Year: 2020
Type: conference-paper
Abstract: Relational reasoning is ability to reason about entities and their interactions, which is not applicable for many deep neural networks. Recurrent relational networks, introduced by Palm et al. (2017), increase the complexity of the reasoning tasks they can address [1]. In this paper, we introduce Stacked Attention Recurrent Relational Networks (SARRN) to answer natural language questions from facts, which fundamentally hinge on multiple steps of relational reasoning and improve the ability of reasoning. Our model is a stacked attention model that use recurrent attention to focus on fine-grained parts of the documents. We apply our model to the BaBi tasks, which have a set of proxy tasks that evaluate reading comprehension via question answering. Our model solve 19/20 tasks and the experimental results on the test sets tasks show that our model yields great improvement. We also use qualitative analysis to show the result intuitively.
Institution: China Telecom
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W3036606616