diff options
Diffstat (limited to 'report/intro_slides/final_slides.tex')
| -rw-r--r-- | report/intro_slides/final_slides.tex | 341 |
1 files changed, 1 insertions, 340 deletions
diff --git a/report/intro_slides/final_slides.tex b/report/intro_slides/final_slides.tex index 521dc003..753f4138 100644 --- a/report/intro_slides/final_slides.tex +++ b/report/intro_slides/final_slides.tex @@ -136,7 +136,7 @@ %\vspace{1.0cm} \begin{exampleblock}{Urdu $\rightarrow$ English} \begin{figure} - {\centering \includegraphics[scale=0.55]{urdu-bw.pdf}} + {\centering \includegraphics[scale=0.55]{urdu.pdf}} \end{figure} \vspace{0.10cm} \end{exampleblock} @@ -458,272 +458,6 @@ We will predominately evaluate using BLEU, but also use automatic structured met \end{frame} -%\begin{frame}[t] -%\frametitle{Inducing a STSG given an observed tree:} -%\only<1>{\frametitle{Inducing a STSG given an observed tree:}} -%\only<2->{\frametitle{Existing approach (Galley et al. 2004):}} -% -%\begin{center} -% \only<1>{\hspace{1mm}\includegraphics[scale=0.45]{full_of_fun_slides_start.pdf}} -% \only<2>{\includegraphics[scale=0.45]{full_of_fun_slides_waligned.pdf}} -% \only<3>{\vspace{-2mm}\includegraphics[scale=0.45]{full_of_fun_slides_waligned_overlay.pdf}} -%% \only<4>{\includegraphics[scale=0.4]{full_of_fun_slides_third.pdf}} -%% \only<5>{\includegraphics[scale=0.4]{full_of_fun_slides_forth.pdf}} -% -% \only<1>{Training instance} -% \only<2>{Step 1: word alignment} -% \only<3>{Step 2: rule extraction heuristic} -%% \only<4>{Step 2: the rules extracted} -%% \only<5>{Step 3: estimate a grammar} -%\end{center} -%\end{frame} - - -% Il ne veut pas travailler - - -%\begin{frame}[t]{Models of translation} -%\begin{block}{Hierarchical} -% \begin{center} -% \includegraphics[scale=0.55]{JeNeVeuxPasTravailler-Hiero-labelled.pdf} -% \hspace{0.3in} -% \includegraphics[scale=0.55]{JeVeuxTravailler-Hiero-labelled.pdf} -% \end{center} -%\end{block} -%\begin{itemize} -%\item \alert{AIM: Implement a large scale open-source synchronous constituent labelling system.} -%\item \alert{AIM: Investigate and understand the relationship between synchronous constituency and SMT performance.} -%\end{itemize} -%\end{frame} -% -%\begin{frame}[t]{Models of translation} -%\begin{block}{Hierarchical} -% \begin{center} -% \includegraphics[scale=0.5]{JeNeVeuxPasTravailler-Hiero-labelled.pdf} -% \includegraphics[scale=0.5]{IlNeVeutPasTravailler-Hiero-labelled.pdf} -% \end{center} -% \vspace{0.001in} -%\end{block} -%\begin{itemize} -%\item \alert{AIM: Implement a large scale open-source synchronous constituent labelling system.} -%\item \alert{AIM: Investigate and understand the relationship between synchronous constituency and SMT performance.} -%\end{itemize} -%\end{frame} - -\begin{frame}[t]{Unsupervised grammar induction} -There has been significant research into monolingual grammar induction: -\vspace{0.1in} -\alert{Constituent context is a prime indicator of constituency.