Features of Big Text Data Visualization for Managerial Decision Making E. A. Makarova1, D. G. Lagerev1, F.Y. Lozbinev2 m4karova.e@yandex.ru | LagerevDG@mail.ru | flozbinev@yandex.ru 1 Bryansk State Technical University, Bryansk, Russia 2 RANEPA, Bryansk, Russia This paper describes text data analysis in the course of managerial decision making. The process of collecting textual data for further analysis as well as the use of visualization in human control over the correctness of data collection is considered in depth. An algorithm modification for creating an “n-gram cloud” visualization is proposed, which can help to make visualization accessible to people with visual impairments. Also, a method of visualization of n-gram vector representation models (word embedding) is proposed. On the basis of the conducted research, a part of a software package was implemented, which is responsible for creating interactive visualizations in a browser and interoperating with them. Keywords: visualization, natural language processing, web application accessibility. of serious human and computational resources, which can nullify 1. Introduction the economic benefits obtained by adding unstructured data to the process of managerial decision making. With the acceleration of scientific and technological Currently, there are various analytical systems that work not progress, as a result, economic growth rates of both global and only with structured data but also with unstructured ones, local markets are rapidly increasing. According to the study [5], including text data downloaded from social media [11]. In these the number of mergers and acquisitions in Russia in 2017 systems, visualization is rarely used at the stage of collecting and increased by 13%. Besides, the number of originated loans is preprocessing big text data. However, collection and growing. According to the United Credit Bureau, the annual preprocessing of data for such systems is still quite time number of loans issued in Russia has increased by 22%, while consuming, and there is a significant risk of using irrelevant lending has increased by 53%. In addition to the accelerated documents as data sources. Usually one of the two approaches is capital turnover, growth is also observed in the labor market. In used: either a fully automatic analysis of collection and Antal Russia’s survey, 27% of employers reported an increase in preprocessing results (faster) or a fully manual review of a large staff turnover in their companies over the past year [12]. array of documents (more qualitative). This article discusses a Higher velocity and number of transactions conducted in hybrid approach based on vector data visualization that allows various spheres of social and economic activity results in greater adding expert assessment of document relevance at the stage of burden on managers at various levels. This requires either an data collection and preprocessing [18]. increase in the decision-making staff or enhancement of information systems supporting managerial decision-making in 2. Extracting information from sources of order to reduce the people’s workload. Beside traditional data used in such systems (e.g., credit history and capital for scoring varying degrees of structuring systems used in loan approval), many researchers and Let us consider in more detail the process of collecting and manufacturers of technological solutions use unstructured analyzing text information from various sources presented in information sources about legal entities and individuals involved Figure 1. in transactions. Examples of such information are data from mass media, social networks, etc. In addition, some studies have shown that adding analysis of text data from social media to prediction models results in greater accuracy. For example, they help to increase the accuracy of legal entity’s bankruptcy prediction [7]. Hence, one of the stages of using managerial decision-making support systems is loading text information into them about an object of socio-economic activity for further use. Objects of socio-economic relations are widely represented on the Internet both through official websites and in the form of digital reputation, i.e., reviews, news, what appears on the network about them without their direct intervention. However, the amount of such data is constantly growing (due to data duplication, borrowing data from another source, etc.), which requires optimization in terms of speed and cost of their Fig. 1. Overview diagram of collecting and processing data collecting and processing. As small and medium-sized businesses have to contact with ever increasing number of people All the processes presented in the diagram are important for in the course of their activity, the risk of a transaction with legal the efficient use of unstructured text data for managerial decision entities or individuals unreliable in terms of tax or other laws making. However, in this paper, the greatest emphasis is placed increases, which may entail long-term consequences, such as on the process of collecting information since accuracy and costs in public image, etc., and may result in legal entity’s resource consumption of further analysis depends on the quality bankruptcy. of the collected data. On the one hand, decisions need to be made faster and faster, In the diagram, DBin is an internal database containing their number is growing, which can lead to more errors and risks. trained models for collecting and analyzing information as well This problem can be solved by the integration of data mining as accumulated information about analysis objects. DBout refers systems into DSS by using large volumes of unstructured data to external databases (structured sources) attached by the user. available for analysis [10]. On the other hand, the process of Conceptual model for collecting text information is