- Abstract
- Motivations
- Aims
- Data Collection
- Data Set
- Code Explanation
- CSV Files
- Findings
- Limitations & Technical Challenges
- Conclusion
- Contributions
- Maia
- Amara
- Sarmad
- Bibliography
As the idiom "birds of a feather flock together" suggests, people of the same ideologies, hobbies, and interests tend to stick together. Will this idiom ring true for American senators and representatives? To find out, our group analyzed the most frequent words used on Twitter by all 535 Members of the 117th United States Congress. Will legislators from the same political party Tweet similar words to each other? We used Twitter's API with elevated access to analyze almost 3 million tweets from all 535 legislators in the US Congress. We then used spacy to extract keywords and then grouped them by Twitter handle, finding the 50 most frequently used keywords by each legislator in their tweets. We grouped the legislators by political party (Republicans and Democrats) and found the most frequently used words by each party collectively along with associated average user engagement statistics for tweets containing certain words of interest. We found significant variations in the language used between the members of Republican and Democratic parties. Furthermore, we found that Tweets with words associated with the opposing parties were used more frequently during times of scandal and high tensions. With an increased polarization in not only Congress but in wider US society, the information we gather from Twitter is an important insight into what American legislators have on their minds and whether or not that aligns with the rest of their party. Finally, we found that discussion of the opposition leads to higher user engagement with Tweets on average.
The first part of this page highlights our group's motivations to choose this topic. Next, we define our aims. Then, we explain our data collection process, followed by our data analysis process. We then highlight our findings and limitations before reaching our conclusion.
This project is very timely as political polarisation in the US Congress continues to get worse (Jeong & Lowry, 2019). Maia is fascinated by how US society will look in the future if polarisation continues. Polarisation erodes democracy by curtailing legislative compromises and leads politicians to pursue their political agendas outside the gridlocked Congress through courts (McCoy & Press, 2022).
As a student studying the philosophy of language, Sarmad is interested in the words chosen by the Members of Congress. He wanted to learn more about the relationship between words, thoughts, and political ideologies.
Amara, who studies politics for half of her degree, believes that language is a powerful tool and that it's difficult to truly understand how much of an effect it has on people and their political opinions. However, people with high amounts of influence, like US politicians, may be aware of this power and therefore may choose their words in a way which will influence people towards their agendas on big platforms like Twitter. By analysing the language that different parties use the most over certain periods, and how this affects user engagement, we may gain an insight into the nature of politicians' tweets and how Twitter is used (or not used) as a political tool.
In doing research for this project, we were surprised to learn that such an analysis had, to our knowledge, not been conducted yet. In contributing to this project, we hope to add to knowledge about political polarisation in the US and grow awareness about the potentially salient role played by words, language, and Twitter in the current American political landscape.
Our first aim in this project was to see which words are most frequently used by each political party as a whole. Then we will look into the words from this top group and analyse the ones which are used exclusively by each party at this level of frequency. Next, we will explore when and why some of these words were frequently used. Lastly, we will investigate the links between these frequently used words and the level of user engagement.
Data Set: We collected recent tweets from all of the current Congresspeople's Twitter accounts, which amounted to a total of almost 3 million tweets, dating between February 2010 to December 2022. We got the list of Twitter handles from an Excel spreadsheet titled Congressional Twitter Accounts created by the University of California San Diego (UCSD) (Smith, 2022). Our data set consists of 223 Democrats (including 4 Delegates) and 215 Republicans (including 1 Delegate and the Resident Commissioner of Puerto Rico), and 3 vacant seats.
This map illustrates the distribution of Congressional representatives throughout all 50 states.
Code Explanation: The code we used to gather our data can be divided into four key sections. First, we implemented the Twitter API to make queries. Second, we converted the Twitter JSON response to a data frame. Third, we extracted keywords from each tweet, and lastly, we grouped and counted keywords per user.
First, we implemented the Twitter API to retrieve Twitter IDs and pages of tweets:
One of the initial obstacles we had to overcome for this project was the limits posed by the Twitter API. The API has three types of access levels. The most basic level allows users to retrieve up to 500,000 tweets per month and have 25 requests per 15 minutes. These limits would hinder our ability to gather the amount of data needed so we decided to apply for elevated access. At this level, we were able to retrieve up to 2 million tweets per month and have 50 requests per 15 minutes. However, since we had to retrieve more than 2 million Tweets for our analysis, we had to wait a full month to finish gathering all of them. Additionally, the maximum number of tweets per request is 100 and it takes 15 minutes to retrieve 5,000 tweets. To put it into context, the average number of tweets per member of Congress in our data set is 2,842 and most politicians tweeted more than 3,000 for the 30-day time period we used. This means it would take around 10 minutes per legislator. To maximise time efficiency and avoid reaching the request limit, we decided to use Comma Separated Values (CSV) files to store our data. This would circumvent the need to ask the Twitter API for data we previously requested, as well as re-running the code more than necessary.
