Pith. sign in

REVIEW 3 major objections 7 minor 49 references

Optimized Text Embedding Models and Benchmarks for Amharic Passage Retrieval

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Amharic-specific dense retrieval models trained on Amharic-pretrained encoders outperform much larger multilingual embedding models for Amharic passage retrieval.

desk verdict Useful benchmark and artifacts for Amharic retrieval, but the abstract's headline claim ignores the paper's own fine-tuned multilingual baseline, which outperforms the proposed model. read the letter →

arxiv 2505.19356 v2 pith:KZEWYUAQ submitted 2025-05-25 cs.IR cs.AIcs.CLcs.LG

Kidist Amde Mekonnen , Yosef Worku Alemneh , Maarten de Rijke This is my paper · ORCID
classification cs.IRcs.AIcs.CLcs.LG
keywords Amharicpassageretrievaldenselow-resourceinformationtokenizationsubwordfertilityColBERTtextembeddingsAfricanlanguages
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that for a morphologically rich low-resource language like Amharic, retrieval models built on Amharic-specific pretrained encoders outperform much larger multilingual embedding models. The authors construct a passage retrieval benchmark from Amharic news data, fine-tune BERT and RoBERTa-based encoders with contrastive learning, and report that their 110M-parameter RoBERTa-Base-Amharic-Embed improves MRR@10 by 17.6% over the strongest 568M-parameter multilingual baseline. A ColBERT late-interaction variant reaches the highest MRR@10 of 0.843. If correct, the result suggests that language-specific adaptation and tokenizer alignment matter more than brute-force model scaling, and it provides reproducible baselines for future Amharic information retrieval.

What carries the argument

The carrying mechanism is contrastive fine-tuning of Amharic BERT and RoBERTa backbones with in-batch negatives, using mean pooling and L2 normalization, trained with multiple negatives ranking loss on roughly 45,000 query-passage pairs. The central explanatory object is subword fertility, the average number of subword tokens per word, which the paper uses to show that Amharic-specific tokenizers preserve word boundaries and produce more semantically coherent embeddings. The second mechanism is ColBERT's late interaction, which scores a query against a passage by summing, over query tokens, the maximum similarity to any passage token, and this token-level matching pushes MRR higher than the bi-encoder on the same backbone.

What would settle it

Build a small Amharic test set with human relevance judgments from multiple annotators, or collect real Amharic search queries with click feedback, and rerun the same models; if the Amharic-specific models no longer beat the multilingual baselines on that set, or if allowing multiple relevant passages changes the model ranking, the central claim fails. A more targeted check is to fine-tune the same multilingual backbone with an Amharic-optimized tokenizer while holding parameters constant; if the retrieval gains vanish, tokenization rather than pretraining data is not the driver.

Watch

Extended reading notes

Core claim

The central discovery is that fine-tuning Amharic-pretrained encoders on Amharic query-passage pairs yields dense retrievers that consistently beat zero-shot multilingual embedding models, despite being far smaller. On the constructed benchmark, RoBERTa-Base-Amharic-Embed (110M parameters) reaches MRR@10 0.775, NDCG@10 0.808, and Recall@10 0.913, versus 0.659, 0.701, and 0.831 for the strongest multilingual baseline Snowflake-Arctic-Embed (568M parameters); even the 42M-parameter RoBERTa-Medium-Amharic-Embed beats all multilingual baselines. The same pattern holds for late interaction: ColBERT built on the Amharic encoder reaches MRR@10 0.843, a 28.31% relative improvement over BM25. The paper attributes the gap largely to tokenization quality, measured by subword fertility, where Amharic-specific tokenizers produce about 1.46 subword tokens per word versus 2.35 for the strongest multilingual model and 13.80 for the weakest.

Load-bearing premise

The entire benchmark rests on treating each news headline as a query and its own article as the only relevant passage, so if headlines are not what real Amharic searchers type, or if other articles are also relevant, the measured rankings and model comparisons may not transfer to actual search behavior.