collecting and preprocessing these data requires the involvement S = < R, M, D; I>, Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). downloaded data is to visualize it. In [17], it was already where R is resources (temporary, material, human); demonstrated that the difference in the content of documents is M is information about previous generations; noticeable in the visualization of an n-gram cloud type (n-gram D is data sent for analysis; refers to a word sequence), and it was noted that this method I is amount of information relevant to the task that is requires further refinement. In the current implementation, available to the system. visualization has undergone a number of changes, such as R = < Rm, Rh, Rp, T> combining word weights that have semantic proximity below a certain threshold, excluding “stop words” and words with small where Rm is money spent on paid services (various APIs and weights from visualization. services); One of the refinement directions for the visualization method Rp is the number of available experts in the subject area; will include its adaptation for use by various groups of people, Rh is hardware limitations. For some problems, hardware including those with disabilities. When developing resource limit will only mean speed of calculations and, visualizations, it is important to consider all user groups, not only accordingly, be considered together with T parameter, but for in terms of compliance with international standards but also in some language processing tasks, for example, when using vector terms of the increasing number of potential users. For example, representation of words, the amount of RAM available will be more than 5% of population suffer from various forms of color key to the usability of these methods; vision deficiency, which can prevent the user from interacting T is time spent on collecting information, which, in turn, can with the visualization to a full extent [2]. be decomposed into the following components: In recent years, the topic of accessible visualization has T = Tu + (Te + Td) + Ta, gained great interest from researchers and software development service providers [14]. For example, Square, Inc. [1] has where Tu is time spent by the main user of the system; published an open-source guide to creating accessible data Te is time spent by the expert who will check and resolve visualizations. Among visualizations they propose there are manually various situations difficult for machine-aided various types of charts and graphs. processing; Visualizations related to the analysis of text information are Td is time delay between expert’s response and continuation comparatively little studied from this side. Next, we will consider of processing (the error to account for the non-round-the-clock two examples of such visualizations that are important for availability of the expert); collecting text data in the described software package. Ta is time spent on automatic processing. Classic works devoted to the construction of an n-gram cloud The task of optimizing data collection consists in reducing R, (or “tag cloud”, “word cloud”) [4, 15], which described H, D parameters while increasing I parameter. algorithms employed by libraries implementing visualization It is also assumed that a number of parameters will decrease data, could not take into account WCAG recommendations on with each subsequent use of the system due to the training of application adaptation for people with visual impairments since users and models, accumulation of useful knowledge about the they had appeared before these guidelines were developed. objects of research. As part of the software package, a client-server architecture In matters of improving the efficiency of information subsystem was implemented as a web application that provides collection, there are two extremes: to make all the work fully interactive visualizations and implementation of user analysis to automatic, thereby saving on human resources, or to make data collection process. So, for example, the developed n-gram process control completely manual. In this paper, an cloud visualization takes into account WCAG 2.1 “intermediate” version is considered when an expert is engaged recommendations. Therefore, the following restrictions and in evaluating the effectiveness of the collection process, but due additions have been introduces to the algorithm: to the use of various tools, such as visualization, his work time is 1) restriction on the contrast of colors significantly reduced [9]. 2) exclusion of vertical text orientation [8] In addition, the following approaches are used in the 3) setting the minimum and maximum text sizes developed software package to optimize information collection 4) adding advanced user settings. before analysis: Considering that the interface of the existing system was 1) refinement of search queries; developed as a web application, it will be reasonable to rely on 2) ignoring duplicate information; the algorithms used to create and display tag clouds [4] adapting 3) preliminary data analysis, etc. them to the problem being solved and WCAG 2.1 recommendations. 