Second, we converted JSON to a data frame:
We had to extract the necessary data from the Twitter JSON response by creating a name-value pair dictionary.
def get_tweet_dict(tweet, handle, name):
metrics = tweet["public_metrics"]
return {"handle": handle,
"name": name,
"tweet_id": tweet["id"],
"author_id": tweet["author_id"],
"lang": tweet["lang"],
"replied_to": ",".join(tweet['edit_history_tweet_ids']),
'created_at': tweet['created_at'],
'tweet_text': tweet['text'],
'possibly_sensitive': tweet['possibly_sensitive'],
'conversation_id': tweet['conversation_id'],
"retweet_count": metrics["retweet_count"],
"reply_count": metrics["reply_count"],
"like_count": metrics["reply_count"],
"quote_count": metrics["quote_count"]}The JSON response is a tree structure and we needed to create columns per tweet. Therefore, this function created a name-value pair dictionary that could be used to create an array of consistent dictionaries to be used creating our panda data frame.
Third, we used spacy to extract keywords from Tweets
nlp = spacy.load("en_core_web_sm")
nlp.disable_pipe("parser")
nlp.add_pipe("sentencizer")To make the code run faster, we used the sentencizer rather than the default parser since we were only using a limited number of functions from Spacy.
include_types = ["ADJ", "NOUN", "PROPN", "VERB", "ADV"]
def get_tokens(doc):
return [token.lemma_.lower() for token in doc if token.is_alpha and token.pos_ in include_types and token.lemma_.lower() not in exclude_words]A second barrier we faced was that most frequently tweeted words were primarily prepositions, interjections, and conjunctions, such as "the", "at", and "in". However, these words do not really give us context as to what the Members of Congress are tweeting and thinking about, and do not offer evidence supporting or opposing our hypothesis. To overcome this, we used Spacy's natural language process to extract only adjectives, nouns, proper nouns, verbs and adverbs. Furthermore, to group past tense, plurals, and similar variables of the same word we used the lemma to extract only the base word. For example, "history", "historical", and "histories" would all be grouped into "history". This would ensure we capture the concepts focused on and thoughts expressed by the Members, rather than the particular word used.
Fourth, we grouped and counted keywords per user and listed all of their tweets
def add_word_count(row):
word_freq = Counter(row["key_word_list"])
common_words = word_freq.most_common(50)
df = pandas.DataFrame(common_words, columns = ['Word', 'Count'])
df["handle"] = row["handle"]
return df[["handle","Word","Count"]]The last major step was to group all the keywords by Twitter handle and to gather all the keywords from each tweet into one array to count. Finally, we used a Counter to count the keywords and then find the 50 most frequently used words per legislator, which we used to create a new data frame and CSV file.
| Handle | Word | Count |
|---|---|---|
| AustinScottGA08 | today | 496 |
| AustinScottGA08 | thank | 361 |
| AustinScottGA08 | georgia | 292 |
| AustinScottGA08 | day | 276 |
| AustinScottGA08 | more | 274 |
This is a snippet of what our word_count.csv file looks like. On the far left is Rep. Austin Scott's Twitter handle. In the middle are five of his top 50 frequently used keywords. Finally, on the far right is how many times each word was used in our time frame.
CSV Files:
Because the CSV files were too large to upload to GitHub, we uploaded them onto Google Drive which can be accessed through these links:
The figures above illustrate that the 50 most common words used by the two parties correspond to the party's respective political ideologies. For instance, we find that the Democrats frequently use words like "health", "protect" and "women". This was expected since the Democratic party's policies aim to reduce the gender pay gap and increase female representation (Horowitz, 2020) and of providing healthcare coverage to Americans (Dunn, 2019). Similarly, we find that Republicans are more likely to use words such as "border" and "business". These words are very consistent with the Republican party's known emphasis on border security (Oliphant & Cerda, 2022) and their economic belief that business is key to a strong economy (Furhmann, 2022).
It makes sense that both parties frequently tweet words that align with their party's ideology since Members of Congress use Twitter as a platform to spread their political ideology. While the findings from the above bar graphs seem obvious, the data is significant because it illustrates that within these two parties, Members of Congress are politically aligned, but not aligned with legislators across the aisle.
Furthermore, one thing that caught our eyes when looking at these bar graphs is that "Biden" is a very frequently used word by the Republicans, and "Trump" is frequently used by the Democrats, rather than vice versa. It is safe to assume that each side is not using the name of the other's leader in a positive light, so we can only gather that each party adopts the tactic of tearing down the other side to gain public support. This is not a completely even split: "Biden" was the number one most frequently used word exclusively by the Republicans, whereas 'Trump' was only eighth most for the Democrats. This shows that, overall, the Republicans have spent more time (or words) on Twitter addressing the opposition than they have spent discussing any singular issue or policy central to their party values. Our data suggests that Republican politicians are more inclined to adopt this method, but nonetheless, both parties generally seem to have engaged in this strategy.
