<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Dense Retrieval | Web IR NLP Group @ NUS</title><link>https://www.comp.nus.edu.sg/~kanmy/wing/tag/dense-retrieval/</link><atom:link href="https://www.comp.nus.edu.sg/~kanmy/wing/tag/dense-retrieval/index.xml" rel="self" type="application/rss+xml"/><description>Dense Retrieval</description><generator>Hugo Blox Builder (https://hugoblox.com)</generator><language>en-us</language><lastBuildDate>Wed, 01 Jul 2026 00:00:00 +0000</lastBuildDate><image><url>https://www.comp.nus.edu.sg/~kanmy/wing/media/icon_hu_ce731d16d7c8719.png</url><title>Dense Retrieval</title><link>https://www.comp.nus.edu.sg/~kanmy/wing/tag/dense-retrieval/</link></image><item><title>Code-Mixed and Multilingual Information Retrieval</title><link>https://www.comp.nus.edu.sg/~kanmy/wing/project/code-mixed-retrieval/</link><pubDate>Wed, 01 Jul 2026 00:00:00 +0000</pubDate><guid>https://www.comp.nus.edu.sg/~kanmy/wing/project/code-mixed-retrieval/</guid><description>&lt;p&gt;Multilingual users often mix languages in real search queries, but dense retrievers are usually evaluated on monolingual ones. This project studies retrieval with code-mixed and mixed-language queries: what happens when the query representation itself is mixed, and when combining languages in a query actually helps retrieval.&lt;/p&gt;
&lt;p&gt;In &lt;a href="https://www.comp.nus.edu.sg/~kanmy/wing/publication/zhu-etal-2026-mixing/"&gt;When Does Mixing Help? Analyzing Query Embedding Interpolation in Multilingual Dense Retrieval&lt;/a&gt; (ACL 2026, oral), we run a ratio-controlled study of mixed-language querying on mMARCO. Given two parallel query translations, we encode both, interpolate their embeddings at a controlled mixing ratio, and retrieve with the normalized mixture. Across 35 language pairs, each with three document-language settings, the best mixed query embedding from BGE-M3 beats the better monolingual endpoint in 88 of 105 settings. The gains are asymmetric and driven by English dominance: mixing helps for non-English document indices, while indices containing English are best served by pure English queries. After controlling for English dominance, larger typological distance between the two languages is associated with smaller mixing gains. The &lt;a href="https://github.com/tongyao-zhu/query-embedding-mix" target="_blank" rel="noopener"&gt;query-embedding-mix repository&lt;/a&gt; provides the code and result tables, and the word-mixed query sets are released on &lt;a href="https://huggingface.co/datasets/hcm777/query-embedding-mix-word-mix" target="_blank" rel="noopener"&gt;Hugging Face&lt;/a&gt;.&lt;/p&gt;</description></item></channel></rss>