Training Set Pruning Sample Clauses

Training Set Pruning. To reduce the time and space requirements for training, we do not always use the full k-best lists. To prune the set T , we rank all the trees in T from 1 to k, according to their log likelihood under the baseline parsing model, and find the rank of the least likely pseudo- gold tree: r∗ = min rank(t) t∈Tˆ Finally, we restrict T based on rank: Tpruned = {t ∈ T|rank(t) ≤ r∗ + ϵ} where ϵ is a free parameter of the pruning procedure. The restricted set Tprruned is con- structed in the same way. When training, we replace the sum over all tree pairs in (T, Tr) in the denominator of (3.6) with a sum over all tree pairs in (Tpruned, Tprruned). The parameter ϵ can be set to any value from 0 to k, with lower values resulting in more efficient training, and higher values resulting in better performance. We set ϵ by empirically determining a good speed/performance tradeoff (see Section 3.5.2).

Training Set Pruning. ‌ To find a good value of the ϵ parameter for training set pruning we tried several different values, using our rapid training setup and testing on the dev set. The results are shown in Table 3.2. We selected 25 as it showed the best performance/speed tradeoff, on average performing as well as if we had done no pruning at all, while requiring only a quarter the memory and CPU time.