alessandrobenedetti commented on PR #12314: URL: https://github.com/apache/lucene/pull/12314#issuecomment-1568397130
> > SUM = the similarity score between the query and each vector is computed, all scores are summed to get the final score > > SUM = every time we find a nearest neighbor vector to be added to the topK, if the document is already there, its score is updated summing the old and new score > > Just a note on the aggregation functions `max` and `sum`. Most commonly it seems that `max` is used as it is length independent. When using `sum`, the longer the original text of a document field, and thus the more passages it will have, the higher the `sum` of all matching passages will be since all passages will "match", thus biasing scoring towards documents with longer text. I'm not sure if it will matter in the end, but my suggestion would be that if `sum` is used, one could optionally use a radius/similarity threshold to limit the advantage of longer texts, and/or allow using just a limited top-k passages of a document for `sum`. > > @alessandrobenedetti Do you have any good references/papers on approaches to re-aggregating passages into documents for SERPs? It seems that the art was abandoned a couple years ago with most approaches settling on `max` passage (which I see is the only method implemented for now). Hi @joshdevins , the dual strategy(MAX/SUM) is implemented in an old commit, I removed it to build an initial smaller and cleaner pull request. Some of the feedback was to introduce strategies later on, and that's fine, I agree with that. I didn't have the time to dive deeper into the aggregation strategies so I don't have references yet. My main focus was to reach a working prototype and then iterate on the various components to make them better/deal. Your observation regarding 'SUM' is correct. In my 'naive' first implementation I used an approach where only the closer vectors you encounter are considered in the SUM when running an approximate search. You can take a look at the commits before the simplification if you are curious, but I believe it would be better to address this discussion when we introduce 'Strategies' again in a separate future PR. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
