introduction: 

In today’s data-driven world, visualization is key to unlocking insights and making informed decisions. With the tools we use, and the way we approach them. This can profoundly impact our results. This is my white paper on how I created a Sentiment analysis, a small word cloud and a scatterplot using Tableau Public.  I found a sizeable dataset on Kaggle.com called “News Detection (Fake or Real) Dataset” which has 9865 unique Values (rows) and one column indicating either this article is real or fake.  

While the dataset lacks official verification, I selected it for this class project due to the limited availability of alternative options at the time. 
 

My digital humanities project was to create this in Tableau Public, which, while powerful, has different strengths and limitations. 

Sentiment Analysis and Visualization: 

In the first chart you see on my project. I went with the approach of weighted scores and distinct categories for strong positive, moderate positive, moderate negative, and strong negative sentiments. This allowed for an exploration of the emotional content/trigger words within the news articles. For the Strong Positive Sentiment I tried to filter out words like Freedom, Democracy, justice among others. For the Moderate Positive Sentiment: I used progress, unity, hope among others. For Moderate Negative Sentiment: I used Scandal, divisive, gridlock among others and Strong Negative Sentiment: I used Tyranny, oppression fascism propaganda.
 

While sentiment can be a useful indicator, Moderate positive sentiment is significantly more common in real news than in fake news. Fake news tends to show  more  negative emotions, with higher values in both moderate negative and strong negative sentiment categories compared to real news. Strong positive sentiment is strong in both fake and real news, but slightly more in real news. While emotions in news stories can be a clue about whether they’re real or fake, it’s not foolproof, so we need to use other methods to be sure. 

Crafting a Word Cloud: 

One of the most intriguing aspects of my project was creating my pseudo “word cloud”, it’s initial propose was to provide a visual snapshot of frequently occurring words. While Tableau Public doesn’t have a dedicated word cloud chart, I found a creative workaround using packed bubbles. 

By meticulously filtering and formatting the packed bubbles, I successfully replicated the word cloud’s appearance and functionality.  
However, there were extreme challenges and takeaways on this part on my end.  

Throughout this process, I encountered challenges ranging from syntax errors in calculated fields to display issues in the word cloud. Each obstacle provided an opportunity for learning and growth, reinforcing the importance of careful attention to detail and perseverance in data visualization. This involved meticulous data preparation, including splitting the text into individual words, filtering out words and neutral terms, and calculating word frequencies. I then created a custom calculated field to ensure the display of individual words rather than aggregated sums. Finally, I  formatted the packed bubbles chart to mimic the appearance of a traditional word cloud, with circular marks.  
 
What’s intriguing is the unexpected classification of words like “France” and “Pope” as positive in a word cloud analysis using distinct sentiment categories further emphasizes how hard it is to accurately identify and categorize sentiment.

Recreating the Scatterplot in Tableau Public: 

I wanted to create a simple scatterplot, showcasing fake news analysis. This plot demonstrated the relationship between word count and sentence length, distinguishing between real and fake news.  

Based on prior experience and looking online for code. I created simple calculated fields to create the scatterplots core. Like ” COUNT([Word])” for the Word count field and “LEN([Text]) – LEN(REPLACE([Text], ‘.’, ”))” for the sentence length field However, I encountered obstacles, with aggregation errors in table calculations but I felt that my scatterplot is robust enough to showcase a bit of data mapping.

Conclusion: 

This project highlighted the importance of precise filtering in sentiment analysis. Just as researchers must safeguard the integrity of digital evidence, data scientists must carefully curate their filtering criteria to avoid bias and ensure accurate results. The choices we make in data preparation, visualization, and analysis have a profound impact on the stories our data tells. While a 2021 survey revealed that many believe in the potential of AI to detect misinformation, it also underscored the importance of human oversight in ensuring accuracy and mitigating bias.

Citations: Kaggle. Nitish Jolly. News Detection: Fake or Real Dataset. 2020. https://www.kaggle.com/datasets/nitishjolly/news-detection-fake-or-real-dataset. Stacy Jo Dixon, Stacy Jo Dixon, Stacy Jo Dixon. “Misinformation on social media – Statistics & Facts | Statista.” Statista, 10 Jan. 2024, https://www.statista.com/topics/9713/misinformation-on-social-media/. Giancarlo Fiorella, Charlotte Godart, Nick Waters, Digital Integrity: Exploring Digital Evidence Vulnerabilities and Mitigation Strategies for Open Source Researchers, Journal of International Criminal Justice, Volume 19, Issue 1, March 2021, Pages 147–161, https://doi.org/10.1093/jicj/mqab022