カンファレンス (国際) An Experimental Comparison of the Voted Perceptron and Support Vector Machines in Japanese Analysis Tasks
The Third International Joint Conferenceon Natural Language Processing (IJCNLP 2008)
We examine various aspects of the voted perceptron and support vector machines in classification tasks in NLP rather than ranking tasks. These aspects include training time, accuracy and learning curves. We used Japanese dependency parsing as a main task for experiments, and Japanese word segmentation and bunsetsu (base phrase in Japanese) chunking as auxiliary tasks. In our experiments we have observed that the voted perceptron is comparable to SVM in terms of accuracy and, in addition, as to learning time and prediction speed the voted perceptron is considerably better than SVM.