Automatic Linking of Short Arabic Text to Wikipedia Articles

Abstract

Given the enormous amount of unstructured texts available on the Web, there has been an emerging need to increase discoverability of and accessibility to these texts. One of the proposed solutions is to annotate texts with information extracted from background knowledge. Wikipedia, the free encyclopedia, has been recently exploited as a background knowledge to dynamically annotate text with complementary information. Given any piece of text the main challenge is how to determine the most relevant information from Wikipedia with the least effort and time. While Wikipedia-based annotation has mainly targeted the English and Latin versions of Wikipedia, little effort has been devoted to annotate Arabic text using the Arabic version of Wikipedia. This work proposes an approach for dynamic linking of Arabic short texts to Wikipedia articles. It reports on the several challenges associated with the design and implementation of the linking approach including the processing and setting up of the Wikipedia's enormous content, the mapping of texts to Wikipedia articles, the problem of article disambiguation, and time efficiency. The proposed approach focuses on short texts because they are generally more difficult to process and annotate than long texts. The proposed approach was assessed over a dataset of 100 short texts gathered from online Arabic articles and then work on it offline. Hyperlinks generated by the approach were compared with the hyperlinks generated by two human raters. The dynamic linking approach achieved 71.79% accuracy, 74.70% average precision, and 82.63 % average recall. A thorough analysis and discussion of the evaluation results are also presented to address the limitations and strengths as well as the recommendations for future improvements. The source code, dataset, and complete experimental results are made available online on: https://github.com/FatoomMFayad/Dynamic-Linking