Teaching Machine How to Think by Natural Language: A study on Machine Reading ComprehensionDownload PDF

Tsung Han Wu, Hung-yi Lee, Yu Tsao, ChaoI, Tuan

27 Sept 2018 (modified: 05 May 2023)ICLR 2019 Conference Withdrawn SubmissionReaders: Everyone
Abstract: Deep learning ends up as a black box, in which how it makes the decision cannot be directly understood by humans, let alone guide the reasoning process of deep network. In this work, we seek the possibility to guide the learning of network in reading comprehension task by natural language. Two approaches are proposed. In the first approach, the latent representation in the neural network is deciphered into text by a decoder; in the second approach, deep network uses text as latent representation. Human tutor provides ground truth for the output of the decoder or latent representation represented by text. On the bAbI QA tasks, we found that with the guidance on a few examples, the model can achieve the same performance with remarkably less training examples.
Keywords: Machine Reading Comprehension
1 Reply

Loading