Preparing glossaries with GPT-4

[Tom Gally]

I got access to a beta version of GPT-4 yesterday that is able to do web searches in order to collect information to use in its responses. My first dozen tests haven't gone well. It is unable to access most sites, maybe because the sites are blocking bots.

I wanted to see if I could get it to search through English and Japanese pages about the same topic (but not translations of each other) and prepare a bilingual glossary of corresponding terms. I couldn't get it to do that from URLs (it couldn't even access Wikipedia pages), but I did succeed in getting it to create a glossary from a couple of short texts I gave it in the prompt. I’ve put that conversation on the following page, including, at the end, its explanation of how much text it can process at one time. Not enough, probably, for any large-scale glossary-preparation tasks.

https://www.gally.net/temp/202212chatgpt/glossarycreation.html

Anthropic has announced that its Claude LLM can accept up to 100,000 tokens (≈ 100,000 words) at one time. That might make automatic glossary-preparation more practical. I’ve applied for access but haven’t received it yet.

Tom Gally