Entry Name: "PKU-Yuan-MC1"

VAST Challenge 2021
Mini-Challenge 1

Team Members:

Shaocong Tan, Peking University, tanshaocong0108@gmail.com PRIMARY

Sihang Li, Peking University, lisihang@pku.edu.cn

Shuai Chen, Peking University, shuai.chen@pku.edu.cn

Jiacheng Yu, Peking University, jiachengyu@pku.edu.cn

Chuanming Huang, QI-ANXIN, huangchuanming@qianxin.com

Liang Tang, QI-ANXIN, tangliang@qianxin.com

Qinghua Shang, QI-ANXIN, shangqinghua@qianxin.com

Zixi Fu, QI-ANXIN, fuzixi@qianxin.com

Zhuo Zhang, QI-ANXIN, zhangzhuo@qianxin.com

Xiaoru Yuan, Peking University, xiaoru.yuan@pku.edu.cn ADVISOR

Student Team: Yes

Tools Used:

D3.js

Approximately how many hours were spent working on this submission in total?

200

May we post your submission in the Visual Analytics Benchmark Repository after VAST Challenge 2021 is complete? YES

Video

https://vis.pku.edu.cn/VASTChallenge2021MC1/VASTChallenge2021MC1.mp4

or

VASTChallenge2021MC1.mp4

Questions

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 use TF-IDF to vectorize each news, calculate cosine distances between these vectors, and generate clusters by the similarities together with manual refinement. For any two news in one cluster, if A is posted after B, then A is possibly derivated from B. The derivation is uncertain if they are posted on the same day. For one news in one cluster, if others are all posted after it, it is called an origin. If some were posted before it, it is called a target. Otherwise, it is uncertain. We draw the graph of the sources according to the relations of news among them and encode the types of news in each source (Figure 1-1).

Figure 1-1: Groups of sources. The primary sources of each group are pointed.

There are 5 groups, each with a news source that mostly publishes original news (Figure 1-1), and we regard it as the primary source. Other sources in each group are derivative sources. All the groups are connected by News Online Today, which has no origins and only reposts others with some cuts (Figure 1-2).

Figure 1-2: Forwarding characteristics of News Online Today. Delete operations are encoded purple.

Notably, Homeland Illumination has some targets in the graph. We then find that it has posted two similar news (297 and 737) to correct mistakes. They are all reposted by News Online Today, which results in news 737 being considered as a target (Figure 1-3).

Figure 1-3: Homeland Illumination reposts another news of it, which changes the number. They are all reposted by News Online Today.

Also, Who What News has a few out edges in the graph, which means it may be a sub-primary source. We then find that it usually cuts some contents from the primary sources, but will then be added in others. The width of the lines also shows that other sources have closer relationships to the primary sources (Figure 1-4). So Who What News is just a derivative source.

Figure 1-4: An example of the relationships among The World, Who What News, and other sources. Delete operations are encoded purple, while add operations are encoded green.

The group of Kronos Star has connections with the group of The Abila Post, implicating the correlation between them. Also, the two main sources have some news with similar meanings (Figure 1-5). So there are also some reposting relationships between the primary sources of different groups.

Figure 1-5: Similar news of primary sources Kronos Star and The Abila Post.

Also, there are three outliers in the graph. Their news is all similar to some others in existing groups posted on the same day. They can be considered derivative sources of certain primary sources (Figure 1-6).

Figure 1-6: Sources with all uncertain news are considered to repost some certain primary sources.

To sum up, we analyze the source groups and identify that The World, Homeland Illumination, The Abila Post, International Times, Kronos Star are the primary sources, while others are derivative sources. News Online Today is derivated from all 5 groups (Figure 1-7).

Figure 1-7: Relations of the sources. Blue ones are primary sources, while green ones are derivative sources.

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.

This question focuses on bias in news reporting. There are various definitions of media bias. We take the common definition of media bias proposed by D'Alessio and Allen [1], which distinguishes between three types: coverage, gatekeeping, and statement. Coverage bias is about the visibility of entities such as specific people or places in media bias. Gatekeeping bias is concerned with what news media choose to report about an entity. Statement bias relates to how to report on an entity.

Guided by the three types of bias, we developed a visual analysis system. As shown in Figure 2-1, the matrix-based visualization shows coverage bias in news sources with respect to people from different groups. Each row represents one news source, which is arranged by connections between them. The thread arcs left represent identified reposting relations between news sources. As there are heavy derivation relationships between news in a group, the entities reported in each group are quite similar. The coverage bais is also clear that Group 4 only report news about people from GASTech. Among these people, Sten Sanjorge Jr., CEO of GASTech is reported the most.

Figure 2-1: A matrix-based visualization to investigate the coverage bias of news sources on different entities.

We find that news sources report different topics about Sten Sanjorge Jr. in the articles (Figure 2-2). Kronos Star, The Truth, The Guide and Wordwise (mostly in Group 1) report the topic "Celebrate" about Sanjorge, which are about two events, i.e., "the opening of new GASTech offices on 1995-10-10" and "the cooperation between GASTech and Kronos". However, these news sources did not report the pollution of the environment due to the company (Figure 2-2 a). Other news sources (Group 2, 4, 5) reported the environmental problem caused by the company.

Figure 2-2: The parallel tags view (a) shows gatekeeping bias against Sten Sanjorge Jr., i.e., what news sources talk about in the articles when mentioning Sanjorge. The timeline view (b, c) shows the temporal distribution of news related to the selected topic in (a).

