Exploration on Media polarity in the News

Data Story: https://rhuangag.github.io/DataStory/

Abstract

We humans have never experienced this era of information explosion before. Every day, thousands of media spread millions of pieces of news with their own remark to some degree. Instead of keeping the purity, objectivity, and neutrality of news, some media are discovered to publish diametrically opponent narratives about the same event. The media no longer work as a convenient information channel, but also the propaganda artifice of political parties and consortiums. The consistency to the truth inspires us to summarize the phenomena of media polarity. In our project, we will mine in the dataset 'News on the Web' (NOW), which contains 6.6 billion words and the worldwide newspaper and magazines. Through sampling and analyzing this authoritative data collection, we will try to find out the internal positions of some media agencies on some sensitive and controversial issues, such as the presidential election. Also, from the chronological aspect, we hope to find out whether there exist time variation for one specific media agency from 2010 to present. We might use some technologies like natural language processing, machine learning, text mining and cloud computing.

Research questions

Will the media report the news objectively?
For the same event, how different media report it? Are there any reasons behind this phenomenon？
For the same media, will it change its view for the same news over time?

Dataset

We will use the dataset 'News On the Web'. In this dataset, there are many .txt files in different folders. The files under 'source' folder contains id, data, country, media, source and title. For example:

460137  455     13-08-17        CA      CityNews        http://www.citynews.ca/2013/08/17/sheila-copps-caught-up-in-cambodian-election-crisis/  Sheila Copps caught up in Cambodian election crisis

If we look at the content of the news, we can find that it is matched with the news on corresponding website. For example:

@@14637662 <h> Willett hears it from the Hazeltine crowd <h> Related Links <p> Chaska -- Danny Willett really had no idea what to expect during his first career match in the pressure-packed Ryder Cup .

What's more, it also provides files containing word, lemma and PoS.

14637200        4738387006      Trailing        trail   vvg_jj

We will extract the useful news related to different media. And we may sort the data on country, date and some key words.
We may use some natural language processing methods and make use of the dataset containing lemma and PoS, to find out the polarity of the words and the whole news.

Milestone 2 timeline

11.11:
- Build subdataset locally and clean the data.
  - The dataset has size more than 300 GB, thus to enable fault tolerant development, we selected the news of the United States from August to October in 2016.
  - By exploring the data content, we found out that there are a lot of meaningless characters, e.g., punctuations which would bring difficulties for further analysis. So we cleaned the data by using regular expression to remove them.
- Apply some natural language processing methods to analyze the data.
  - We found out that some type of news are more popular than others and thus more suitable for analysis. To select those popular topics out, we apply tf-idf to the news documents.
  - To enable further investigation on polarity of media, the polarities of pieces of news are necessary. Therefore, we utilized nltk and textblob sentimental analyzers to obtain those numbers of news.
11.18:
- Get preliminary result from the sample dataset
  - Through investigating into the popularity of topics, we found out that Trump was far more popular than others in our sampled dataset. So we select this topic as a starting point of analysis on polarity.
  - For simplicity we started our analysis on the data in Oct 2016.
  - Then we obtained the polarity of each piece of news. Multiple methods were tried in this section and they will be discussed in details in the Discussion part C.
  - We then got polarity of each media.
11.25:
- Do the data visualization
  - We used boxplot to illustrate the polarity distribution of medias.
- Extension to bigger datasets
  - After validating the results on one month basis, we took three months into account both individually and collectively. The polarity distribution of each month and the change through months were visualized.

Response to the questions about milestone1

The news are supposed to be written without a polarity. What should we do? Ans: First of all, we use our analyzing tools to calculate the polarity of randomly sampled news. Then, we read and check the news and find out that they do have different polarity. One possible explantion resides in the nature of media: to gain attention from possible readers. Thus they may take different stands to adapt to different classes of readers. This assumption enables us to conduct the further analysis of polarity of media.

Discussion

The problems we encountered during working on milestone2

A. Select Topic

We choose the top 5 keywords to select the topic without any priority. Therefore, sometimes these 5 keywords may not be related to each other. We might choose relative keywords from the most frequent keywords in the future work.

