A Primer in BERTology: What we know about how BERT works

Rogers, Anna Backman; Ковалева, Ольга Николаевна; Rumshisky, Anna

doi:10.48550/arxiv.2002.12327

Cited by 51 publications

(65 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…The distribution of linguistic information differs not only by model and input data (Ethayarajh 2019;Gao et al 2019), but even for different hyperparameters, training runs, or contexts (Tenney, Das, and Pavlick 2019). Indeed, it may not be uniquely identifiable at all (Rogers, Kovaleva, and Rumshisky 2020). For that reason, an aggregation of layer representations that provides semantic similarity or distances between tokens currently requires exogenous labels for a probing procedure (Hewitt and C. D. Manning 2019).…”

Section: A1 Challenges In the Application Of Deep Language Modelsmentioning

confidence: 99%

“…For that reason, an aggregation of layer representations that provides semantic similarity or distances between tokens currently requires exogenous labels for a probing procedure (Hewitt and C. D. Manning 2019). Even if available, this procedure introduces a gap between the training objective of the model and the analysis procedure used for inference, possibly leading to bias (Rogers, Kovaleva, and Rumshisky 2020).…”

Section: A1 Challenges In the Application Of Deep Language Modelsmentioning

confidence: 99%

“…Furthermore, we do not require alignment of embedding spaces, because our output is a network with mathematically well-defined relations that are independent of the corpus. Since our training and inference task are identical, we do not risk a gap between training and analysis (Rogers, Kovaleva, and Rumshisky 2020) and the accuracy of our approach directly depends on the model's native training performance. On the other hand, any technique to increase performance of lexical substitutions (e.g., Amrami and Goldberg 2019;Laicher et al 2021;Schick and Schütze 2019) apply to our model as well.…”

Section: A1 Challenges In the Application Of Deep Language Modelsmentioning

confidence: 99%

“…However, the contextual representations of transformer models are distributed across model layers in a non-isotonic and non-Euclidian fashion, which is also corpus-, training-and instance-dependent (Rogers, Kovaleva, and Rumshisky 2020). Consequently, there is no obviously correct, generally accepted approach to working with word representations of deep language models.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Text analysis and deep learning: A network approach

Marquart¹

2021

Preprint

View full text Add to dashboard Cite

Much information available to applied researchers is contained within written language or spoken text. Deep language models such as BERT have achieved unprecedented success in many applications of computational linguistics. However, much less is known about how these models can be used to analyze existing text. We propose a novel method that combines transformer models with network analysis to form a self-referential representation of language use within a corpus of interest. Our approach produces linguistic relations strongly consistent with the underlying model as well as mathematically well-defined operations on them, while reducing the amount of discretionary choices of representation and distance measures. It represents, to the best of our knowledge, the first unsupervised method to extract semantic networks directly from deep language models. We illustrate our approach in a semantic analysis of the term "founder". Using the entire corpus of Harvard Business Review from 1980 to 2020, we find that ties in our network track the semantics of discourse over time, and across contexts, identifying and relating clusters of semantic and syntactic relations. Finally, we discuss how this method can also complement and inform analyses of the behavior of deep learning models.

show abstract

Section: A1 Challenges In the Application Of Deep Language Modelsmentioning

confidence: 99%

Section: A1 Challenges In the Application Of Deep Language Modelsmentioning

confidence: 99%

Section: A1 Challenges In the Application Of Deep Language Modelsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Text analysis and deep learning: A network approach

Marquart¹

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…There is a growing field of study concerned with investigating the inner working of large-scale transformers like BERT. A new word BERTology has even been defined to describe the related research work carried out around BERT [3]. Our work is focused on BERT input embedding that is both critical to the BERT model and easy to overlook.…”

Section: Introductionmentioning

confidence: 99%

Learning to Remove: Towards Isotropic Pre-trained BERT Embedding

Liang¹,

Cao²,

Zheng³

et al. 2021

Preprint

View full text Add to dashboard Cite

Pre-trained language models such as BERT have become a more common choice of natural language processing (NLP) tasks. Research in word representation shows that isotropic embeddings can significantly improve performance on downstream tasks. However, we measure and analyze the geometry of pre-trained BERT embedding and find that it is far from isotropic. We find that the word vectors are not centered around the origin, and the average cosine similarity between two random words is much higher than zero, which indicates that the word vectors are distributed in a narrow cone and deteriorate the representation capacity of word embedding. We propose a simple, and yet effective method to fix this problem: remove several dominant directions of BERT embedding with a set of learnable weights. We train the weights on word similarity tasks and show that processed embedding is more isotropic. Our method is evaluated on three standardized tasks: word similarity, word analogy, and semantic textual similarity. In all tasks, the word embedding processed by our method consistently outperforms the original embedding (with average improvement of 13% on word analogy and 16% on semantic textual similarity) and two baseline methods. Our method is also proven to be more robust to changes of hyperparameter.

show abstract

Evaluation of taxonomic and neural embedding methods for calculating semantic similarity

Yang

Yin

2021

Nat. Lang. Eng.

View full text Add to dashboard Cite

Modelling semantic similarity plays a fundamental role in lexical semantic applications. A natural way of calculating semantic similarity is to access handcrafted semantic networks, but similarity prediction can also be anticipated in a distributional vector space. Similarity calculation continues to be a challenging task, even with the latest breakthroughs in deep neural language models. We first examined popular methodologies in measuring taxonomic similarity, including edge-counting that solely employs semantic relations in a taxonomy, as well as the complex methods that estimate concept specificity. We further extrapolated three weighting factors in modelling taxonomic similarity. To study the distinct mechanisms between taxonomic and distributional similarity measures, we ran head-to-head comparisons of each measure with human similarity judgements from the perspectives of word frequency, polysemy degree and similarity intensity. Our findings suggest that without fine-tuning the uniform distance, taxonomic similarity measures can depend on the shortest path length as a prime factor to predict semantic similarity; in contrast to distributional semantics, edge-counting is free from sense distribution bias in use and can measure word similarity both literally and metaphorically; the synergy of retrofitting neural embeddings with concept relations in similarity prediction may indicate a new trend to leverage knowledge bases on transfer learning. It appears that a large gap still exists on computing semantic similarity among different ranges of word frequency, polysemous degree and similarity intensity.

show abstract

A Primer in BERTology: What we know about how BERT works

Cited by 51 publications

References 0 publications

Text analysis and deep learning: A network approach

Text analysis and deep learning: A network approach

Learning to Remove: Towards Isotropic Pre-trained BERT Embedding

Evaluation of taxonomic and neural embedding methods for calculating semantic similarity

Contact Info

Product

Resources

About