Abstract
This paper presents an image recommendation system called “News2Image”, which takes as input a news text and outputs a small set of images selected from the database such that the selected images present contexts similar to the keywords of the input text. The proposed framework starts by preprocessing and summarizing the news for the selection of its keywords, which are then converted to vectors using word2vec. The framework associates each database image with a set of sentences describing its content using image captioning. After preprocessing the sentences and translating them into the word vectors as performed for input news texts, the word subspaces are compared using the Mutual Subspace Method (MSM). Experimental evaluation of the proposed framework including a comparison with the previous approach demonstrates its overall effectiveness.