Vive la Petite Différence: Exploiting Stylistic Nuances for Polish Gender Attribution

TL;DR

Is a text’s gender defined by what is said or how it is said? This paper tackles gender attribution in Polish short texts using a rigorously balanced version of the "He Said She Said" (HSSS) corpus. By filtering out "thematic leaks" and combining logistic regression with language models via neural networks, the authors achieved an accuracy of 71.06%, proving that subtle stylistic differences persist even when the subject matter is normalized.

Background: The Problem of "Thematic Bias"

Gender classification isn't new, but it is often "cheating." In many datasets, a model might correctly guess a female author simply because the text mentions "pregnancy" or "cosmetics," or a male author because of "engineering" or "sports." This is topic bias, not linguistic style.

The authors argue that to truly understand gendered language, we must:

1. Use Short Texts: Harder to classify but more representative of the modern web (Common Crawl).
2. Neutralize the Content: Filter the data so that for every male text from a specific website, there is a corresponding female text from the same domain.
3. Anonymize Indicators: Mask the obvious Polish verb endings (like -łem vs -łam) that explicitly reveal gender in the first person.

Methodology: Fusing Statistics with Style

The core of the paper’s success lies in its hybrid architecture. Rather than relying on a single "silver bullet" algorithm, the researchers combined two different perspectives on the text.

1. The Statistical Approach (Vowpal Wabbit)

Using Logistic Regression within the Vowpal Wabbit framework, they analyzed lower-cased tokens to find statistical correlations between word usage and gender. This achieved a baseline of 67.54%.

2. The Probabilistic Approach (KenLM)

The team built two separate 3-gram language models (one for each gender). During testing, the model calculates which "identity" (male or female) is more likely to have produced the sequence of words. This yielded 67.98%.

3. The Neural Fusion

The true breakthrough came from feeding the outputs of both the statistical and probabilistic models into a Neural Network layer. This ensemble approach pushed the performance to the peak 71.06%.

(Note: This diagram illustrates the fusion of Vowpal Wabbit and KenLM outputs into a final neural classifier.)

Experimental Results: Morphology over Vocabulary

One of the most profound experiments in the paper involved stripped-down features. The authors trained a classifier using only morphosyntactic tags (POS tags, person, number, etc.), ignoring the actual words themselves.

• Morphosyntactic N-grams Accuracy: ~60.58%

While lower than the full-text model, this is significantly higher than the 50% random baseline. It proves that there is a "gendered syntax" in Polish—a unique way of arranging parts of speech—that exists independently of the topic being discussed.

(Note: The table above, sourced from the paper's submission to Gonito.net, highlights the steady increase in accuracy as models move from simple regular expressions to neural ensembles.)

Critical Insights: The "Leaks" and Future Work

The authors candidly discuss the difficulty of "cleaning" a language. Even after rigorous filtering, they found "leaks"—subtle clues like adjectives in the form jestem zadowolony (I am happy, masculine) that were not caught by automated masking.

Comparison to SOTA

In the context of the era, reaching >70% on short, topic-balanced snippets is more impressive than reaching 85% on long-form blogs where thematic clues are abundant. The work emphasizes reproducibility by hosting all code and data on the Gonito.net platform, a standard that current AI research often fails to meet.

Conclusion

"Vive la Petite Différence" serves as a reminder that linguistic identity is deeply woven into the fabric of grammar and style. For developers and researchers in NLP, the takeaway is clear: when building classifiers, one must distinguish between content (what is said) and style (how it is expressed).

Future directions include refining the "leak" detection and exploring how Transformers like BERT might capture even more subtle "differences" in the Polish language landscape.