diff options
Diffstat (limited to 'report/pr-clustering')
| -rw-r--r-- | report/pr-clustering/posterior.tex | 20 |
1 files changed, 19 insertions, 1 deletions
diff --git a/report/pr-clustering/posterior.tex b/report/pr-clustering/posterior.tex index ea8560c1..8a199e63 100644 --- a/report/pr-clustering/posterior.tex +++ b/report/pr-clustering/posterior.tex @@ -183,7 +183,12 @@ for each phrase $\textbf{p}$: \sum_i \lambda_{\textbf{p}iz}\leq \sigma. \] This dual objective can be optimized with projected gradient -descent. +descent. Notice that each phrase has its own objective and +constraint. The $\lambda$s are not shared acrossed +phrases. Therefore we can optimize the objective +for each phrase separately. It is convenient for parallelizing +the algorithm. It also makes the objective easier to optimize. + The $q$ distribution we are looking for is then \[ q_i(z)\propto P_{\theta}(z|\textbf{p},\textbf{c}_i) @@ -382,3 +387,16 @@ 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. \section{Conclusion} +The posterior regularization framework has a solid theoretical foundation. +It is shown mathematically to balance between constraint and likelihood. +In our experiments, +we used it to enforce sparsity constraint and agreement constraint and +achieved results comparable to non-parametric method that enforces +sparcity through priors. The algorithm is fairly fast if the constraint +can be decomposed into smaller pieces and compute separately. In our case, +the sparsity constraint for phrases can be decomposed into one small optimization +procedure for each phrase. In practice, our algorithm is much +faster than non-parametric models with Gibbs sampling inference. +The agreement +models are even faster because they are performing almost the same amount +of computation as the simple models trained with EM. |