Idea: Use embeddings / similarity search to provide passive review options

General
1

Overall I like the new /practice feature. The multiple-choice responses are a bit too easy for passive review, though. I wonder if you could use the OpenAI embeddings API to generate embeddings for every answer, then when selecting multiple choice responses, choose responses based on similarity search results. This would make the tests a bit more challenging an realistic. It might be possible to use not only multiple choice responses from the current article / current user account, but perhaps response choices from any public article, thereby giving a more diverse set of possible responses.

Not sure how well this idea would work in practice, but I thought I’d share it.

1 Like
2

Thanks for the suggestion. Someone else suggested something similar, I’ll consider it.

One concern I have is that it could lead to clashes where it finds words that are so close that one word’s translation could be a legitimate translation of another word thus leading to the user being graded as “incorrect” when they what they selected was actually correct which would be very frustrating.

2 Likes