PhD Dissertation Proposal: Neha Nayak Kennard, Peer Review in the Age of Large Language Models: Discourse Structure and Sociological Implications
Content
Speaker:
Abstract:
Peer review is trusted by the international scientific community for the validation and curation of scientific knowledge. Yet in recent decades, the increasing volume of submitted manuscripts has led to untenable reviewer workloads. Since this problem involves large amounts of text, Natural Language Processing (NLP) methods to approach it seem apt. For example, discourse parsing could be used to characterize the complex arguments made in peer reviews and rebuttals. Peer reviews tend to reflect deep insights and highly specialized knowledge, presenting a challenge for many NLP systems. The advent of highly performant large language models (LLMs) enabled simple and elegant solutions to previously open problems. This advance has, however, been a double-edged sword: although peer reviews are now more easily analyzed, plausible review text can also be generated with minimal effort on the reviewer’s part, leading to new, sociotechnical problems which demand urgent, principled responses rooted in sociology of science. This thesis develops datasets, methods and theory to support the peer review process in both the pre- and post-LLM paradigms.
First, I present DISAPERE, a dataset of intertextual discourse structure between peer reviews and rebuttals. This dataset consists of 506 review-rebuttal pairs, with each of their 20k sentences labeled with discourse actions. I develop a taxonomy of rebuttal actions, characterizing the authors’ attitude towards reviewers’ comments. I show how these labels can be used to calculate metrics such as agreeability that characterize peer review discussions and could support area chairs in triaging between the manuscripts they oversee. I propose analyses in which models trained on DISAPERE are used to study correlations between discourse structure and attributes like reviewer confidence, metareviewer sentiment, and score changes.
Next, I present ongoing work aiming to produce high-quality Rhetorical Structure Theory (RST) parses of peer review text.
RST parses go beyond the sentence-level speech acts labeled by existing models for peer review discourse, providing a detailed characterization of the argumentative strategies used by reviewers. However, directly producing RST parses is not yet within the capabilities of contemporary LLMs, and annotating new data to train state-of-the-art neural parsers is prohibitively expensive. To bridge this gap, we present a data augmentation method for RST. Using LLMs, we generate texts conditioned on partial RST parses to produce an order of magnitude more labeled data than previously available, with substantially more diverse discourse structure, at a lower cost than that of annotating more data. I propose two types of analyses using these models. First, I plan to compare the nature of arguments made in peer reviews in different fields. Second, I will compare the argument structure of traditional and LLM-generated peer reviews to gauge whether peer reviews suffer the collapse in discourse structure diversity observed in other genres.
Finally, I address the role of NLP in peer review in the post-LLM paradigm, by developing a framework to account for the consequences of scientific reviews generated in the absence of traditional reviewers (Socially Untethered PEer Reviews, or SUPeRs). Since LLM 'reviewers' violate important assumptions used in sociological theory, I draw on concepts from philosophy of language to show how SUPeRs are fundamentally distinct from traditional peer reviews in their ability to support objectivity and efficiency in science. I then derive the conditions – in terms of a scientific community’s attitudes towards LLMs – in which the two become equivalent. I then present a set of open questions to guide future research on NLP in peer review, and show how this framework could be extended to other LLM 'speech' with effects in the social world.
Advisor:
Andrew McCallum