To further explore this tactic of politicians tearing down their opposition, we created time series plots showing the frequency of the use of the word 'Biden' for Republicans and 'Trump' for Democrats in 2022. See below:
We were interested to see whether temporal factors play a role in the usage of words by legislators. We observed that the usage of the word "Biden" spiked between February 20th and 28th 2022 for Republicans. We conducted research into why this may be and found that Biden was under considerable criticism by Republicans in the US Congress for his handling of the Ukrainian crisis (Morgan, 2022). Similarly, we found that Democrats were more likely to use the word "Trump" in June 2022. We conducted similar research into this and found that June was when the hearings for the January 6th Capitol Attack began, putting Trump under significant criticism as a threat to US democracy (Breuninger, 2022). This corroborates our theory that politicians may choose Twitter as a method of attacking the opposition. The time series illustrates that when an opportunity arises (i.e. the opposition makes a mistake) there is no hesitation in publicising this on Twitter with a possible aim to maximise reach. This does raise interesting questions for future research that may build on ours. For example, do legislators echo what they say in Congress on their Twitter? What is the distinction between what Members of Congress say through an official versus through an unofficial lens?
Regardless, to find further evidence for our case, we examined how the usage of certain words affected user engagement on Twitter.
As you can see in the figures above, for Democrats, tweets that included the word 'Trump' had a far higher level of user engagement than any other most frequent words exclusive to the Democrats. The vast majority of engagement with the tweets from Democrats including the word "Trump" were retweets, which ultimately means that these tweets were able to reach the highest number of users through the social network. Similarly, the tweets with the highest engagement for the Republicans included the word 'Democrats', closely followed by 'Biden'. This is consistent with the findings discussed above and further emphasises that attacking the opposition on Twitter is an effective way to expand reach and therefore increase support. This raises the question of whether or not politicians make this choice consciously, i.e. is attacking the opposition a strategic move that they know will lead to higher numbers of retweets and expand their online reach? This would be an interesting topic for further research, but would be hard to answer without speaking to politicians in person.
Another interesting observation that we were able to make from these statistics was that the Democrats seem to have much higher user engagement with their tweets. However, we have only compared these statistics for the most frequently tweeted words from each party, which means we cannot rule out the possibility of a heavily skewed distribution of engagement in favour of more frequent words for the Democrats or less frequent words for the Republicans. Therefore, further research would be required to substantiate this observation.
Our observations may also tell us something about the users of Twitter. Perpetuated hate speech and echo chambers have been an issue on Twitter for quite some time (Frenkel & Conger, 2022) and we expect that users will interact more with Tweets using words like "Trump", "Democrat", and "Biden" because, as previously mentioned, these three words would be used in a negative light and we expect that users are more likely to engage with negative Tweets as they are more controversial and possibly spark more debate. But we are not convinced that this is a totally one-way street; Twitter's algorithms may also have something to do with the increased user engagement for tweets of this nature.
So, is it the nature of Twitter and its algorithms which cause this perpetuated negativity, meaning that this tactic of attacking the opposition leads to higher user engagement? Is this why this may be considered a successful strategy for gaining support from the two main political parties in the US? Or is it perhaps that these politicians and the language they use are part of the cause of Twitter's problem? This may be an interesting topic for further research, however, for now, we have found that this approach is an efficient way for US politicians to expand their online reach on Twitter and, whether they are doing this consciously or not, it is one of the most frequent occurrences in their tweets.
Our project has a few limitations and faced some technical challenges.
Firstly, to replicate this code (specifically extract_tweets.py), data scientists would need elevated access to Twitter API, which they may not always have. We used elevated access to ensure we were able to access a statistically significant sample of tweets. Once one receives elevated access from Twitter, all code used is replicable. However, on the Github repository, there is a file called csv_cache which includes csv files of all legislator's collected tweets. With these csv files, it is possible for someone to replicate my code without using the Twitter API.
Secondly, we found that Spacy was not completely successful in grouping similar words as one, for example 'nation' and 'national'. However, after consideration, we realised that this is not an issue as words like 'nation' and 'national' are actually used in completely different contexts; one being a noun referring to the actual country and the other being an adjective which could be describing any other noun related to the country as a whole. We concluded that the Spacy software was right not to group these kinds of words together and that this would not significantly affect our findings.
Lastly, we ran into a few technical, pre-processing errors involving the number of times tweets were liked or replied to. We found that the number of likes and replies was identical, suggesting that one of them was incorrect. To verify this, we cross-referenced with actual tweets on Twitter (keeping in mind that the values will have changed since our data collection) and found that the figures for replies and retweets were accurate but figures for likes were inaccurate. We were unable to diagnose a solution in time for this project, and therefore excluded likes from our analysis. However, this did not prevent us from gaining useful insights regarding user engagement, as highlighted above.
