north) {\tiny{$e_y()$}};
......@@ -1307,7 +1307,7 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\node[wordnode,right=0.4\base of demb3] (end1) {$\cdots$};
\node[wordnode,right=0.4\base of dec3] (end2) {$\cdots$};
\node[wordnode,right=0.4\base of softmax3] (end3) {$\cdots$};
% Decoder input words
\node[wordnode,below=0pt of demb1] (decwordin) {$\langle$sos$\rangle$};
\ExtractX{$(demb2.south)$}
......@@ -1316,7 +1316,7 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\ExtractX{$(demb3.south)$}
\ExtractY{$(decwordin.base)$}
\node[wordnode,anchor=base] () at (\XCoord,\YCoord) {you};
% Decoder output words
\node[wordnode,above=0pt of softmax1] (decwordout) {Do};
\ExtractX{$(softmax2.north)$}
......@@ -1325,7 +1325,7 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\ExtractX{$(softmax3.north)$}
\ExtractY{$(decwordout.base)$}
\node[wordnode,anchor=base] () at (\XCoord,\YCoord) {know};
% Connections
\draw[-latex'] (init1.east) to (enc1.west);
\draw[-latex'] (dec3.east) to (end2.west);
......@@ -1340,20 +1340,20 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\draw[-latex'] (enc\x.east) to (enc\y.west);
\draw[-latex'] (dec\x.east) to (dec\y.west);
}
\coordinate (bridge) at ([yshift=0.4\base]enc2.north west);
\draw[-latex'] (enc3.north) .. controls +(north:0.3\base) and +(east:\base) .. (bridge) .. controls +(west:2.7\base) and +(west:0.3\base) .. (dec1.west);
\end{scope}
\begin{scope}
\coordinate (start) at (8.5\base,0.1\base);
\node [anchor=center,minimum width=5.7em,minimum height=1.3em,draw,rounded corners=0.3em] (hidden) at (start) {};
\node [anchor=west,minimum width=1em,minimum size=1em,fill=ugreen!20] (cell01) at ([xshift=0.2em]hidden.west) {\scriptsize{.2}};
\node [anchor=west,minimum width=1em,minimum size=1em,fill=ugreen!10] (cell02) at (cell01.east) {\scriptsize{-1}};
\node [anchor=west,minimum width=1em,minimum size=1em,fill=white] (cell03) at (cell02.east) {\scriptsize{$\cdots$}};
\node [anchor=west,minimum width=1em,minimum size=1em,fill=ugreen!50] (cell04) at (cell03.east) {\scriptsize{5}};
\visible<2->{
\node [anchor=south,minimum width=10.9em,minimum height=1.3em,draw,rounded corners=0.3em] (target) at ([yshift=1.5em]hidden.north) {};
\node [anchor=west,minimum width=1em,minimum size=1em,fill=ugreen!10] (cell11) at ([xshift=0.2em]target.west) {\scriptsize{-2}};
......@@ -1365,7 +1365,7 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\node [anchor=west,minimum width=1em,minimum size=1em,fill=ugreen!10] (cell17) at (cell16.east) {\scriptsize{-1}};
\node [anchor=west,minimum width=1em,minimum size=1em,fill=ugreen!20] (cell18) at (cell17.east) {\scriptsize{.2}};
}
\visible<3->{
\node [anchor=south,minimum width=1em,minimum height=0.2em,fill=ublue!80,inner sep=0pt] (label1) at ([yshift=2.5em]cell11.north) {};
\node [anchor=west,rotate=90,font=\tiny] (w1) at (label1.north) {$\langle$eos$\rangle$};
