MaltParser and LIBLINEAR Transition-based dependency parsing with linear classification for feature model optimization

In this thesis, MaltParser has been extended with an interface to a software for large-scale linear classification (LIBLINEAR). This combination was then used for learning and parsing with four different treebanks (Slovene, Danish, Arabic, and Turkish). The initial tests compared the accuracy of MaltParser using LIBLINEAR to that when using LibSVM (MaltParser’s default classifier) with its linear algorithm. The results are significantly different in favor of LibSVM for two of the four treebanks (Danish and Arabic). The LIBLINEAR classifier was then used for optimizing a feature model for each of the four treebanks. These results showed a significant improvement compared to the baseline feature model both when using only LIBLINEAR, and when using the optimized feature model with LibSVM. Learning and parsing times are much smaller when using LIBLINEAR, though accuracy is not quite as high as that of LibSVM’s polynomial kernel. In conclusion, using LIBLINEAR for feature model optimization is a good idea, since it takes advantage of the speed of LIBLINEAR while not sacrificing accuracy when using LibSVM’s polynomial kernel for the final feature model. This also indicates that the MaltParser/LIBLINEAR combination should be explored further, and preferably with larger treebanks.