Zoltán Kmetty

Bendegúz Váradi, Zoltán Kmetty

We present a Procrustes-conditioned Joint End-to-end Top-K Sparse Autoencoder (SAE) for extracting cross-seed universal features from independently trained BERT models. Cross-seed feature universality is a fundamental challenge in mechanistic interpretability: because dictionary learning is non-convex, independently trained networks learn misaligned feature spaces, so apparently identical features may differ by random initialization. We address this by computing an orthogonal Procrustes rotation between seeds' activation spaces before joint SAE training, combining Top-K sparsity, end-to-end downstream optimization, and an auxiliary dead-feature revival loss based on previous SAE literature. Evaluating on five independent seed pairs (ten BERT models) across three benchmark datasets (SST-2, Stanford Politeness, TweetEval Emotion), our full pipeline produces more universal features (Pearson r $\geq$ 0.70 across seeds) than post-hoc alignment baselines on all three datasets. A minimal qualitative analysis confirms that high-universality features encode interpretable sociolinguistic patterns.

Zoltán Kmetty, Bence Kollányi, Krisztián Boros

Transformer-based machine learning models have become an essential tool for many natural language processing (NLP) tasks since the introduction of the method. A common objective of these projects is to classify text data. Classification models are often extended to a different topic and/or time period. In these situations, deciding how long a classification is suitable for and when it is worth re-training our model is difficult. This paper compares different approaches to fine-tune a BERT model for a long-running classification task. We use data from different periods to fine-tune our original BERT model, and we also measure how a second round of annotation could boost the classification quality. Our corpus contains over 8 million comments on COVID-19 vaccination in Hungary posted between September 2020 and December 2021. Our results show that the best solution is using all available unlabeled comments to fine-tune a model. It is not advisable to focus only on comments containing words that our model has not encountered before; a more efficient solution is randomly sample comments from the new period. Fine-tuning does not prevent the model from losing performance but merely slows it down. In a rapidly changing linguistic environment, it is not possible to maintain model performance without regularly annotating new text.

Zoltán Kmetty, Julia Koltai, Tamás Rudas

Research on social stratification is closely linked to analysing the prestige associated with different occupations. This research focuses on the positions of occupations in the semantic space represented by large amounts of textual data. The results are compared to standard results in social stratification to see whether the classical results are reproduced and if additional insights can be gained into the social positions of occupations. The paper gives an affirmative answer to both questions. The results show fundamental similarity of the occupational structure obtained from text analysis to the structure described by prestige and social distance scales. While our research reinforces many theories and empirical findings of the traditional body of literature on social stratification and, in particular, occupational hierarchy, it pointed to the importance of a factor not discussed in the main line of stratification literature so far: the power and organizational aspect.