Are the existing training corpora unnecessarily large?

Miguel Ballesteros, Jesús Herrera, Virginia Francisco, Pablo Gervás


This paper addresses the problem of optimizing the training treebank data because the size and quality of the data has always been a bottleneck for the purposes of training. In previous studies we realized that current corpora used for training machine learning--based dependency parsers contain a significant proportion of redundant information at the syntactic structure level. Since the development of such training corpora involves a big effort, we argue that an appropriate process for selecting the sentences to be included in them can result in having parsing models as accurate as the ones given when training with bigger -- non optimized corpora (or alternatively, bigger accuracy for an equivalent annotation effort). This argument is supported by the results of the study we carried out, which is presented in this paper. Therefore, this paper demonstrates that the training corpora contain more information than needed for training accurate data--driven dependency parsers.

Texto completo: