When classifying grammatical role, BERT doesn’t care about word order... except when it matters

Papadimitriou, Isabel; Futrell, Richard; Mahowald, Kyle

doi:10.18653/v1/2022.acl-short.71

Cited by 9 publications

(11 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The fact that sentence representations of later model layers are more suitable for decoding plausibility than those of earlier layers is consistent with previous results showing that semantic information tends to be encoded more strongly in later layers (Belinkov et al, 2017;Papadimitriou, Futrell, & Mahowald, 2022;Tenney, Das, & Pavlick, 2019). The trend we observed in one of the models, MPT, where mid-layer performance exceeded late-layer performance, should be examined further in other large (multi-billion parameter) models.…”

Section: Internal Representations Of Event Plausibility Generalize Ac...supporting

confidence: 91%

“…A probe that is trained on a mix of active and passive sentences performs as successfully as the probe trained and tested on only one voice type, suggesting that plausible and implausible sentence embeddings in late LLM layers are linearly separable by the same hyperplane across syntactic structures. This finding aligns with recent computational work showing that even though most sentences in the language input describe prototypical events , LLMs are able to correctly represent the argument structure of nonprototypical event descriptions in late layers (Papadimitriou et al, 2022).…”

Section: Llms Can Infer Thematic Rolessupporting

confidence: 90%

See 1 more Smart Citation

Event Knowledge in Large Language Models: The Gap Between the Impossible and the Unlikely

Kauf,

Ivanova,

Rambelli

et al. 2023

Cognitive Science

View full text Add to dashboard Cite

Word co‐occurrence patterns in language corpora contain a surprising amount of conceptual knowledge. Large language models (LLMs), trained to predict words in context, leverage these patterns to achieve impressive performance on diverse semantic tasks requiring world knowledge. An important but understudied question about LLMs’ semantic abilities is whether they acquire generalized knowledge of common events. Here, we test whether five pretrained LLMs (from 2018's BERT to 2023's MPT) assign a higher likelihood to plausible descriptions of agent−patient interactions than to minimally different implausible versions of the same event. Using three curated sets of minimal sentence pairs (total n = 1215), we found that pretrained LLMs possess substantial event knowledge, outperforming other distributional language models. In particular, they almost always assign a higher likelihood to possible versus impossible events (The teacher bought the laptop vs. The laptop bought the teacher). However, LLMs show less consistent preferences for likely versus unlikely events (The nanny tutored the boy vs. The boy tutored the nanny). In follow‐up analyses, we show that (i) LLM scores are driven by both plausibility and surface‐level sentence features, (ii) LLM scores generalize well across syntactic variants (active vs. passive constructions) but less well across semantic variants (synonymous sentences), (iii) some LLM errors mirror human judgment ambiguity, and (iv) sentence plausibility serves as an organizing dimension in internal LLM representations. Overall, our results show that important aspects of event knowledge naturally emerge from distributional linguistic patterns, but also highlight a gap between representations of possible/impossible and likely/unlikely events.

show abstract

Section: Internal Representations Of Event Plausibility Generalize Ac...supporting

confidence: 91%

Section: Llms Can Infer Thematic Rolessupporting

confidence: 90%

Event Knowledge in Large Language Models: The Gap Between the Impossible and the Unlikely

Kauf,

Ivanova,

Rambelli

et al. 2023

Cognitive Science

View full text Add to dashboard Cite

show abstract

“…Word order is a crucial aspect of natural language, and studies have investigated its impact on language models by perturbing word order (Sinha et al, 2021b;Pham et al, 2021;Gupta et al, 2021;Hessel and Schofield, 2021;Clouatre et al, 2022;Yanaka and Mineshima, 2022;Papadimitriou et al, 2022) Negheimish et al (2023) tries to preserve the importance of word order by forcing the model to identify permuted sequences as invalid samples. In summary, existing works have coherently found that breaking word order do not result in a significant decrease in task performance.…”

Section: Related Workmentioning

confidence: 99%

“…Word order, referring to the sequential order of individual words within a text, is a fundamental concept in natural language. Previous works have investigated word order's impacts by altering order, and they find that no matter altering order in the pre-training or the training/inference data, the performance of downstream tasks drops marginally (Sinha et al, 2021a,b;Pham et al, 2021;Gupta et al, 2021;Hessel and Schofield, 2021;Clouatre et al, 2022;Yanaka and Mineshima, 2022;Papadimitriou et al, 2022), which demonstrates a counter-intuitive and unnatural phenomenon.…”

Section: Introductionmentioning

confidence: 98%

“…Regarding the experimental results, specifically, the marginal performance drops induced by altering word order, existing works provide varied explanations or hypotheses. For example, Papadimitriou et al (2022) argues that word order may be redundant with lexical semantics, exemplified by the bagof-words model. In other words, while they believe word order matters, they assert that order information can be derived from bag-of-words, rendering it redundant.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Materials Science at Beihang: A Special Issue of Materials Lab Dedicated to the 70th Anniversary of Beihang University

Zhao¹,

Guo²,

Jiang³

2022

MatLab

View full text Add to dashboard Cite

Beihang is the first university in China focusing on aeronautical and astronautical education and research, also covering diverse fields in the natural sciences, high technology, economics, management, the liberal arts, law, philosophy, foreign languages and education. School of Materials Science and Engineering at Beijing University was founded in 1954 as the Department of Aeronautical Metallurgy. In 2001, it was named School of Materials Science and Engineering, consisting

show abstract

Meaning creation in novel noun-noun compounds: humans and language models

Chen,

Poeppel,

Zuanazzi

2023

Language, Cognition and Neuroscience

View full text Add to dashboard Cite

The interpretation of novel noun-noun compounds (NNCs, e.g., "devil salary") requires the combination of nouns in the absence of syntactic cues, an interesting facet of complex meaning creation. Here we examine unconstrained interpretations of a large set of novel NNCs, to investigate how NNC constituents are combined into novel complex meanings. The data show that words' lexical-semantic features (e.g., material, agentivity, imageability, semantic similarity) differentially contribute to the grammatical relations and the semantics of NNC interpretations.Further, we demonstrate that passive interpretations incur higher processing cost (longer interpretation times and more eye-movements) than active interpretations.Finally, we show that large language models (GPT-2, BERT, RoBERTa) can predict whether a NNC is interpretable by human participants and estimate differences in processing cost, but do not exhibit sensitivity to more subtle grammatical differences. The experiments illuminate how humans can use lexicalsemantic features to interpret NNCs in the absence of explicit syntactic information.

show abstract

When classifying grammatical role, BERT doesn’t care about word order... except when it matters

Cited by 9 publications

References 8 publications

Event Knowledge in Large Language Models: The Gap Between the Impossible and the Unlikely

Event Knowledge in Large Language Models: The Gap Between the Impossible and the Unlikely

Materials Science at Beihang: A Special Issue of Materials Lab Dedicated to the 70th Anniversary of Beihang University

Meaning creation in novel noun-noun compounds: humans and language models

Contact Info

Product

Resources

About