} -\begin{unpacked_itemize} -\item Alexander Clark. Unsupervised induction of stochastic context-free grammars using distributional clustering, 2001 -\item Dan Klein and Chris Manning. A Generative Constituent-Context Model for Improved Grammar Induction, 2002 -\end{unpacked_itemize} -\vspace{0.1in} -\alert{We can formalise this notion in algebraic structures} -\begin{itemize} -\item Alexander Clark. A learnable representation for syntax using residuated lattices, 2009 -\end{itemize} -\vspace{0.1in} -Deep connections to unsupervised word sense disambiguation, thesaurus extraction etc. -\end{frame} - -%\begin{frame}[t]{Monolingual grammar induction} -%Induce bracketing phrase-structure grammars: -% \includegraphics[scale=1]{klein_ccm.pdf} -% -%\vspace{2ex} -%And dependency trees: \\ -% \includegraphics[scale=1]{klein_dependency.pdf} -% -%\vspace{2ex} -%Informed by constituent context: surrounding words are a good indicator of substitutability -%\end{frame} - - -\begin{frame}[t]{SCFG Grammar Induction} -%\vspace{1.0cm} -\begin{exampleblock}{Distributional Hypothesis} -\begin{quote} -\emph{Words that occur in the same contexts tend to have similar meanings} -\end{quote} -\hfill (Zellig Harris, 1954) -\end{exampleblock} - -\vspace{3ex} - -We will leverage this in a translation setting: -\begin{itemize} - \item Use the contexts to \alert{cluster} translation units into groups - \item Units in the same group expected to be semantically and syntactically similar - \item Then use these cluster labels to guide translation - \begin{itemize} - \item lexical selection: translating ambiguous source word/s - \item reordering: consistent syntactic patterns of reordering - \end{itemize} -\end{itemize} -\end{frame} - -\begin{frame}[t]{Monolingual Example} -Task: cluster words into their parts-of-speech. \\ - -\vspace{1ex} -Illustrate by starting with the word `deal' (noun or verb): - -\only<1>{\includegraphics[width=\columnwidth]{deal_first.pdf} \\ Step 1: Find contexts for `deal'} -\only<2->{\includegraphics[width=\columnwidth]{deal.pdf} \\ Step 2: Find other words which occur in these contexts} -%\only<3>{\includegraphics[width=\columnwidth]{deal_more.pdf} \\ \ldots continue to expand} - -\only<3>{ -\vspace{1ex} -Notice that the instances of deal can be split into two connected sub-graphs: -\begin{itemize} - \item noun: the left two contexts ``a \ldots with'' and ``a \ldots that'' - \item verb: the right two contexts ``to \ldots with'' and ``not \ldots with'' - \item neighbouring words of these contexts share the same PoS -\end{itemize} -} - -\end{frame} - -%\begin{frame}[t]{More Formally} -% -%Construct a bipartite graph -%\begin{itemize} -% \item Nodes on the top layer denote word types (bilingual phrase pairs) -% \item Nodes on the bottom layer denote context types (monlingual/bilingual words) -% \item Edges connect words and their contexts -%\end{itemize} -% -%\includegraphics[width=\columnwidth]{bipartite.pdf} -% -%\end{frame} - -\begin{frame}[t]{Clustering} - -Task is to cluster the graph into sub-graphs. Nodes in the sub-graphs should be -\begin{itemize} -\item strongly connected to one another -\item weakly connected to nodes outside the sub-graph -\item could formulate as either \emph{hard} or \emph{soft} clustering -\end{itemize} -Choose \alert{soft clustering} to allow for syntactic and semantic ambiguity - -\centering -\includegraphics[width=0.7\columnwidth]{bipartite_lda.pdf} - -\end{frame} - -\begin{frame}[t]{Constituency and context} -\vspace{0.25in} -\begin{center} -\only<1>{ - \includegraphics[scale=0.5]{WantTo_Veux_context.pdf} - \includegraphics[scale=0.5]{WantTo_Veux_context2.pdf} -} -\only<2>{ - \includegraphics[scale=0.5]{WantTo_Veux_context_split.pdf} - \includegraphics[scale=0.5]{WantTo_Veux_context2_split.pdf} -} -\only<3>{ - \includegraphics[scale=0.5]{WantTo_Veux_context_split_mono.pdf} - \includegraphics[scale=0.5]{WantTo_Veux_context2_split_mono.pdf} -} -\end{center} -\vspace{0.1in} -%\only<1>{ -% There has been significant research into monolingual grammar induction: -% \vspace{0.1in} -% \begin{unpacked_itemize} -% \item Alexander Clark. Unsupervised induction of stochastic context-free grammars using distributional clustering, 2001 -% \item Dan Klein and Chris Manning. A Generative Constituent-Context Model for Improved Grammar Induction, 2002 -% \end{unpacked_itemize} -% \alert{Constituent context is a prime indicator of constituency.