Can BERT eat RuCoLA*? Topological Data Analysis to Explain

This repository contains code and data for the EACL 2023 SlavNLP workshop paper Can BERT eat RuCoLA? Topological Data Analysis to Explain.

In the paper, we investigate how Transformer language models (LMs) fine-tuned for acceptability classification capture linguistic features. Our approach uses the best practices of topological data analysis (TDA) in NLP: we construct directed attention graphs from attention matrices, derive topological features from them, and feed them to linear classifiers. We introduce new topological features, suggest a new TDA-based approach for measuring the distance between pre-trained and fine-tuned LMs with large and base configurations, and determine the roles of attention heads in the context of LA tasks in Russian and English.

*Arugula or rocket salad in English.

---

Attention graphs for acceptable (left) and unacceptable (right) sentences with corresponding attention matrices, extracted from BERT (Layer, Head: [9,5]).

Usage

• Use 1_Fine-tuning_example.ipynb for training LM on LA (Russian or English) task.
• Use 2_1_Topological_and_template_features_calculation.ipynb and 2_2_Ripser_features.ipynb for computing features.
• Use 3_1_Topological_features_distance.ipynb and 3_2_JS_divergence_distance.ipynb to measure fine-tuning effect with TDA feature distance and JS Shennon divergence.
• Use 4_TDA_acceptability_classification.ipynb for acceptability judgements classification with TDA features.
• Use 5_Head_importance.ipynb to evaluate per-sample confidence and estimate head and feature importances per violation group and per-sample.

Features extracted from fine-tuned LMs can be found here: https://huggingface.co/iproskurina (models with prefix tda-*).
Classifiers can be downloaded here.
We conduct all the experiments on monolingual encoders fine-tuned on grammatical acceptability corpora in English CoLA and Russian RuCoLA🍃.

Other notebooks directory contains notebooks for acceptability judgements classification with TDA features with linear feature selection, principal components importance estimation with Shapley values, and autosklearn classification example.

Consider the following related work introducing TDA-based approaches in NLP:

• Kushnareva, L., Cherniavskii, D., Mikhailov, V., Artemova, E., Barannikov, S., Bernstein, A., Piontkovskaya, I., Piontkovski, D., & Burnaev, E. (2021). Artificial Text Detection via Examining the Topology of Attention Maps. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (pp. 635–649). Association for Computational Linguistics. [paper][code]
• Kushnareva, Laida, Dmitri Piontkovski, and Irina Piontkovskaya. "Betti numbers of attention graphs is all you really need." arXiv preprint arXiv:2207.01903 (2022). [paper]
• Barannikov, Serguei. "The framed Morse complex and its invariants." Advances in Soviet Mathematics 21 (1994): 93-116.* [paper] (The algorithm for the calculation of persistent barcodes (or canonical forms))

Feature importance

Important features sorted by Mann-Whitney test p-value are presented here.

Remark: Results in that folder for Swedish were obtained when fine-tuning Swe-BERT on DaLAJ dataset. For Italian, we report feature importances extracted from Ita-BERT fine-tuned on ItaCoLA dataset.

Contact

If you have any questions about the project and/or implementation, you can reach us here.

Cite

Our paper has been accepted to the EACL 2023 Slavic NLP workshop.

Irina Proskurina, Ekaterina Artemova, and Irina Piontkovskaya. 2023. Can BERT eat RuCoLA? Topological Data Analysis to Explain. In Proceedings of the 9th Workshop on Slavic Natural Language Processing 2023 (SlavicNLP 2023), pages 123–137, Dubrovnik, Croatia. Association for Computational Linguistics.

@inproceedings{proskurina-etal-2023-bert,
    title = "Can {BERT} eat {R}u{C}o{LA}? Topological Data Analysis to Explain",
    author = "Proskurina, Irina  and
      Artemova, Ekaterina  and
      Piontkovskaya, Irina",
    booktitle = "Proceedings of the 9th Workshop on Slavic Natural Language Processing 2023 (SlavicNLP 2023)",
    month = may,
    year = "2023",
    address = "Dubrovnik, Croatia",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.bsnlp-1.15",
    pages = "123--137",
    abstract = "This paper investigates how Transformer language models (LMs) fine-tuned for acceptability classification capture linguistic features. Our approach is based on best practices of topological data analysis (TDA) in NLP: we construct directed attention graphs from attention matrices, derive topological features from them and feed them to linear classifiers. We introduce two novel features, chordality and the matching number, and show that TDA-based classifiers outperform fine-tuning baselines. We experiment with two datasets, CoLA and RuCoLA, in English and Russian, which are typologically different languages. On top of that, we propose several black-box introspection techniques aimed at detecting changes in the attention mode of the LM{'}s during fine-tuning, defining the LM{'}s prediction confidences, and associating individual heads with fine-grained grammar phenomena. Our results contribute to understanding the behaviour of monolingual LMs in the acceptability classification task, provide insights into the functional roles of attention heads, and highlight the advantages of TDA-based approaches for analyzing LMs.We release the code and the experimental results for further uptake.",
}