Reputation: 111
The spelling and grammar corrections that appear to take place within the ChatGPT system are pretty great... but it got me wondering if language model based question answering systems like these leave it entirely to the language model to perform the grammar and spelling fixes? There are many other more performant approaches to spelling and grammar correction, so I thought it might be possible that question answering systems like these first run queries through a lighter system that just does the spelling and grammar. For those who know more of the system internals of GPT and other QA systems, if there is a preliminary query parsing step at work, would this step use bottom up parsing, top down, statistical or something else entirely?
Upvotes: -2
Views: 73
Reputation: 21
I also think similar to your thought. Not also for the grammar or spelling checks, such LLMs are very heavy for small tasks even they perform really well. You could still find many researches using old fashioned but giving amazing performance with those small tasks without using large language models.
Upvotes: -1