} -%} -%\only<1>{ -\begin{unpacked_itemize} -\item Design and apply large scale scale clustering and topic modelling algorithms (LDA, HDPs, HPYPs etc), -\item identify sets of frequent contexts that distinguish synchronous constituent properties. -\item Motivated by successful models of monolingual grammar induction, -\item deep connections to unsupervised word sense disambiguation, thesaurus extraction etc. -\end{unpacked_itemize} -%} -\end{frame} - -\begin{frame}[t]{Latent Dirichlet Allocation (LDA)} - -LDA is a generative model which treats documents as bags of words -\begin{itemize} - \item each word is assign a \alert{topic} (cluster tag) - \item words are generated from a topic-specific multinomial - \item topics are \alert{tied} across a document using a Dirichlet prior - \item $\alpha < 1$ biases towards \alert{sparse} distributions, i.e., topic reuse - \item inferred $\theta_d$ describes a document and $\phi_t$ describes a topic -\end{itemize} - -\vspace{-3ex} -\includegraphics[scale=0.55]{lda.pdf} - -\end{frame} - -\begin{frame}[t]{LDA over Contexts} - -Generative story: -\begin{itemize} - \item for each word type $w$ - \item for each of the $L$ contexts - \item first we draw a topic $t$, then generate the context $\vec{c}$ given the topic - \item the Dirichlet prior ties the topics for each $w$ - \item we're primarily interested in the learnt $\theta$ values -\end{itemize} - -\includegraphics[scale=0.4]{context_lda.pdf} - -\end{frame} - -\begin{frame}[t]{Scalable grammar extraction with MapReduce} -\begin{itemize} -\item Divide and conquer approach to...counting -\begin{itemize} -\item map function $\mathcal{M}(x) \rightarrow \langle k_1, v_1 \rangle, \langle k_2, v_2 \rangle, \ldots$ -\item write a reduce function $\mathcal{R}(k_i : v_7, v_{13} , \ldots) \rightarrow \langle k_i, \overline{v} \rangle$ -\end{itemize} -\end{itemize} -\begin{center} - \includegraphics[scale=0.4]{mroutline.pdf} -\end{center} -\end{frame} -\begin{frame}[t]{Scalable grammar extraction with MapReduce : mapper} -\begin{center} - \includegraphics[scale=0.4]{mapper.pdf} -\end{center} -\end{frame} - -\begin{frame}[t]{Scalable grammar extraction with MapReduce : reducer} -\begin{center} - \includegraphics[scale=0.4]{reducer.pdf} -\end{center} -\end{frame} - -\begin{frame}[t]{Scalable grammar extraction with MapReduce : Hadoop} -\begin{center} - \includegraphics[scale=0.4]{hadoop-extract.pdf} -\end{center} -\end{frame} - -\begin{frame}[t]{Scalable grammar extraction with MapReduce : Hadoop} -\begin{center} - \includegraphics[scale=0.4]{hadoop-extract-arrows.pdf} -\end{center} -\end{frame} - - -%\begin{frame}[t]{Discriminative training} -%\begin{unpacked_itemize} -%\item MIRA -%\item Expected loss minimisation. -%\end{unpacked_itemize} -%\end{frame} - \begin{frame}[t]{Language pairs (small)} \begin{itemize} @@ -770,79 +504,6 @@ Generative story: \end{itemize} \end{frame} -%\begin{frame}[t]{Draft