It is suitable for text matching tasks?

[xuanzebi]

In the text matching task, colbert and sentence-bert sbert are both representation-based models. I would like to ask how the comparison of colbert and sentence-bert compares to effect, will colbert be better?

[okhat]

Hi @xuanzebi! Good question. Yes, the fine-grained interaction of ColBERT is designed to be stronger than single-vector models like SBERT, DPR, etc. When you say text matching, as you referring to retrieval tasks (e.g., MS MARCO Ranking, Natural Questions)? For this, check our ColBERT and ColBERT-QA papers. Or do you have semantic similarity tasks (e.g., STS) in mind? We haven't tested on those, but it would certainly be a cool experiment.

[xuanzebi]

Hi @xuanzebi! Good question. Yes, the fine-grained interaction of ColBERT is designed to be stronger than single-vector models like SBERT, DPR, etc. When you say text matching, as you referring to retrieval tasks (e.g., MS MARCO Ranking, Natural Questions)? For this, check our ColBERT and ColBERT-QA papers. Or do you have semantic similarity tasks (e.g., STS) in mind? We haven't tested on those, but it would certainly be a cool experiment.

Thank you very much for such a quick reply. The text matching I mean here refers to the similarity of two texts, such as judging whether two texts are on the same subject. I will also conduct experimental comparisons afterwards. thanks~

[okhat]

Awesome!

Do you have specific benchmarks in mind? I might be able to help you set things up if you like.

[okhat]

Closing for lack of activity. Please feel free to re-open if you have any other questions!