# From Balustrades to Pierre Vinken: Looking for Syntax in Transformer Self-Attentions

arXiv: https://arxiv.org/abs/1906.01958 · HTML: https://mlsanity.com/paper/1906.01958/

## Key facts

- Cited by 22 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 5 Jun 2019.

## Built on by (most cited first, 22 of 22)

- [Probing Classifiers: Promises, Shortcomings, and Advances](https://mlsanity.com/paper/2102.12452/) (2021)
- [Incorporating Residual and Normalization Layers into Analysis of Masked Language Models](https://mlsanity.com/paper/2109.07152/) (2021)
- [Cross-modal Attention Congruence Regularization for Vision-Language Relation Alignment](https://mlsanity.com/paper/2212.10549/) (2022)
- [Predicting Discourse Trees from Transformer-based Neural Summarizers](https://mlsanity.com/paper/2104.07058/) (2021)
- [Linguistic Dependencies and Statistical Dependence](https://mlsanity.com/paper/2104.08685/) (2021)
- [What does Transformer learn about source code?](https://mlsanity.com/paper/2207.08466/) (2022)
- [Transformer Language Models Handle Word Frequency in Prediction Head](https://mlsanity.com/paper/2305.18294/) (2023)
- [GradTS: A Gradient-Based Automatic Auxiliary Task Selection Method Based on Transformer Networks](https://mlsanity.com/paper/2109.05748/) (2021)
- [Linguistic Interpretability of Transformer-based Language Models: a systematic review](https://mlsanity.com/paper/2504.08001/) (2025)
- [Towards Understanding Large-Scale Discourse Structures in Pre-Trained and Fine-Tuned Language Models](https://mlsanity.com/paper/2204.04289/) (2022)
- [Measuring Affinity between Attention-Head Weight Subspaces via the Projection Kernel](https://mlsanity.com/paper/2601.10266/) (2026)
- [Discourse Structure Extraction from Pre-Trained and Fine-Tuned Language Models in Dialogues](https://mlsanity.com/paper/2302.05895/) (2023)
- [CUE: An Uncertainty Interpretation Framework for Text Classifiers Built on Pre-Trained Language Models](https://mlsanity.com/paper/2306.03598/) (2023)
- [Embedded Named Entity Recognition using Probing Classifiers](https://mlsanity.com/paper/2403.11747/) (2024)
- [Emergent Linguistic Structures in Neural Networks are Fragile](https://mlsanity.com/paper/2210.17406/) (2022)
- [Revisiting the Practical Effectiveness of Constituency Parse Extraction from Pre-trained Language Models](https://mlsanity.com/paper/2211.00479/) (2022)
- [Semantic Composition in Visually Grounded Language Models](https://mlsanity.com/paper/2305.16328/) (2023)
- [Where exactly does contextualization in a PLM happen?](https://mlsanity.com/paper/2312.06514/) (2023)
- [Multilingual Vision-Language Models, A Survey](https://mlsanity.com/paper/2509.22123/) (2025)
- [The Grammar of Transformers: A Systematic Review of Interpretability Research on Syntactic Knowledge in Language Models](https://mlsanity.com/paper/2601.19926/) (2026)
- [Focused Contrastive Training for Test-based Constituency Analysis](https://mlsanity.com/paper/2109.15159/) (2021)
- [On Eliciting Syntax from Language Models via Hashing](https://mlsanity.com/paper/2410.04074/) (2024)

## Abstract (arXiv)

We inspect the multi-head self-attention in Transformer NMT encoders for three source languages, looking for patterns that could have a syntactic interpretation. In many of the attention heads, we frequently find sequences of consecutive states attending to the same position, which resemble syntactic phrases. We propose a transparent deterministic method of quantifying the amount of syntactic information present in the self-attentions, based on automatically building and evaluating phrase-structure trees from the phrase-like sequences. We compare the resulting trees to existing constituency treebanks, both manually and by computing precision and recall.