NeuCLIRTech Topics and Queries

NeuCLIRTech is an evaluation benchmark for monolingual and cross-language retrieval on technical documents. The documents are Chinese academic abstracts from the Chinese Scientific Literature (CSL) dataset [Li et al., 2022]. The topics were developed for the TREC NeuCLIR track in 2023 and 2024 [Lawrie et al., 2024; Lawrie et al., 2025].

The document collection can be found at neuclir/csl.

Paper: NeuCLIRTech: Chinese Monolingual and Cross-Language Information Retrieval Evaluation in a Challenging Domain

Supporting Tasks and Corresponding Data

NeuCLIRTech supports two types of tasks: Chinese monolingual and English-Chinese cross-language retrieval. The following specifies the documents, queries, and qrels (labels) that should be used for each task.

Please report nDCG@20 for all tasks.

We use : to indicate different subset under the dataset.

Task	Documents	Queries	Qrels
Monolingual	neuclir/csl:csl	zho split of neuclir/tech:queries	neuclir/tech:qrels
Cross-Language	neuclir/csl:csl	eng split of neuclir/tech:queries	neuclir/tech:qrels

Files

data/

File	Content
topic.jsonl	The 110 evaluation topics, i.e., all topics with relevance judgments. Each topic has the original English version, a human translation into Chinese, and a Google Translate translation into Chinese, each with title, description, and narrative fields.
eng.tsv	English queries for the 110 evaluation topics, one per line as topic_id<TAB>title description. Used by all CLIR runs except BM25 with query translation.
zho.tsv	Chinese queries from the human translations, in the same format. Used by all monolingual runs.
qrels.gains.txt	Relevance judgments for the 110 evaluation topics in TREC qrels format: 35,962 judgments, of which 840 have gain 3 and 1,250 have gain 1.
2023.jsonl, 2024.jsonl	All topics developed in 2023 (41 topics, IDs 1–41) and 2024 (106 topics, IDs 200–305) in the same format as topic.jsonl, including the 37 topics that have no relevance judgments.

runs/

Run files are in TREC run format (topic_id Q0 doc_id rank score tag) and follow the naming scheme {run_handle}_{task}.trec, where task is mono or clir. Monolingual runs use the Chinese queries in zho.tsv; CLIR runs use the English queries in eng.tsv. First-stage runs return up to 1,000 documents per topic. Reranker runs return the 100 documents they rerank, which are the top 100 documents of the fusion run (fusion3_mono or fusion3_clir) of the same task.

Evaluation

pip install ir_measures
ir_measures data/qrels.gains.txt runs/plaidx_clir.trec 'nDCG@20 Judged@20'

Baseline Retrieval Results and Run Files

We also provide all baseline retrieval results reported in the paper. Detailed descriptions of each model are in Baseline Systems. The table below lists nDCG@20 over the 110 evaluation topics in the same order as Table 1 of the paper.

Run Handle	Model Type	Model Name	Monolingual nDCG@20	CLIR nDCG@20
bm25dt	Lexical	English BM25 on Document MT	–	0.237
bm25qt	Lexical	Chinese BM25 with Translated Queries	0.290	0.274
bgem3sparse	Bi-Encoder	BGE-M3 Sparse	0.307	0.044
me5large	Bi-Encoder	e5 Large	0.233	0.151
repllama	Bi-Encoder	RepLlama	0.269	0.245
arctic	Bi-Encoder	Arctic-Embed Large v2	0.359	0.262
milco	Bi-Encoder	MILCO	0.253	0.264
jinav3	Bi-Encoder	JinaV3	0.382	0.305
plaidx	Bi-Encoder	PLAID-X	0.356	0.362
qwen600m	Bi-Encoder	Qwen3 0.6B Embed	0.402	0.377
fusion3	Bi-Encoder	Fusion (PLAID-X + MILCO + Qwen3 8B Embed)	0.438	0.431
qwen4b	Bi-Encoder	Qwen3 4B Embed	0.469	0.450
qwen8b	Bi-Encoder	Qwen3 8B Embed	0.480	0.472
mt5	Pointwise Reranker	Mono-mT5XXL	0.456	0.407
searcher-rerank	Pointwise Reranker	SEARCHER Reranker	0.351	0.419
jina-rerank	Pointwise Reranker	Jina Reranker	0.489	0.446
qwen3-0.6b-rerank	Pointwise Reranker	Qwen3 0.6B Rerank	0.494	0.485
qwen3-8b-rerank	Pointwise Reranker	Qwen3 8B Rerank	0.521	0.508
rank1	Pointwise Reranker	Rank1	0.531	0.512
qwen3-4b-rerank	Pointwise Reranker	Qwen3 4B Rerank	0.522	0.526
rankzephyr	Listwise Reranker	RankZephyr 7B	0.434	0.432
firstqwen	Listwise Reranker	FIRST Qwen3 8B	0.539	0.520
rq32b	Listwise Reranker	RankQwen-32B	0.541	0.526
qwq	Listwise Reranker	Rank-K (QwQ)	0.542	0.533

