key: cord-0501969-4svuz97w
authors: Kastner, John; Wei, Hong; Samet, Hanan
title: Viewing the Progression of the Novel Corona Virus (COVID-19) with NewsStand
date: 2020-02-28
journal: nan
DOI: nan
sha: 9f91b4f341c19000badf20d72c9728df1bd0a5c7
doc_id: 501969
cord_uid: 4svuz97w

With the continuing spread of COVID-19, it is clearly important to be able to track the progress of the virus over time to be better prepared to anticipate its emergence in new regions. Officially released numbers of cases will likely be the most accurate means by which to track this, but they will not necessarily paint a complete picture. We have developed an application 1 usable on desktop and mobile devices that allows users to explore geographic spread in discussion about the virus through analysis of keyword prevalence in geotaged news articles .

With the continuing spread of COVID-19, it is clearly important to be able to track the progress of the virus over time to be be er prepared to anticipate its emergence in new regions. ere are application to monitor o cially released numbers of cases [1] which are likely to be the most accurate method for tracking the progress of the virus; however, they will not necessarily paint a complete picture. To begin lling any gaps in o cial reports, we have developed an application 1 usable on desktop and mobile devices that allows users to explore the geographic spread in discussion about the virus through analysis of keyword prevalence in geotagged news articles. e application we describe here relies heavily on the existing NewsStand system [4] . NewsStand is a well suited basis for this application because it is a spatio-textual search engine. at is, it has the capacity to query a database of news articles in terms of both the raw textual content of the articles and the implicit spatial information encoded by toponyms mentioned in the article. In terms of what this paper requires, textual queries are needed to identify articles that contain keywords of interest to the user while spatial queries are needed to display the geographic distribution of articles that contain the keywords.

Two steps are required to obtain geocoded news keywords: important keywords must be extracted from news articles, and the news articles must be geocoded by identifying and resolving toponyms to speci cs geographic coordinates. Once we have identi ed a set of keywords and a set of geographic locations in an article, we take the cross product of these sets as the set of geocoded keywords. In other words, we create a pair associating every keyword in the article with every location mentioned in the article. Each of the is work was sponsored in part by the NSF under Grant iis-18-16889. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for pro t or commercial advantage and that copies bear this notice and the full citation on the rst page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permi ed. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior speci c permission and/or a fee. Request permissions from permissions@acm.org. Conference'17, Washington, DC, USA © 2016 ACM. 978-x-xxxx-xxxx-x/YY/MM. . . $15.00 DOI: 10.1145/nnnnnnn.nnnnnnn keyword location pairs is also associated with the time of publication of the article in order to enable the temporal component of our application.

e extraction of news keywords is handled by the original News-Stand system. NewsStand identi es keywords based on their TF-IDF scores. Words with high TF-IDF scores, words that appear much more frequently in a given document than in other documents in the database, are likely to be important to the document in which the words appear. e word with the highest TF-IDF score for a document is therefore the best keyword. More keywords can be obtained by selecting words with progressively lower scores.

For the purpose of this application, once keywords have been extracted from article text, we examine only those that we believe will be most relevant to the disease we are tracking. For instance, we may speci cally look for news articles containing the keyword "coronavirus". is approach should achieve high precision, since it is unlikely that an article unrelated to the virus will contain this keyword, but it may su er from low recall.

Geocoding is the process of associating concrete geographic information (i.e. latitude longitude pairs) with a piece of text. Geocoding can be framed as a speci c variant of the keyword extraction task in the sense that we must rst nd keywords that are likely to refer to geographic locations and then decide which of these locations are important enough to associate with the documents [2] . e third step in geocoding which is not required for general keyword extraction is toponym resolution. In toponym resolution, a toponym must be assigned a single latitude longitude pair to be its nal location.

is is nontrivial because there many ambiguous toponyms that can be used to refer to multiple distinct locations (e.g. Paris, London, etc.) [3] .

Our application is implemented as an interactive website using HTML, CSS, and JavaScript that communicates with our database through a Ruby On Rails web server. It consists of an interactive map query interface that is used to display our data and animations. In addition,a collection of controls are used to download data from our server and to select how it is displayed on the map. A screenshot of our application is show in Figure 1 .

