On the evolution of syntactic information encoded by BERT’s contextualized representations

Pérez-Mayos, Laura; Carlini, Roberto; Ballesteros, Miguel; Wanner, Leo

doi:10.18653/v1/2021.eacl-main.191

Cited by 4 publications

(3 citation statements)

References 30 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A combination of techniques may also be used to bring downstream tasks back into the picture. In a recent study, Pérez-Mayos et al ( 2021 ) uses structural probing, not to assess whether a single static model has learned syntax or not, but to track how syntactic capabilities evolve as a pre-trained model is fine-tuned for different tasks. One could imagine a similar experimental design using TSE instead of (or together with) probing.…”

Section: Discussionmentioning

confidence: 99%

Schrödinger's tree—On syntax and neural language models

Kulmizev

Nivre

2022

Front. Artif. Intell.

View full text Add to dashboard Cite

In the last half-decade, the field of natural language processing (NLP) has undergone two major transitions: the switch to neural networks as the primary modeling paradigm and the homogenization of the training regime (pre-train, then fine-tune). Amidst this process, language models have emerged as NLP's workhorse, displaying increasingly fluent generation capabilities and proving to be an indispensable means of knowledge transfer downstream. Due to the otherwise opaque, black-box nature of such models, researchers have employed aspects of linguistic theory in order to characterize their behavior. Questions central to syntax—the study of the hierarchical structure of language—have factored heavily into such work, shedding invaluable insights about models' inherent biases and their ability to make human-like generalizations. In this paper, we attempt to take stock of this growing body of literature. In doing so, we observe a lack of clarity across numerous dimensions, which influences the hypotheses that researchers form, as well as the conclusions they draw from their findings. To remedy this, we urge researchers to make careful considerations when investigating coding properties, selecting representations, and evaluating via downstream tasks. Furthermore, we outline the implications of the different types of research questions exhibited in studies on syntax, as well as the inherent pitfalls of aggregate metrics. Ultimately, we hope that our discussion adds nuance to the prospect of studying language models and paves the way for a less monolithic perspective on syntax in this context.

show abstract

Section: Discussionmentioning

confidence: 99%

Schrödinger's tree—On syntax and neural language models

Kulmizev

Nivre

2022

Front. Artif. Intell.

View full text Add to dashboard Cite

show abstract

“…BERT and related models have been investigated extensively with different strategies over the last years, finding that BERT learns syntactic (Hewitt and Manning, 2019;Tenney et al, 2019) and semantic representations (Ettinger, 2020) that are generally preserved through fine-tuning for semantic tasks like paraphrasing (Pérez-Mayos et al, 2021). However, Hessel and Schofield (2021) find that on the GLUE tasks (Wang et al, 2018), BERT is relatively insensitive to shuffling of the input sentences, which removes many syntactic clues in English.…”

Section: Related Workmentioning

confidence: 99%

A Transformer for SAG: What Does it Grade?

Willms

Padó

2022

Proceedings of the 11th Workshop on Natural Language Processing for Computer-Assisted Language Learning (NLP4CALL 2022)

View full text Add to dashboard Cite

Automatic short-answer grading aims to predict human grades for short free-text answers to test questions, in order to support or replace human grading. Despite active research, there is to date no wide-spread use of ASAG in real-world teaching. One reason is a lack of transparency of popular methods like Transformer-based deep neural networks, which means that students and teachers cannot know how much to trust automated grading. We probe one such model using the adversarial attack paradigm to better understand their reliance on syntactic and semantic information in the student answers, and their vulnerability to the (easily manipulated) answer length. We find that the model is, reassuringly, likely to reject answers with missing syntactic and semantic information, but that it picks up on the correlation between answer length and correctness in standard training. Thus, real-world applications have to safeguard against exploitation of answer length.

show abstract

“…One body of work approaches the problem by applying heuristic rules of perturbation to input sequences (Wallace et al, 2019;Jia and Liang, 2017;, while another uses neural models to construct adversarial examples (Li et al, 2020(Li et al, , 2018 or manipulate inputs in embedding space (Jin et al, 2020). Our work also contributes to efforts to understand impacts and outcomes of the fine-tuning process (Miaschi et al, 2020;Mosbach et al, 2020;Wang et al, 2020;Perez-Mayos et al, 2021).…”

Section: Related Workmentioning

confidence: 99%

On the Interplay Between Fine-tuning and Composition in Transformers

Ettinger

2021

Preprint

View full text Add to dashboard Cite

Pre-trained transformer language models have shown remarkable performance on a variety of NLP tasks.However, recent research has suggested that phrase-level representations in these models reflect heavy influences of lexical content, but lack evidence of sophisticated, compositional phrase information (Yu and Ettinger, 2020). Here we investigate the impact of fine-tuning on the capacity of contextualized embeddings to capture phrase meaning information beyond lexical content. Specifically, we fine-tune models on an adversarial paraphrase classification task with high lexical overlap, and on a sentiment classification task. After fine-tuning, we analyze phrasal representations in controlled settings following prior work. We find that fine-tuning largely fails to benefit compositionality in these representations, though training on sentiment yields a small, localized benefit for certain models. In follow-up analyses, we identify confounding cues in the paraphrase dataset that may explain the lack of composition benefits from that task, and we discuss potential factors underlying the localized benefits from sentiment training.

show abstract

On the evolution of syntactic information encoded by BERT’s contextualized representations

Cited by 4 publications

References 30 publications

Schrödinger's tree—On syntax and neural language models

Schrödinger's tree—On syntax and neural language models

A Transformer for SAG: What Does it Grade?

On the Interplay Between Fine-tuning and Composition in Transformers

Contact Info

Product

Resources

About