Overall, our data illustrate that within political parties, Members of Congress frequently use words that align with their respective parties' political agendas. However, when they do deviate from this, it is to criticise the other political party. We found that this happens most during moments where the Democrats can criticise Republican leadership and vice versa and leads to increased user engagement. For this reason, we believe that attacking the opposition may be the best tactic to gain support from constituents. If we were to continue our investigation, we would be interested to investigate whether or not politicians are aware of this link, and how often they consciously adopt this strategy. This has also made us curious about the relationship between the nature of Twitter and of its users (e.g. US politicians, as in this case), in terms of which influences the other more strongly leading to this observed success of online negativity (like the chicken or the egg question), or whether it is a mutual influence. Additionally, we would also examine the distinctions, if any, between what politicians tweet online versus what they say in Congress itself.
Maia: Maia created the code to collect the Twitter dataset showing the 50 most commonly used words per Twitter handle and the code to create the data frames for the Republican and Democratic parties. She also wrote the Index, Abstract, Motivations, Aims, and Data Collection sections of the README.md, and did final edits and grammar checks. Lastly, she created the repository and webpage and added a theme.
Amara: Amara analysed the data, transforming it into multiple more easily usable data frames. She then created some more easily interpretable bar graphs, time series graphs and word clouds. She interpreted the data and drew conclusions based on our observations. She wrote the Motivations, Findings and Conclusion sections of the README.md.
Sarmad: Sarmad wrote the Motivations, Limitations and Bibliography sections of the README.md, while also copy-editing the rest of it to ensure cohesiveness and consistency. He also did empirical research for all the substantive claims made about the Democrats and Republicans throughout the README.md.
Breuninger, K. (2022) Jan. 6 hearing highlights 'carnage' of capitol riot in new video footage, Trump officials concede he lost election, CNBC. CNBC. Available at: https://www.cnbc.com/2022/06/09/trump-capitol-riot-hearing-jan-6-investigators-release-new-findings.html (Accessed: February 2, 2023).
Brush, M. (2010) White House not concerned about new census numbers, Michigan Radio. Michigan Radio. Available at: https://www.michiganradio.org/politics-government/2010-12-21/white-house-not-concerned-about-new-census-numbers (Accessed: January 21, 2023).
Dunn, A. (2019) Democrats differ over best way to provide health coverage for all Americans, Pew Research Center. Pew Research Center. Available at: https://www.pewresearch.org/fact-tank/2019/07/26/democrats-differ-over-best-way-to-provide-health-coverage-for-all-americans/ (Accessed: February 2, 2023).
Frenkel, S. and Conger, K. (2022) Hate Speech’s Rise on Twitter Is Unprecedented, Researchers Find. Available at: https://www.nytimes.com/2022/12/02/technology/twitter-hate-speech.html (Accessed: January 26, 2023).
Furhmann, R. (2022) Republican and Democratic approaches to regulating the economy, Investopedia. Investopedia. Available at: https://www.investopedia.com/ask/answers/regulating-economy.asp (Accessed: January 27, 2023).
Horowitz, J.M. (2020) Wide partisan gaps in U.S. over how far the country has come on gender equality, Pew Research Center's Social & Demographic Trends Project. Pew Research Center. Available at: https://www.pewresearch.org/social-trends/2017/10/18/wide-partisan-gaps-in-u-s-over-how-far-the-country-has-come-on-gender-equality/ (Accessed: January 27, 2023).
Jeong, G.-H. and Lowry, W. (2019) “The polarisation of energy policy in the US congress,” Journal of Public Policy, 41(1), pp. 17–41. Available at: https://doi.org/10.1017/s0143814x19000175.
Morgan, D. (2022) Republicans target Biden for blame over Putin's Ukraine invasion, Reuters. Thomson Reuters. Available at: https://www.reuters.com/world/us/republicans-target-biden-blame-over-putins-ukraine-invasion-2022-02-24/ (Accessed: February 2, 2023).
Oliphant, J.B. and Cerda, A. (2022) Republicans and Democrats have different top priorities for U.S. immigration policy, Pew Research Center. Pew Research Center. Available at: https://www.pewresearch.org/fact-tank/2022/09/08/republicans-and-democrats-have-different-top-priorities-for-u-s-immigration-policy/ (Accessed: January 21, 2023).
Smith, K.L. (2022) Libguides: Congressional twitter accounts: Home, Home - Congressional Twitter Accounts - LibGuides at University of California San Diego. University of California San Diego. Available at: https://ucsd.libguides.com/congress_twitter (Accessed: January 21, 2023).