Nowadays, companies are convinced that putting people at the heart of their businesses is vital for their competition performances and success. To achieve this goal, a deep understanding of people's published content is mandatory to evaluate their satisfaction, frustration, and interestingness, and eventually recommend them other items and services. Media Analytics or News analytics is the main solution used to collect, process and analyze the diferent news and content published by people on media, social networks, etc. to provide the required insights and metrics which help in data-driven decision making. Press newspapers, precisely, PDF newspapers are one of the valuable and rich generated content that should be collected and analyzed by companies and organizations due to the capital of information they contain. Newspaper Analytics is a promising field of study, however, a reduced number of works are interested in this topic. The lack of work dealing with this kind of content (PDF newspapers) for data analytics motivates us to propose: (1) An end-to-end approach which allows conducting a newspaper analytics applied to PDF newspapers; (2) A detailed approach for PDF newspaper pattern recognition, for Latin and Arabic PDF, is presented. This approach consists in identifying the diferent blocks of press articles in a PDF page, and their associated metadata (authors, title). Once boundings of each article is detected, the recognized articles and metadata can be extracted using the OCR techniques (which is out of the scope of this paper). This approach is validated through diferent experiments applied on our proper constructed dataset. This later contains more than 1.500 PDFs collected from diferent Algerian newspapers in Latin and Arabic languages, during five months. The obtained results are promising and allowed us to develop a tool which presents the results of PDF newspaper analytics to end-users through dashboards and KPIs (key performance indicators) to keep an eye on their presence in the media and their reputation.
News analytics, Media analytics or Media Intelligence is a really promising field which needs more attention regarding its added value for companies, industries, organizations and governments. News analytics refers to the diferent tools, solutions, metrics and indicators used to analyze the huge amount of published news, comments, posts, reviews, etc., in their diferent formats mainly on websites and social networks. The advents made by web2.0 have encouraged people to share more content on the web. According to Datareportal1 special report published in July 2020, important raise is experienced in social media activities during the beginning of the COVID-19.
These active users on social media have become more powerful and companies become more careful with this users content. This content expresses satisfaction, frustration, or delights user sentiments [1] and has a big impact on influencing customers’ decisions [ 2]. As a consequence, companies are spending a lot of efort in making and improving their business by deeply analyzing and understanding this content [3].
Gathering and analyzing this kind of data is very important mainly for top management to keep an eye on the evolution of the market, build competitive advantages, analyze the evolution of business, understand the requirements of their customers or their prospects, etc.
News analytics is based on complex techniques of natural language processing, advanced linguistic analysis to process and understand the context of text, then derive interesting qualitative and quantitative insights about targeted audience, opinion analysis, fake news detection, context analysis, etc. This information is valuable to help managers to build their data-driven strategies, improve their risk management tools, and make better business decisions.
Newspaper analytics is a sub-field of news analytics which focuses on the analysis of data published in press newspapers by journalists in the form of articles. For many years, specific entities are created inside companies and organizations to collect, read and analyze the diferent articles published in plenty of newspapers. Specific roles are hired just to perform this burden task of manually selecting, then cutting the portions of newspaper that may interest the managers. The emergence of the web and the development of technology has reduced the dificulty of this task thanks to the availability of electronic newspapers in HTML or PDF formats.
Many solutions are provided to collect, process and analyze the content of web news based on advanced analytics techniques like machine learning and deep learning, NLP. Hundreds of paid solutions are provided by vendors to facilitate this task like Refinitiv Machine Readable News Analytics2. However, a lack of works and solutions dealing with newspapers, mainly the PDF format, is identified in industrial and academic fields. Even if some open source and paid solutions allow extracting the whole content of this type of documents like (pdf2text python library) in a raw shufled format, but this is not really suitable for PDF Newspaper. This latter is characterized by its specific format of articles blocks. Where in each page, many blocks are dumped and each block represents the text of the article, its authors and its title.
Identifying the parts and the boundings of each article like illustrated in figure 1 with its related metadata like title and author is an imperative requirement, in order to be able in the future to extract articles one by one in text format from each PDF page. Once extracted, diferent NLP and text processing techniques can be applied on text to understand and analyze it. Plenty of usage cases can be proposed like sentiment analysis [4] of a given article, opinion mining, Topic modeling, etc.
The lack of academic and industrial works dealing with this issue motivates our proposal.
