Deep neural networks based on gating mechanism for open-domain question answering

No hay miniatura disponible
Arch Tijera, Drake Christian
Título de la revista
Revista ISSN
Título del volumen
Universidad Católica San Pablo
Proyectos de investigación
Unidades organizativas
Número de la revista
Nowadays, Question Answering is being addressed from a reading comprehension approach. Usually, Machine Comprehension models are poweredby Deep Learning algorithms. Most related work faces the challenge by improving the Interaction Encoder, proposing several architectures strongly based on attention. In Contrast, few related work has focused on improving the Context Encoder. Thus, our work has explored in depth the Context Encoder. We propose a gating mechanism that controls the ow of information, from the Context Encoder towards Interaction Encoder. This gating mechanism is based on additional information computed previously. Our experiments has shown that our proposed model improved the performance of a competitive baseline model. Our single model reached 78.36% on F1 score and 69.1% on exact match metric, on the Stanford Question Answering benchmark.
I would like to thank in a special way the National Council of Science, Technology and Technological Innovation (CONCYTEC) and the National Fund for Scientific, Technological development and Technological Innovation (FONDECYT-CIENCIACTIVA), which through the Management Agreement N 234-2015-FONDECYT, they have allowed the grant and financing of my studies of Master in Computer Science at the Universidad Cat´olica San Pablo (UCSP).
Palabras clave
Question Answering, Machine Comprehension, Natural Language, Processing, Deep Learning