Speakers
Description
The Oxford English Dictionary (OED) has long been recognized for its rigorous scholarship, yet it is equally notable for its readiness to adopt new technologies. As the volume and availability of linguistic evidence has grown in the digital era, so too have the challenges of efficiently revising a dictionary of this size and scope: in response, the OED has been undertaking a systematic investigation into the potential role of AI in supporting historical lexicography, aspects of which have been reported at previous conferences (Hawkes, Nicholson, and Rogers 2025; Wild 2024). In this paper we present findings to date on successful and less successful experiments, and describe the OED’s approach to scaling up the successes.
Much has been written about the capabilities of generative AI to draft dictionary definitions; in this paper, we report on two experiments that instead explore how LLMs can support the adaptation or updating of existing definitions for different purposes or audiences. The first experiment involved adapting definitions from the OED (a large historical dictionary of English) into a style suitable for the Oxford Dictionary of English (a general-purpose dictionary of current English). This requires removing obsolete or minor senses, making technical or formal language more accessible, simplifying long definitions, and similar transformations. We tested whether prompting an LLM to perform this task, followed by editorial review, could yield consistent and time-saving results; however, the output proved too variable to be implemented at scale.
The second experiment (initially reported in Wild 2024) focused on modernizing unrevised OED definitions, many of which were written in the 19th century, into a style appropriate for contemporary readers. With over half a million entries and more than 800,000 senses, the OED is one of the largest dictionaries in existence. This presents significant challenges in determining which entries and senses are most in need of updating. We first discuss methods of identifying outdated definitions, including traditional computational methods (such as finding definitions containing words which have declined in frequency in a corpus), and AI-supported methods (including diachronic embeddings and LLM verification). We then present the results of an experiment using an LLM to 'translate' archaic wording into modern English, when supplied with the unrevised OED definition, its associated illustrative quotations, and a prompt with examples. The results showed that AI-modernized definitions can serve as an aid for editors, along with other sources, although editorial input remains essential.
We have also investigated the capacity for AI to assist with retrieving and analyzing quotation evidence. OED's rich historical sense inventory is supported by typical, illustrative examples of a word or meaning, including the earliest use of a given sense and a recent example if the sense is still in current use. Research activity on the OED in this area involves laborious sifting of a large quantity of data from primary source databases, and the manual nature of this activity has limited the efficiency of how we work, and the volume of entries we can update. We evaluate the extent to which AI-assisted word sense disambiguation can support this task, firstly in a project using LLMs to retrieve quotations from a corpus of historical English and match them to OED senses to create a ‘Quotations Finder’ tool that aims to support the work of OED editors, and supplement OED quotation paragraphs for external users (initially reported in Hawkes, Nicholson & Rogers 2025). As part of this project, we explored the performance of 7 different LLM models and their comparative ability to carry out word sense disambiguation, and we will describe the relative merits of the model selected for the production of this tool. We also discuss a similar project for modern English which matches OED senses with examples from large corpora of 21st-century English. Our analysis shows that AI-assisted word sense disambiguation (in both historical and modern texts) has a relatively high accuracy rate for certain types of entry and sense, but that there are limitations and areas where editorial intervention is essential in selecting suitable quotations.
Taken together, these projects demonstrate both the promise and the limits of applying AI to historical lexicography. While current models cannot replace the nuanced judgement required for OED revision, they can meaningfully augment editorial work by accelerating manual tasks, and improve the discoverability of evidence. Our findings suggest that, when deployed selectively and with rigorous oversight, AI can help the OED scale its research and modernize its content, while preserving the scholarly standards that define the dictionary.