An analysis of the user occupational class through Twitter content

Daniel Preotiuc-Pietro, Vasileios Lampos, Nikolaos AletrasACL 2015 · OpenReview

Abstract

Social media content can be used as a complementary source to the traditional methods for extracting and studying collective social attributes. This study focuses on the prediction of the occupational class for a public user profile. Our analysis is conducted on a new annotated corpus of Twitter users, their respective job titles, posted textual content and platform-related attributes. We frame our task as classification using latent feature representations such as word clusters and embeddings. The employed linear and, especially, non-linear methods can predict a user’s occupational class with strong accuracy for the coarsest level of a standard occupation taxonomy which includes nine classes. Combined with a qualitative assessment, the derived results confirm the feasibility of our approach in inferring a new user attribute that can be embedded in a multitude of downstream applications.

Introduction

In sociolinguistics, it has long been theorized that social status—often reflected through occupation—profoundly influences language use. Traditionally, studying this relationship required labor-intensive surveys or interviews. This research explores whether professional identity can be automated at scale using social media, specifically Twitter. Despite the informal and "noisy" nature of tweets, the study aims to determine if latent features in a user's post history are strong enough to map them to a standardized professional hierarchy, even when they are not explicitly discussing their work.

Methodology and Data

The researchers framed the problem as a multi-class classification task, comparing three primary models: Logistic Regression (LR), Support Vector Machines (SVM), and Gaussian Processes (GP). The GP approach is particularly noteworthy for its use of Automatic Relevance Determination (ARD), which identifies the most relevant features for prediction, providing a layer of interpretability.

For features, the study moved beyond simple word counts to more sophisticated representations, including word embeddings (SVD & Word2Vec) and spectral clusters. Word clusters group similar words into thematic "topics" (e.g., a "Health" cluster containing terms like chemotherapy and diagnosis).

Dataset

The study introduced the Twitter Occupational Dataset, consisting of 5,191 users mapped to the UK’s Standard Occupational Classification (SOC) taxonomy. This taxonomy is hierarchical, but the analysis focuses on the nine "Major Groups" (C1-C9), ranging from Managers to Elementary Occupations.

Standard Occupational Classification (SOC) taxonomy used to categorize Twitter users into nine major groups (C1-C9).

Users were identified via self-disclosed job titles in their bios, which were manually verified to filter out corporate accounts. The final dataset includes approximately 10.8 million tweets. A 10% sample of the Twitter "Gardenhose" stream from 2011 served as a reference corpus for building initial embeddings and clusters.

Results and Discussion

The best-performing model, a Gaussian Process with Word2Vec clusters, achieved 52.7% accuracy on the 9-way classification task, significantly outperforming the most-frequent-class baseline of 34.4%. GPs consistently outperformed SVMs, demonstrating that Bayesian approaches are well-suited for high-dimensional, non-linear social media text.

Key Findings

  • Content vs. Metadata: User-level metadata (e.g., follower counts, tweet frequency) proved largely irrelevant. Professional identity is encoded in the content of speech rather than platform mechanics.
  • Clusters vs. Embeddings: Word clusters were more effective than raw embeddings, likely because they prevent common words from drowning out specific thematic signals.
  • Linguistic Skill Gradient: The model revealed a clear divide in interests across occupational classes. Higher-skilled classes (Groups 1-2) tweet more about Politics, Corporate Business, and Higher Education. In contrast, lower-skilled classes (Groups 6-9) more frequently use elongated words (e.g., "yaaay") and discuss Beauty Care.
Cumulative distribution functions of topic proportions across occupational classes for selected discriminative topics.

Analysis of the confusion matrix showed that errors typically involved "adjacent" classes (e.g., confusing a professional with an associate professional), suggesting the model captures an underlying skill level or social status gradient. Using ARD, the researchers identified that the "Arts" topic was the most discriminative feature across the entire dataset.

Conclusion

This study demonstrates that Twitter content is a viable mirror of professional life. While the model is strong at the 9-class "Major Group" level, accuracy declines at finer granularities (e.g., 29.2% for 55 "Minor Groups"). Future work should address self-selection biases in bio-disclosed occupations and explore hierarchical models to improve granular predictions. Despite these limitations, the ability to infer occupational class from text opens new avenues for large-scale sociolinguistic research and downstream applications.

Summary updated

Abstract

Social media content can be used as a complementary source to the traditional methods for extracting and studying collective social attributes. This study focuses on the prediction of the occupational class for a public user profile. Our analysis is conducted on a new annotated corpus of Twitter users, their respective job titles, posted textual content and platform-related attributes. We frame our task as classification using latent feature representations such as word clusters and embeddings. The employed linear and, especially, non-linear methods can predict a user’s occupational class with strong accuracy for the coarsest level of a standard occupation taxonomy which includes nine classes. Combined with a qualitative assessment, the derived results confirm the feasibility of our approach in inferring a new user attribute that can be embedded in a multitude of downstream applications.
BibTeX
@inproceedings{preotiucpietro2015analysis,
  author = {Daniel Preotiuc-Pietro and Vasileios Lampos and Nikolaos Aletras},
  title = {{An analysis of the user occupational class through Twitter content}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2015},
  url = {https://openreview.net/forum?id=S148PjlubB},
}
Figure