diff options
Diffstat (limited to 'report/pr-clustering')
| -rw-r--r-- | report/pr-clustering/posterior.tex | 44 |
1 files changed, 33 insertions, 11 deletions
diff --git a/report/pr-clustering/posterior.tex b/report/pr-clustering/posterior.tex index eb53e915..ea8560c1 100644 --- a/report/pr-clustering/posterior.tex +++ b/report/pr-clustering/posterior.tex @@ -316,23 +316,43 @@ distribution is obtained from Collins parser trained on Penn Treebank. Since not all phrases are constituents, we ignored -phrases that don't correspond any constituents. +phrases that don't correspond to any constituents. + +We conducted experiments with various data pre-processing +and tried different models. Number of phrase categories +is fixed at $25$. We chose to only look at +the target side language. The context is set +to be $1$ word to the left and to the right of the phrase. +We chose such a setting because it emperically works better +in the pipeline than other variations. This is also +the case for non-parametric methods. The models as we discussed +in previous sections are EM, EM with sparsity constraint, +agreement of two models in reverse directions and agreement +of two models trained on two languages. We tried our models +with word classes as well. In the context, each word is +replaced with a word class unsupervisedly learned from the data. +The results are shown in Table \ref{tab:results}. \begin{table}[h] \centering - \begin{tabular}{ |*{3}{c|} } + \begin{tabular}{ |*{4}{c|} } \hline - model & BLEU & H(Gold$|$Predicted)\\ + \multicolumn{2}{|c|}{model} & BLEU & H(Gold$|$Predicted)\\ \hline - hiero & 21.1 & 5.77\\ - hiero+POS & 22.3 & 1.00 \\ - SAMT & 24.5 & 0.00 \\ + \multicolumn{2}{|c|}{hiero} & 21.1 & 5.77\\ + \multicolumn{2}{|c|}{hiero+POS} & 22.3 & 1.00 \\ + \multicolumn{2}{|c|}{SAMT} & 24.5 & 0.00 \\ \hline - EM & 20.9 & 2.86 \\ - PR $\sigma=100$ & 21.7 & 2.36 \\ - agree language & 21.7 & 2.68 \\ - agree direction & 22.1 & 2.35\\ - non-parametric & 22.2 & ?\\ + \multirow{2}{*}{EM} & words & 20.9 & 2.85 \\ + & word classes & 21.54 & 2.86 \\ \hline + \multirow{2}{*}{PR $\sigma=100$}&words & 21.1 & 2.56 \\ + &word classes & 21.7 & 2.36 \\ \hline + \multirow{2}{*}{agree language}&word & 21.7 & 2.80 \\ + &word classes & 21.4 & 2.69\\ \hline + \multirow{2}{*}{agree direction}&word & 21.6 & 2.48\\ + &word classes &22.1 &2.36 \\ \hline + \multirow{2}{*}{non-parametric}&word & 22.0 & 2.86\\ + & word classes&22.3&2.27\\ \hline \end{tabular} \caption @@ -356,6 +376,8 @@ in the beginning of this chapter. PR $\sigma=100$ is posterior regularization model with sparsity constraint explained in Section \ref{sec:pr-sparse}. $\sigma$ is the constant controls strongness of the constraint. +We picked $\sigma$ by trying different values ranging from +$1$ to $100$. Agree language and agree direction are models with agreement constraints mentioned in Section \ref{sec:pr-agree}. Non-parametric is non-parametric model introduced in the previous chapter. |