1https://datareportal.com/reports/digital-2020-july-global-statshot
2https://www.refinitiv
.com/en/financial-data/financial-news-coverage/political-news-feeds-analysis/newsanalytics
Figure 1: Newspaper PDF Blocks -Akhbar Alyoum; May 08,
2021In this paper, we aim to propose: (
To validate our proposal, some experiments are conducted on our own constructed dataset composed of more than 1.500 PDF newspapers in Arabic and Latin languages and collected from many Algerian newspapers like Echourouk, Ennahar, El Watan, Jeune indépendant, etc. Interesting results are achieved which encourage us to consider this solution in our newspaper analytics tool which returns many KPIs and graphs via an interactive dashboard to be used by many companies and institutions.
This paper is organized as follows, section 2 summarizes the related work, section 3 presents the detailed approach and deep learning model, section 4 details the conducted experiments and achieved results, section 5 concludes the paper.
The literature is abundant of works related to news and media analytics [5], computer vision and image patterns recognition [6], in both academic and industrial fields. However, a restricted number of works are interested in press newspaper data analysis despite their added value [7]. In this section, we highlight the main works and solutions related to press newspaper analysis approaches (news discourse analysis is out of the scope of our paper), and the main pattern recognition used solutions.
Press News analysis passes through many steps principally subscribed in NLP discipline. Many frameworks and approaches are proposed to analyze the news as a type of text [8] characterized by its language, structure and context. Linguistic and grammatical operators are applied on press news to extract valuable semantic properties [9] that may be used for diferent analytics purposes like sentiment analysis [4], financial reporting [ 10], Crime analytics [11] where data mining and lexicon based techniques are used.
In general, analyzing news requires the creation of a rich dataset [12], this dataset resulted from manual or automatic collection of articles, headlines [13], content [14] from archives to be used for diferent analytic purposes. The direct use of these datasets, in most cases, is not advised, text preprocessing is an imperative step to be performed in order to enhance the quality of text, extract features and tokens and vectorize extracted text to facilitate machine understanding. [13] proposed a complete text mining approach to analyze the text headlines extracted from the newspaper front pages. Word cloud, word sentiment analysis and word clustering case studies are performed after manual collection and preprocessing of headlines. Newspaper articles are also used to visually explore social networks [14]. The text articles were extracted manually using a German articles corpora. [15] proposed a complete tool to analyze and visualize the newspaper front pages. This tool downloaded images of newspaper front pages, imported them into a local software application, and hand-coded them for coverage measures. However, this tool did not extract the content and cannot identify the blocks of articles in the whole newspaper. It focuses on coloring the areas and calculating their surfaces to show the coverage rate.
We found one work [7] which proposed a mechanism to identify photos areas and text areas in electronic newspapers. However, this work did not focus on identifying the block of each article and its metadata, it colored the whole text area with a given color and image area with another color.
Image pattern recognition is a field of computer vision which studies how automated systems can process and understand digital images or videos with the aim of making them work as similarly as possible to humans [16]. Many solutions are proposed for image pattern recognition. On one hand, some approaches are traditional and called feature descriptors for the extraction of image features like: edges, corners, colors. SIFT (Scale-Invariant Feature Transform) [17], SURF (Speeded-Up Robust Features) [18] and BRIEF (Binary Robust Independent Elementary Features) [19]. They use a series of mathematical approximations to learn a representation of the image. However, these solutions are complex and need more expertise [20]. On the other hand, many deep learning based solutions are proposed in literature where artificial neural networks (ANN) and Convolution Neural networks (CNN) are designed to train needed models which automatically extract image features .
To summarize, many works proposed approaches to analyze news or media content. However, few works are interested in press newspapers for data analytics. Moreover, the analysis of the literature shows the absence of works dealing with PDF newspapers or proposing pattern recognition solutions to identify the blocks of press articles in each PDF page. These findings motivate our proposal.
To fill-in the gap in academic and industrial fields, we propose in this section, a complete approach which allows identifying in each PDF newspaper page the diferent articles associated with their authors and titles. This solution is widely required in the context of newspaper analytics.
This approach allows us to extract and analyze the content of press PDF Newspapers, understand their purpose, then use some metrics to present a dashboard for top management to help them making decisions. This solution focuses on press newspaper pattern recognition applied to PDF newspapers in Latin or Arabic languages.
