poster.tex

% Written by Daina Chiba (daina.chiba@gmail.com).
% It was mostly copied from two poster style files:
% beamerthemeI6pd2.sty written by
%	 	Philippe Dreuw <dreuw@cs.rwth-aachen.de> and 
% 		Thomas Deselaers <deselaers@cs.rwth-aachen.de>
% and beamerthemeconfposter.sty written by
%     Nathaniel Johnston (nathaniel@nathanieljohnston.com)
%		http://www.nathanieljohnston.com/2009/08/latex-poster-template/
\documentclass[final]{beamer}
\usepackage[orientation=landscape,size=a0,scale=1.2,debug]{beamerposter}
\mode<presentation>{\usetheme{OpenNMT}}
\usepackage[english]{babel}
\usepackage[latin1]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{amsmath,amsthm, amssymb, latexsym}
\usepackage{exscale}
\usepackage{array,booktabs,tabularx}
\newcolumntype{Z}{>{\centering\arraybackslash}X} % centered tabularx columns

% comment
\newcommand{\comment}[1]{}

% (relative) path to the figures
\graphicspath{{figs/}}

\newlength{\columnheight}
\setlength{\columnheight}{105cm}
\newlength{\sepwid}
\newlength{\onecolwid}
\newlength{\twocolwid}
\newlength{\threecolwid}

\setlength{\sepwid}{0.024\paperwidth}
\setlength{\onecolwid}{0.24\paperwidth}
\setlength{\twocolwid}{0.4\paperwidth}
\setlength{\threecolwid}{0.19\paperwidth}

\title{\huge OpenNMT: Open-Source Toolkit for Neural Machine Translation}
\author{Guillaume Klein$^\dagger$, Yoon Kim$^*$, Yuntian Deng$^*$, Jean Senellart$^\dagger$, Alexander M. Rush$^*$}
\institute[]{Harvard University$^*$, SYSTRAN $^\dagger$}
\def\conference{2017 Annual Meeting of the Association for Computational Linguistics}
\def\yourEmail{opennmt.net}

\begin{document}
\begin{frame}[t]
  \begin{columns}[t]
    \begin{column}{\onecolwid}


      \begin{block}{Features}
        {\bf OpenNMT implements many additional features on top of the standard sequence-to-sequence model:}

        \vskip1ex

        \begin{itemize}
        \item Model variants: bidirectional encoder, convolutional encoder, variational dropout, attention parameterization, copying mechanisms, etc.
        \item Factored input representation for richer features.
        \item Tokenization and data preparation tools.
          \item Multi-GPU training.
%        \item learning rate decay strategies.
%        \item advanced model retraining and adaptation.
        \item Beam search length normalization.
        \item \ldots~and many more!
        \end{itemize}

      \end{block}

      \vskip2ex

      \begin{block}{Competitive System}
        OpenNMT achieves competitive results against other systems, e.g. in the recent WMT 2017 translation task:

        \vskip2ex

        \begin{table}
          \begin{tabular}{l r}
            {\bf System} & {\bf BLEU-cased} \\
            \midrule
            uedin-nmt-ensemble & 28.3 \\
            LMU-nmt-reranked-wmt17-en-de & 27.1 \\
            {\bf SYSTRAN-single} & {\bf 26.7} \\
            \bottomrule
          \end{tabular}
          \vskip1ex
          \caption{Top 3 on English-German \emph{newstest2017}.}
        \end{table}

        \vskip2ex

        OpenNMT is optimized for training time and memory efficiency.

        \vskip2ex

        \begin{table}
          \begin{tabular}{l c c r}
            {\bf System} & {\bf Training (WPS)} & {\bf Inference (WPS)} & {\bf BLEU} \\
            \midrule
            Nematus & 3221 & 252 & 18.25 \\
            OpenNMT & 5254 & 457 & 19.34 \\
            \bottomrule
          \end{tabular}
          \vskip1ex
          \caption{Comparison of source words per second (WPS) on similar model size.}
        \end{table}

      \end{block}

      \vskip2ex

      \begin{block}{Extensions}
        {\bf OpenNMT supports other tasks than machine translation:}

        \vskip1ex

        \begin{itemize}
        \item Sequence tagging.
        \item Language modeling.
        \item Speech-to-text, using a pyramidal RNN encoder.
        \item Image-to-text, using a combination of CNN and RNN layers:


        \emph{Im2Text} (\alert{github.com/OpenNMT/Im2Text}) is an extension that can be used for image captioning, optical character recognition, or \LaTeX~decompilation:
\end{itemize}
        \vskip2ex

        \begin{figure}
          \begin{center}
            \includegraphics[width=1.0\onecolwid]{im2text}
          \end{center}
        \end{figure}

      \end{block}

    \end{column}

    \begin{column}{\twocolwid}
      \vskip3ex

      \begin{figure}
        \begin{center}
          \includegraphics[width=\twocolwid]{seq2seq}
          \caption{Sequence-to-sequence model with attention.}
        \end{center}
      \end{figure}

