Time-Aware Word Embeddings for Three Lebanese News Archives

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4717–4725 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4717 Time-Aware Word Embeddings for Three Lebanese News Archives Jad Doughman, Fatima K. Abu Salem, Shady Elbassuoni Computer Science Department American University of Beirut jad17@mail.aub.edu, fa21@aub.edu.lb, se58@aub.edu.lb Abstract Word embeddings have proven to be an effective method for capturing semantic relations among distinct terms within a large corpus. In this paper, we present a set of word embeddings learnt from three large Lebanese news archives, which collectively consist of 609,386 scanned newspaper images and spanning a total of 151 years, ranging from 1933 till 2011. The diversified ideological nature of the news archives alongside the temporal variability of the embeddings offer a rare glimpse onto the variation of word representation across the left-right political spectrum. To train the word embeddings, Google’s Tesseract 4.0 OCR engine was employed to transcribe the scanned news archives, and various archive-level as well as decade-level word embeddings were learnt. To evaluate the accuracy of the learnt word embeddings, a benchmark of analogy tasks was used. Finally, we demonstrate an interactive system that allows the end user to visualize for a given word of interest, the variation of the top-k closest words in the embedding space as a function of time and across news archives using an animated scatter plot. Keywords: lebanese news archives, word embeddings, optical character recognition 1. Introduction The public’s perception is capable of shaping a nation’s social and economic demeanour, thus making the source of that perception a vital matter. In the postmodern era, news media were restricted to local stations covering regional and international news, thus limiting the frequency and variety by which people obtained information. Citizens of under-developed countries relied predominantly on their primary local newspaper stations for the latest insight on political, economic, and cultural updates. Despite the vast amount of newspaper archives that has been generated throughout the years, historians have attempted to manu- ally analyze this data for decades. The daunting process of transcribing such large archives propels both historians and computational linguists to try to find ways for automating the analysis of such a corpora through modern feature learning techniques in natural language processing (NLP), mainly word embeddings. In this paper, we describe Arabic word embeddings trained using three large Lebanese news archives, namely: Assafir1, Alhayat2, and Annahar3. The diversified history of the three newspaper archives provides a more repre- sentative sample of ideologies from both the center-right and center-left spectrum. The three archives combined consist of 609,386 scanned newspaper images, which span a period of 151 years, ranging from 1933 till 2011. To transcribe our large corpus, an adequate OCR engine had to be selected. To facilitate this process, a ground truth consisting of a manual transcription of 5 newspaper images was generated and tested against various OCR engines and numerous configurations. Google’s Tesseract 4.0 OCR engine, which returned the highest F-measure, was selected and used to transcribe …

Content truncated to 500 words