Hi Everyone,
I have a documents search engine, and the users have ability to Rate the search result for any query they make. For the first versions of Search engine, I am using Universal Sentence Encoder to generate document embeddings, and at the time of search User search queries are also embedded and the documents with most closest embeddings are presented in search result.
User can rate a documents from the search result, on some scale, say 0 to 5 (0 being Not Relevant and 5 is Very relevant)
Using this kind of feedback is there a way we can fine tune the search results?
One idea is using BERT with triplet loss, where we can use:
Anchor : User Search Query
Contradiction : Document which User found Not Relevant
Entailment: Document user found very relevant
Anybody experience in doing this? or any other ideas, suggestions , papers are welcome.
If the cheating thing is under control you can use several proposed methods to include relevance feedback like:
Conference Paper Applying BERT to Document Retrieval with Birch
Article Evaluating Sentence-Level Relevance Feedback for High-Recall...
Pratik Ramchandra Erande ,
I would be careful in any implementation of relevance feedback unless your user profile has no incentive to cheat the engine. If you have a considerable percentage of users that have no incentive on cheating, maybe you can filter those out, but if it is a general search engine, it will be difficult to establish any user profile.
Thanks Arturo Geigel , I have control over choosing the feedbacks I want to use based on some User Reputation, in a nutshell, Lets assume No Cheating.
If the cheating thing is under control you can use several proposed methods to include relevance feedback like:
Conference Paper Applying BERT to Document Retrieval with Birch
Article Evaluating Sentence-Level Relevance Feedback for High-Recall...
You ask if you have any ideas. I answer-there are ideas
Eugene Veniaminovich Lutsenko haha, can you provide those ideas here in answer section please. :)
You ask me if I could share these ideas here on RG? Of course you could!
I have control over choosing the feedbacks I want to use based on some User Reputation .
- Badges
- Science topic