Date created: | Last Updated:

Category: Project

Description: Large Language Models (LLMs) have the potential to play a pivotal role in corpus linguistics research, providing a deep and nuanced view of language use in a variety of domains and registers. Current implementations of LLMs, such as ChatGPT, are built using massive amounts of language data and can be easily accessed through web interfaces and application programming interfaces (APIs). However, when these models are prompted for insights on language, it can be difficult to evaluate what specific data informs the responses or even how accurate the responses are due to 'hallucinations' (i.e., errors in the model output). Here, we show how LLMs can be integrated into a traditional corpus analysis toolkit, allowing them to be linked directly to target corpora. When used in this way, LLMs can be prompted to query individual corpus files or the entire corpus. They can also be prompted for insights on the results of traditional corpus tools, such as KWIC concordancers, offering a completely new view of the data. The risk of 'hallucinations' is still present, but this phenomenon can be greatly reduced through prompt engineering, and the insights offered by the LLM can be checked using direct links back to the target corpus.