Xilun Chen

@jobergum Thanks for the prompt response! A couple of quick clarification questions regarding the MS MARCO latencies. In the blog post, I see the following table regarding the latency: It...

Thank you for the detailed explanation, @jobergum! It would be great if you plan to re-run the experiments on ODQA with real latency numbers for the hybrid model! It would...

Yes, I'm primarily interested in the the retrieval stage of ODQA.

And since you mentioned the number of dimensions, I wonder if you have any insights in how the latency of ANN search scales with the number of dimensions? For instance,...

Thank you for the prompt response. I see that this `compute_query_document_score` method was implemented in a way that adds "an additional filter clause that restricts consideration to only the docid...

I have zero experience with Anserini. Could you kindly provide some pointers on how to add query constraints to restrict the search to multiple docids? https://github.com/castorini/anserini/blob/bafa0edd552e295fed22e2a947bdcbb08feed961/src/main/java/io/anserini/index/IndexReaderUtils.java#L654 I don't expect duplicating...

@MXueguang Thanks for the suggestion. This sounds very promising! I'm using the pre-trained "wikipedia-dpr" index provided by Pyserini, but I ran into this exception: ``` JavaException: JVM exception occurred: Document...

@MXueguang @lintool Thank you both for your help!

@MXueguang I tried this approach, but it is actually slower than computing individual BM25 scores one by one.. I experimented with 128 queries and 256 documents: Calling `compute_query_document_score` 128*256 times...

Hi @yuanzilin, It's been 5 years since I ran the experiments in this paper, and I don't remember how long exactly it would take. Half an hour does sound a...