Abstract
Research in author profiling has primarily focused on English-speaking users and attributes like age, gender and occupation. We present first experiments on automatic profiling Dutch Twitter users for two less-studied attributes, namely their political preference and income level (low vs high). We create two novel corpora using distant supervision, evaluate the corpus creation approach, and train predictive models for each attribute. Our empirical evaluation shows that distant supervision is surprisingly reliable and political preference and income level of Dutch users can be predicted relatively accurately from the linguistic input. We also discuss which features are predictive for income and political preference, respectively.
Introduction
While automated detection of user traits—known as author profiling—is a well-established field, research has historically centered on English-language datasets and easily observable demographics such as age or gender. This research, conducted at the University of Groningen, expands the scope to the Dutch language and tackles two "latent" attributes: political preference and income level. Unlike age, these traits are rarely stated explicitly in tweets, requiring sophisticated modeling to uncover the digital breadcrumbs left by users.
Methodology: Distant Supervision
A significant hurdle in profiling Dutch social media is the lack of large-scale, manually labeled datasets. To overcome this, the authors employed distant supervision. This technique involves using heuristics or metadata to automatically label data. Specifically, the researchers parsed Twitter biographies for keywords indicating political affiliation (e.g., mentions of parties like VVD, CDA, or D66) and job titles.
For income prediction, job titles extracted from bios were mapped to average salary data provided by Statistics Netherlands (CBS). Users were then categorized into "high" or "low" income brackets based on the Dutch modal income of €34,500.
Experimental Setup and Datasets
The study introduced two novel corpora:
- $D_P$ (Political Preference): Comprising 2,000 users across five major Dutch parties (VVD, PvdA, CDA, GroenLinks, D66), with 500 tweets per user.
- $D_I$ (Income Level): Comprising 2,000 users split into binary income classes.
Manual verification of a data sample confirmed the efficacy of this approach: political labels reached 98.4% accuracy, while income labels (based on occupation) achieved 74.5% accuracy. Predictive models were then built using Logistic Regression, testing word n-grams, surface features (e.g., word length), and readability metrics (e.g., Flesch Reading Ease).
Key Findings
Linguistic Indicators of Politics
The research found that simple word choice (unigrams) is the most powerful predictor of political ideology. For instance, supporters of the VVD (a liberal-conservative party) frequently used terms related to entrepreneurship (ondernemers) and lifestyle (shoppen, lunchen). In contrast, terms like vluchtelingen (refugees) and armoede (poverty) served as negative indicators for VVD, appearing more often in the discourse of opposing parties.
Income and Readability
Interestingly, readability scores—often strong predictors of income in English-language studies—proved less effective for Dutch users. This suggests that the link between writing complexity and socioeconomic status may be culturally or linguistically specific, rather than a universal trait of social media communication.
Results
The predictive models demonstrated significant performance over baseline measures:
- Political Preference: Achieved 66% accuracy in a 5-way classification task (random baseline: 20%).
- Income Level: Achieved an F1-score of 0.72 (random baseline: 0.50), with the best results coming from a combination of unigrams, bigrams, and trigrams.
Conclusion and Future Work
This study confirms that distant supervision is a viable method for creating "silver-standard" datasets for low-resource languages. However, limitations remain. The "politician bias"—where the model may learn the specific jargon of activists rather than average voters—and the noise introduced by corporate accounts are areas for future refinement. Future research could explore more granular multi-class income prediction and the inclusion of syntactic features to further enhance profiling accuracy.
Summary updated