@inproceedings{ng-2026-escaping,
title = "Escaping the Tokenizer Trap: Probabilistic Semantic Routing for Dense Retrieval in Specialized Domains",
author = "Ng, Siu Hin",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.16/",
pages = "320--332",
ISBN = "979-8-89176-439-2",
abstract = "Dense retrievers can blur the distinctions between short, jargon-heavy records that share most of their vocabulary. We study Lex-TM, a training-free layer that interpolates dense cosine similarity with the Jensen{--}Shannon similarity of topic distributions from an LDA model whose topic{--}word prior is scaled by a Herfindahl{--}Hirschman (HHI) exclusivity score. On NTCIR-13 MedWeb, cast as retrieving tweets that share a query{'}s exact symptom-label set, Lex-TM reaches an English MRR of 0.8151 against 0.7081 for BM25+dense reciprocal rank fusion, a significant paired difference; in Japanese and Chinese the differences from fusion are not significant. With Qwen3-Embedding-0.6B, routing improves on dense retrieval significantly in English and Chinese. Three findings bound these gains. SPLADE++ and BGE-M3{'}s multi-vector mode match or exceed Lex-TM. In English and Chinese the gain depends on the label strings in the indexed tweets: without them, routing no longer improves on dense retrieval. The HHI prior brings no benefit: on short texts it becomes a rarity weight, and a symmetric prior performs as well across five Gibbs seeds. Outside MedWeb, routing matches dense retrieval on Chinese medical QA and lowers nDCG@10 on ChatDoctor, FiQA and long biomedical abstracts. Code, seeds and a changelog of fixed bugs are released."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ng-2026-escaping">
<titleInfo>
<title>Escaping the Tokenizer Trap: Probabilistic Semantic Routing for Dense Retrieval in Specialized Domains</title>
</titleInfo>
<name type="personal">
<namePart type="given">Siu</namePart>
<namePart type="given">Hin</namePart>
<namePart type="family">Ng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-439-2</identifier>
</relatedItem>
<abstract>Dense retrievers can blur the distinctions between short, jargon-heavy records that share most of their vocabulary. We study Lex-TM, a training-free layer that interpolates dense cosine similarity with the Jensen–Shannon similarity of topic distributions from an LDA model whose topic–word prior is scaled by a Herfindahl–Hirschman (HHI) exclusivity score. On NTCIR-13 MedWeb, cast as retrieving tweets that share a query’s exact symptom-label set, Lex-TM reaches an English MRR of 0.8151 against 0.7081 for BM25+dense reciprocal rank fusion, a significant paired difference; in Japanese and Chinese the differences from fusion are not significant. With Qwen3-Embedding-0.6B, routing improves on dense retrieval significantly in English and Chinese. Three findings bound these gains. SPLADE++ and BGE-M3’s multi-vector mode match or exceed Lex-TM. In English and Chinese the gain depends on the label strings in the indexed tweets: without them, routing no longer improves on dense retrieval. The HHI prior brings no benefit: on short texts it becomes a rarity weight, and a symmetric prior performs as well across five Gibbs seeds. Outside MedWeb, routing matches dense retrieval on Chinese medical QA and lowers nDCG@10 on ChatDoctor, FiQA and long biomedical abstracts. Code, seeds and a changelog of fixed bugs are released.</abstract>
<identifier type="citekey">ng-2026-escaping</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.16/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>320</start>
<end>332</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Escaping the Tokenizer Trap: Probabilistic Semantic Routing for Dense Retrieval in Specialized Domains
%A Ng, Siu Hin
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F ng-2026-escaping
%X Dense retrievers can blur the distinctions between short, jargon-heavy records that share most of their vocabulary. We study Lex-TM, a training-free layer that interpolates dense cosine similarity with the Jensen–Shannon similarity of topic distributions from an LDA model whose topic–word prior is scaled by a Herfindahl–Hirschman (HHI) exclusivity score. On NTCIR-13 MedWeb, cast as retrieving tweets that share a query’s exact symptom-label set, Lex-TM reaches an English MRR of 0.8151 against 0.7081 for BM25+dense reciprocal rank fusion, a significant paired difference; in Japanese and Chinese the differences from fusion are not significant. With Qwen3-Embedding-0.6B, routing improves on dense retrieval significantly in English and Chinese. Three findings bound these gains. SPLADE++ and BGE-M3’s multi-vector mode match or exceed Lex-TM. In English and Chinese the gain depends on the label strings in the indexed tweets: without them, routing no longer improves on dense retrieval. The HHI prior brings no benefit: on short texts it becomes a rarity weight, and a symmetric prior performs as well across five Gibbs seeds. Outside MedWeb, routing matches dense retrieval on Chinese medical QA and lowers nDCG@10 on ChatDoctor, FiQA and long biomedical abstracts. Code, seeds and a changelog of fixed bugs are released.
%U https://aclanthology.org/2026.aacl-long.16/
%P 320-332
Markdown (Informal)
[Escaping the Tokenizer Trap: Probabilistic Semantic Routing for Dense Retrieval in Specialized Domains](https://aclanthology.org/2026.aacl-long.16/) (Ng, AACL-IJCNLP 2026)
ACL