Emergent linguistic structure in artificial neural networks trained by self-supervision

By Christopher D. Manning, Kevin Clark, John Hewitt, Urvashi Khandelwal, Omer Levy. 2020

402 citations recorded by the source corpus.

This paper explores the knowledge of linguistic structure learned by large artificial neural networks, trained via self-supervision, whereby the model simply tries to predict a masked word in a given context. Human language communication is via sequences of words, but language understanding requires constructing rich hierarchical structures that are never observed explicitly. The mechanisms for this have been a prime mystery of human language acquisition, while engineering work has mainly proceeded by supervised learning on treebanks of sentences hand labeled for this latent structure. However, we demonstrate that modern deep contextual language models learn major aspects of this structure, without any explicit supervision. We develop methods for identifying linguistic hierarchical structure emergent in artificial neural networks and demonstrate that components in these models focus on syntactic grammatical relationships and anaphoric coreference. Indeed, we show that a linear transformation of learned embeddings in these models captures parse tree distances to a surprising degree, allowing approximate reconstruction of the sentence tree structures normally assumed by linguists. These results help explain why these models have brought such large improvements across many language-understanding tasks.

Subfields: Artificial Intelligence, Natural Language Processing, Representation & Self-Supervised Learning

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 Natural Language Processing

In Representation & Self-Supervised Learning

Citations in the corpus

Emergent Abilities of Large Language Models (2022, neutral): “It is now well-known that increasing the scale of language models (e.g., training compute, model parameters, etc.) can lead to better performance and sample efﬁciency on a range of downstream NLP tasks (Devlin et al., 2019; Brown et al., 2020, inter alia ).”
Pre-Trained Models: Past, Present and Future (2021, background): “Furthermore, Manning et al. (2020) explore to reconstruct the sentence tree structures given by linguists using a linear transformation of PTMs ’ embeddings and achieve promising results.”
UnNatural Language Inference (2020, background): “…phenomena (Linzen et al., 2016) that transformer-based models pretrained on LM do acquire some knowledge of natural language syntax (Gulordava et al., 2018; Chrupała and Alishahi, 2019; Jawahar et al., 2019; Lin et al., 2019; Manning et al., 2020; Hawkins et al., 2020; Linzen and Baroni, 2021).”
Large Language Models Can Be Strong Differentially Private Learners (2021, background): “Pretrained language models tend to contain general knowledge of language (Manning et al., 2020) and thus should make the downstream private learning problem easier.”
HittER: Hierarchical Transformers for Knowledge Graph Embeddings (2020, background): “Although prior work shows Trans-formers can learn relational knowledge from large amounts of unstructured textual data (Jiang et al., 2020; Manning et al., 2020), HittER explicitly operates over structured inputs using a hierarchical architecture .”
Low-Complexity Probing via Finding Subnetworks (2021, background): “1 Clark et al. (2019), Hewitt and Manning (2019), and Manning et al. (2020) found that BERT captures various properties of syntax.”
Acquisition of chess knowledge in AlphaZero (2021, background): “Certain language models reproduce sentence parse trees (7).”
Understanding Natural Language Understanding Systems. A Critical Analysis (2023, background): “The key aspect of language modeling is that it does not require annotated data, because the learning objective of the network is to reproduce the input word sequence itself, hence the name of self-supervised leaning (Manning, Clark, Hewitt, Khandelwal and Levy 2020).”
Brains and algorithms partially converge in natural language processing (2022, background): “Further inspection of arti ﬁ cial 8,68 and biological networks 10,28,69 remains necessary to further decompose them into interpretable features.”
Shared computational principles for language processing in humans and deep language models (2022, background): “Remarkably, the internal contextual embedding space in DLMs can capture many aspects of the latent structure of human language, including syntactic trees, voice, co-references, morphology and long-range semantic and pragmatic dependencies 1,63,64 .”
Masked Language Modeling and the Distributional Hypothesis: Order Word Matters Pre-training for Little (2021, background): “In this work, we try to uncover how much of MLM’s success comes from simple distributional information, as opposed to “the types of syntactic and semantic abstractions traditionally believed necessary for language processing” (Tenney et al., 2019; Manning et al., 2020).”
Extending Machine Language Models toward Human-Level Language Understanding (2019, methodology): “The Query and RepVs were made up for illustration, inspired by (24).”