2016/07/20 by Janez Starc, Starc, Janez, Dunja Mladenić +1
Computer Science · Social Sciences · #Artificial Intelligence (cs.AI) #Computation and Language (cs.CL) #Computational and Text Analysis Methods #FOS: Computer and information sciences #Natural Language Processing Techniques #Neural and Evolutionary Computing (cs.NE) #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1607.06025
openalex publication_date 2016/07/20 · openalex created_date 2022/10/06 · openalex updated_date 2026/07/28
Natural Language Inference is an important task for Natural Language\nUnderstanding. It is concerned with classifying the logical relation between\ntwo sentences. In this paper, we propose several text generative neural\nnetworks for generating text hypothesis, which allows construction of new\nNatural Language Inference datasets. To evaluate the models, we propose a new\nmetric -- the accuracy of the classifier trained on the generated dataset. The\naccuracy obtained by our best generative model is only 2.7% lower than the\naccuracy of the classifier trained on the original, human crafted dataset.\nFurthermore, the best generated dataset combined with the original dataset\nachieves the highest accuracy. The best model learns a mapping embedding for\neach training example. By comparing various metrics we show that datasets that\nobtain higher ROUGE or METEOR scores do not necessarily yield higher\nclassification accuracies. We also provide analysis of what are the\ncharacteristics of a good dataset including the distinguishability of the\ngenerated datasets from the original one.\n