B. Get Original Data from Website

When we try to get the content of the news, sometimes we will also get some irrelevant data. We need to find an efficient way to get only the data we want.

C. Conduct Sentimental Analysis

Now we use sentiment analyzer module of NLTK and TextBlob to analyse the polarity of each sentence in an article. And we set positive threshold and negative threshold to classify the polarity of the sentences. By calculating the percentage of sentences with different polarities, we determine the polarity of the whole article. One of the problems is that some of the content of the news might be irrelevant to the topic (e.g., Thank you for your reading) because of the incorrect data extraction. And this irrelevant sentences might have a very high or low polarity score, which has a bad influence on our results. Our solution is that we subtract the median polarity score from the mean score to avoid the bias.

Milestone 3 timeline

12.1
- Build up our classifier on polarity of article
- Extend our range to bigger datasets using spark
12.9
- Modify our classifier (if needed) to better decide the polarity about various topics
- Explore the polarity comparisons from different perspective:
  - Variance of polarity of same media on same topic overtime
  - Variance of polarity of different media on same topic
  - Variance of polarity of same media on different topics
12.15
- Finish analysis and visualization of polarity from above viewpoints

Response to the questions about milestone2

1.How to show that we correctly classifying the news?

Ans: First of all, we randomly chose some articles as samples and classified them manually. All the team members read those sample articles and tagged them as "positive", "negative", or "neutral" news with an integer value 1, -1 or 0. Then, we chose the majority tag to be the final polarity tag for an article. Finally, we used our sentimental classifier to tag those articles and calculated the accuracy.

2.Is the detected polarity score significant?

Ans: The noise in the sentimental analysis is quite difficult to deal with, and we found out the quantitative result could not distinguish the media polarity very much. And the time variation was not very obvious. Therefore, we switched our method from quantitative to qualitative. Now we found the threshold and classify the articles to three classes which are 'positive', 'negative', and 'neural', based on the polarity score. The results became better after changing the method.

3.Is the polarity change over time significant?

Ans: As we mentioned above, we classify the articles to three classes instead of polarity scores. The results showed that the proportion of three classes of medias almost remained the same except October. Therefore, we have checked the content of some articles of October. Surprisingly, we found many pieces of news about scandals in the dataset of October. After searching on the internet, we found the phrase “October surprise”. An October surprise is news specifically (and cynically) timed by one side or another to sway an election, which explains the increasing "negative" tags in October and showed that our results were reliable.

Main workflow in milestone3

Improvement of Milestone2

Media Grouping and Selection: for one media, it can be presented in multiple formats, such as 'Fox News', 'fox6', and 'www.fox5sendiego.com'. Therefore, we need to collect and group them together as one media. Also, as there are thousands of medias on the website, we decided to choose the mainstream medias based on the reference(https://www.allyoucanread.com/american-newspapers/).
Qualitative method: We switched our method from quantitative to qualitative, which is tagging the articles with "positive", "negative" and "neutral" labels based on the polarity score.

Topic Expansion

Based on the keywords we got in milestone2 with ti-idf, we found more topics related to Trump by three methods which are simple frequency check, cosine similarity check, and Latent Dirichlet allocation.

Media Polarity Analysis for New Topics

We tagged the articles based on the polarity score we got in milestone2, and then calculated the proportion of three classes for each media. With the results, we compare the polarity of different medias about the same topic, and then compare the porlarity of different topics for the same media.

Visualization

We visualized the results we got above to show the polarity of medias in different topics.

Contribution

Fengyu Cai: Coming up with the research questions, topic selection and expansion, writing up the data story;

Liangwei Chen: Visualization and plotting the graphs, code optimization;

Ruibin Huang: Coming up with the sentiment analysis method, data analysis for extended topics.

Special thanks:

Dr. Periklis Chrysogelos, Data-Intensive Applications and Systems Laboratory (DIAS), École polytechnique fédérale de Lausanne, gave us feedback and suggestions during the whole project process.

TRUMANCFY/Project_ADA_2018