copyedit: Using Cross-Encoders as reranker

[dandv]

Type of change:

• Documentation updates (non-breaking change which updates documents)

https://weaviate.io/blog/2022/08/Using-Cross-Encoders-as-reranker-in-multistage-vector-search.html

Higher-level feedback on the article

• would help to explain on the first use why "Bi-Encoder" and "Cross-Encoder" models are named that way
• would it help to explain "pooling" from the Bi-Encoder diagram?
• the article refers to a "Multistage search pipeline", but only two stages are listed. Would "two-stage" be more accurate, or can other stages be added, so the general term is preferred?

CC @laura-ham

[sebawita]

hey @laura-ham do you have any ideas for Dan's questions?

[laura-ham]

Thanks for the suggestions and improvements @dandv!

Here are my thoughts about your additional comments:

1. would help to explain on the first use why "Bi-Encoder" and "Cross-Encoder" models are named that way

Sure! I don't know if it adds value, but if you do think so then let's add it :)

2. would it help to explain "pooling" from the Bi-Encoder diagram?

Yeah good point, let's add a line or 2 about that.

3. the article refers to a "Multistage search pipeline", but only two stages are listed. Would "two-stage" be more accurate, or can other stages be added, so the general term is preferred?

Hmm I'll keep it to "Multistage", exactly for that reason that the pipeline can (and will) be longer in the future. So for consistency I would use "Multistage".

Regarding 1 and 2, do you feel comfortable adding that or do you want me to do that?

[dandv]

Thanks for the quick reply!

I was asking about "bi-encoders" because the term is unintuitive to me, and the SBERT article on Bi-Encoder vs. Cross-Encoder doesn't explain it either. From what I've asked around, it doesn't refer to transformer encoders placing attention both before and after the token, either. Does "bi" mean anything here (perhaps the fact that for scoring sentence similarity, bi-encoders work on pairs of inputs?), or are the names "bi-encoder" and "cross-encoder" just not that descriptive?

Regarding 1 and 2, do you feel comfortable adding that or do you want me to do that?

Given my insufficient knowledge about the topic at the moment, I'd rather you make those additions :)

[databyjp]

Hi @dandv - I am merging the copyedit and opened a new issue #316 to capture your feedback & discussion.