3. Visualization of big text data for data mining Many ready-made visualization tools do not take into optimization. account contrast for different groups of people, including those suffering from visual impairment and color deficiency. However, Let us consider some features which require human it should be understood that the purpose of creating a tag cloud interference for more efficient work and for which various is often to effectively illustrate an array of information rather than visualization methods have been studied and refined as part of a detailed analysis [15]. the work on the system [19]. As a data source in this example, Color contrast according to WCAG 2.1 is: we will use web-based media, but the methods being developed (L1 + 0.05) / (L2 + 0.05) > Cmin are applicable, with some adjustment, to all sources of a similar L1, L2 are relative brightness of compared colors. structure. Since all words in visualization will be interactive, the When setting up uploading of text documents from a certain required contrast for them should be calculated as for controls, source, by which the search is possible, users of the system may i.e., Cmin = 3 for n-grams located separately. In addition, contrast encounter the fact that the query does not correspond to the for each individual color compared to the background should be required result, e.g., if the request has turned out to be too equal to Cmin = 4,5 [16]. Calculations show that it is possible to “general” or information about homonymous objects is present find only 2 colors that will be simultaneously contrastive with the in the same sources. A way out of this situation may be to view background and between themselves. a part of the collected text documents, their brief contents or Also there appear restrictions on the font size. On the one some metadata. It is time consuming for the user (subject matter hand, the minimum size of n-grams should not be less than 16pt expert or employee). Another way to familiarize the user with the [16]. On the other hand, the same standard imposes the condition that all texts on a page can be magnified to 200% maintaining Table 1. Average time spent by the user on one document their readability, which constrains the maximum possible font per document (in seconds). size when displaying a page at the size of 100%. In order to Task BMZ BMZ + BMZ + Ecofrio Ecofrio + maintain the approximate position of containers in which the text Bryansk Bryansk potatoes will be when enlarged, CSS Grid technology [3] and slicing +Industry floorplan algorithm [4] were used for the interface design. Group Besides, the user should be able to add custom settings for Group 1 12,5 13 11,5 13 12 colors and sizes of visualization. Let us consider a specific (one task) Group 2 12 13,5 13 14 13 example. In [17], it is described in detail how visual analysis of (three task) a part of text documents on a search query allows understanding Group 3 17,5 19 18,5 16 14 whether various search entities need to be added or excluded (one task) from the query. Figure 2 shows visualization implementation for Group 4 17 15 14,5 17,5 15 adjusting data collection for “BMZ” object (AO UK BMZ – (three task) Bryansk Machine-Building Plant). Presented figures Group 5 11 10,5 11 12 11,5 demonstrate the work with the texts in Russian. User’s task is to (three task) assess reputation of this legal entity. To do this, it is necessary to collect data on this object. The goal of this visualization is to The work on this topic [17] demonstrates how word track whether the context of the request, that implied the search embedding models [6] pre-trained on different collections of text for an enterprise located in the Bryansk region, was transmitted documents group words differently in terms of their semantic correctly. As the user can see from the visualization, the search proximity. Also, errors related to the content of the source data settings were incorrect, which resulted in occurrence in the occur in models built on word embedding. In the described collected data of many documents related to the activity of a system, these models are used not only to simplify visualization similar enterprise in the Republic of Belarus. Exclusion of text but also to remove duplicate documents during their further documents containing the word “Belarus” from the search results processing. Canvas-based visualization was developed [13] to increased significantly the accuracy of the collection by give the user an opportunity to edit acceptable boundaries of discarding also documents with references to such objects as semantic proximity (or cancel n-gram combining if options “Africa”, “Chad”, etc. proposed are unacceptable for the problem being solved). In the center of visualization there is a word position of which in the word embedding model is being explored. Distances from an n- gram are defined so that a two-dimensional vector would be equal to the similarity indicator of this n-gram to the one under study (by default, this value is 0.4). Further, the algorithm selects positions for n-grams in such a way as to ensure the readability of n-grams, including the recommendations described above (no intersections with other elements, a horizontal text of an acceptable size). An example of this visualization for n-grams having maximum semantic affinity with the word "industry" is presented in Figure 3. Fig. 2. N-gram cloud visualization of a text document collection in Russian In addition, since we are talking about displaying in the browser, all elements will have the “tabindex” attribute in ascending order as the significance decreases in the sample and the “aria-label” attribute with the weight of this element to facilitate the perception by people who have vision deficiency and use special programs for reading from the screen. Typically, to solve a data collection configuration tasks required to view (or use visualization) 20-30 random documents from the collection of documents, depending on the amount of Fig. 3. N-gram based nearest neighbor visualization in the available data. An experiment was conducted on the effect of the word embedding model method of solving configuration tasks in which five groups of users participated: users of groups 1 and 2 to solve the data Table 2 demonstrates how using n-gram cloud visualization collection configuration tasks using the visualization with and applying analysis results to the search query parameters standard settings, users of groups 3 and 4 used quick skimming increase the number of relevant documents received during data of documents, users of group 5 used visualization with