Natural Syntax, Artificial Intelligence and Language Acquisition

Round 1

Reviewer 1 Report

The manuscript makes an interesting proposal of testing contemporary large language models (LLMs) to predict the acquisition of negation in English as a two-step developmental process. English is chosen as a test case because although negative concord is not found in standard English, English-speaking children are known to go through a stage in development when they show a preference for negative concord. I enjoyed reading it and I have a comment and two clarification requests. I recommend to accept the manuscript for publication.

 

1.     The main motivation for the proposed study is the poverty of the stimulus argument which in its absolute sense diminishes the role of input in language acquisition. Given the current state of research, it seems necessary to acknowledge that even researchers who work within the generative perspective no longer operate within the principles and parameters framework. Take for example a long history of research on the Null Subject phenomenon where a cross-linguistic approach brought us to the conclusion that children are sensitive to fine-grained distinctions in the input and the behavior of the English-speaking children matches the input they receive. For the LLMs, a very interesting question to address is how much input is necessary for a learner to acquire a linguistic phenomenon. This question is particularly vital in the context of multilingual language learning when input is reduced and distributed. Perhaps the authors could comment on this.

2.     In Section 3.1 the authors mention an early preference for negative concord and refer to Thornton et al. (2016).  Can you clarify whether this preference is found in production, comprehension, or both?

3.     It would be helpful if the authors could explain the induction bias in Section 3.2

Author Response

We are very grateful to the reviewer for their comments.

1. Yes, the issue of input in the context of multilingual acquisition is a very important one, which we have now acknowledged in footnote 1, along with a reference documenting that even in the case of bilingual acquisition, there are delays in the acquisition of each language due to the reduced input.
2. We have revised our discussion of the Thornton et al. study to make it clear that it involved a comprehension experiment. [line 237]
3. The term ‘induction bias’ was a mistake and has been removed.

Reviewer 2 Report

The current paper's format seems to be more of a literature review and theoretical proposal rather than a report on empirical research conducted by the authors. The section described as an "experiment to consider" suggests an experiment that could be carried out, but does not report on actual results from an experiment carried out by the authors.

If the authors have conducted research, that should be clarified in the paper. They need to describe their methodology, the research participants or data sources, how the data was analyzed, and what the findings were.

If no research was conducted, the paper should be revised to present it as a theoretical or literature review. The "experiment to consider" section should be more clearly framed as a proposal for future research.

I would propose a few changes to improve its clarity and flow:

Line 235: Consider changing "mundane acquisition" to "common acquisition." The term "mundane" may be interpreted negatively.

Line 245: Change "learn" to "develop." Saying children 'learn' a pattern they do not encounter could be confusing; suggesting they develop it might be clearer.

Line 278: Use "These findings are surprising" instead of "These are surprising findings." This makes the sentence easier to read.

Lines 319-320: Consider changing "yet to be precisely identified, but presumably complete before adulthood" to "that typically occurs before adulthood."

Lines 330-338: The quotes from other works could be integrated more smoothly. Perhaps introduce them by noting that there's a consensus or common argument that...

Line 342: Remove "(say)" and simply state the ages. It's clearer and more authoritative.

Lines 355-363: This could be a little clearer. Perhaps: "One potential outcome is that the Type A language model accurately mirrors children's two-stage acquisition process, suggesting this result can be achieved without regard to processing cost. This finding could suggest that humans could do the same, despite their more limited computational resources. Alternatively, only the Type B language model might mirror the two-stage process. This would suggest that a processing-cost bias is necessary even in machine learning, a conclusion that likely extends to humans as well."

Line 367: Consider changing "According to the recollection of pioneering psychologist and neuroscientist George Miller [34]" to "As pioneering psychologist and neuroscientist George Miller [34] recalls."

Line 381: It could be clearer what "the mistakes of the past" were. Consider expanding on this.

Line 386: "Already there are signs that a similar mis-step has been made" – it would be clearer if you specified what this mis-step is.

Line 392: Conclude with a clear and direct statement summarizing the main point or purpose of the paper. This will make it easier for readers to remember your argument.

 

Overall, the language quality of the document is good and professional.

Author Response

We deeply appreciate the reviewer’s comments and suggestions.

Yes, the reviewer is right that the leading idea of our paper involves a thought experiment, which we have now made clear from the outset, including in the abstract.

RE: the stylistic changes:

Line 235 [now 215]: We retained the word ‘mundane’ as it is a technical term in the work we do.

Line 245 [now 225-26]: Revised as recommended.

Line 278 [now 258]: Revised as recommended.

Lines 319-20 [now 330]: Revised

Lines 330-38 [now 308]: Revised as recommended

Line 342 [now 326]: Revised as recommended

Lines 355-363 [now 337ff]: Revised as recommended

Line 367 [now 351]: Revised, but in a slightly different way

Lines 381 & 386: Removed

Line 392 [now 363ff]: Two new final paragraphs were composed in order to comply with the reviewer’s wise advice.

Round 2

Reviewer 2 Report

The authors improved the manuscript as suggested.