Common use of Comparison with Supervised Parsing Models Clause in Contracts

Comparison with Supervised Parsing Models. ‌ The headline MT findings from the three annotation systems described in this thesis are summarized in Table 5.6. Once we include the tree transformation system, the trend from Table 5.1 still holds, with larger agreement score improvements resulting in larger syntactic MT improvements. However, given that the joint models improve agreement in a fairly different way from the tree transformation system, one obvious next step is to see if these methods can be combined. To test whether combination of these systems is helpful, we tried learning tree transfor- mations from the output of each of the joint models. We then applied the appropriate set of learned transformations to reannotate the output of these models. The results are shown in Table 5.7. Unfortunately, combination is not as successful as we had hoped. The output of the bilingual reranker does benefit from learned transformations, but the aggregate result is not quite as good for MT performance as simply learning to transform the output of the baseline monolingual English parser. Learning transformations for the output of the joint parsing and alignment model yields the highest total agreement score by far, but though tuning set MT performance improves, test set performance does not.