DEEP BELIEF NETWORK BASED QUESTION ANSWERING SYSTEM USING ALTERNATE SKIP-N GRAM MODEL AND NEGATIVE SAMPLING APPROACHES

摘要

The Question Answering (QA) system becomes essential owing to the increasing amount of web content and high demand for the right and short information. Intending to enhance QA results towards the Natural Language Processing (NLP) community, most of the Question answering system exploits machine learning algorithms to generate an appropriate answer to the user query. Even though, it lacks to predict accurately over the large-scale data by itself, needs an external force to make adjustments in the answer prediction. With the recent evolution in deep learning, the neural network architecture reflects its potential for QA. The deep learning model can determine the issues in answer prediction on their own and resolves it. The class of deep neural network such as Deep Belief Network (DBN) is widely applied in question answering especially for text processing. Moreover, most of the works on the text processing exploits the skip-gram model for representing the relevant words in the vector over a massive volume of unstructured text data. However, it results in inefficient outcomes, especially when processing the combination of frequent data and stop words. To resolve these issues, this paper introduces the Deep Neural network for Answering user queries (DNA). The proposed DNA approach performs the QA system over DBN by applying alternate skip-N gram and negative sampling. The conditional probability measurement develops an alternate skip-N gram model and alternatively applying the normal N-gram and skip-N gram model. It improves the efficiency of relevant word-pair detection without increasing the computational complexity. By only using samples, the negative sampling reduces the impact of noise on the accuracy of alternate skip-N gram model and improves the efficiency of the QA system. Finally, the DNA is evaluated using Java and compared with the existing Unified model for Document-Based Question Answering (UDBQA). The results show the efficiency of DNA, for instance, the UDBQA approach reduces the F-measure by 16.3%, compared to the DNA approach with 2000 number of queries.