Our approach is illustrated in figure 2, where the diferent steps are defined as follows:
1- Requirements Analysis: In any data-driven decision making or data analytics solution,
defining and understanding the business requirements is an imperative step. These steps allow
us to define the needs and expectations of this product. The developers should analyze the
analytics requirements expressed by the end-users. Generally, these requirements are given as:
(
5.1. Image Collection: in order to develop a deep learning model subscribed to the computer
vision field, we need a voluminous dataset of annotated images. The diferent boundings of
articles, titles, authors should be drawn manually and extract their positions.
5.2. Image Preprocessing: to achieve best results of accuracy and precision, the collected images
which will be used in the training phase should be of a good quality. The image quality criteria
that we defined are: clearness, luminosity, contrast, zoom. To enhance the image quality, we
developed a python script based on the PIL library which curates images based on these criteria.
5.3. Image Annotation: Once the quality is treated, these images are annotated using the image
annotation tool. This later allows drawing the boundaries of the article content, title and authors
(as described in Figure 1). Then yolo files are exported to be used in the next step. The yolo file
contains the 4 positions of each drawn boundary for each image extracted from the newpaper.
5.4. Deep learning model development: this is the core of the solution. At this phase, a python
code is written to train, evaluate and tune the parameters of the deep learning model used to
learn the article patterns obtained from the previous step. The deep learning model is based on
Keras Tensorflow, where a CNN model is trained on the annotated dataset. We use some metrics
to evaluate this model like precision, recall, loss. The CNN model we developed is composed
of many layers. Three main layers should be present in any CNN model: (
Our motivation for these experiments is guided by a real project of newspaper analytics in the R&D&I of our company. We have been working on this project for more than three years. Many deep experiments were conducted in these years before obtaining the accepted results presented in this paper.
We collected our newspapers from 10 Arabic and Latin Algerian PDF newspapers like echourouk, ennahar, el watan, el massa, depeche kabylie, etc. during 5 months. At the end, we obtained more than 1.500 PDF. Each PDF was converted into a list of JPG images to construct a dataset of more than 150.000 images (each newspaper contains an average number of 10 pages). These images should be manually annotated to train our model of pattern detection. For this end, we hired a group of 4 annotators to do this work, we obtained 10.000 annotated images. The process of annotation is detailed in the previous section. In our case, we used an online tool which helped us to annotate the boundaries of articles, authors, titles of each image and convert this annotation into a yolo file containing the pixel positions of these boundaries.
We developed our CNN model using Tensorflow in order to allow the detection of patterns in newspapers. This model was composed of many layers mainly (convolution layer, pooling layer and fully connected layer). The obtained yolo files were splitted into training, testing and validation datasets.
To obtain better performance of processing, we used our internal servers equipped with many GPUs. The training phase takes around 16 hours.
We obtained a model with a training accuracy of 85 % and inference accuracy of 81 %. We used this model to infer results from unseen images. The example of detected patterns on new images is given in figure 5 and figure 4.
The obtained results are promising. More experiments are conducted now to improve the accuracy of the model. New images are also annotated for better results. The finality of this work is: after detecting these blocks of data (article, title, authors) and an OCR is developed (out of the scope) to transform them into text then analyze it using sentiment analysis models, topic modeling, opinion mining, etc. Moreover, the obtained results helped us to create dashboards given to the end-user to evaluate the presence and reputation of the media.
A limitation, even if our work returned good results in terms of precision but still many enhancements are needed. In some cases the boundaries are not detected exactly which make the OCR not really performing. In case of complex articles where the text is splitted into many positions, our model cannot detect the article as one block. All these issues are the objects of our improvements.
Using an OCR model, we extracted the text from images to feed a database of newspaper data using MongoDB database. This latter is used to feed a dashboard containing many metrics and graphs useful for newspaper analytics in diferent companies.
News / Press data are valuable sources of information which are widely considered in academic and industrial worlds. The correct exploitation of these data helps in improving many services and satisfying the end-users. PDF newspapers are one of these important sources which require more attention for efective analytics.
The exploitation and analysis of PDF newspapers provide companies and institutions with many elements of information about their brand image, their presence in media, their coverage by newspapers, etc. It may also have other advantages like: security management, risk analysis, competitors tracking, etc. All these data help in proactive decision-making and prevent actions.
