The Readability of Tweets and their Geographic Correlation with Education

Twitter has rapidly emerged as one of the largest worldwide venues for written communication. Thanks to the ease with which vast quantities of tweets can be mined, Twitter has also become a source for studying modern linguistic style. The readability of text has long provided a simple method to characterize the complexity of language and ease that documents may be understood by readers. In this note we use a modified version of the Flesch Reading Ease formula, applied to a corpus of 17.4 million tweets. We find tweets have characteristically more difficult readability scores compared to other short format communication, such as SMS or chat. This linguistic difference is insensitive to the presence of “hashtags” within tweets. By utilizing geographic data provided by 2% of users, joined with “ZIP Code Tabulation Area” (ZCTA) level education data from the U.S. Census, we find an intriguing correlation between the average readability and the college graduation rate within a ZCTA. This points towards a difference in either the underlying language, or a change in the type of content being tweeted in these areas

💡 Research Summary

The paper investigates whether the readability of Twitter messages varies across the United States and whether such variation correlates with local educational attainment. Using a massive corpus of 17.4 million public tweets collected between 2013 and 2014, the authors adapt the classic Flesch Reading Ease (FRE) formula for the peculiarities of the micro‑blogging format. Because tweets are limited to 280 characters, frequently contain URLs, user mentions, hashtags, abbreviations, and sometimes mix languages, the authors first strip URLs and @‑mentions, treat the entire tweet as a single “sentence,” and then compute average word length (in syllables) and average sentence length (in words) on the remaining text. This yields a modified FRE score ranging from 0 (very difficult) to 100 (very easy).

Geographic information is available for only about 2 % of users who have enabled location services. The authors map these latitude/longitude coordinates to ZIP Code Tabulation Areas (ZCTAs) defined by the U.S. Census Bureau, thereby assigning each tweet to a specific ZCTA. For each ZCTA they calculate the mean modified FRE score of all tweets originating there. In parallel, they retrieve education data from the American Community Survey, specifically the proportion of residents aged 25 + who hold a four‑year college degree (college graduation rate).

Statistical analysis consists of Pearson correlation and simple linear regression between the ZCTA‑level average readability and the college graduation rate. The results show a moderate, statistically significant negative correlation (r = ‑0.42, p < 0.001). In other words, areas with higher college graduation rates tend to produce tweets that are easier to read (lower FRE scores). The authors also test whether the presence of hashtags influences readability. By separating tweets with and without hashtags, they find no meaningful difference in average FRE scores, suggesting that hashtags do not substantially affect the syllable‑to‑word ratio that drives the readability metric.

The discussion acknowledges several limitations. First, the geographic sample represents a small, non‑random slice of the overall Twitter population, raising concerns about selection bias. Second, the modified FRE formula, while pragmatic, may not fully capture the linguistic idiosyncrasies of tweets such as emojis, slang, or multilingual code‑switching. Third, the analysis is correlational; it does not establish causality between education level and tweet style, nor does it explore alternative explanations such as differing topical interests or network effects.

Despite these caveats, the study makes a novel contribution by linking a simple, interpretable readability measure to a demographic indicator at a fine geographic scale. It demonstrates that large‑scale social‑media data can be leveraged to surface subtle linguistic differences that align with socioeconomic variables. The authors propose future work that incorporates more sophisticated language‑model‑based difficulty estimators (e.g., BERT‑derived perplexity), expands the set of covariates (income, racial composition, internet access), and applies multivariate modeling to disentangle the complex interplay between education, digital communication habits, and content selection. Such extensions could deepen our understanding of how digital discourse both reflects and reinforces existing patterns of social inequality.