1 – Characterize the news data sources provided. Which are primary sources and which are derivative sources? What are the relationships between the primary and derivative sources?
Please limit your answer to 8 images and 300 words.
We differentiate sources based on both the novelty and the number of their articles.
Primary sources report novel events timely, while derivative sources plagiarize and aggregate articles from primary sources. Besides, primary sources should cover most events.
Figure 1-1. The image presents the pipeline to analyze the information flow of different sources.
Firstly, we use the plagiarism detection algorithm to extract the news plagiarism network. This algorithm examines the similarity of sentences in different articles. We then aggregate the news plagiarism network as the media plagiarism network and visualize it to show how information flow through different sources.
We propose a novel pipeline (Fig. 1-1) to analyze the information reuse among sources. The pipeline leverages plagiarism detection to measure the similarity between articles. Similar articles above the similarity threshold (0.4, chosen empirically) have plagiarism connections. This process results in a news plagiarism network containing lots of connected subgraphs. For each subgraph, we identify the media which post earliest and view it as the origin of the subgraph (see Fig. 1-3 as an example). We then aggregate all connected subgraphs into a summarized media plagiarism network (Fig. 1-1). Each edge in the network stands for how many articles origin from the source media. We also consider the explicit citation relationship to construct the network (see Fig. 1-4).
Figure 1-2. One connected subgraph in the news plagiarism network.Figure 1-3. News articles in the connected subgraph shown in Fig. 1-2.
There are five articles in the subgraph. The Abila Post is the origin of the subgraph. It reports the event on October 1st in 2010, while other media report in the latter days. Besides, other media make up the news by copying and pasting with minor changes from the news 233.
Figure 1-4. We also consider the explicit reference statements between news.
For example, articles 342 and 154 contain explicit referral sentences, such as "reprinted with permission from the Homeland Illumination".
We use the Sugiyama algorithm [1] to visualize the result as a flowchart (Fig. 1-5). Compared with a traditional force-directed graph, this method places the original media at the beginning layers and maintains most edges in the same direction.
In Fig. 1-5, we find Kronos Star, International Times, The World, Homeland Illumination , and The Abila Post are five primary sources. In contrast, others are derivative because they plagiarize most news from primary sources. All the five media locate at the top-3 layers in the flowchart. Other media either have limited articles or have lots of non-original news. Note that News Online Today has the most articles.
However, it is a news collector from the five primary media. Therefore, we consider News Online Today as a derivative source.
Figure 1-5. Information flow of different media sources.
We use the Sugiyama algorithm to layout the media. The Sugiyama algorithm reverses the minimum edges to transform the media plagiarism network into a layered directed graph. Besides, it places nodes at different layers with minimum crossings. Finally, we recover the reversed edges shown as the red edges. Each blue node stands for the media, and its width represents the number of articles. Gray edges show the information flow from left to right, while the red ones from right to left. We can conclude that the five media (Kronos Star, International Times, The World, Homeland Illumination, and The Abila Post) are primary sources due to their broad coverage of novel articles.
2 – Characterize any biases you identify in these news sources, with respect to their representation of specific people, places, and events. Give examples.
Please limit your answer to 6 images and 500 words.
We consider two kinds of biases: commission and omission of information,
and event selection [2]. We analyze the bias of commission and omission by comparing the information contained in news articles in different sources. Event selection needs analysts to uncover how various sources differ in the news coverage at the event level.
Figure 2-1. The pipeline to analyze media bias:
we leverage named entity recognition to extract people, places, and organizations from sources first. Then we cluster the articles based on aggregated metrics considering keyword distance, time distance, and entity distance. Each cluster of the articles reports the same event. Finally, we use correspondence analysis to visualize the media bias with respect to specific people, places, and events.
We first use named entity recognition to extract keywords of people, places, and organizations and then count the frequency of the keywords for different sources. This process results in a source-entity contingency table. For event selection bias, we need to extract events from sources first. We leverage a similar technique [3] to clustering articles based on an aggregated metric, including keyword distance, entity distance, and temporal distance. Fig. 2-1 presents the definitions of the three distances. Then we cluster the articles based on the aggregated similarity metric and visualize them using dimension reduction technique t-SNE. The resulted map enables analysts to explore the relationship between each event visually. Finally, we count the frequency of the events for different sources. This process results in a source-event contingency table.
