2020
DOI: 10.31234/osf.io/97qcg
|View full text |Cite
Preprint
|
Sign up to set email alerts
|

Neural Language Models Capture Some, But Not All, Agreement Attraction Effects

Abstract:

The number of the subject in English must match the number of the corresponding verb (dog runs but dogs run). Yet in real-time language production and comprehension, speakers often mistakenly compute agreement between the verb and a grammatically irrelevant non-subject noun phrase instead. This phenomenon, referred to as agreement attraction, is modulated by a wide range of factors; any complete computational model of grammatical planning and comprehension would be expected to derive this rich empirical pic… Show more

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
2
1
1

Citation Types

2
12
2

Year Published

2021
2021
2024
2024

Publication Types

Select...
5
2
2

Relationship

0
9

Authors

Journals

citations
Cited by 13 publications
(16 citation statements)
references
References 15 publications
2
12
2
Order By: Relevance
“…Yet, a major gap remains between humans and these algorithms: current language models are still poor at story generation and summarization as well as dialogue and question answering (10)(11)(12)(13)(14); they fail to capture many syntactic constructs and semantics properties (15)(16)(17)(18)(19), and their linguistic understanding is often superficial (16,(18)(19)(20).…”
mentioning
confidence: 99%
“…Yet, a major gap remains between humans and these algorithms: current language models are still poor at story generation and summarization as well as dialogue and question answering (10)(11)(12)(13)(14); they fail to capture many syntactic constructs and semantics properties (15)(16)(17)(18)(19), and their linguistic understanding is often superficial (16,(18)(19)(20).…”
mentioning
confidence: 99%
“…Just as importantly, we want to understand which kinds of phenomena do not emerge from such an account, suggesting additional learning mechanisms or inductive biases may be responsible. Our approach is shared by a number of other recent works using state-of-the-art architectures from machine learning to provide insights into human cognition, such as using pre-trained models in categorization (Lake et al, 2015;Peterson et al, 2018) or in language (Arehalli and Linzen, 2020;Manning et al, 2020).…”
Section: Discussionmentioning
confidence: 99%
“…A prime example of the pure prediction approach can be found in Gulordava et al ( 2018 ): a vanilla LSTM is trained on a Language Modeling task, under the argument that the predictive mechanism is sufficient for the network to predict long-distance number agreement. The authors conclude that “LM-trained RNNs can construct abstract grammatical representations.” In a more ambivalent study, Arehalli and Linzen ( 2020 ) consider how real-time human comprehension and production do not always follow the general grammatical constraint of subject-verb agreement, due to a variety of possible syntactic or semantic factors. They replicate six experiments from the agreement attraction literature using LSTMs as subjects, and find that the model, despite its relatively simple structure, captures human behavior in at least three of them.…”
Section: Neural Language Models and Language Developmentmentioning
confidence: 99%