......@@ -1389,15 +1389,15 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\node [anchor=south,minimum width=1em,minimum height=0.4em,fill=ublue!80,inner sep=0pt] (label8) at ([yshift=2.5em]cell18.north) {};
\node [anchor=west,rotate=90,font=\tiny] (w8) at (label8.north) {have};
}
\visible<2->{
\filldraw [fill=red!20,draw=white] (target.south west) -- (target.south east) -- ([xshift=-0.2em,yshift=0.1em]hidden.north east) -- ([xshift=0.2em,yshift=0.1em]hidden.north west);
\draw [->,thick] ([xshift=0.2em,yshift=0.1em]hidden.north west) -- (target.south west);
\draw [->,thick] ([xshift=-0.2em,yshift=0.1em]hidden.north east) -- (target.south east);
\node [anchor=south] () at ([yshift=0.3em]hidden.north) {\scriptsize{$\hat{s}=Ws$}};
}
\visible<3->{
\node [rounded corners=0.3em] (softmax) at ([yshift=1.25em]target.north) {\scriptsize{$p(\hat{s}_i)=\frac{e^{\hat{s}_i}}{\sum_j e^{\hat{s}_j}}$}};
\filldraw [fill=blue!20,draw=white] ([yshift=0.1em]cell11.north west) {[rounded corners=0.3em] -- (softmax.west)} -- (label1.south west) -- (label8.south east) {[rounded corners=0.3em] -- (softmax.east)} -- ([yshift=0.1em]cell18.north east) -- ([yshift=0.1em]cell11.north west);
......@@ -1407,11 +1407,11 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\visible<4->{
\draw [-latex'] (w5.east) to ([yshift=0.3cm]w5.east);
}
\coordinate (tmp) at ([yshift=-3pt]w5.east);
\node [draw=red,thick,densely dashed,rounded corners=3pt,inner sep=5pt,fit=(cell01) (cell11) (label1) (label8) (target) (hidden) (tmp)] (output) {};
\end{scope}
\draw [->,thick,densely dashed,red] ([yshift=-0.2em]softmax3.east) .. controls +(east:2\base) and +(west:\base) .. (output.west);
\end{tikzpicture}
\end{center}
......@@ -1600,7 +1600,7 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
\end{center}
{\scriptsize\begin{tabular}{l}
*$x_t$: 上一层的输出，$h_t$: 同一层上一时刻的隐藏状态\\
*$x_t$: 前一层的输出，$h_t$: 同一层上一时刻的隐藏状态\\
*$c_t$: 同一层上一时刻的记忆
\end{tabular}}
\end{frame}
......@@ -2546,7 +2546,7 @@ $\textrm{``you''} = \argmax_{y} \textrm{P}(y|\textbf{s}_1, \alert{\textbf{C}})$
\item 翻译出``world''的时候``世界''的权重很大
\end{enumerate}
\item 互译的词通常都会产生较大的注意力权重
\item 注意力的权重包含了词对齐的信息
\item 注意力的权重一定程度上反应了词语间的对应关系
\end{itemize}
\begin{center}
\hspace*{\fill}
......@@ -2823,8 +2823,8 @@ $\textrm{``you''} = \argmax_{y} \textrm{P}(y|\textbf{s}_1, \alert{\textbf{C}})$
\begin{itemize}
\item LSTM遗忘门偏置初始为1，也就是始终选择遗忘记忆$c$，可以有效防止初始时$c$里包含的错误信号传播后面所有时刻
\item 网络的其他偏置一般都初始化成0，可以有效防止加入过大或过小的偏置后使得激活函数的输出跑到``饱和区''，也就是梯度接近0的区域，使得训练一开始就无法跳出局部极小
\item 网络的权重矩阵$W$一般使用Xavier参数初始化方法，可以有效稳定训练过程，特别是对于比较``深''的网络$$W \sim \mathcal{U}(-\sqrt{\frac{6}{d_{\mathrm{in}}+d_{\mathrm{out}}}},\sqrt{\frac{6}{d_{\mathrm{in}}+d_{\mathrm{out}}}})$$
\item $d_{\mathrm{in}}$和$d_{\mathrm{out}}$分别是$W$的输入和输出的维度大小，经典的论文\\
\item<2-> 网络的权重矩阵$W$一般使用Xavier参数初始化方法，可以有效稳定训练过程，特别是对于比较``深''的网络$$W \sim \mathcal{U}(-\sqrt{\frac{6}{d_{\mathrm{in}}+d_{\mathrm{out}}}},\sqrt{\frac{6}{d_{\mathrm{in}}+d_{\mathrm{out}}}})$$