Another example is the bias against "Henk Bodrogi", one of the founders of POK. Only seven news sources reported him (Figure 2-3 a) and Homeland Illumination is the primary source (Figure 2-3 b). For news reporting Bodrogi, the topic "Health", relating to the contaminated war, was more covered.

Figure 2-3: Bias against bias against "Henk Bodrogi", one of the founders of POK.

When reporting places, there are also biases in the news sources (Figure 2-4). Tiskele is where the gas field is located. On one hand, some news sources like (Kronos Star, The Guide, WorldWide) report the benefit of tax income. On the other hand, water pollution by gas drilling is reported by news sources like The Wrap, News Desk.

Figure 2-4: Bias against the place Tiskele, at which the gas field is located.

When reporting the same event "GASTECH ADOPTS NEW LOGO", News Online Today ignored the ongoing protest of POK (Figure 2-5).

Figure 2-5: Statement bias that when reporting the same event.

When reporting the event "TISKELE BEND EXPLOSION AND FIRE", there are three news sources. The Ablia Post is the primary source while Central Bulletion and News Online Today are derivative sources. News Online Today deleted the detailed information of explosion when reposting. But Central Bulletion deleted the sentence "in contrast to previous reports there were no deaths and none missing as a result of the incident. The deleted sentence could relieve the public concerns about the explosion.

Figure 2-6: Bias against reporting explosion and fire at the TIskele Bend.

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.

We combined entity extraction and affective computing technology to extract two main categories of relationship, ally and opponent, from news collection and historical reports, which are represented by different colored edges in the entity graph (Figure 3-1). We can see that GASTech and Government are allies, while the POK and the APA are allies.

Figure 3-1: Relationships among the entities and individuals belonging to them. Green edges for allies, while yellow edges for opponents.

Then we filter the nodes to further see the relationship between GASTech, POK, APA, or Government. It can be seen that the relationships between GASTech and Government and their employees are almost allies (Figure 3-2).

Figure 3-2: Relationships between GASTech and Government.

The relationship between POK and GASTech is mostly antagonistic (Figure 3-3). However, we find two exceptions: Isia Vann and Hank Fluss. Isia Vann works for GASTech as Perimeter Control and he is the brother of Juliana Vann. The next possible leader of POK, Mandor Vann, is his uncle. Another person Hank Fluss is the former COO of GASTech, who used to meet with Carmine Osvaldo and Henk Bodrogi, the two founders of POK, on the issue of contamination.

Figure 3-3: Relationships between GASTech and POK. GASTech COO Hank Fluss is highlighted.

The relationships between POK and the government are also mostly adversarial (Figure 3-4), with the exception of one case, Cesare Nespola, Minister of Health, who has had many meetings with the founders of POK. While the successor of Cesare Nespola, Vincent Kapelou is hostile to POK.

Figure 3-4: Relationships between POK and Government.

In the email data of GASTech employees, we annotate the email type by the semantics of the title of the email data and classify the emails into work-related and non-work-related, and by filtering on the number of emails on the edges, we can see that the emails related to work constitute a network in which GASTech employees can be clustered into several groups, which are employees of different departments of GASTech (Figure 3-5).

Figure 3-5: Email network of GASTech company. Work-related emails.

We further filtered the non-work-related emails and adjusted the threshold value of the edge filter. It is shown that there are two departments in GASTech with more non-work-related communication in emails from their employees (Figure 3-6). These two groups are the employees in the Administration and Facilities departments, respectively. We further observe the content of the emails on the side and the employee titles of the employees and find that there are some problems with our labeling. For the employees with the position of truck drivers, "Traffic advice for today" and "Route suggestion for next shift" were marked as non-work related. We further adjust the labeling of the data to get the following email network.

Figure 3-6: Email network of GASTech company. Non-work-related emails.

4 –– If you solved this mini-challenge in 2014, how did you approach it differently this year?

The mini-challenge this year requires us to explore the relationship among news sources, entities, and people, which is more complex and ambiguous than that of 2014. The first question and second question are not explicitly proposed in 2014. In addition to using the above visualization to show different aspects of the relationship, we also introduced various machine learning methods to lighten the burden of data processing and improve the performance of real-time interaction.

In the first question, we use Bert as the encoder to calculate the similarity of the news and align two paragraphs sentence by sentence. It avoids manual comparison and classification one by one, making our system more scalable.

In the second question, we adapt the topic calculation model to extract significant events in the news collection and assign topics to each news based on its semantics. The topic labels on the visual interface speed up the efficiency of analyzing the news.

In the third question, unsupervised entity extraction helps us quickly locate sentences in news and background materials that can recognize the relationship between people and organizations.

Although machine learning has made great progress in these years, it is still not efficient for analyzing complex text documents automatically.

For example, when identifying biases in news sources, we try to find out the attitude of each article with the sentiment computing model. However, due to the complexity and diversity of semantics, we cannot train a good model under the current technology and data volume. When analyzing the primary and derivative sources, it is still necessary for users to interactively correct wrong connections, or enhances the utilization of clues based on multiple knowledge. Human interaction and feedback are needed in the process of analysis. Our method can reveal complex relationships interactively by linking overviews to detailed views.

Reference

[1] Dave D'Alessio, Mike Allen. Media Bias in Presidential Elections: A Meta-Analysis. Journal of Communication. 50(4): 133–156, 2000.