Deep neural networks for semantic alignments and recognizing textual entailment

• Grant number: 16/02466-5
• Support Opportunities: Scholarships abroad - Research Internship - Doctorate
• Start date: April 11, 2016
• End date: April 10, 2017
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Sandra Maria Aluísio
• Grantee: Erick Rocha Fonseca
• Supervisor: Bernardo Magnini
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Institution abroad: Fondazione Bruno Kessler (FBK), Italy
• Associated to the scholarship: 13/22973-0 - Textual inference applied to Question and Answering Systems, BP.DR

Abstract

Recognizing Textual Entailment (RTE) is a topic of growing interest in the Natural Language Processing (NLP) community. It consists in determining whether one piece of text can be entailed by another one, considering how people would interpret them, and is useful for many NLP applications that deal with texts from different sources. Lately, methods based on long short-term memories (LSTM), a kind of recurrent neural network, have achieved state-of-the-art results in a benchmark dataset. However, research on such methods is still incipient; there is no architecture agreed upon to the be best suited to this task. Also, the dataset in which LSTMs obtain their best results is composed mostly of simple sentences, describing visual scenes, and does not reflect a representative scenario of language in general. This internship proposal aims at studying RTE methods, especially ones based on neural networks, and investigating ways to incorporate more knowledge to them in order to deal with more complex sentences, which can be found in other datasets widespread in the NLP community. The capability of such neural models to align words and phrases of two texts, shown to be very promising in recent research, is a point of great interest to this proposal. (AU)