$d_{\mathrm{in}}$和$d_{\mathrm{out}}$分别是$W$的输入和输出的维度大小，参考论文\\
\textbf{Understanding the difficulty of training deep feedforward neural networks}\\
\textbf{Glorot, X., \& Bengio, Y., 2010, In Proc of AISTATS}
\end{itemize}
......@@ -2844,7 +2844,7 @@ $\textrm{``you''} = \argmax_{y} \textrm{P}(y|\textbf{s}_1, \alert{\textbf{C}})$
\end{tabular}
\end{center}
\item 因此需要快速得到模型看一下初步效果，选择Adam
\item 若是需要在一个任务上得到最优的结果，选择SGD
\item<2-> 若是需要在一个任务上得到最优的结果，选择SGD
\begin{itemize}
\item 需要注意的是，训练RNN的时候，我们通常会遇到梯度爆炸的问题，也就是梯度突然变得很大，这种情况下需要使用``梯度裁剪''来防止梯度$\pi$超过阈值$$\pi'=\pi \cdot \frac{\mathrm{threshold}}{\max(\mathrm{threshold},\parallel \pi \parallel_2)}$$
\item 其中$\mathrm{threshold}$是手工设定的梯度大小阈值，$\parallel \cdot \parallel_2$是L2范数
......@@ -2858,9 +2858,11 @@ $\textrm{``you''} = \argmax_{y} \textrm{P}(y|\textbf{s}_1, \alert{\textbf{C}})$
\item 不同优化器需要的学习率不同，比如Adam一般使用$0.001$或$0.0001$，而SGD则在$0.1\sim 1$之间挑选
\item 但是无论使用哪个优化器，为了保证训练又快又好，我们通常都需要根据当前的更新次数来调整学习率的大小
\begin{itemize}
\item 学习率预热：模型训练初期，梯度通常很大，直接使用很大的学习率很容易让模型跑偏，因此需要学习率有一个从小到大的过程
\item 学习率衰减：模型训练接近收敛的时候，使用大学习率会很容易让模型错过局部极小，因此需要学习率逐渐变小来逼近局部最小
\item<2-> 学习率预热：模型训练初期，梯度通常很大，直接使用很大的学习率很容易让模型跑偏，因此需要学习率有一个从小到大的过程
\item<2-> 学习率衰减：模型训练接近收敛的时候，使用大学习率会很容易让模型错过局部极小，因此需要学习率逐渐变小来逼近局部最小
\end{itemize}
\visible<2->{
\begin{center}
\begin{tikzpicture}
\footnotesize{
......@@ -2880,14 +2882,15 @@ $\textrm{``you''} = \argmax_{y} \textrm{P}(y|\textbf{s}_1, \alert{\textbf{C}})$
}
\end{tikzpicture}
\end{center}
}
\end{itemize}
\end{frame}
\begin{frame}{训练 - 加速}
\begin{itemize}
\item 万事俱备，只是为什么训练这么慢？\visible<2>{\alert{- RNN需要等前面所有时刻都完成计算以后才能开始计算当前时刻的输出}}
\item 我有钱，是不是多买几台设备会更快？\visible<2>{\alert{- 可以，但是需要技巧，而且也不是无限增长的}}
\item<2> 使用多个设备并行计算进行加速的两种方法
\item 万事俱备，只是为什么训练这么慢？\visible<2->{\alert{- RNN需要等前面所有时刻都完成计算以后才能开始计算当前时刻的输出}}
\item 我有钱，是不是多买几台设备会更快？\visible<2->{\alert{- 可以，但是需要技巧，而且也不是无限增长的}}
\item<3> 使用多个设备并行计算进行加速的两种方法
\begin{itemize}
\item 数据并行：把``输入''分到不同设备上并行计算
\item 模型并行：把``模型''分到不同设备上并行计算
......@@ -2901,7 +2904,8 @@ $\textrm{``you''} = \argmax_{y} \textrm{P}(y|\textbf{s}_1, \alert{\textbf{C}})$
模型并行 & \specialcell{l}{可以对很大的模型进行\\运算} & \specialcell{l}{只能有限并行，比如\\多少层就多少个设备} \\
\end{tabular}
\end{center}
\item<2> 这两种方法可以一起使用！！！
\vspace{0.5em}
\item<3> 这两种方法可以一起使用！！！
\end{itemize}
\end{frame}
......@@ -4578,7 +4582,7 @@ PE_{(pos,2i+1)} = cos(pos/10000^{2i/d_{model}})
}
\visible<3->{
\filldraw [fill=blue!20,draw,thick,fill opacity=0.85] ([xshift=-0.9em,yshift=0.5em]a15.north west) -- ([xshift=0.5em,yshift=-0.9em]a51.south east) -- ([xshift=0.5em,yshift=0.5em]a55.north east) -- ([xshift=-0.9em,yshift=0.5em]a15.north west);
\node[anchor=west] (labelmask) at ([xshift=0.3em,yshift=0.5em]a23.north east) {Mask};
\node[anchor=west] (labelmask) at ([xshift=0.3em,yshift=0.5em]a23.north east) {Masked};
\node [rounded corners=0.3em,anchor=west,fill=blue!20] (mask) at ([xshift=0.1em]add.east) {\large{$Mask$}};
}
......
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