Figure 2-2. Correspondence analysis of the sources with people and places:
(a) The star coordinates places sources in the coordinate and entities (people and places) at the ring. Media at the center of the coordinate are fairer with most coverage of the entities while others will bias towards specific entities. (b) The corresponding source-entity contingency table of the star coordinate.
We leverage correspondence analysis [4] of the source-entity table and source-event table to analyze the two kinds of biases. The results are two star coordinates [5] (or biplots[6]) (see Fig. 2-2 and Fig. 2-3). Each star coordinate projects the media in the circle and projects entities or the events on the ring of the star coordinate. The size of each source represents the number of its articles. The orders of the entities and the events along the ring are determined by the correspondence analysis automatically. Sources that close to the center of the star coordinate are fairer, which cover the most information. In contrast, sources towards specific entities or events mean that they bias towards them.
In Fig. 2-2, we find entities for different organizations cluster on the ring. News Online Today is fairer than other sources because it collects most events from the five primary sources, which is consistent with the conclusion from the first question. Besides, sources cluster at different fans of the star coordinate. For example, The World, The Tulip, The General Post, The Light of Truth, and the Who What News cluster together and bias towards the persons of GAStech and APA. While Homeland Illumination, The Wrap, All News Today, Daily Pegasus, and The Orb bias towards persons of POK. Similarly, in Fig. 2-3, correspondence analysis groups the events into four main clusters and the sources bias towards different event clusters.
Figure 2-3. Correspondence analysis of the sources with events:
(a) The star coordinates places sources in the coordinate and events at the ring. Media at the center of the coordinate are fairer, with most coverage of the events while others will bias towards specific events. (b) The corresponding source-event contingency table of the star coordinate. (c) The summarized titles of every event.
Fig. 2-3 provides an overview of how sources bias towards different events. To discover more details of the bias for each event, we provide an interactive tool to explore the events by moving a lens on events. For example, in Fig. 2-4, sources split their opinions into two groups. One group bias towards POK, the other one bias towards the police. Another interesting example is shown in Fig. 2-5.
Figure 2-4. Analysts can leverage a lens to discover specific biases on events:
(a) This lens reveals two clusters (C1 and C2) of news articles for the selected event. (b) The cluster C1 comes from the primary media Homeland Illumination with five derivative sources. This cluster is consistent with the pattern in Fig. 2-2. (c) The other cluster C2 comes from the primary source Kronos Star with five derivative sources. This cluster is also consistent with the pattern in Fig. 2-2. (d) We select one article from Homeland in C1 as the representative news. Other articles in this cluster copy and paste the article with minor changes. We can find articles in C1 hold a positive attitude towards POK. (e) We show a representative article in C2 from Kronos Star. Notice that sources in C2 bias towards police and pay more attention to traffic blocking due to the movement of POK.
Figure 2-5. Another example to reveal media bias on events:
(a) This lens reveals two clusters (C1, C2 and C3) of news articles for the selected event. (b) C1 contains the primary source Homeland Illumination and four derivative sources. (c) C2 includes the primary source Kronos Star and four derivative sources. (d) Homeland Illumination strongly emphasizes how the defeated tax policy influences POK and even states that the POK is a grassroots environmental group. (c) However, Kronos Star says nothing with POK. (d) While News Online Today, just copy one sentence from Homeland Illumination.
3 – Given the data sources provided, use visual analytics to identify potential official and unofficial relationships among GASTech, POK, the APA, and Government. Include both personal relationships and shared goals and objectives. Provide evidence for these relationships.
Please limit your answer to 6 images and 400 words.
Figure 3-1. Pipeline to discover the relationships among organizations.
We propose a visual analytics method to discover the relationships among entities using named entity recognition and neural relation extraction [7] with human-in-the-loop (Fig.3-1). We visualize the extracted knowledge graph using a force-directed layout (Fig.3-2). Besides, to discover the potential relationships among GAStech staff, we leverage both their resumes and emails. We first cluster the 1175 emails into 278 clusters, each contains the same group of staff. There are two kinds of clusters: one has staff in the same department, the other includes staff in different departments. We pay more attention to the cross-department emails as they are more suspicious. Finally, there are only 61 clusters of cross-department emails (Fig.3-1 (b)), which greatly reduces the human labor to examine the email contents.
Figure 3-2. The knowledge graph extracted using named entity recognition and neural relation extraction after human fix.
Discovery 1: POK may lead the APA.
Evidence: POK may use sales of the drug to finance their activities (Fig.3-5 (c)), while APA is also engaged in drug trafficking (Fig.3-3(a)). Therefore, it is reasonable to believe POK leads the APA for drug selling.
