What Does BERT Learn about the Structure of Language?

By Ganesh Jawahar, Benoît Sagot, Djamé Seddah. 2019

1,594 citations recorded by the source corpus.

BERT is a recent language representation model that has surprisingly performed well in diverse language understanding benchmarks. This result indicates the possibility that BERT networks capture structural information about language. In this work, we provide novel support for this claim by performing a series of experiments to unpack the elements of English language structure learned by BERT. Our findings are fourfold. BERT’s phrasal representation captures the phrase-level information in the lower layers. The intermediate layers of BERT compose a rich hierarchy of linguistic information, starting with surface features at the bottom, syntactic features in the middle followed by semantic features at the top. BERT requires deeper layers while tracking subject-verb agreement to handle long-term dependency problem. Finally, the compositional scheme underlying BERT mimics classical, tree-like structures.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Large Language Models

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Large Language Models

Citations in the corpus

How Can We Know What Language Models Know? (2019, background): “…a broad suite of NLP tasks are used to analyze both syntactic and semantic properties, providing evidence that contextualized representations encode linguistic knowledge in different layers (Hewitt and Manning, 2019; Tenney et al., 2019a; Tenney et al., 2019b; Jawahar et al., 2019; Goldberg, 2019).”
MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers (2020, background): “Some works show that self-attention distributions of pre-trained LMs capture a rich hierarchy of linguistic information (Jawahar et al., 2019; Clark et al., 2019).”
Of Non-Linearity and Commutativity in BERT (2021, methodology): “The authors of [31] and [13] show that features are extracted in BERT following the same order as in the classical NLP pipeline, with syntactic features extracted in earlier layers and semantic features in later layers.”
Transformer Feed-Forward Layers Are Key-Value Memories (2020, background): “The study of the transformer architecture has focused on the role and function of self-attention layers (Voita et al., 2019; Clark et al., 2019; Vig and Belinkov, 2019) and on inter-layer differences (i.e. lower vs. upper layers) (Tenney et al., 2019; Jawahar et al., 2019).”
A Primer in BERTology: What We Know About How BERT Works (2020, background): “Jawahar et al. (2019) include a brief illustration of a dependency tree extracted directly from self-attention weights, but provide no quantitative evaluation.”
BERTology Meets Biology: Interpreting Attention in Protein Language Models (2020, neutral): “These results are consistent with prior work in NLP that suggests deeper layers in text-based Transformers attend to more complex properties (Vig & Belinkov, 2019) and encode higher-level representations (Raganato & Tiedemann, 2018; Peters et al., 2018; Tenney et al., 2019; Jawahar et al., 2019).”
Continual Sequence Generation with Adaptive Compositional Modules (2022, positive): “This is consistent with previous ﬁndings (Jawahar et al., 2019) that bag-of-word information is mainly captured by higher transformer layers, while lower transformer layers capture surface and syntactic information.”
ECONET: Effective Continual Pretraining of Language Models for Event Temporal Reasoning (2020, background): “Previous research report that upper layers of pre-trained language models focus more on complex semantics as opposed to shallow surface forms or syntax on the lower layers (Tenney et al., 2019; Jawahar et al., 2019).”
Leveraging Pre-trained Checkpoints for Sequence Generation Tasks (2019, neutral): “in “Initializing a Subset of Layers”, this would mean that the DiscoFuse task proﬁts the most from pre-trained information about POS, constituents, dependencies and semantic roles. A similar study by Jawahar et al. (2019) found that BERT captures phrase-level information in the lowerlayers and lingu”
MixText: Linguistically-Informed Interpolation of Hidden Space for Semi-Supervised Text Classification (2020, background): “Based on (Jawahar et al., 2019), the { 3,4,5,6,7,9,12 } are the most informative layers in BERT based model and each of them captures different types of informa-Figure tion (e.g., surface, syntactic, or semantic).”
jiant: A Software Toolkit for Research on General-Purpose Text Understanding Models (2020, methodology): “Early versions of jiant have already been used in multiple works, including probing analyses (Tenney et al., 2019b,a; Warstadt et al., 2019; Lin et al., 2019; Hewitt and Manning, 2019; Jawahar et al., 2019), transfer learning experiments (Wang et al.”
MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices (2020, background): “Attention Transfer (AT) The attention mechanism greatly boosts the performance of NLP and becomes a crucial building block in Transformer and BERT (Clark et al., 2019a; Jawahar et al., 2019).”
Behind the Scene: Revealing the Secrets of Pre-trained Vision-and-Language Models (2020, methodology): “ion (NMI), an unsupervised metric for evaluating dierences between clusters. A larger NMI value implies that the distinction between two clusters is more signicant, indicating a lower fusion degree [14]. For example, when NMI is equal to 1.0, the two clusters represent the original visual tokens and”
BERTs of a feather do not generalize together: Large variability in generalization across models with similar test set performance (2019, background): “…of BERT, whether to assess its encoding of sentence structure (Lin et al., 2019; He-witt and Manning, 2019; Chrupała and Alishahi, 2019; Jawahar et al., 2019; Tenney et al., 2019b); its representational structure more generally (Ab-nar et al., 2019); its handling of speciﬁc linguistic…”
Intermediate-Task Transfer Learning with Pretrained Language Models: When and Why Does It Work? (2020, methodology): “Many recent works have attempted to understand the knowledge and linguistic skills BERT learns, for instance by analyzing the language model surprisal for subject–verb agreements (Goldberg, 2018), identifying specific knowledge or phenomena encapsulated in the representations learned by BERT using p”