summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
-rw-r--r--report/introduction.tex4
1 files changed, 4 insertions, 0 deletions
diff --git a/report/introduction.tex b/report/introduction.tex
index adcd15b0..21e0e907 100644
--- a/report/introduction.tex
+++ b/report/introduction.tex
@@ -115,6 +115,10 @@ We were able to show that each of these techniques could lead to faster decoding
Chapter \ref{chap:decoding} describes this work.
\paragraph{3) Discriminative training labelled SCFG translation models}
+The third stream of the workshop focussed on implementing discriminative training algorithms for the labelled SCFG translation models produced by our unsupervised grammar induction algorithms.
+Though the existing MERT \cite{och02mert} training algorithm is directly applicable to these grammars, it doesn't allow us to optimise models with large numbers of fine grained features extracted from the labels we've induced.
+In order to maximise the benefit from our induced grammars we explored and implemented discriminative training algorithms capable of handling thousands, rather than tens, of features.
+The algorithms we explored were Maximum Expected Bleu \cite{smith,li} and MIRA \cite{chiang}.
Chapter \ref{chap:training} describes this work.
The remainder of this introductory chapter provides a formal definition of SCFGs and describes the language pairs that we experimented with.