      \vskip3ex

      \begin{alertblock}{}
        \baselineskip=.7\baselineskip
        \vskip1ex
        {\bf OpenNMT is an industrial-strength, open-source neural machine translation ecosystem featuring:}
        \vskip1ex
        \begin{itemize}
        \item Ready-to-use and highly configurable implementations in \emph{Torch} and \emph{PyTorch}.
        \item State-of-the-art translation accuracy and competitive training efficiency.
        \item Implementations of latest advances in neural machine translation and sequence-to-sequence learning.
        \item Standalone and dependency-free inference engine in C++.
        \item Rich set of model and training options covering a large set of needs of academia and industry.
        \item Extensions to allow other sequence generation tasks such as summarization, image-to-text, and speech-to-text.
        \end{itemize}
        \vskip1ex
      \end{alertblock}

      \vskip3ex

      \begin{columns}[t,totalwidth=\twocolwid]
        \begin{column}{\threecolwid}
          \begin{block}{Approach}

            Neural machine translation (NMT) is a simple approach for machine translation based on neural networks that has led to remarkable improvements---particularly in terms of human evaluation---compared to rule-based and statistical machine translation (SMT) systems.


            \vskip2ex

            OpenNMT implements the attention-based encoder-decoder architecture that models the probability of a target sentence $w_{1:T}$ given a source sentence $x_{1:S}$ as:

            \begin{equation*}
              p(w_{1:T}| x) =  \prod_{t=1}^T p(w_t| w_{1:t-1}, x; \theta)
            \end{equation*}

            \vskip2ex

            The encoder and decoder are usually parameterized with LSTM recurrent neural networks.
          \end{block}
        \end{column}

        \begin{column}{\threecolwid}
          \begin{block}{Platforms}
            {\bf OpenNMT is an ecosystem based on multiple technologies and frameworks:}
            \vskip1ex
            \begin{itemize}
            \item \alert{OpenNMT}: the original full-featured project in \emph{LuaTorch}, focusing on maintainability, user support, and production.
            \item \alert{OpenNMT-py}: a \emph{PyTorch} clone of OpenNMT, focusing on research and modularity.
            \item \alert{CTranslate}: an inference engine for OpenNMT models in C++ and \emph{Eigen}, focusing on embedded and production environments.
            \end{itemize}
          \end{block}
        \end{column}
      \end{columns}

    \end{column}

    \begin{column}{\onecolwid}

      \begin{block}{Additional Resources}
        {\bf OpenNMT provides additional resources including:}
        \vskip1ex

        \begin{itemize}
        \item A documentation portal (\alert{opennmt.net/OpenNMT}) for beginners to advanced users describing data preparation, models, training strategies, command line options, pre-trained models, etc.
        \item Visualization tools for debugging or understanding.

        \begin{figure}
          \begin{center}
            \includegraphics[width=0.5\onecolwid]{beam_search}
            \caption{Beam search visualization}
          \end{center}
        \end{figure}
        \end{itemize}

      \end{block}

      \vskip2ex

      \begin{block}{Industry Deployment}
        \begin{figure}
          \begin{center}
            \includegraphics[width=0.45\onecolwid]{demo}
            \caption{Live demo of OpenNMT}
          \end{center}
        \end{figure}

        OpenNMT is robust and has been deployed for production by SYSTRAN, a major translation services provider. SYSTRAN is using OpenNMT for its Pure Neural\texttrademark~Machine Translation offering which enables higher translation quality in existing services.
      \end{block}

      \vskip2ex

      \begin{block}{Active Community}
        OpenNMT is also a community around machine translation and language modeling. The forum (\alert{forum.opennmt.net}) counts more than 200 users with daily questions on how to improve or adapt their systems and training procedures.

        \vskip1ex

        \begin{figure}
          \begin{center}
            \includegraphics[width=0.45\onecolwid]{stars}
            \caption{GitHub statistics as of July 18\textsuperscript{th}, 2017.}
          \end{center}
        \end{figure}

        A user survey showed that users are equally coming from industry and academia, suggesting that the framework is well-suited for many use cases.
      \end{block}

      \vskip2ex

      \begin{block}{Open-source Landscape}
        Landscape for neural machine translation (and sequence-to-sequence learning in general) software has been growing and several excellent open-source alternatives exist, such as:

        \vskip1ex

        \begin{itemize}
        \item \emph{Nematus} (\alert{github.com/EdinburghNLP/nematus})
        \item \emph{Marian NMT} (\alert{marian-nmt.github.io})
        \item \emph{Neural Monkey} (\alert{github.com/ufal/neuralmonkey})
        \item \emph{Google seq2seq} (\alert{google.github.io/seq2seq})
        \end{itemize}
      \end{block}

    \end{column}
  \end{columns}

\end{frame}
\end{document}