Schedule} -%\begin{itemize} -%\item Pre-workshop: -% \begin{itemize} -% \item Collect existing open-source tools for synchronous grammar induction, -% \item Collect corpora across a range of translations conditions: small, large, low-density languages etc. -% \item Implement phrase and context extraction algorithms. -% \item Design the integration of various existing approaches into the decoders. -% \end{itemize} -%\item Week 1: -% \begin{itemize} -% \item Optimise and reconfigure decoders to handle labelled synchronous grammars, -% \item Perform a empirical study of synchronous constituency models. -% \end{itemize} -%\end{itemize} -%\end{frame} - -%\begin{frame}[t]{Draft Schedule} -%\begin{itemize} -%\item Week 2-3: -% \begin{itemize} -% \item Continue optimising decoder to handle labelled synchronous grammars, -% \item Implement unsupervised label induction algorithms, initially inducing a single label per-phrase. -% \item Extend to ''topic"-modelling style representation where a phrase may have multiple labellings. -% \item Perform experimental comparison of existing synchronous grammar translation models. -% \end{itemize} -%\item Week 3-6: -% \begin{itemize} -% \item Perform experimental comparison of unsupervised synchronous grammar translation models. -% \item Extend the evaluation to small/big data sets, hi-density vs. low-density language pairs. -% \item Create ``semi-supervised'' models combining knowledge from treebank parser into the unsupervised algorithms. -% \item Wrap-up and write final report. -% \end{itemize} -%\end{itemize} -%\end{frame} - - -\begin{frame}[t]{Pre-workshop experiments} -\vspace{0.25in} -We have implemented a baseline constituent modelling and distrbuted grammar extraction pipeline. Initial results on the small BTEC corpora: - -\vspace{0.25in} -\begin{exampleblock} -\footnotesize -\centering -\begin{tabular}{lcccccc} -\toprule -Categories & \small 1-gram & \small 2-grams & \small 3-grams & \small 4-grams & \small BP & BLEU \\ -\midrule -1 & \small 84.7 & \small 62.0 & \small 47.2 & \small 36.4 & \small 0.969 & \textcolor{blue}{53.10} \\ -10 & \small 84.0 & \small 60.9 & \small 46.4 & \small 35.9 & \small 0.979 & \textcolor{red}{52.88} \\ -25 & \small 84.4 & \small 61.8 & \small 47.6 & \small 36.7 & \small 0.973 & \textcolor{blue}{53.47} \\ -50 & \small 84.8 & \small 61.2 & \small 46.6 & \small 36.2 & \small 0.971 & \textcolor{red}{52.83} \\ -100 & \small 83.5 & \small 60.1 & \small 45.7 & \small 35.3 & \small 0.972 & \textcolor{red}{51.86} \\ -\bottomrule -\end{tabular} -\end{exampleblock} -\end{frame} - - -%{\centering -%A unique opportunity to bring together researchers operating at the coal face of SMT development with leading theoreticians in the field of formal grammar induction. -%} -%\begin{unpacked_itemize} -%\item Understand the relationship between constituent labels and performance in SMT, -%\item Compare monolingual and bilingual induced grammars against parser output in terms of translation quality, -%\item Produce a large scale implementation of the label induction algorithms, -%\end{unpacked_itemize} -%\begin{unpacked_itemize} -%\item \alert{Learn language-pair dependent structure that produces translation performance gains across all language pairs,} -%\item \alert{Initiate a research program that redirects the SMT research community back to language neutral unsupervised systems.} -%\end{unpacked_itemize} - \begin{frame}[t]{Summary} \begin{itemize} |