Discovery 2: Some staff in GAStech might be potential members of POK, including Rachel Pantanal, Ruscella Mies Haber, Inga Ferro, Isia Vann, Loreto Bodrogi, Hennie Osvaldo, and Minke Mies (Fig.3-2).
Figure 3-3. Evidence that Ruscella Mies, Ruscella Mies Haber, Inga Ferro, Isia Vann, Loreto Bodrogi, Hennie Osvaldo, and Minke Mies are potential POK members in GAStech: (a) These news articles explain the APA has been engaged in drug trafficking, and it has sought to build brand through a high-quality magazine called "Arise". (b) Some employees in GAStech mentioned "Arise" through emails.
Evidence 2-1: APA has sought to build their brand through a high-quality magazine called Arise (Fig.3-3 (a)). While people in this group have mentioned "Arise" through emails (Fig.3-3 (b) and Fig.3-4 (b)).
Evidence 2-2: Loreto Bodrogi, Hennie Osvaldo, and Minke Mies have the same last names as Henk Bodrogi, Carmine Osvaldo, and Valentine Mies respectively, members in POK during 1997-2001, which may imply these people have family connections.
Figure 3-4. Evidence that Rachel Pantanal, Isia Vann and Ruscella Mies Haber are potential POK members in GAStech : (a) These news articles explain the APA has been engaged in drug trafficking, and it has sought to build the brand through a high-quality magazine called "Arise". Besides, the drug has a nickname called flower tulip. (b) Rachel Pantanal and Isia Vann mentioned flowers in their emails. Rachel Pantanal and Ruscella Mies Haber mentioned "Arise" in their emails.
Evidence 2-3: APA has been engaged in drug trafficking, and the drug's nicknames include the tulip flower (Fig.3-4 (a)). While Isia Vann and Rachel Pantanal have mentioned "flowers" in their emails. This evidence may be implicit. However, it implies the two persons engage in a personal relationship.
Figure 3-5.
(a) Loreto Bodrogi, Isia Vann, Hennie Osvaldo, Minke Mies, and Inga Ferro have worked in the Armed Forces of Kronos. (b) Isia Vann is the older brother of Juliana Vann, the 10-year-old girl who died in 1988 after drinking contaminated water from the Tiskele River. (c) POK may use sales of the drug to finance their activities.
Evidence 2-4
: Most of them have worked in the Armed Forces of Kronos (Fig.3-5 (a)).
Evidence 2-5
: Isia Vann is the older brother of Juliana Vann, the 10-year-old girl who died in 1988 after drinking contaminated water from the Tiskele River. This event also brought serious schism in the POK (Fig.3-5 (b)).
Figure 3-6. The relationships between the government with other organizations.
Discovery 3: In 1999-2001, the government supports POK for their shared objective to protect the environment (Fig.3-6 (a)).
Discovery 4: In 2012, the relationship between the government and POK deteriorates. Vincent Kapelou, minister of health, received letters threaten from POK. While government spokesmen claim the POK as gangs of criminals (Fig.3-6 (b)).
Discovery 5: The government may ally with GAStech (Fig.3-6 (c)).
4 – If you solved this mini-challenge in 2014, how did you approach it differently this year?
We did not solve the mini-challenge in 2014.
Reference
K. Sugiyama, S. Tagawa and M. Toda, "Methods for Visual Understanding of Hierarchical Systems". IEEE Transactions on Systems, Man and Cybernetics 11(2):109-125, 1981.
Hamborg F, Donnay K, Gipp B. Automated identification of media bias in news articles: an interdisciplinary literature review[J]. International Journal on Digital Libraries, 2019, 20(4): 391-415.
Mishra A, Ginjpalli S, Bryan C. News Kaleidoscope: Visual Investigation of Coverage Diversity in News Event Reporting[J]. arXiv preprint arXiv:2105.03839, 2021.
Greenacre M. Correspondence analysis in practice[M]. chapman and hall/crc, 2007.
Rubio-Sánchez M, Raya L, Diaz F, et al. A comparative study between radviz and star coordinates[J]. IEEE transactions on visualization and computer graphics, 2015, 22(1): 619-628.
Greenacre M J. Biplots in practice[M]. Fundacion BBVA, 2010.
Han X, Gao T, Yao Y, et al. OpenNRE: An open and extensible toolkit for neural relation extraction[J]. Proceedings of EMNLP-IJCNLP: System Demonstrations:169-174, 2019.