Baseline Systems

Sparse Retrieval

Sparse retrieval performs lexical matches between the query and the document and produces a score for each retrieved document based on these matches. We run BM25 with RM3 using the Patapsco framework [Costello et al., 2022], a framework built on top of Pyserini [Lin et al., 2021] specifically for experiments in many languages. Given the reliance on lexical matches, we provide results on BM25 with both document translation (matching in English) and query translation (matching in Chinese). We set k1=0.9 and b=0.4 following standard conventions. RM3 query expansion [Jaleel et al., 2004] adds ten words based on the top ten documents, weighting terms from the original query and the expansion terms equally.

• Chinese BM25 with Translated Queries (bm25qt_mono, bm25qt_clir). For monolingual retrieval, the queries are the human translations produced by the assessors during topic development. For CLIR, the English queries are translated into Chinese with Google Translate, which treats Google Translate as part of the CLIR system. All Chinese characters in queries and documents are tokenized and normalized using spaCy [Honnibal et al., 2020].
• English BM25 on Document MT (bm25dt_clir) matches the English queries against Google Translate translations of the documents into English. The translated documents are preprocessed with the spaCy tokenizer and stemmer, with stopwords removed.

Dense Retrieval

Bi-encoders represent queries and documents as a single dense vector in a latent space [Lin et al., 2022]. For efficient first-stage retrieval, these representations can be stored in a vector index that supports approximate nearest neighbor search, such as FAISS [Douze et al., 2025]. We use fast scan product quantization with FAISS, where the number of codes is set to half the embedding dimensionality and each code is 4 bits.

• Arctic Embed v2 [Yu et al., 2025] (arctic_*) is a multilingual model based on an XLM-RoBERTa backbone with 0.6B parameters. We use the snowflake-arctic-embed-l-v2.0 checkpoint, which is the largest available Arctic Embed model.
• E5 [Wang et al., 2024] (me5large_*) is a multilingual model based on an XLM-RoBERTa backbone with 0.6B parameters. We use the multilingual-e5-large-instruct checkpoint.
• Jina v3 [Sturua et al., 2024] (jinav3_*) is a multilingual model based on an XLM-RoBERTa backbone with 0.6B parameters.
• Qwen3-Embedding [Zhang et al., 2025] (qwen600m_*, qwen4b_*, qwen8b_*) is a family of embedding models trained with synthetic data on top of Qwen3 backbones [Yang et al., 2025a]. We use all three model sizes: 0.6B, 4B, and 8B.
• RepLlama [Ma et al., 2023] (repllama_*) is a dense retrieval model based on LLaMA-2-7B. We use the repllama-v1-7b-lora-passage checkpoint trained on MS MARCO.

Multi-Dense Vector Retrieval

• PLAID-X [Yang et al., 2024a; Yang et al., 2024b] (plaidx_*) is a cross-language variant of the ColBERT model. PLAID-X is fine-tuned from an XLM-RoBERTa 0.6B model with cross-language distillation from an mT5 reranker (described below) using MS MARCO v1 training queries and passages. Each document is separated into passages using a sliding window of 180 tokens with a stride of 90. At inference time, document scores are aggregated from the passage scores using MaxP [Dai and Callan, 2019].

Learned Sparse Retrieval

Learned sparse retrieval (LSR) represents queries and documents as sparse vectors where each dimension is associated with a term in a vocabulary. Some LSR methods produce vectors containing only terms from the input text, while others use a masked language modeling head to also perform term expansion [Nguyen et al., 2023]. For efficient first-stage retrieval, these representations can be stored in an inverted index. We run all LSR methods with Anserini [Yang et al., 2017], which is built on Lucene.

• BGE-M3-Sparse [Chen et al., 2024] (bgem3sparse_*) is an LSR method trained as part of the multi-granularity M3-Embedding model. It is based on an XLM-RoBERTa backbone with 0.6B parameters. This method does not perform expansion, so it can only be used for monolingual retrieval.
• MILCO [Nguyen et al., 2025] (milco_*) is a multilingual LSR method that performs query expansion and document expansion. MILCO performs Chinese retrieval by producing sparse vectors tied to an English vocabulary, so the model implicitly translates the non-English input. It is based on an XLM-RoBERTa backbone with 0.6B parameters.

First-Stage Retrieval Fusion

To provide strong first-stage retrieval for the rerankers, we use reciprocal rank fusion (RRF) [Cormack et al., 2009] to combine the rankings of three strong first-stage retrieval models from the three families above: PLAID-X, MILCO, and Qwen3-Embedding 8B (fusion3_*). Following common practice and Cormack et al. [2009], we set k=60 and do not use per-method weights.

Rerankers

We rerank the top 100 results of the first-stage retrieval fusion using a range of pointwise and listwise reranking models. Pointwise models take a query-document pair as input and output a relevance score. Listwise models take a query and a set of documents as input. Some listwise models output one relevance score for each document, whereas others output an ordering of their input documents without producing relevance scores. For models that output an ordering, we use the partial sorting algorithm from RankGPT [Sun et al., 2023] to produce a ranking of all 100 results, reranking 20 documents at a time with a sliding window stride of 10 from the bottom of the top 100 candidate documents.

• FIRSTQwen-8B [Chen et al., 2025] (firstqwen_*) is a listwise reranking model built on top of Qwen3 8B [Yang et al., 2025a] and trained with the FIRST objective [Reddy et al., 2024].
• jina-reranker-v3 [Wang et al., 2025] (jina-rerank_*) is a reranking model built on top of Qwen3 0.6B that can perform pointwise or listwise reranking. In both configurations, the model outputs one relevance score for each document. We use it in a pointwise configuration.
• mT5 [Bonifacio et al., 2022] (mt5_*) is a pointwise reranking model built on top of a multilingual T5 backbone. We use the mt5-13b-mmarco-100k checkpoint with 13B parameters.
• Qwen3-Reranker [Zhang et al., 2025] (qwen3-0.6b-rerank_*, qwen3-4b-rerank_*, qwen3-8b-rerank_*) is a family of pointwise reranking models built on top of Qwen3. We use all three model sizes: 0.6B, 4B, and 8B.
• Rank1 [Weller et al., 2025] (rank1_*) is a pointwise reranking model fine-tuned to perform reasoning before reranking. Rank1 is built on top of the Qwen2.5 base model [Yang et al., 2024c]. We use the 32B variant.
• Rank-K [Yang et al., 2025b] (qwq_*) is a listwise reranking model that performs reasoning before reranking using the QwQ 32B reasoning model [Qwen Team, 2025].
• RankQwen-32B (rq32b_*) is a zero-shot listwise reranker that uses Qwen3-32B without any training to rerank the top 100 documents. This system does not use any reasoning.
• RankZephyr-7B [Pradeep et al., 2023] (rankzephyr_*) is a listwise reranking model built on top of Mistral 7B [Jiang et al., 2023].
• SEARCHER [Agarwal et al., 2025] (searcher-rerank_*) is a pointwise reranking model built on top of Mistral-Nemo-Base-2407 with 12B parameters. It was trained on Persian, Russian, and Chinese translations of MS MARCO passages [Bajaj et al., 2018], keeping the queries in English. Documents in each training batch were a mixture of all three languages.

Citation

@article{lawrie2026neuclirtech,
  title={NeuCLIRTech: Chinese Monolingual and Cross-Language Information Retrieval Evaluation in a Challenging Domain},
  author={Lawrie, Dawn and Mayfield, James and Yang, Eugene and Yates, Andrew and MacAvaney, Sean and Pradeep, Ronak and Miller, Scott and McNamee, Paul and Soldaini, Luca},
  journal={arXiv preprint arXiv:2602.05334},
  year={2026}
}

References

• Agarwal et al., 2025. ISI's SEARCHER II System for TREC's 2024 NeuCLIR Track.
• Bajaj et al., 2018. MS MARCO: A Human Generated MAchine Reading COmprehension Dataset. arXiv:1611.09268
• Bonifacio et al., 2022. mMARCO: A Multilingual Version of the MS MARCO Passage Ranking Dataset. arXiv:2108.13897
• Chen et al., 2024. M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. ACL Anthology
• Chen et al., 2025. Accelerating Listwise Reranking: Reproducing and Enhancing FIRST.
• Cormack et al., 2009. Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods.
• Costello et al., 2022. Patapsco: A Python Framework for Cross-Language Information Retrieval Experiments.
• Dai and Callan, 2019. Deeper Text Understanding for IR with Contextual Neural Language Modeling.
• Douze et al., 2025. The Faiss Library. IEEE Transactions on Big Data.
• Honnibal et al., 2020. spaCy: Industrial-Strength Natural Language Processing in Python. doi:10.5281/zenodo.1212303
• Jaleel et al., 2004. UMass at TREC 2004: Novelty and HARD.
• Jiang et al., 2023. Mistral 7B. arXiv:2310.06825
• Lawrie et al., 2024. Overview of the TREC 2023 NeuCLIR Track.
• Lawrie et al., 2025. Overview of the TREC 2024 NeuCLIR Track. arXiv:2509.14355
• Li et al., 2022. CSL: A Large-scale Chinese Scientific Literature Dataset. ACL Anthology
• Lin et al., 2021. Pyserini: A Python Toolkit for Reproducible Information Retrieval Research with Sparse and Dense Representations.
• Lin et al., 2022. Pretrained Transformers for Text Ranking: BERT and Beyond.
• Ma et al., 2023. Fine-Tuning LLaMA for Multi-Stage Text Retrieval.
• Nguyen et al., 2023. A Unified Framework for Learned Sparse Retrieval.
• Nguyen et al., 2025. MILCO: Learned Sparse Retrieval Across Languages via a Multilingual Connector. arXiv:2510.00671
• Pradeep et al., 2023. RankZephyr: Effective and Robust Zero-Shot Listwise Reranking is a Breeze!
• Qwen Team, 2025. QwQ-32B: Embracing the Power of Reinforcement Learning. Blog
• Reddy et al., 2024. FIRST: Faster Improved Listwise Reranking with Single Token Decoding. doi:10.18653/v1/2024.emnlp-main.491
• Sturua et al., 2024. jina-embeddings-v3: Multilingual Embeddings With Task LoRA.
• Sun et al., 2023. Is ChatGPT Good at Search? Investigating Large Language Models as Re-Ranking Agents.
• Wang et al., 2024. Multilingual E5 Text Embeddings: A Technical Report. arXiv:2402.05672
• Wang et al., 2025. jina-reranker-v3: Last but Not Late Interaction for Document Reranking.
• Weller et al., 2025. Rank1: Test-Time Compute for Reranking in Information Retrieval.
• Yang et al., 2017. Anserini: Enabling the Use of Lucene for Information Retrieval Research.
• Yang et al., 2024a. Translate-Distill: Learning Cross-Language Dense Retrieval by Translation and Distillation. doi:10.1007/978-3-031-56060-6_4
• Yang et al., 2024b. Distillation for Multilingual Information Retrieval.
• Yang et al., 2024c. Qwen2.5 Technical Report. arXiv:2412.15115
• Yang et al., 2025a. Qwen3 Technical Report. arXiv:2505.09388
• Yang et al., 2025b. Rank-K: Test-Time Reasoning for Listwise Reranking. arXiv:2505.14432
• Yu et al., 2025. Arctic-Embed 2.0: Multilingual Retrieval Without Compromise.
• Zhang et al., 2025. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models. arXiv:2506.05176