report.tex.svn-base

moses开源的机器翻译系统

model trained over morphological features.

We extend this model to address the problems of poor generalization by
decomposing the surface-form translation process into two parallel
processes: translation of lemmas and morphological features.  Lemmas
and morphological features are then recombined in a generation process
to create target surface forms. Details of both these models and
experiments we conducted with them are described below.

For both sets of experiments {\tt EuroMini} data was preprocessed to
using FreeLing~\cite{Atserias:2006} for both part of speech tags and
morphological analysis.

\subsection{Explicit Agreement and Coherence Models}

In these experiments, we applied factored models as an extension to
standard phrase-based MT.  In these experiments standard
surface-to-surface translation is performed, but we add stochastic
constraints on possible target hypotheses to limit enforce (in a soft
way) agreement relations.  This is done by generating a latent factor
(in this case morphological features) from each target word that is
hypothesized during decoding.  Hypotheses are then scored with both
standard MT model components and language models trained over
agreement features.  Figure~\ref{fig:latent-factor-check} shows the
configuration of two models described below: Verb/Noun/Preposition
(VNP) and Noun/Determiner/Adjective (NDA).

\begin{figure}[t]
\begin{center}
\includegraphics[width=4cm]{wade-latent-factors} 
\caption{Latent Factor Checking using Agreement Features}
\label{fig:latent-factor-check}
\end{center}
\end{figure}

This configuration was also used with POS tags to ensure long term
coherence.  Table~\ref{tab:LM-models} summarizes different models to explicitly
check for agreement and ensure coherence.

\begin{table}[h]
  \begin{center}
    \begin{tabular}{|l|l|}
      \hline
      \bf Problems Addressed & \bf Model Type \\ \hline \hline
      Explicit Agreement & -- LMs over verbs + subjects \\
                         & -- LMs over nouns + determiners + adjectives \\
      \hline
      Long Span Coherence & -- LMs over POS Tags \\
      \hline
    \end{tabular}
  \end{center}
  \caption{LM-based Agreement and Coherence Models}
  \label{tab:LM-models}
\end{table}
\subsubsection{}

\subsubsection{Verb/Subject and Noun-Phrase Agreement Models}

In a first set of experiments we used features derived from a
morphological analysis of the training data to create n-gram language
models.

We produced two models for experimentation.  In one, NDA, number and
gender features were generated for each noun, determiner and adjective
in a hypothesis during decoding.  Non-NDA words deterministically
generated ``don't care'' features.

In a second model, VNP, features required for Spanish verb-subject
agreement were chosen in addition the identity of preposition in
hypothesized sentences.  The inclusion of prepositions allows us
potentially to model the selection relationship between verbs and
prepositions (though this is not strictly an agreement phenomenon).

Table~\ref{tab:nda-vnp} shows the features used for both these models
and their possible values.

\begin{table}[h]
  \begin{center}
    \begin{tabular}{|l|}
      \hline
      \bf NDA Features \\ \hline \hline
      {\bf Gender:} {\it masc, fem, common, none} \\ \hline
      {\bf Number:} {\it sing, plural, invariable, none} \\ \hline \hline
      \bf VNP Features \\ \hline \hline
      {\bf Number:} {\it sing, plural, invariable, none} \\ \hline
      {\bf Person:} {\it 1p, 2p, 3p, none} \\ \hline
      {\bf Prep-ID:} {\it preposition, none} \\ \hline
    \end{tabular}
  \end{center}
  \caption{Latent Features used for NDA and VNP models}
  \label{tab:nda-vnp}
\end{table}

Since ``don't care'' values can intervene between words with NDA and
VNP features, we also experimented with language models that skip
words lacking features of interest.  This effectively increases the
context length for our n-gram based models and should yield more
robust estimation.  This is shown schematically in
Figure~\ref{fig:skipped-lm-nda-vnp}.  Factors marked with ``X'' are not
scored in VNP or NDA language models.

\begin{figure}[t]
\begin{center}
\includegraphics[width=10cm]{wade-skipped} 
\caption{Latent Factor Checking using Agreement Features}
\label{fig:skipped-lm-nda-vnp}
\end{center}
\end{figure}

Results with models {\tt EuroMini} corpus using the standard 60k-word
3-gram language model and evaluated against the 2005 ACL Workshop
Shared Task are shown in Table~\ref{tab:nda-vnp-perf}.  Both NDA and
VPN models improve performance over the baseline system.  We ran
additional experiments that incorporated all morphological features
from our analyzer and this too improved the performance of the system,
though inclusion of part of speech information did not.  The use of
all morphological features closes the gap between the baseline system
and it's large LM counterpart by 74\%, suggesting that better target
language modeling could compensate for the sparsity of target language
data.

\begin{table}[h]
  \begin{center}
    \begin{tabular}{|l|r|}
      \hline
      \bf Model & \bf BLEU \\ \hline \hline
      {\bf Baseline} & 23.41 \\ \hline
      {\bf Baseline + 950k LM} & 25.10 \\ \hline
      {\it NDA} & 24.47 \\ \hline
      {\it VPN} & 24.33 \\ \hline
      {\bf BOTH} & {\bf 24.54} \\ \hline \hline
      {\it NDA w/skipping} & 24.03 \\ \hline
      {\it VPN w/skipping} & 24.16 \\ \hline \hline
      {\bf All Morph Features} & {\bf 24.66} \\ \hline
      {\it All Morph Features + POS Tag} & 24.25 \\ \hline
    \end{tabular}
  \end{center}
  \caption{Latent Features used for NDA and VNP models}
  \label{tab:nda-vnp-perf}
\end{table}

Unfortunately, the use of skipping models didn't improve performance.
Further experiments will be necessary as interactions with pruning
during decoding may have limited the performance of these systems.

\subsubsection{Lemma-based Models for Translation}

The agreement models describe above use a factored approach to add
statistical constraints on target language sequences.  We attempted to
extend this model by improving the underlying translation modeling.
To do this, we created a parallel translation model in which source
words are decomposed into base lemma and morphological features.  Both
lemmas and morphological features are translated from the source to
the target language.  Target words are then re-composed through a
statistically trained generation process on the target side.  This
process is shown schematically in Figure~\ref{fig:parallel-trans}.
Language models are then applied to both morphological features and
surface forms to constrain the output.

\begin{figure}[t]
\begin{center}
\includegraphics[width=10cm]{wade-parallel-trans} 
\caption{A Factored Model for Parallel Translation of Lemmas and Morphology}
\label{fig:parallel-trans}
\end{center}
\end{figure}

Table~\ref{tab:lemma+morph-results} shows the performance of this
system on a 2005 ACL Workshop Shared Task when a large language model
was used (trained with 950k sentences).  As these
experiments show, the lemma+morphology can lead to improvements through better translation modeling and with the addition of a morphology LM.


\begin{table}[h]
  \begin{center}
    \begin{tabular}{|l|r|}
      \hline
      \bf Model & \bf BLEU \\ \hline \hline
      {\it Baseline} & 25.10 \\ \hline
      {\it Lemma + Morph  } & 25.93 \\ 
      {\it Lemma + Morph plus Morph LM} & 26.11 \\ 
      \hline
    \end{tabular}
  \end{center}
  \caption{Results from Lemma+Morphology Models}
  \label{tab:lemma+morph-r