Memory Network Based Deep Learning Models for Question Answering System

Abstract

Question Answering (QA) system is a domain of Natural Language Processing (NLP) wherein the users can ask questions in natural human language and obtain brief answers rather than a list of all relevant documents. The earlier NLP techniques based on statistical approaches and Machine Learning (ML) techniques were unable to handle ambiguities in language and failed to produce optimal results for typical NLP tasks. newlineIn recent times, the remarkable growth of Deep Learning (DL) networks and other related technologies have made human level accuracy and efficiency a possibility for most NLP related tasks. DL networks solve NLP tasks without the need for understanding grammar and other aspects of the language as they have the ability to learn from given examples. newlineA ML system can be trained to map a story and question text pair with the corresponding answer text. DL approaches to QA tasks avoid the tedium of feature extractions used in traditional linguistic tools. They use bidirectional Long Short-Term Memory (LSTM) models to generate embeddings of question and answers, measure cosine similarity and compute the distance between question-answer pairs to determine the appropriate answer. Deep learning networks are capable of learning any QA task represented in any language. However, they fail to produce acceptable accuracy while solving complex QA tasks. newlineThe main objective of this research is to examine the existing Recurrent Neural Networks (RNNs) and memory network based deep learning models, apply them to the state of the art QA systems and develop an improved QA model that can handle complex QA tasks efficiently. newlineThis study explores the key aspects of a deep learning network using state of the art deep learning network models and evaluates their performance using different language text corpora with the intent of improving the performance of DL networks when handling complex QA tasks. The scope of this research is to develop a Bi-model MemN2N network to handle complex QA tasks with improved accuracy.