Performance of Structural Generalization in Connectionist Japanese Complex Sentence Parser

We propose the connectionist Japanese complex sentence parser JSPAC. JSPAC is composed of three modules, the simple sentence parser, the segmenter and the stack. The segmenter and the stack segment a complex sentence into simple sentences. The simple sentence parser generates a case structure from the segmented simple sentence. After training JSPAC with the learning data which we generated, we evaluate the JSPAC correctness, especially the ability of the structural generalization for unrestricted recursive structures.