Arabic Question Answering System for Information Retrieval on Large-scale Image Objects

Abstract

The image objects retrieval modeling focused on discovering the images contained within each document based on a specific "object" statistic. In this study, we introduce the potential of classifying images into 17 classes to extract objects using Deep Neural Networks such as Convolutional Neural Networks (CNN), Visual Geometry Group-16 (VGG-16), and Sequential Minimal Optimization (SMO) algorithms. Moreover, we show the objects extraction process from the text (user's queries) in Arabic language using NLP algorithms. Building the proposed model started by collecting 2649 images from multiple sources that were entered into three image classifiers to extract objects: CNN, VGG-16, and SMO. Afterward, building an index for image URLs, images objects, and classes. The retrieved against our index was tested using user's queries that will go through multiple NLP approaches to extract objects to retrieve the required images. The most frequent object names types were Nouns (NNs), Noun Phrases (NPs), and a combination of NN and NPs within 110 queries as a total. The retrieving process will start by testing queries against the Image Index. For this query level, the retrieved images information was evaluated in Mean Average Precision (MAP) with multiple retrieved values for retrieving images: 1,5,10,100,1000 as the Recall and F1 were calculated for the top 1000 retrieved images. Our proposed system obtained the best results when searching with expanded queries in terms of MAP where the retrieval from Image Index scored an 88% MAP for retrieving 1000 images for object types as NNs and NPs.