TL;DR
This paper investigates the reliability of authorship profiling (AP) models for predicting gender and age. It reveals that while models can achieve decent accuracy within a specific dataset, their performance drops significantly when applied to new, unseen data from different domains or languages, suggesting that learned patterns are often dataset-specific and prone to stereotyping.
What Problem Does This Paper Address?
Authorship Profiling (AP) aims to infer demographic attributes like gender and age from text. While AP holds promise for applications in social science, forensics, and marketing, there's a growing concern about its potential for harm. This includes privacy breaches, the oversimplification of complex demographic identities (like treating gender as binary), and the reinforcement of stereotypes. A key assumption in AP research is that demographic signals are consistent across different datasets and languages. This paper challenges this assumption by investigating how well AP models generalize across domains and languages, and what factors contribute to their errors, particularly focusing on the interplay between topics and demographic prediction.
Methods
The authors employed a combination of established machine learning techniques for authorship profiling. They used both white-box models (like Logistic Regression and Random Forest) which offer interpretability, and black-box models (like BERT, RoBERTa, and XLNet) which are often more powerful but less transparent. For feature extraction, they utilized a range of linguistic features, including surface (e.g., token length), syntactic (e.g., part-of-speech tags), textual complexity, and sentiment features. To analyze the impact of topics, they employed BERTopic to extract topics from the text. Their analysis involved training models on specific datasets and testing them on others to assess cross-domain generalization.
Datasets
The study utilized two well-established datasets for authorship profiling:
- PAN13: This dataset originates from a shared task on plagiarism detection and authorship verification. It includes conversational texts in both English (PAN13-EN) and Spanish (PAN13-ES).
- BLOG: This is the Blog Authorship Corpus, containing longer blog posts.
Both datasets were annotated with self-reported gender and age. The authors ensured data balancing for gender and adapted age categories for comparability.
Key Insights and Ideas
- Dataset-Specific Signals: AP models learn patterns that are highly specific to the dataset they are trained on. These patterns don't reliably transfer to other datasets, even if they are from the same language.
- Stereotype Traps: Models are prone to errors when authors write about topics stereotypically associated with a different demographic group. This can lead to misclassifications that reinforce societal biases.
- Topic vs. Style: While stylistic features (like word choice and sentence structure) are important, topics play a significant role in AP performance. However, relying too heavily on topics can lead to biased decisions.
- Interpretability Matters: The modest performance gains of black-box models over white-box models in cross-domain scenarios suggest that the loss of interpretability might not always be justified, especially when accuracy gains are marginal.
Key Results
- In-Domain Accuracy: Standard classifiers achieved accuracies generally above the majority baseline for both gender and age prediction within their respective training datasets. For example, RoBERTa achieved accuracies up to 0.79 for age prediction on the BLOG dataset.
- Cross-Domain Performance Drop: When models trained on one dataset were tested on another, accuracy dropped significantly. For instance, cross-dataset accuracy for gender prediction was often only slightly better than the majority baseline.
- Language Differences: Accuracy on the Spanish PAN13-ES dataset was generally higher than on the English PAN13-EN dataset, suggesting language might play a role in signal consistency.
- Topic Influence on Errors: The analysis of topics revealed that common topics like "website and marketing" and "love, life, and Jesus" contributed to higher error rates, particularly when authors discussed themes stereotypically linked to another gender. For example, men discussing "shoes and handbags" were more likely to be mispredicted as female.
Limitations and Future Work
The authors acknowledge several limitations:
- Binary Gender Framework: The datasets used treat gender as a binary variable (male/female), which fails to capture the complexity and nuances of gender identity. Future work should aim for more inclusive datasets.
- Limited Languages and Features: The study was primarily conducted on English and Spanish, and some feature categories were limited to English. Expanding to more languages and feature sets is recommended.
- Age Ranges: The age categories were limited to three ranges (10s, 20s, 30s), which restricts the analysis of age-related patterns.
Future work should focus on developing more robust and generalizable AP models, potentially by exploring domain adaptation techniques, reducing topic bias, and investigating state-of-the-art Large Language Models that might better capture generalizable features. The authors emphasize the need for careful consideration of AP model deployment to avoid marginalizing and misrepresenting demographic groups.
Summary updated