World of Wordle
Final Visualization: Observable Notebook
Daisy Choi (jc4490), Eve Washington (esw2175), Samia Menon (sm4788), Rifqi Luthfan (rl3154)
After being released in October 2021, it did not take long for Wordle to become a world-wide phenomenon. The word-guessing game garners over 300,000 players every day, with users coming back again and again to see if they can use color-cues to determine the day’s five-letter word. Despite being months old, Wordle has had its fair share of discussion, history, and controversy. We created a data visualization for the Wordle-wise to discover the trends in difficulty over the game’s lifespan, hone their game, and determine how their strategies align with others’. There are four sections to our work:
- The Meteoric Rise
- Is It Really Getting Harder?
- What Makes a Wordle Hard?
- Pathways to Victory
Other work, like “Wordle, 15 Million Tweets Later” has focused on summarizing and explaining wordle tweets, rather than using them as a tool for decision making.
Those that do think about strategy aren’t particularly interactive, like “How to play Wordle.” While we also plan to format this as a narrative, we hope to use interactive opportunities to explain the analysis step-by-step, prioritizing discovery and exploration rather than explanation.
This approach will allow us to address questions like, “What Makes A Wordle Word Hard” but because we are focusing on tweets, focus on general patterns rather than the specific solutions.
It is possible to analyze a specific Wordle at the letter and word level. At the letter level, we identified two factors that influence gameplay, the frequency of that letter in the English language, and how many unique letters there are (for instance whether or not there are double letter occurrences). At the word level, we looked at the placement of vowels and the number of similar words if one letter is changed (rhyming words, like MATCH and BATCH, for example). We also considered the word outside of the context of the game, by looking at word frequency in the English language and knowledge as measured by a 2018 study on Word Prevalence.
Word prevalence norms for 62,000 English lemmas - Behavior Research Methods
We present word prevalence data for 61,858 English words. Word prevalence refers to the number of people who know the…
In order to see how hard a Wordle is we needed to define heuristics that reflect difficulty and see if any of these Wordle word characteristics influence those heuristics. We defined three measures of difficulty, and conducted linear regression analysis to see what word characteristics had impacted each difficulty heuristic.
The first measure of difficulty was the average number of guesses needed to complete any given Wordle, more guesses for a particular Wordle means it’s more difficult. This factor was most heavily influenced by the number of double letters and the number of common letters. More double letters and more common letters meant more guesses.
We can also use semantic data to measure difficulty. Because we are using twitter data, often in addition to tweeting the word results users include words like, “hard” or “tough”, we looked for mentions of difficulty as another definition of challenge. A higher perceived difficulty correlated with more double letters, and words less frequently used.
Lastly, we looked at the probability of success as another measure of difficulty — of all those who tweeted their Wordle results, what percentage of them failed. It is critical to note that this measure is slightly biased; we found that when a world is more challenging (as defined by the other heuristics) people tend to tweet less. Still there was a strong correlation between more failures, less common letters present, and the word having many similar words differentiated by only one letter.
Our design is focused on providing insights to the avid Wordle player while being accessible to anyone who stumbles by. As a result, we decided to keep our brief ‘popularity’ introduction visualizations before delving into the ‘difficulty’ portion of the project (section 1 in Observable Notebook). Often, there was a significant amount of data to be conveyed, and it was difficult to portray them all on the same figure cleanly; to solve this, we incorporated tool-tips and toggle mechanics to streamline our design (section 2,3 in Observable Notebook). In our final graphic, we leaned deeply towards interactivity, creating a custom graphic allowing the user to submit their Wordle guess and compare it to others’. This graph went through many iterations — many focusing on a ‘tree’-like visualization — however, we found those extremely noisy or data intensive. Our final visualization is simple to understand, intuitive to use, but we did have to sacrifice precision by using color as a channel. (section 4 in Observable Notebook)
Here is the link to our original Figma wireframe.
We retrieved our data from available Kaggle sources [data source 1, data source 2] for sample player tweets and we also manually scrape tweets from @wordlestats to get high level daily statistics of all Wordle players, both using the available Python libraries. Additionally, we keep track of the daily Wordle answers to be mapped to the Wordle tweets data, which is manual human input.
After gathering all the datasets (>200 MB per day, the longer the period the bigger), we do text data cleaning, extraction, and transformation (in Python + Pandas + Regex) and load the structured data into multiple small CSV files (<1 MB each). This CSV is then uploaded to Observable.
We had several rounds of user feedback and updates. We first got our feedback from the classmates to create a storyline using our data by grouping the data. For instance, showing how the difficulty of Wordle changed since the NYT acquisition–decrease in the number of Twitter posts, increased number of trials to get the Wordle word correct, etc. So we divided the variables for difficulty measure into three parts: Avg Guesses Taken, People Saying it’s hard, and Percentage People Fail. Then we set the variables that make such difficulty measures vary and made a separate bar graph so that the user can toggle between each difficulty measure. If a user wants to take a deeper look into each variable, they can select such variable and our regression model with scatter plot is given below. The update made the story clear that although the Wordle words have become more difficult after the NYT acquisition, the users’ percentage of getting the word right hasn’t decreased much.
Moreover, we also received negative feedback on our previous (currently abandoned alternatives) Pathway to Success. We explored with too much data so the visualization was quite unclear and hard to follow. So we explored using only a selected few twitter posts to make the data more static. By making a table to show the “most common word first guess”, it is easier to see the rank of the words guessed. We also added an interactive visualization so that the users can compare their day’s Wordle guesses with other people’s result for the day.
Our visualization confusion came from not creating a clear storyline using the data and using too much data. Being more specific on how and what we choose the data helped us clear the storyline of our data and to make the visualization easier to understand.
Some improvement points that need to be addressed are:
- Eliminating manual work: create pipeline to connect visualization with data processing service, possibly by building API to serve data access so that Observable can just request data to the API
- Improving difficulty approximations: use more features e.g. benefit of experienced players
- Additional interactivity and visual encoding: as opposed to manually inputting the guess, users can just input their Twitter username and the visualization will automatically draw their path from their tweets.
References and Acknowledgements