user collection (for 20 random documents from a search sample). settings, pre-setting time is also included in the final calculation. Table 2. Impact of manual adjustment of the request on the The test results for some tasks are presented in table 1. Prior number of relevant document search tasks to working with the tasks presented, all users were trained on a Number of Object 1 Object 2 Object 3 Object 4 relevant test task. Some user groups performed only one group of tasks “BMZ” “Isoterm” “Ecofrio” “Spetsstroy” documents (for example, analyzing entities associated with “BMZ”) at one Before user time, while others immediately started the next task after solving 20% 30% 85% 10% adjustment the current task. After 85% 45% 90% 20% On average, time saved using visualization, compared to a adjustment quick skimming of texts, varies depending on user's familiarity On average, there has been registered an increase in the with the system and ranges from 18 to 42%. number of relevant documents by about 24%. The number of relevant documents in the experiment was determined by the method of expert viewing of 20 random documents from a search CFP1561Y-ART, pp. 02-38-NSAP. sample. doi: 10.1109/MEACS.2015.741490 11. Prangnawarat N., Hulpus I., Hayes C. (2015) Event 4. Conclusion Analysisin Social Media using Clustering of Heterogeneous Information Networks. The 28th Adding textual data to analyzed ones in the process of International FLAIRS Conference (AAAI Publications) managerial decision making can increase the efficiency. In this (AAAI) paper, special attention is paid to the process of collecting text 12. Staff turnover has started to grow. Available by link: data from various sources. It is shown that visualization of big https://www.antalrussia.com/news/staff-turnover-has- text data can significantly reduce time spent on its human started-to-grow/ processing: time savings compared to skimming of texts is from 13. The canvas elements. Available by link: 18 to 42%, and the number of relevant documents found https://html.spec.whatwg.org/multipage/canvas.html#the- increases by about 24%. Besides, a part of a software package canvas-element has been developed, which allows for visualization of text data 14. The Future of Data Visualization: Predictions for 2019 and and models of vector representation of words. When developing Beyond А. Available by link: visualization algorithms, it is necessary to take into account https://depictdatastudio.com/the-future-of-data- international standards for creating web applications for people visualization-predictions-for-2019-and-beyond/ with disabilities, thus making them [applications] accessible to a 15. Viégas B., Wattenberg M., Feinberg J. (2009) Participatory wide range of users. visualization with Wordle. IEEE Transactions on In the future, it is planned to continue the study of efficient Visualization and Computer Graphics 15, no. 6, pp. 1137– data collection methods for analysis to support managerial 1144. doi:10.1109/TVCG.2009.17 decision-making. In particular, it is planned to study in more 16. Web Content Accessibility Guidelines (WCAG) 2.1. detail n-gram vector representation and its use for identifying and Available by link: https://www.w3.org/TR/WCAG21/ deleting duplicate data. 17. Zakharova A.A., Lagerev D.G., Makarova E.A. (2019) Evaluation of the semantic value of textual information for 5. References the development of management decisions. CPT2019 The 1. Accessible Colors for Data Visualization. Available by Conference Proceedings, TzarGrad, Moscow region, link: https://medium.com/@zachgrosser/accessible-colors- Russia for-data-visualization-2ad64ac4ee7e 18. Zakharova A.A., Vekhter E.V., Shklyar A.V. (2017) 2. Causes of Colour Blindness. Available by link: Methods of Solving Problems of Data Analysis Using http://www.colourblindawareness.org/colour- Analytical Visual Models. Scientific Visualization, vol. 9, blindness/causes-of-colour-blindness/ no. 4, pp. 78-88. doi: 10.26583/sv.9.4.08 3. CSS Grid – Table layout is back. Be there and be square. 19. Zhao J., Zhao G., Zhao L., Zhao W., (2014). PEARL: An Available by link: Interactive Visual Analytic Tool for Understanding https://developers.google.com/web/updates/2017/01/css- Personal Emotion Style Derived from Social Media. IEEE grid Conference on Visual Analytics Science and Technology, 4. Kaser O., Lemire D. (2007). Tag-Cloud Drawing: VAST 2014 - Proceedings. Algorithms for Cloud Visualization. Tagging and Metadata doi: 10.1109/VAST.2014.7042496. for Social Information Organization. A workshop at WWW2007, pp 1086-1087. 5. KPMG presents the results of a survey of Russia's mergers and acquisitions market in 2017. Available by link: https://home.kpmg/ru/en/home/media/press- releases/2018/03/ma-survey-2017.html 6. Kutuzov A, Kutuzov I. (2015) Texts in, meaning out: neural language models in semantic similarity task for Russian. Proceedings of the Dialog 2015 Conference, Moscow, Russia 7. Mai F., Mai T., Ling C., Ling M. (2018). Deep Learning Models for Bankruptcy Prediction using Textual Disclosures. European Journal of Operational Research. doi: 10.1016/j.ejor.2018.10.024. 8. Make your information more accessible. National Disability Authority. Available by link:http://nda.ie/Resources/Accessibility-toolkit/Make- your-information-more-accessible/ 9. Podvesovskii A.G., Isaev R.A. (2018) Visualization Metaphors for Fuzzy Cognitive Maps. Scientific Visualization, vol. 10, no. 4, pp. 13-29. doi: 10.26583/sv.10.4.02 10. Podvesovskii A.G., Gulakov K.V., Dergachyov K.V., Korostelyov D.A., Lagerev D.G. (2015) The choice of parameters of welding materials on the basis of fuzzy cognitive model with neural network identification of nonlinear dependence. Proceedings of the 2015 International Conference on Mechanical Engineering, Automation and Control Systems (MEACS) (Tomsk, Russia, December 1-4, 2015), IEEE Catalog Number: