KSE-Web: An Analysis of Hybrid Retrieval and LLM-Assisted Query Expansion for Low-Resource Khmer Semantic Search

KSE-Web: An Analysis of Hybrid Retrieval and
LLM-Assisted Query Expansion for Low-Resource
Khmer Semantic Search
Nimol Thuon
No Institute Given
Abstract. As a low-resource language, Khmer presents several retrieval
challenges, including limited annotated data, ambiguous word bound-
aries, weak support in multilingual embedding models, and frequent
mixed Khmer–English usage. This paper presents KSE-Web, an anal-
ysis of hybrid retrieval and LLM-assisted query expansion for Khmer
semantic search. We construct the dataset from approximately 17K can-
didate Khmer titles and retain 3K cleaned full-text Khmer documents
after filtering, normalization, deduplication, and document-length con-
trol. The dataset includes 300 manually reviewed user-style Khmer search
queries and silver relevance labels with partial human verification. We
evaluate character n-gram BM25, multilingual dense retrieval, hybrid
BM25+dense retrieval, and LLM-assisted query expansion using Qwen2.5
models. Experimental results show that BM25 achieves the strongest
overall performance, reaching 0.943 Recall@10 and 0.876 nDCG@10.
Hybrid BM25+dense retrieval performs comparably, achieving 0.929 Re-
call@10 and 0.871 nDCG@10, while dense retrieval alone performs lower.
LLM-assisted query expansion does not outperform non-expanded re-
trieval; however, Qwen2.5-3B produces substantially stronger expanded-
query results than Qwen2.5-0.5B, suggesting that LLM size and expan-
sion quality matter for low-resource Khmer retrieval. Our analysis fur-
ther shows that direct LLM expansion can introduce topic drift, generic
terms, and noisy reformulations, while simple filtering may remove useful
semantic cues. These findings highlight both the potential and limita-
tions of LLM-assisted retrieval for Khmer semantic search and provide
a foundation for future Khmer retrieval datasets with stronger human-
verified annotations and Khmer-aware retrieval models. The dataset and
documentation will be made available at

mantic cues. These findings highlight both the potential and limita-
tions of LLM-assisted retrieval for Khmer semantic search and provide
a foundation for future Khmer retrieval datasets with stronger human-
verified annotations and Khmer-aware retrieval models. The dataset and
documentation will be made available at github.com/back-kh/Khmer-
Semantic-Search.
Keywords: Khmer semantic search, Low-resource language retrieval,
Khmer web retrieval, Hybrid retrieval, LLM-assisted query expansion
1 Introduction
Large language models (LLMs) have recently achieved strong performance across
natural language processing, computer vision, and document understanding tasks.
arXiv:2608.21365v1 [cs.CL] 17 Jun 2026

-- 1 of 17 --

2 Nimol Thuon
Fig. 1. Overview of the KSE-Web retrieval framework. The offline stage constructs
sparse and dense indexes from cleaned Khmer web documents. At inference time, a
Khmer query is normalized and optionally expanded using an LLM. The original or ex-
panded query is then used for sparse BM25 retrieval, dense retrieval with Multilingual-
E5, and hybrid score fusion to produce top-k ranked Khmer documents.
Their ability to generate paraphrases, reformulate questions, and provide con-
textual expansions has motivated their use in retrieval-oriented applications,
including semantic search and retrieval-augmented systems [22, 21, 5]. However,
the effectiveness of LLM-assisted retrieval remains uneven across languages. In
particular, low-resource languages often lack large-scale retrieval datasets, high-
quality relevance judgments, robust tokenization tools, and strong pretrained
representations. As a result, methods that perform well for English or other high-
resource languages may not directly transfer to languages with limited digital
resources [1, 12, 13]. Khmer is one such language. Although it is widely used in
Cambodia, Khmer remains underexplored in information retrieval and semantic
search research in the modern era. Khmer web retrieval

l for English or other high-
resource languages may not directly transfer to languages with limited digital
resources [1, 12, 13]. Khmer is one such language. Although it is widely used in
Cambodia, Khmer remains underexplored in information retrieval and semantic
search research in the modern era. Khmer web retrieval presents several practical
challenges [18, 17]. First, existing Khmer semantic search work has mainly fo-
cused on domain-specific retrieval and conventional preprocessing, while limited
attention has been given to hybrid retrieval strategies or LLM-assisted semantic
search models [16, 19]. Second, Khmer does not use whitespace as a reliable word
boundary, making standard word-level tokenization and retrieval less straightfor-
ward. Third, real-world Khmer web queries often include informal expressions,
spelling variations, named entities, and mixed Khmer–English usage. Finally,
multilingual dense retrieval models may not fully capture Khmer semantic sim-
ilarity because Khmer is underrepresented in many multilingual training and

-- 2 of 17 --

Khmer Semantic Search: Hybrid Retrieval and LLM-Assisted Query 3
evaluation resources. These challenges make Khmer a valuable case for studying
retrieval in low-resource language settings [3, 9].
This paper presents KSE-Web, an analysis of hybrid retrieval and LLM-
assisted query expansion for low-resource Khmer semantic search using a Khmer
web retrieval dataset. The overall framework is shown in Fig. 1. KSE-Web sepa-
rates offline corpus processing from online query processing. In the offline stage,
cleaned Khmer web documents are used to construct both sparse and dense in-
dexes. In the online stage, a user query is normalized and optionally expanded
using an LLM before being passed to sparse retrieval, dense retrieval, and hybrid
score fusion.
To support this analysis, we construct a Khmer web retrieval dataset from
approximately 17K candidate Khmer web titles and retain 3K cleaned full-text
Khmer

e online stage, a user query is normalized and optionally expanded
using an LLM before being passed to sparse retrieval, dense retrieval, and hybrid
score fusion.
To support this analysis, we construct a Khmer web retrieval dataset from
approximately 17K candidate Khmer web titles and retain 3K cleaned full-text
Khmer documents after filtering, normalization, deduplication, and document-
length control. The dataset construction pipeline is summarized in Fig. 2. The
dataset includes 300 manually reviewed user-style Khmer queries and 5,412
query–document relevance labels generated through silver labeling with par-
tial human verification. The queries cover multiple search styles, including short
keyword queries, question-style queries, informal user-style queries, and mixed
Khmer–English queries.
We evaluate character n-gram BM25, multilingual dense retrieval, hybrid
BM25+dense retrieval, and LLM-assisted query expansion using Qwen2.5 mod-
els. Our results show that character n-gram BM25 remains the strongest overall
method, achieving 0.943 Recall@10 and 0.876 nDCG@10. Hybrid retrieval per-
forms comparably, reaching 0.929 Recall@10 and 0.871 nDCG@10, while dense
retrieval alone performs substantially lower. The main performance trends are
summarized in Fig. 3. LLM-assisted query expansion does not outperform non-
expanded retrieval, but Qwen2.5-3B produces substantially stronger expanded-
query results than Qwen2.5-0.5B. Qualitative examples in Fig. 4 further show
that LLM expansion may introduce topic drift, generic terms, and noisy refor-
mulations, while simple filtering can remove useful semantic cues together with
noise.
The main contributions of this work are as follows:
– We construct a Khmer web retrieval dataset containing 3K cleaned full-text
Khmer web documents, 300 manually reviewed user-style queries, and 5,412
silver relevance labels with partial human verification.
– We present KSE-Web, a unified analysis framework for low-resource Khmer
semantic

of this work are as follows:
– We construct a Khmer web retrieval dataset containing 3K cleaned full-text
Khmer web documents, 300 manually reviewed user-style queries, and 5,412
silver relevance labels with partial human verification.
– We present KSE-Web, a unified analysis framework for low-resource Khmer
semantic search covering sparse, dense, hybrid, and LLM-assisted retrieval
settings.
– We show that character n-gram BM25 is a strong and necessary baseline for
Khmer web retrieval, while hybrid BM25+dense retrieval performs closely
but does not surpass BM25.
– We analyze LLM-assisted query expansion using Qwen2.5 models and show
that larger LLMs produce more useful expansions than smaller LLMs, but
direct query expansion still remains below non-expanded retrieval.

-- 3 of 17 --

4 Nimol Thuon
– We provide qualitative analysis of LLM expansion behavior, including helpful
expansion, topic drift, over-expansion, generic-term insertion, and filtering
effects.
2 Related Work
2.1 Low-Resource and Multilingual Information Retrieval
Information retrieval for low-resource languages remains challenging because
many languages lack large-scale corpora, standardized query sets, relevance judg-
ments, and language-specific preprocessing tools. Recent benchmarks such as
BEIR, Mr. TyDi, mMARCO, and MIRACL have advanced retrieval evaluation
across diverse domains and languages [15, 23, 2, 24]. However, Khmer remains
much less represented in retrieval research compared with high-resource and
better-studied multilingual settings. This motivates the need for Khmer-specific
retrieval datasets and evaluation protocols.
Khmer retrieval presents both linguistic and practical challenges [3, 16]. Khmer
does not use whitespace as a reliable word boundary, making word-level tokeniza-
tion difficult. In addition, real-world Khmer web content often includes mixed
Khmer–English terms, transliterated names, named entities, informal expres-
sions, and spelling variations. These issues

nd practical challenges [3, 16]. Khmer
does not use whitespace as a reliable word boundary, making word-level tokeniza-
tion difficult. In addition, real-world Khmer web content often includes mixed
Khmer–English terms, transliterated names, named entities, informal expres-
sions, and spelling variations. These issues affect both sparse retrieval methods,
which depend on lexical overlap, and dense retrieval methods, which depend
on learned semantic representations [11, 10]. Recent hybrid and LLM-assisted
retrieval methods offer promising directions for low-resource retrieval [21, 6, 8,
14]. In this work, we focus on Khmer web retrieval and construct a dataset for
controlled analysis of sparse, dense, hybrid, and LLM-assisted retrieval methods.
2.2 Sparse, Dense, and Hybrid Retrieval
Sparse lexical retrieval remains a strong baseline in information retrieval. BM25
is widely used because it is efficient, interpretable, and effective for ranking
documents using lexical evidence [11]. For languages such as Khmer, however,
word-level BM25 is affected by tokenization difficulty. Character n-gram BM25
offers a practical alternative because it does not require explicit word segmen-
tation and can capture partial lexical overlap between queries and documents.
This makes character-level sparse retrieval suitable for Khmer web search.
Dense retrieval represents queries and documents as continuous vectors and
ranks documents by vector similarity. Dense Passage Retrieval (DPR) showed the
effectiveness of dual-encoder retrieval for open-domain question answering [10].
More recent embedding models, such as E5, provide general-purpose text rep-
resentations for retrieval, clustering, and classification [20]. Multilingual embed-
ding models are useful for low-resource languages because they can be applied
without training a language-specific retriever. However, their performance de-
pends on how well the target language is represented during training. For Khmer,
off-the-shelf

retrieval, clustering, and classification [20]. Multilingual embed-
ding models are useful for low-resource languages because they can be applied
without training a language-specific retriever. However, their performance de-
pends on how well the target language is represented during training. For Khmer,
off-the-shelf multilingual embeddings may still struggle with semantic similarity,
named entities, and mixed-script usage.

-- 4 of 17 --

Khmer Semantic Search: Hybrid Retrieval and LLM-Assisted Query 5
Hybrid retrieval combines sparse and dense retrieval signals. Sparse retrieval
is effective for exact lexical matching, named entities, and rare terms, while dense
retrieval can capture paraphrases and semantically related expressions. In this
paper, we use a simple score-fusion strategy between character n-gram BM25
and multilingual-E5 dense retrieval. This allows us to examine whether dense
retrieval adds useful semantic signals beyond character-level lexical matching
for Khmer search.
2.3 LLM-Assisted Query Expansion and Relevance Labeling
Query expansion improves retrieval by reformulating a query or adding related
terms to better match relevant documents. Traditional methods include rel-
evance feedback, pseudo-relevance feedback, and term selection [4, 5]. Recent
studies have also explored LLM-based query expansion. Query2doc generates
pseudo-documents to expand queries for sparse and dense retrieval [21], while
HyDE generates hypothetical documents for zero-shot dense retrieval without
relevance labels [7]. These methods suggest that generated text can provide use-
ful retrieval context, especially for short or ambiguous queries.
However, LLM-assisted expansion can introduce errors. Generated text may
broaden the original intent, add generic terms, or include related but non-
discriminative words. These risks are more important for low-resource languages,
where LLMs may have weaker language-specific knowledge and generated terms
may not match the target

LM-assisted expansion can introduce errors. Generated text may
broaden the original intent, add generic terms, or include related but non-
discriminative words. These risks are more important for low-resource languages,
where LLMs may have weaker language-specific knowledge and generated terms
may not match the target collection [12, 13]. In this work, we analyze Qwen2.5-
based query expansion for Khmer retrieval [22] by comparing Qwen2.5-0.5B and
Qwen2.5-3B. The larger model produces more useful expansions than the smaller
model, but direct expansion still does not outperform strong non-expanded re-
trieval baselines.
Reliable relevance labels are necessary for retrieval evaluation, but full man-
ual annotation is costly, especially for low-resource languages. Many retrieval
datasets therefore use pooling, where candidate documents are retrieved by one
or more systems and then judged for relevance. In resource-constrained settings,
silver labels provide a practical starting point, although they may introduce bias
toward the systems used for pooling. In KSE-Web, we construct silver relevance
labels using source-document matching, BM25-based candidate retrieval, rule-
based relevance assignment, and partial human verification. Therefore, we view
the experiments as a controlled analysis of Khmer web retrieval behavior rather
than a final gold-standard benchmark.
3 KSE-Web Dataset Construction
This section describes the construction of KSE-Web, a Khmer web retrieval
dataset designed for controlled analysis of sparse, dense, hybrid, and LLM-
assisted retrieval methods. Since publicly available Khmer retrieval datasets with
user-style queries and relevance labels are limited, we construct a new resource
consisting of cleaned Khmer web documents, manually reviewed queries, and

-- 5 of 17 --

6 Nimol Thuon
Fig. 2. Construction pipeline of the KSE-Web retrieval dataset. The dataset is built
from approximately 17K candidate Khmer web titles, filtered and cleaned into 3K

ce labels are limited, we construct a new resource
consisting of cleaned Khmer web documents, manually reviewed queries, and

-- 5 of 17 --

6 Nimol Thuon
Fig. 2. Construction pipeline of the KSE-Web retrieval dataset. The dataset is built
from approximately 17K candidate Khmer web titles, filtered and cleaned into 3K full-
text Khmer web documents, and paired with 300 manually reviewed user-style queries.
Candidate retrieval, source-document matching, silver relevance labeling, and partial
human verification produce 5,412 query–document relevance labels.
silver relevance labels with partial human verification. The overall pipeline is
shown in Fig. 2.
3.1 Document Collection, Categories, and Preprocessing
We construct the document collection from publicly available Khmer web con-
tent. Starting from approximately 17K candidate Khmer web titles, we collect
the corresponding pages and process them into a clean full-text corpus. Noisy,
duplicated, very short, and non-document-like pages are removed through title
filtering, Khmer-content filtering, duplicate removal, and basic quality control.
The extracted text is further cleaned using HTML removal, Unicode normaliza-
tion, whitespace normalization, and noise filtering. Long documents are trun-
cated after cleaning to keep document-level retrieval efficient and comparable.
After filtering and cleaning, the final KSE-Web corpus contains 3K full-text
Khmer web documents. Each document includes a document identifier, title,
cleaned body text, source URL, category metadata, and document-length in-
formation. Each document is assigned to one of four broad categories: public
service, education, tourism and culture, and general news/information. These
categories reflect common Khmer web search needs and support domain-level
retrieval analysis. Category assignment is based on rule-based keywords over ti-
tles and cleaned text, followed by partial manual inspection to reduce obvious
errors.

-- 6 of 17 --

Khmer Semantic

d culture, and general news/information. These
categories reflect common Khmer web search needs and support domain-level
retrieval analysis. Category assignment is based on rule-based keywords over ti-
tles and cleaned text, followed by partial manual inspection to reduce obvious
errors.

-- 6 of 17 --

Khmer Semantic Search: Hybrid Retrieval and LLM-Assisted Query 7
Table 1. Statistics of the KSE-Web dataset.
Component Count
Candidate Khmer web titles ∼17K
Khmer web documents 3000
Document categories 4
User-style queries 300
Labeled query-document pairs 5412
Relevance scale 0, 1, 2
3.2 Query Construction and Silver Relevance Labeling
We create 300 user-style Khmer search queries for document-level retrieval evalu-
ation. Candidate queries are generated from source document titles using template-
based transformations, allowing each query to be linked to at least one relevant
source document. The query set covers four types: short, question-style, infor-
mal, and mixed Khmer–English. To improve quality, all queries are manually
reviewed by a Khmer speaker to remove duplicated, unnatural, overly broad, or
noisy template-generated queries.
Fully manual relevance annotation is costly in low-resource settings, so KSE-
Web uses silver relevance labels with partial human verification. For each query,
we retrieve the top-20 candidate documents using character n-gram BM25. The
source document is added to the candidate pool when it is not retrieved in
the top-20 results. Each query–document pair is assigned a three-level relevance
label: 2 for highly relevant, 1 for partially relevant, and 0 for non-relevant. Label
assignment uses source-document matching, title-query overlap, BM25 ranking,
category agreement, rule-based scoring, and partial human inspection.
3.3 Dataset Statistics and Scope
Table 1 summarizes the main statistics of KSE-Web. The dataset contains 3K
cleaned Khmer web documents, 300 manually reviewed queries, and 5,412 labeled
query–document pairs. The

g, title-query overlap, BM25 ranking,
category agreement, rule-based scoring, and partial human inspection.
3.3 Dataset Statistics and Scope
Table 1 summarizes the main statistics of KSE-Web. The dataset contains 3K
cleaned Khmer web documents, 300 manually reviewed queries, and 5,412 labeled
query–document pairs. The three-level relevance scale supports both binary rel-
evance evaluation and graded ranking metrics such as nDCG.
KSE-Web is intended as a practical Khmer web retrieval dataset rather than
a final large-scale gold-standard benchmark. The current version focuses on web-
extracted Khmer text and does not include PDFs, scanned pages, or OCR-based
documents. In addition, some queries are derived from document titles, and the
silver labels partly rely on BM25-based candidate retrieval, which may introduce
lexical bias. These limitations are considered in the experimental analysis and
motivate future work on larger human-verified Khmer retrieval datasets.
4 Overall Retrieval Framework
This section describes the overall retrieval framework used to evaluate low-
resource Khmer semantic search in KSE-Web. Given a Khmer user query, the

-- 7 of 17 --

8 Nimol Thuon
goal is to retrieve a ranked list of relevant documents from the KSE-Web doc-
ument collection. As shown in Fig. 1, the framework contains two main stages:
an offline indexing stage and an online retrieval stage. The offline stage prepares
sparse and dense indexes from cleaned Khmer web documents, while the online
stage processes user queries through optional LLM-assisted expansion, sparse
retrieval, dense retrieval, and hybrid score fusion.
4.1 Task Definition
Let D = {d1, d2, . . . , dN } denote the KSE-Web document collection, where each
document di contains a title and cleaned Khmer web text. Given a user query
q, the retrieval task is to produce a ranked list of documents:
R(q) = [d(1), d(2), . . . , d(k)],
where documents appearing earlier in the list are expected to be more relevant
to the

, dN } denote the KSE-Web document collection, where each
document di contains a title and cleaned Khmer web text. Given a user query
q, the retrieval task is to produce a ranked list of documents:
R(q) = [d(1), d(2), . . . , d(k)],
where documents appearing earlier in the list are expected to be more relevant
to the query. The task is evaluated at the document level rather than the passage
level. Each method retrieves documents from the same KSE-Web corpus and is
evaluated using the same silver relevance labels described in Section 3. Relevance
is represented using a three-level scale: 2 for highly relevant, 1 for partially
relevant, and 0 for non-relevant.
4.2 Sparse Retrieval with Character n-gram BM25
As the sparse lexical baseline, we use BM25 over character n-gram representa-
tions. BM25 is a strong retrieval method for lexical matching, but word-level
BM25 requires reliable tokenization. This is difficult for Khmer because whites-
pace does not consistently indicate word boundaries. To avoid reliance on Khmer
word segmentation, both queries and documents are represented using overlap-
ping character n-grams.
For each query and document, we remove unnecessary whitespace and extract
character n-grams with n ∈ {2, 3, 4}. The document representation is created
from the concatenation of the document title and cleaned body text. The sparse
retrieval score is computed as:
SBM25(q, d) = BM25(G(q), G(d)),
where G(·) denotes the character n-gram representation. This design allows the
sparse retriever to capture partial lexical overlap, named entities, spelling vari-
ants, and short Khmer phrases without requiring explicit word segmentation.
4.3 Dense Retrieval with Multilingual Embeddings
For dense retrieval, we use a multilingual sentence embedding model to encode
queries and documents into vector representations. In our experiments, we use
multilingual-E5-small as the dense retriever. Following the E5 input format,

-- 8 of 17 --

Khmer Semantic Search: Hybrid

e Retrieval with Multilingual Embeddings
For dense retrieval, we use a multilingual sentence embedding model to encode
queries and documents into vector representations. In our experiments, we use
multilingual-E5-small as the dense retriever. Following the E5 input format,

-- 8 of 17 --

Khmer Semantic Search: Hybrid Retrieval and LLM-Assisted Query 9
queries are prefixed with query:, while documents are prefixed with passage:.
Each document embedding is computed from the concatenation of its title and
cleaned text.
Let eq denote the embedding of query q, and let ed denote the embedding of
document d. Dense retrieval ranks documents using cosine similarity:
Sdense(q, d) = cos(eq , ed).
Dense retrieval is intended to capture semantic similarity beyond exact lexi-
cal overlap. However, because Khmer is less represented in many multilingual
training resources, the effectiveness of off-the-shelf multilingual embeddings for
Khmer retrieval remains uncertain. This motivates our comparison between
dense retrieval, sparse retrieval, and hybrid retrieval.
4.4 Hybrid Score Fusion
Hybrid retrieval combines sparse and dense retrieval signals. Sparse retrieval
is effective for exact lexical matching, named entities, and rare terms, while
dense retrieval can capture broader semantic relatedness. To combine these two
signals, we first normalize BM25 and dense scores for each query using min–max
normalization:
ˆS(q, d) = S(q, d) − mind′∈D S(q, d′)
maxd′∈D S(q, d′) − mind′∈D S(q, d′) .
The final hybrid score is then computed as:
Shybrid(q, d) = α ˆSBM25(q, d) + (1 − α) ˆSdense(q, d),
where α controls the contribution of sparse retrieval. In this study, we set α =
0.5 to give equal weight to BM25 and dense retrieval. Documents are ranked
according to Shybrid(q, d).
4.5 LLM-Assisted Query Expansion
We analyze whether LLM-assisted query expansion can improve Khmer seman-
tic search. Given an original query q, an instruction-tuned LLM generates an
expanded query q+ by adding related

5 to give equal weight to BM25 and dense retrieval. Documents are ranked
according to Shybrid(q, d).
4.5 LLM-Assisted Query Expansion
We analyze whether LLM-assisted query expansion can improve Khmer seman-
tic search. Given an original query q, an instruction-tuned LLM generates an
expanded query q+ by adding related terms, paraphrases, or commonly used
Khmer–English expressions. In this work, we evaluate Qwen2.5-0.5B-Instruct
and Qwen2.5-3B-Instruct to study the effect of LLM size on expansion quality.
The LLM is prompted to preserve the original query meaning, keep named
entities unchanged, and add only retrieval-useful terms. The expanded query
is then used as input to BM25, dense retrieval, or hybrid retrieval. Formally,
retrieval with an expanded query is computed as:
S(q+, d),
where S may correspond to sparse, dense, or hybrid retrieval.

-- 9 of 17 --

10 Nimol Thuon
In addition to raw LLM expansion, we evaluate a filtered expansion variant.
The filtering step removes repeated terms, overly generic English words, broad
Khmer terms, and excessive expansion tokens. The filtered query is represented
as:
qf = q ⊕ F (q+),
where F (·) denotes the filtering function and ⊕ denotes concatenation with the
original query. This variant tests whether simple filtering can reduce expansion
noise and topic drift.
4.6 Evaluated Retrieval Variants
Using the components above, we evaluate the following retrieval variants:
– BM25-char-ngram: sparse retrieval using Khmer character n-grams.
– Dense: dense retrieval using multilingual-E5-small embeddings.
– Hybrid-BM25-Dense: score fusion of BM25-char-ngram and dense re-
trieval.
– BM25 + LLM Expansion: sparse retrieval using LLM-expanded queries.
– Dense + LLM Expansion: dense retrieval using LLM-expanded queries.
– Hybrid + LLM Expansion: hybrid retrieval using LLM-expanded queries.
– Filtered LLM Expansion: retrieval using filtered expanded queries.
This framework allows us to examine not only which retrieval method per-
forms

val using LLM-expanded queries.
– Dense + LLM Expansion: dense retrieval using LLM-expanded queries.
– Hybrid + LLM Expansion: hybrid retrieval using LLM-expanded queries.
– Filtered LLM Expansion: retrieval using filtered expanded queries.
This framework allows us to examine not only which retrieval method per-
forms best, but also how LLM-assisted query expansion interacts with sparse,
dense, and hybrid retrieval in low-resource Khmer semantic search.
5 Experiments
This section describes the experimental setup used to evaluate sparse, dense,
hybrid, and LLM-assisted retrieval methods on KSE-Web. The goal is not only
to identify the strongest retrieval method, but also to analyze how different
retrieval strategies behave in a low-resource Khmer semantic search setting.
5.1 Dataset and Evaluation Setting
We conduct experiments on the KSE-Web retrieval dataset described in Sec-
tion 3. The dataset contains 3,000 cleaned full-text Khmer web documents and
300 manually reviewed user-style Khmer queries. The queries are balanced across
four broad document categories and cover four query types: short, question-style,
informal, and mixed Khmer–English queries. Each retrieval method searches over
the same document collection and returns the top-20 ranked documents for each
query.
Evaluation is performed using the silver relevance labels described in Sec-
tion 3. Each query–document pair is assigned a three-level relevance label: 2 for
highly relevant, 1 for partially relevant, and 0 for non-relevant. For Recall, Pre-
cision, and MRR, labels 1 and 2 are treated as relevant. For nDCG, the original
graded relevance labels are used.

-- 10 of 17 --

Khmer Semantic Search: Hybrid Retrieval and LLM-Assisted Query 11
5.2 Compared Methods
We compare the following retrieval methods.
BM25-char-ngram. This is the main sparse retrieval baseline. Queries and doc-
uments are represented using Khmer character n-grams with n ∈ {2, 3, 4}, and
documents are ranked using BM25. This method is

Search: Hybrid Retrieval and LLM-Assisted Query 11
5.2 Compared Methods
We compare the following retrieval methods.
BM25-char-ngram. This is the main sparse retrieval baseline. Queries and doc-
uments are represented using Khmer character n-grams with n ∈ {2, 3, 4}, and
documents are ranked using BM25. This method is designed to avoid dependence
on Khmer word segmentation.
Dense retrieval. Dense retrieval uses multilingual-E5-small to encode queries
and documents into vector representations. Queries are prefixed with query:,
and documents are prefixed with passage:. Documents are ranked by cosine
similarity.
Hybrid BM25+dense retrieval. Hybrid retrieval combines normalized BM25 and
dense retrieval scores using linear interpolation with α = 0.5. This setting gives
equal weight to sparse and dense retrieval signals.
LLM-assisted query expansion. We evaluate query expansion using Qwen2.5-
0.5B-Instruct and Qwen2.5-3B-Instruct. Given an original Khmer query, the
LLM generates an expanded query while being instructed to preserve the original
meaning, keep named entities unchanged, and add only retrieval-useful terms.
The expanded queries are then used with BM25, dense retrieval, and hybrid
retrieval.
Filtered LLM expansion. We also evaluate a filtered expansion variant. The
filtering step removes repeated terms, generic English words, broad Khmer terms,
and excessive expansion tokens. This variant tests whether simple filtering can
reduce expansion noise and topic drift.
5.3 Evaluation Metrics
We evaluate retrieval performance using standard top-k metrics at cutoffs 5 and
10, including Recall@k, Precision@k, MRR@k, and nDCG@k. Recall@k mea-
sures how many relevant documents are retrieved, Precision@k measures the
relevance ratio among the top-k results, MRR@k measures the rank of the first
relevant document, and nDCG@k evaluates ranking quality with graded rele-
vance labels. All metrics are averaged over the full query set.
5.4 Implementation Details
All experiments

y relevant documents are retrieved, Precision@k measures the
relevance ratio among the top-k results, MRR@k measures the rank of the first
relevant document, and nDCG@k evaluates ranking quality with graded rele-
vance labels. All metrics are averaged over the full query set.
5.4 Implementation Details
All experiments are implemented in Python. BM25 retrieval uses character n-
gram tokenization over the concatenation of document title and cleaned body
text. Dense retrieval uses multilingual-E5-small with normalized embeddings
and cosine similarity. For hybrid retrieval, BM25 and dense scores are min–
max normalized per query before score fusion. LLM-assisted query expansion

-- 11 of 17 --

12 Nimol Thuon
Fig. 3. Summary of retrieval performance on KSE-Web. BM25 achieves the strongest
overall performance, while hybrid retrieval remains close. LLM-assisted expansion im-
proves with Qwen2.5-3B relative to Qwen2.5-0.5B, but does not surpass non-expanded
BM25 or hybrid retrieval. Full quantitative results are reported in Table 2.
is performed using Qwen2.5-0.5B-Instruct and Qwen2.5-3B-Instruct with deter-
ministic decoding. The maximum number of generated tokens is limited to keep
the expanded queries short and retrieval-focused. For each method, the top-20
documents are retrieved, and metrics are reported at cutoffs 5 and 10.
6 Results and Analysis
This section reports the retrieval results on KSE-Web and analyzes the behavior
of sparse, dense, hybrid, and LLM-assisted retrieval methods. The main retrieval
trends are summarized visually in Fig. 3, while Table 2 reports the full quanti-
tative results.
6.1 Overall Retrieval Performance
Table 2 shows that character n-gram BM25 achieves the strongest overall per-
formance, reaching 0.943 Recall@10 and 0.876 nDCG@10. This confirms that
character-level sparse retrieval is a highly competitive baseline for Khmer web
search. Since Khmer does not reliably mark word boundaries with whitespace,
character n-grams provide a

that character n-gram BM25 achieves the strongest overall per-
formance, reaching 0.943 Recall@10 and 0.876 nDCG@10. This confirms that
character-level sparse retrieval is a highly competitive baseline for Khmer web
search. Since Khmer does not reliably mark word boundaries with whitespace,
character n-grams provide a robust way to capture partial lexical overlap, named
entities, and short phrases without relying on explicit word segmentation.
Hybrid BM25+dense retrieval performs very closely to BM25, reaching 0.929
Recall@10 and 0.871 nDCG@10. This suggests that dense representations pro-
vide some complementary retrieval signal, but they do not surpass the sparse
lexical baseline in the current KSE-Web setting. Dense retrieval alone performs
substantially lower, with 0.563 Recall@10 and 0.523 nDCG@10. This indicates

-- 12 of 17 --

Khmer Semantic Search: Hybrid Retrieval and LLM-Assisted Query 13
Table 2. Retrieval performance on KSE-Web at cutoffs 5 and 10. BM25 denotes char-
acter n-gram BM25, Dense denotes multilingual-E5-small retrieval, and Hybrid denotes
BM25+dense score fusion. Qwen-0.5B and Qwen-3B indicate LLM-assisted query ex-
pansion using Qwen2.5-0.5B and Qwen2.5-3B, respectively. F denotes filtered expan-
sion. Best results are shown in bold.
Method 	R@5 P@5 MRR@5 nDCG@5 R@10 P@10 MRR@10 nDCG@10
BM25 	0.895 0.291 0.905 0.863 0.943 0.166 0.906 0.876
Dense 	0.504 0.151 0.549 0.504 0.563 0.089 0.558 0.523
Hybrid 	0.873 0.279 0.911 0.854 0.929 0.157 0.914 0.871
BM25 + Qwen-0.5B 0.774 0.238 0.794 0.753 0.824 0.133 0.798 0.770
BM25 + Qwen-0.5B-F 0.761 0.229 0.768 0.730 0.810 0.128 0.772 0.747
Dense + Qwen-0.5B 0.312 0.087 0.324 0.299 0.365 0.052 0.332 0.318
Dense + Qwen-0.5B-F 0.319 0.087 0.306 0.291 0.361 0.051 0.312 0.304
Hybrid + Qwen-0.5B 0.732 0.220 0.763 0.710 0.788 0.124 0.767 0.729
Hybrid + Qwen-0.5B-F 0.707 0.208 0.745 0.688 0.767 0.118 0.749 0.709
Dense + Qwen-3B 0.373 0.109 0.381 0.350 0.440 0.066 0.391 0.372
Dense + Qwen-3B-F 0.349 0.100

365 0.052 0.332 0.318
Dense + Qwen-0.5B-F 0.319 0.087 0.306 0.291 0.361 0.051 0.312 0.304
Hybrid + Qwen-0.5B 0.732 0.220 0.763 0.710 0.788 0.124 0.767 0.729
Hybrid + Qwen-0.5B-F 0.707 0.208 0.745 0.688 0.767 0.118 0.749 0.709
Dense + Qwen-3B 0.373 0.109 0.381 0.350 0.440 0.066 0.391 0.372
Dense + Qwen-3B-F 0.349 0.100 0.350 0.325 0.410 0.060 0.360 0.347
Hybrid + Qwen-3B 0.798 0.245 0.806 0.766 0.868 0.143 0.810 0.788
Hybrid + Qwen-3B-F 0.769 0.231 0.773 0.733 0.838 0.134 0.779 0.755
that multilingual-E5-small captures some Khmer semantic similarity, but re-
mains limited compared with character n-gram BM25.
The gap between BM25 and dense retrieval suggests that off-the-shelf multi-
lingual embeddings are not yet sufficient for Khmer semantic retrieval. This may
be due to Khmer’s low-resource status, mixed-script usage, named entities, and
domain-specific vocabulary. At the same time, the strong hybrid performance
shows that dense retrieval is not useless: it provides complementary signals, al-
though the lexical signal remains dominant.
6.2 Impact of LLM-Assisted Query Expansion
LLM-assisted query expansion does not outperform non-expanded retrieval. For
example, BM25 with Qwen2.5-0.5B expansion achieves 0.824 Recall@10 and
0.770 nDCG@10, which is lower than the original BM25 baseline. Similarly, hy-
brid retrieval with Qwen2.5-0.5B expansion reaches 0.788 Recall@10 and 0.729
nDCG@10, below the non-expanded hybrid baseline.
The same pattern holds for Qwen2.5-3B. Hybrid retrieval with Qwen2.5-3B
expansion achieves 0.868 Recall@10 and 0.788 nDCG@10. This is better than
hybrid retrieval with Qwen2.5-0.5B expansion, but still below the non-expanded
hybrid baseline. These results suggest that direct LLM query expansion may
dilute the original lexical signal or introduce terms that shift the query away
from the most relevant documents.
6.3 Effect of LLM Size
Although LLM expansion does not surpass the non-expanded baselines, model
size clearly affects expansion quality.

baseline. These results suggest that direct LLM query expansion may
dilute the original lexical signal or introduce terms that shift the query away
from the most relevant documents.
6.3 Effect of LLM Size
Although LLM expansion does not surpass the non-expanded baselines, model
size clearly affects expansion quality. Qwen2.5-3B consistently outperforms Qwen2.5-

-- 13 of 17 --

14 Nimol Thuon
Fig. 4. Qualitative examples of LLM-assisted query expansion behavior. While some
expansions preserve the original intent and add useful related terms, other cases in-
troduce topic drift, generic terms, or over-expanded reformulations. These examples
illustrate why direct LLM expansion does not consistently improve Khmer retrieval
performance.
0.5B under both dense and hybrid retrieval settings. For dense retrieval, Qwen2.5-
3B improves Recall@10 from 0.365 to 0.440 and nDCG@10 from 0.318 to 0.372
compared with Qwen2.5-0.5B. For hybrid retrieval, Qwen2.5-3B improves Re-
call@10 from 0.788 to 0.868 and nDCG@10 from 0.729 to 0.788.
This indicates that larger instruction-tuned LLMs generate more retrieval-
useful query expansions than smaller models. However, the improvement is not
enough to outperform the original-query BM25 or hybrid baselines. Therefore,
LLM size helps, but direct expansion remains insufficient for robust Khmer re-
trieval.
6.4 Effect of Filtering Expanded Queries
The filtered expansion variants generally perform worse than the raw expanded-
query variants. For example, Hybrid + Qwen2.5-3B decreases from 0.868 Re-
call@10 and 0.788 nDCG@10 to 0.838 Recall@10 and 0.755 nDCG@10 after
filtering. A similar decrease is observed for Qwen2.5-0.5B. This suggests that
simple filtering removes not only noisy terms but also useful semantic cues.
These results show that filtering LLM expansions is not straightforward.
Generic stopword-like filtering may reduce topic drift in some cases, but it can
also weaken useful reformulations. More advanced filtering strategies

This suggests that
simple filtering removes not only noisy terms but also useful semantic cues.
These results show that filtering LLM expansions is not straightforward.
Generic stopword-like filtering may reduce topic drift in some cases, but it can
also weaken useful reformulations. More advanced filtering strategies should pre-
serve named entities, maintain query intent, and adapt to different query types.

-- 14 of 17 --

Khmer Semantic Search: Hybrid Retrieval and LLM-Assisted Query 15
6.5 Qualitative Analysis of LLM Expansion
Figure 4 illustrates common behaviors of LLM-assisted query expansion. Some
expansions preserve the original query intent and add useful related terms. How-
ever, other cases introduce topic drift, generic terms, or overly broad reformu-
lations. For example, a specific entity query may be expanded into a general
tourism or culture query, causing the retriever to return broad documents rather
than the target document. In other cases, the LLM adds general terms such as in-
formation, history, document, or Cambodia, which may match many documents
and reduce ranking precision.
These qualitative examples help explain the quantitative results. LLM expan-
sion can be helpful when it adds precise related terms, but it can hurt retrieval
when it weakens the original lexical signal or changes the query focus. This is
particularly important for Khmer, where exact entity matching and character-
level overlap remain strong retrieval signals.
6.6 Discussion, Limitations, and Future Work
The results lead to three main observations. First, character n-gram BM25 is
a strong baseline for Khmer web retrieval. Second, hybrid retrieval performs
close to BM25, suggesting that dense semantic representations provide com-
plementary but limited signals. Third, LLM-assisted query expansion depends
strongly on model size and expansion quality. Larger LLMs produce better ex-
pansions than smaller LLMs, but direct expansion still does not outperform
strong

performs
close to BM25, suggesting that dense semantic representations provide com-
plementary but limited signals. Third, LLM-assisted query expansion depends
strongly on model size and expansion quality. Larger LLMs produce better ex-
pansions than smaller LLMs, but direct expansion still does not outperform
strong non-expanded retrieval baselines. These findings do not suggest that
LLMs are ineffective for Khmer retrieval. Rather, they show that unrestricted
query expansion alone is insufficient. Future LLM-assisted retrieval should con-
sider query-type-aware prompting, entity-preserving expansion, multi-query fu-
sion, retrieval-aware reranking, and stronger human-verified evaluation. The rele-
vance labels are silver labels constructed from source-document matching, BM25
candidate retrieval, rule-based scoring, category agreement, and partial human
verification, which may favor lexical retrieval. Finally, dense retrieval is evalu-
ated with multilingual-E5-small, and LLM-assisted retrieval uses direct expan-
sion with simple filtering. Future work will extend KSE-Web to PDF and OCR-
based documents, collect more natural user queries, add larger human-verified
relevance judgments, and evaluate stronger multilingual, Khmer-adapted, and
LLM-based retrieval methods.
7 Conclusion
This paper presented KSE-Web, an analysis of hybrid retrieval and LLM-assisted
query expansion for low-resource Khmer semantic search using a Khmer web re-
trieval dataset. We constructed the dataset from approximately 17K candidate
Khmer web titles and retained 3K cleaned full-text Khmer documents after fil-
tering, normalization, deduplication, and document-length control. The dataset

-- 15 of 17 --

16 Nimol Thuon
includes 300 manually reviewed user-style Khmer queries and 5,412 silver rele-
vance labels with partial human verification. Our experiments compared char-
acter n-gram BM25, multilingual dense retrieval, hybrid BM25+dense retrieval,
and LLM-assisted query expansion using

l. The dataset

-- 15 of 17 --

16 Nimol Thuon
includes 300 manually reviewed user-style Khmer queries and 5,412 silver rele-
vance labels with partial human verification. Our experiments compared char-
acter n-gram BM25, multilingual dense retrieval, hybrid BM25+dense retrieval,
and LLM-assisted query expansion using Qwen2.5 models. The results show that
character n-gram BM25 achieves the strongest overall performance, with 0.943
Recall@10 and 0.876 nDCG@10. Hybrid retrieval performs closely, reaching 0.929
Recall@10 and 0.871 nDCG@10, while dense retrieval alone performs substan-
tially lower. LLM-assisted query expansion does not outperform non-expanded
retrieval, although Qwen2.5-3B produces stronger expanded-query results than
Qwen2.5-0.5B. These findings show that LLMs should not be assumed to improve
low-resource retrieval automatically. For Khmer semantic search, character-level
lexical retrieval remains a strong baseline, dense retrieval provides complemen-
tary but limited semantic signals, and LLM-assisted query expansion requires
careful control to avoid topic drift and loss of lexical specificity. We hope this
study provides a useful foundation for future research on Khmer retrieval, low-
resource semantic search, and LLM-assisted document understanding.
References
1. Abe, K., Takeoka, K., Kato, M.P., Oyamada, M.: Llm-based query expansion fails
for unfamiliar and ambiguous queries. In: Proceedings of the 48th International
ACM SIGIR Conference on Research and Development in Information Retrieval.
pp. 3035–3039 (2025)
2. Bonifacio, L., Jeronymo, V., Abonizio, H.Q., Campiotti, I., Fadaee, M., Lotufo,
R., Nogueira, R.: mmarco: A multilingual version of the ms marco passage ranking
dataset. arXiv preprint arXiv:2108.13897 (2021)
3. Buoy, R., Taing, N., Chenda, S.: Khmer word search: challenges, solutions, and
semantic-aware search. arXiv preprint arXiv:2112.08918 (2021)
4. Carpineto, C., Romano, G.: A survey of automatic query expansion in

: A multilingual version of the ms marco passage ranking
dataset. arXiv preprint arXiv:2108.13897 (2021)
3. Buoy, R., Taing, N., Chenda, S.: Khmer word search: challenges, solutions, and
semantic-aware search. arXiv preprint arXiv:2112.08918 (2021)
4. Carpineto, C., Romano, G.: A survey of automatic query expansion in information
retrieval. ACM Computing Surveys 44(1), 1–50 (2012)
5. Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X.,
Wang, C., Wang, Y., et al.: A survey on evaluation of large language models. ACM
transactions on intelligent systems and technology 15(3), 1–45 (2024)
6. Coleman, J., Rosales, R., Toal, K., Cuadros, D., Leeds, N., Krishnamachari,
B., Iskarous, K.: Comparing llm-based translation approaches for extremely low-
resource languages. In: Proceedings for the Ninth Workshop on Technologies for
Machine Translation of Low Resource Languages (LoResMT 2026). pp. 49–68
(2026)
7. Gao, L., Ma, X., Lin, J., Callan, J.: Precise zero-shot dense retrieval without rele-
vance labels. Proceedings of the 61st Annual Meeting of the Association for Com-
putational Linguistics pp. 1762–1777 (2023)
8. Jian, P., Yu, D., Zhang, J.: Large language models know what is key visual entity:
An llm-assisted multimodal retrieval for vqa. In: Proceedings of the 2024 Con-
ference on Empirical Methods in Natural Language Processing. pp. 10939–10956
(2024)
9. Kaing, H., Dabre, R., Song, H., Tran, V.H., Tanaka, H., Utiyama, M.: Prahokbart:
A pre-trained sequence-to-sequence model for khmer natural language generation.

-- 16 of 17 --

Khmer Semantic Search: Hybrid Retrieval and LLM-Assisted Query 17
In: Proceedings of the 31st International Conference on Computational Linguistics.
pp. 1309–1322 (2025)
10. Karpukhin, V., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., Yih,
W.t.: Dense passage retrieval for open-domain question answering. In: Proceedings
of the 2020 Conference on Empirical Methods in Natural Language Processing
(EMNLP). pp.

nal Conference on Computational Linguistics.
pp. 1309–1322 (2025)
10. Karpukhin, V., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., Yih,
W.t.: Dense passage retrieval for open-domain question answering. In: Proceedings
of the 2020 Conference on Empirical Methods in Natural Language Processing
(EMNLP). pp. 6769–6781 (2020)
11. Robertson, S., Zaragoza, H.: The probabilistic relevance framework: Bm25 and
beyond. Foundations and Trends in Information Retrieval 3(4), 333–389 (2009)
12. Shen, L., Tan, W., Chen, S., Chen, Y., Zhang, J., Xu, H., Zheng, B., Koehn, P.,
Khashabi, D.: The language barrier: Dissecting safety challenges of llms in mul-
tilingual contexts. In: Findings of the Association for Computational Linguistics:
ACL 2024. pp. 2668–2680 (2024)
13. Sindhujan, A., Kanojia, D., Orasan, C., Qian, S.: When llms struggle: Reference-
less translation evaluation for low-resource languages. In: Proceedings of the First
Workshop on Language Models for Low-Resource Languages. pp. 437–459 (2025)
14. Takehi, R., Voorhees, E.M., Sakai, T., Soboroff, I.: Llm-assisted relevance assess-
ments: When should we ask llms for help? In: Proceedings of the 48th International
ACM SIGIR Conference on Research and Development in Information Retrieval.
pp. 95–105 (2025)
15. Thakur, N., Reimers, N., Rücklé, A., Srivastava, A., Gurevych, I.: Beir: A heteroge-
nous benchmark for zero-shot evaluation of information retrieval models. arXiv
preprint arXiv:2104.08663 (2021)
16. Thuon, N.: Khmer semantic search engine (kse): Digital information access and
document retrieval. arXiv preprint arXiv:2406.09320 (2024)
17. Thuon, N., Du, J.: Khmerformer: Multi-scale cnns-transformer with ex-
ternal attention for ancient khmer palm leaf isolated glyph classifica-
tion. In: 2024 Asia Pacific Signal and Information Processing Association
Annual Summit and Conference (APSIPA ASC). pp. 1–6. IEEE (2024).
https://doi.org/10.1109/APSIPAASC63619.2025.10849201
18. Thuon, N., Du, J., Zhang, J.:

-transformer with ex-
ternal attention for ancient khmer palm leaf isolated glyph classifica-
tion. In: 2024 Asia Pacific Signal and Information Processing Association
Annual Summit and Conference (APSIPA ASC). pp. 1–6. IEEE (2024).
https://doi.org/10.1109/APSIPAASC63619.2025.10849201
18. Thuon, N., Du, J., Zhang, J.: Syllable analysis data augmentation for khmer an-
cient palm leaf recognition. In: 2022 Asia-Pacific Signal and Information Processing
Association Annual Summit and Conference (APSIPA ASC). pp. 1855–1862. IEEE
(2022). https://doi.org/10.23919/APSIPAASC55919.2022.9980217
19. Thuon, N., Zhang, W., Thuon, S.: Ksw: Khmer stop word based dictionary for
keyword extraction. arXiv preprint arXiv:2405.17390 (2024)
20. Wang, L., Yang, N., Huang, X., Jiao, B., Yang, L., Jiang, D., Majumder, R., Wei,
F.: Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint
arXiv:2212.03533 (2022)
21. Wang, L., Yang, N., Wei, F.: Query2doc: Query expansion with large language
models. In: Proceedings of the 2023 Conference on Empirical Methods in Natural
Language Processing. pp. 9414–9423 (2023)
22. Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F.,
Wei, H., et al.: Qwen2.5 technical report. arXiv preprint arXiv:2412.15115 (2024)
23. Zhang, X., Ma, X., Shi, P., Lin, J.: Mr. tydi: A multi-lingual benchmark for dense
retrieval. In: Proceedings of the 1st workshop on multilingual representation learn-
ing. pp. 127–137 (2021)
24. Zhang, X., Thakur, N., Ogundepo, O., Kamalloo, E., Alfonso-Hermelo, D., Li,
X., Liu, Q., Rezagholizadeh, M., Lin, J.: Miracl: A multilingual retrieval dataset
covering 18 diverse languages. Transactions of the Association for Computational
Linguistics 11, 1114–1131 (2023)

-- 17 of 17 --