Hi Semantic Scholar team. I am using Semantic Scholar for our product. One major problem I encounter is that the paper search results are very often empty. Due to the nature of my product, most of the search queries are longer and have a lot of keywords. I suspect this is he main reason why the empty results are frequent.
I have read your blog post about the search algorithm currently used. Although I have found many problems with this approach (one of which is to train the reranker on fairly limited ranking data without training updates), this shouldn't be the sole reason of empty results. I suspect you are using a very low (or high) threshold for the reranker that eliminates all of the results for certain queries.
So I wonder if any of these features can be added for the API:
Allowing getting the original search results, bypassing the reranker.
Allowing user to adjust the reranker threshold to make it more "lenient".
Thank you so much for your help and the wonderful product.
Hi Semantic Scholar team. I am using Semantic Scholar for our product. One major problem I encounter is that the paper search results are very often empty. Due to the nature of my product, most of the search queries are longer and have a lot of keywords. I suspect this is he main reason why the empty results are frequent. I have read your blog post about the search algorithm currently used. Although I have found many problems with this approach (one of which is to train the reranker on fairly limited ranking data without training updates), this shouldn't be the sole reason of empty results. I suspect you are using a very low (or high) threshold for the reranker that eliminates all of the results for certain queries. So I wonder if any of these features can be added for the API: Allowing getting the original search results, bypassing the reranker. Allowing user to adjust the reranker threshold to make it more "lenient". Thank you so much for your help and the wonderful product.