Our map query interface is an interactive web map provided by the Lea et JavaScript Library. Data is rendered onto this map using a technique called marker clustering, which is implemented by an Rather than rendering each point individually, points are clustered into aggregate markers. As the user increases the zoom level of the map, focusing on a smaller area, these aggregate markers are split into two or more new markers that together represent all the points represented by the original marker. is decomposition allows more detail to be displayed when examining a small area without showing too much detail at lower zoom levels.

Each marker represents some collection of points, so the size and color of the markers are scaled to represent the magnitude of this collection. Markers representing a single point are set to have a radius of 40 pixels. From there, the marker's radius scales with the squared logarithm of the number of points ( r M = 40 + log 2 |M |) Marker color follows a linear scaling from green for markers representing a single point, up to red for those that have more than 100 points. ese functions are chosen to be visually appealing, but there is no other concrete reason for their selection. e map query interface also gives access to the news articles from which the markers are derived. When a marker that represents more than one physical location is clicked, the zoom level is increased to a point where the constituent points are represented by separate markers, but when a marker that represents a single location is clicked, a pop up listing the articles from within the speci ed time range that mention the location is opened.

e rst control located below the map interface is a slider bar that can be manipulated to control the range of dates in the data displayed on the map. Moving either knob on the slider causes the map interface to be updated. e exact dates selected are displayed below the slider.

Below the date range slider there are three panels labeled "Download Controls, " "Animation Controls, " and "Layer Controls. " e download controls are used to retrieve data from our database. First, there is a text entry box labeled "keyword" in which the user types a keyword that they want to visualize. ere are then two date entry elds labeled "Data Start Date" and "Data End Date". ese control the range of dates for data that is downloaded from our server. Note that this is not necessarily the same as the range of dates for data that is rendered on the map. Finally, there is a bu on labeled "Download Data" that, when clicked, sends a request to our server for data within the speci ed time range matching the provided keyword.

is data is rendered on the map query interface, replacing any previous data. e animation controls deal with using downloaded data to generate time series animations. Most prominently, there is the "Start Animation" bu on that initiates the animation when clicked. Clicking this bu on second time will stop an ongoing animation. e properties of this animation are controlled by the next three inputs. e rst input controls the size of the range of dates displayed in each frame of the animation. It can be set to hour, day, week, or month. It can also be set to a custom size through use of the "Displayed Data Range" controls discussed above. Next, their is a selection input for the interval advanced between each frame of the animation. is can be set to hour, day, week, or month. e nal input eld is used to set the speed of the animation to either "slow", "normal", or "fast". ese values correspond to a minimum number of milliseconds waited between frames of the animation: 500, 100, and 0 milliseconds respectively. While these are minimum delays, the actual interval may be longer due to time required to compute the subsequent frame.

Finally, the layer controls are used to select between the possible data layers that can be rendered on the map interface. By default, only the news keyword data layer is selected. When the con rmed cases layer is selected, there are marker representing the total number of con rmed cases in a location rendered on the map. Similarly there are layers for recoveries and deaths that contain markers representing these quantities. e data used to generate these layers was originally gathered by Dong et al. [1] .

is section provides a walk through demonstrating how our application can be used to obtain an animation to show the geographic change in the discussion of the Novel Corona Virus over time. e instructions here are applicable to other similar uses of our application.

To begin, navigate to coronaviz.umiacs.io to access our website. Once there the "Keyword" entry eld should be initially populated with "coronavirus". If this is not the case, select the eld and type in this term. e next step is selecting the range of dates to be viewed. A reasonable range for visualizing Corona Virus might start in December 2019 and proceed until the current date. To make this selection, click on and select a date for the "Data Start Date" eld. "Data End Date" should have been initialized the current date, but, if it is not, a date should also be selected here. Once the keyword and date range have been selected, click "Download Data" to download this data from our server. is may take a noticeable amount of time depending on network speed and the size of the query result. Once the download is complete, data will be rendered on the map. e next step is to con gure the animations controls and initiate an animation sequence. e animation options ("Animation Window" and "Animation Step") are set to week and day respectively by default. ese should be reasonable values but, at this point, they can be changed. Clicking the "Start Animation" bu on starts an animation on the map query interface that begins by displaying downloaded data with the earliest publication date and proceeds until the nal frame that contains the most recent data. Each frame of the animation shows a period of time de ned by "Animation Window" and the start of the frame is advanced by an interval de ned by "Animation

Step" between each frame. e animation can be terminated early by clicking the bu on, now labeled "Stop Animation", a second time.