Having these motivations in mind, and because of the lack identified in academic and industrial worlds, we propose a complete end-to-end approach for newspaper analytics. Moreover, a special focus is given to the newspaper pattern recognition module, which consists of recognizing in each PDF newspaper page, the diferent blocks of articles and their associated metadata (authors and title) using advanced deep learning techniques. This information is required for future use, after applying OCR techniques to transform the detected patterns into text.
Interesting experiments are conducted on our own constructed dataset which is composed of thousands of images extracted from diferent Algerian Latin and Arabic collected PDF newspapers. These results allowed us to trust our approach and used it in our developed tool dedicated to our customers. Our newspaper analytics tool provides an interactive dashboard with very interesting KPI required by the end user to analyze their coverage by newspapers.
As a perspective, we are working on the OCR module which allows extracting text from each block of the detected article and their associated metadata. Interesting metrics can be derived from text analysis of the extracted articles to be used for a rich newspaper analytics experience.
We would like to thank Mr Farid GHANEM and Mr Samir TAGZOUT for their eforts for the success of this project and to achieve the defined objectives.
We would also like to thank our intern students [ABOUD Ibrahim, KETFI Hibet Allah, BOUGUESSA Wail] who helped us to prepare and annotate the dataset of images. [4] S. Taj, B. B. Shaikh, A. F. Meghji, Sentiment analysis of news articles: a lexicon based approach, in: 2019 2nd international conference on computing, mathematics and engineering technologies (iCoMET), IEEE, 2019, pp. 1–5. [5] D. Zeng, H. Chen, R. Lusch, S.-H. Li, Social media analytics and intelligence, IEEE
Intelligent Systems 25 (2010) 13–16. [6] J. Wright, Y. Ma, J. Mairal, G. Sapiro, T. S. Huang, S. Yan, Sparse representation for computer vision and pattern recognition, Proceedings of the IEEE 98 (2010) 1031–1044. [7] P. Fyfe, Q. Ge, Image analytics and the nineteenth-century illustrated newspaper, Journal of Cultural Analytics 3 (2018) 11032. [8] T. A. Van Dijk, News as discourse, Routledge, 2013. [9] T. A. Van Dijk, News analysis: Case studies of international and national news in the press,
Routledge, 2013. [10] G. Mitra, L. Mitra, The handbook of news analytics in finance, John Wiley & Sons, 2011. [11] I. Jayaweera, C. Sajeewa, S. Liyanage, T. Wijewardane, I. Perera, A. Wijayasiri, Crime analytics: Analysis of crimes through newspaper articles, in: 2015 Moratuwa Engineering Research Conference (MERCon), IEEE, 2015, pp. 277–282. [12] M. Reason, B. García, Approaches to the newspaper archive: content analysis and press coverage of glasgow’s year of culture, Media, Culture & Society 29 (2007) 304–331. [13] A. Hossain, M. Karimuzzaman, M. M. Hossain, A. Rahman, Text mining and sentiment analysis of newspaper headlines, Information 12 (2021) 414. [14] A. Kochtchi, T. v. Landesberger, C. Biemann, Networks of names: Visual exploration and semi-automatic tagging of social networks from newspaper articles, in: Computer graphics forum, volume 33, Wiley Online Library, 2014, pp. 211–220. [15] S. Costanza-Chock, P. Rey-Mazon, Pageonex: New approaches to newspaper front page analysis, International Journal of Communication 10 (2016) 28. [16] C. Orhei, M. Mocofan, S. Vert, R. Vasiu, End-to-end computer vision framework, in: 2020 International Symposium on Electronics and Telecommunications (ISETC), IEEE, 2020, pp. 1–4. [17] D. G. Lowe, Distinctive image features from scale-invariant keypoints, International journal of computer vision 60 (2004) 91–110. [18] H. Bay, T. Tuytelaars, L. V. Gool, Surf: Speeded up robust features, in: European conference on computer vision, Springer, 2006, pp. 404–417. [19] M. Calonder, V. Lepetit, C. Strecha, P. Fua, Brief: Binary robust independent elementary features, in: European conference on computer vision, Springer, 2010, pp. 778–792. [20] S. Khan, H. Rahmani, S. A. A. Shah, M. Bennamoun, A guide to convolutional neural networks for computer vision, Synthesis lectures on computer vision 8 (2018) 1–207. [21] S. Mostafa, F.-X. Wu, Diagnosis of autism spectrum disorder with convolutional autoencoder and structural mri images, in: Neural Engineering Techniques for Autism Spectrum Disorder, Elsevier, 2021, pp. 23–38.