Trustpilot: Bridging the Gap Between Big Data and Sociolinguistics

TL;DR

Sociolinguistics has long been caught between a rock and a hard place: the precision of small, hand-curated interviews versus the massive but "identity-blind" data of the web. This paper introduces the Trustpilot Corpus, a massive resource covering 13 languages and millions of users. By using the profile data users provide when reviewing businesses, the researchers have created a dataset that includes age, gender, and location—allowing for statistically powerful studies on how our identity shapes our language.

The Problem: The Statistical Power Crisis

For decades, sociologists studying language (how gender, age, or class affects speech) had to do it the hard way: transcribing interviews. This results in tiny samples, sometimes fewer than five people. In the era of "Big Data," researchers turned to Twitter, but Twitter users rarely state their age or gender. We were left guessing who was behind the text.

Methodology: Turning Reviews into Research

The authors identified Trustpilot as a unique source. Unlike Twitter, Trustpilot users often provide:

• Birth Year: Allowing for precise age-grading studies.
• Gender: Either explicitly stated or inferable via high-confidence name analysis.
• Location: Geocodable to specific cities and regions.

Data Augmentation

To boost the dataset's utility, the team:

1. Gender Inference: Used names (e.g., "Simone" in Italy vs. Germany) with a 95% confidence threshold to double the pool of gendered data.
2. Geocoding: Cleaned town names and mapped them to latitude/longitude to study regional dialects.

The study validated the data's representativeness by comparing it to official census data, showing an impressively close match in age distribution.

Key Pilot Studies & Insights

1. The Death of the "Nose" Emoticon

One of the most striking findings is the evolution of the emoticon. The team found a near-perfect anti-correlation (ρ = -0.99) between age and emoticon frequency. Furthermore, there is a "generational divide" regarding the use of "noses" in smileys (e.g., :-) vs :)). Users over 34 are significantly more likely to include the hyphen "nose," while younger users have almost entirely abandoned it.

2. Gendered Lexicons

Using tf-idf to find salient words, the researchers found that:

• Women tend to express satisfaction (e.g., "loved", "pleased", "satisfaite").
• Men focus on technicality and problems (e.g., "problems", "parts", "conforme").

3. Regional Dialects in Action

The study looked at the Danish word "træls" (meaning unpleasant/annoying), a hallmark of the Jutland dialect. By mapping reviews, they proved that 84% of its usage remains west of the "Great Belt" divide, confirming that even in a globalized internet, regional identity remains digitally sticky.

Scientific Impact: Why This Matters

This work isn't just about smileys or Danish slang. It provides a roadmap for Fairness in NLP. Most AI models are trained on newswire or general web crawls, which skew towards specific demographics. By using the Trustpilot Corpus, researchers can now test if a sentiment analyzer works as well for a 70-year-old woman in rural France as it does for a 20-year-old man in London.

Conclusion

The Trustpilot Corpus represents a significant leap forward. It proves that the "noise" of the internet isn't just trash—it's a rich, structured signal of human identity. The authors have released their harvesting and preprocessing scripts, inviting the community to explore how demographics continue to drive the evolution of global languages.

Takeaway: Your online reviews are more than just a complaint about a late delivery; they are a digital fingerprint of your sociolinguistic background.