In this section we evaluate the e ectiveness of tracking keyword usage in news article to be er understand the progression of COVID-19. To do this we compare the con rmed case numbers collected by Dong et al. [1] to the cumulative number of articles recognized by our system. We include geographic information in our analysis by limiting it to con rmed cases and news articles in a speci c location. We are able to restrict the analysis in this manner because the dataset of con rmed cases we use already organizes con rmed case numbers by location, and our news article dataset is geocoded as described in Section 2.2. To further investigate the utility of our technique, we perform the comparison at three geographic scales: world, nation, and region. e results of this evaluation are summarized in Table 1 and Figures 2, 3 , and 4. 2 We rst ignore speci c geographic information extracted from articles and perform this analysis on a global scale. is does not utilize the geographic information we collect, but it is interesting as a point of comparison to other scales. Figure 2a plots the daily number of articles using the keyword "coronavirus" as a time series 2 In each plot of article counts, the daily number of articles collected is zero from Febuary 14 to Febuary 17, March 9 to March 11, and on March 22. is was caused by problems in our data collection system, so the data points are removed when computing the correlation coe cients in Table 1 . between January 16, 2020 and March 19, 2020. To demonstrate the utility of news articles in tracking the progress of the virus, this gure also shows a time series for the daily number of new con rmed cases world wide. It is clear that both datasets trend upwards at similar rates. To determine if correlation exists between article counts and case numbers, we calculated the Pearson correlation coe cient between the datasets. We found coe cient to be 0.68, indicating that the datasets are linearly correlated. To investigate the e ect of our keyword lter, we performed a similar comparison where we do not lter news articles for any speci c terms. Figure 2b show the time series plot for this data. is time, the article counts clearly do no correlate with case numbers. is is realized in a −0.05 correlation coe cient. Since the magnitude of the correlation coe cient is signi cantly larger when applying the keyword lter, we can see that it is useful in this application. e plots in Figure 3 are forma ed the same as Figure 2 , but they focus on smaller geographic scales. In Figure 3a , the plot is restricted to articles geocoded to China and con rmed cases in China. e correlation coe cient between these datasets is −0.09, indicating that there is no correlation a er applying these geographic lters. Figure 3c further restricts the data to the Hubei province of China. e correlation coe cient at this scale is −0.11, similar to found for all of China. Figures 3b, and 3d show the plots for China and Hubei when the keyword lter is not applied. Since there was no correlation found with the lter applied we do not expect anything to change when it is removed. As expected, the new correlation coe cients are almost unchanged with −0.08 and −0.11 for China and Hubei respectively.

In our nal set of evaluations, we shi geographic focus from China to the United States. found for the United States (0.71) is greater than that found in the global analysis. While the correlation coe cient calculated for the Washington data is much lower (0.24), it is notably higher than the values computed for either China or Hubei. e correlation coe cient for the United States and Washington when not applying the keyword lter (0.06 and 0.10 respectively) is comparable to that for that of the three previous datasets.

In our evaluations we found that the number of news articles using the term "coronavirus" is correlated with the number of new cases of COVID-19 when correlation is computed at a global scale. When restricting the analysis to a country wide scale, we found correlation when using data for the United States, but we were not able nd any correlation in our dataset for China. ere is a similar but less pronounced relationship at the region level. While the Washington state dataset does not display particularly strong correlation with the con rmed cases dataset, it is notably higher than the correlation we found for Hubei. A possible explanation for this inconsistency is that many articles discussing COVID-19 will mention China at some point since it is the origin of the virus. To overcome this issue we may be able to develop a technique to determine which term, toponym pairs in a news article should be included in our dataset. Our dataset currently contains an entry for every toponym in an article if the article also contains the query term.

An interactive web-based dashboard to track COVID-19 in real time

Multifaceted toponym recognition for streaming news

Adaptive context features for toponym resolution in streaming news

Reading news with maps by exploiting spatial synonyms