Editorial extensions

If this is right

  • Amharic retrieval systems can use a 42M-parameter model that beats the 568M multilingual baseline while being over 13x smaller, cutting compute cost substantially.
  • Subword fertility becomes a diagnosable factor in low-resource retrieval: models with fertility near 1.5 outperform models with fertility above 2, so tokenizer choice should be reported and optimized as part of retrieval benchmarks.
  • Multilingual embedding models gain sharply from Amharic-specific fine-tuning, with MRR@10 rising from 0.659 to 0.827, indicating that zero-shot evaluation understates what such models can do with in-language supervision.
  • Late interaction with ColBERT improves precision over the bi-encoder on the same Amharic encoder, pointing to token-level matching as a promising direction for morphologically rich languages.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the headline-as-query setup may reward retrievers that match the article's own wording, so on real user queries with greater lexical divergence the gap between sparse and dense models could widen rather than shrink.
  • Editorial inference: the fertility-retrieval correlation suggests a controlled experiment, taking one multilingual checkpoint and swapping only its tokenizer for an Amharic-optimized one; if gains do not reproduce, then pretraining data, not tokenization alone, is the active ingredient.
  • Editorial inference: the reported negation failure, where a model ranks the same passage for affirmative and negated queries, implies a concrete test of whether Amharic embeddings capture polarity; a polarity-reversed query set would show whether this failure is systematic.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper introduces Amharic-specific dense passage retrieval models based on Amharic pre-trained BERT and RoBERTa encoders, fine-tuned with multiple negatives ranking loss on a new benchmark derived from AMNEWS news articles (headlines as queries, article bodies as passages). It evaluates bi-encoder models of varying sizes, a ColBERT late-interaction model, and sparse (BM25) and multilingual dense baselines, reporting MRR@10, NDCG@10, and Recall@K. The main claimed contribution is that Amharic-optimized models outperform the strongest multilingual zero-shot baseline (Arctic Embed 2.0) by 17.6% in MRR@10, with the ColBERT variant reaching 0.843 MRR@10; the authors also include a supplementary evaluation on the external 2AIRTC collection, hyperparameter sensitivity analysis, and release their code, data, and models.

Significance. The paper provides a useful resource for low-resource IR: a public Amharic retrieval dataset, fine-tuned models, a tokenization-fertility analysis, and reproducibility artifacts. The evaluation on an external collection (2AIRTC) and the hyperparameter sensitivity analysis are additional strengths. However, the central claim that language-specific adaptation outperforms multilingual scaling is not supported by the full set of results: the headline comparison is zero-shot multilingual versus fine-tuned Amharic models, and the fine-tuned multilingual baseline (Section 6.5) actually outperforms the proposed models. The paper's contribution would be better framed as a demonstration that in-language fine-tuning is crucial, with language-specific models offering a parameter-efficient alternative in zero-shot regimes rather than a general superiority over multilingual models. With revised claims, the resource and analysis would be valuable to the community.

major comments (3)
  1. [6.5 (Table 3) and Abstract] The fine-tuned multilingual baseline snowflake-arctic-embed-l-v2.0-AM (Table 3) achieves MRR@10 0.827 and Recall@10 0.942, both higher than the proposed RoBERTa-Base-Amharic-Embed's 0.775 and 0.913 (Table 1). The abstract's claim of a 17.6% relative MRR@10 improvement over the 'strongest multilingual baseline' holds only for zero-shot baselines; when the same Amharic supervision is applied to the multilingual model, it outperforms the proposed model. This directly undermines the conclusion in Sections 1 and 7 that language-specific adaptation outperforms multilingual scaling. The evidence instead shows that supervised fine-tuning on in-language data is the dominant factor, and the larger multilingual model benefits more from it. Please revise the headline claims and conclusions to include the fine-tuned multilingual comparison and discuss the implications for the 'language-specific vs. multilingual' narrative.
  2. [Appendix A.1 (Table 4)] On the external 2AIRTC collection (Table 4), the proposed Amharic-specific models do not outperform multilingual baselines: multilingual-e5-large-instruct achieves MRR@100 0.905, NDCG@100 0.808, and Recall@100 0.853 versus 0.861, 0.770, and 0.830 for RoBERTa-Base-Amharic-Embed. The text in A.1 states that the model is 'just one point below' the strongest multilingual baseline, which is inaccurate for NDCG and MRR (differences of 3.8 and 4.4 points, respectively). This external evaluation is important evidence that the 'consistently outperform' claim (Sections 1 and 7) is too strong; the paper should incorporate this finding into the main narrative and qualify the conclusion to specify the conditions under which language-specific models are competitive.
  3. [5.1 and Section 8] The main evaluation in Table 1 and following sections relies on a self-built benchmark with heuristic relevance labels: each headline is assumed relevant only to its own article, and the test split is only about 10% of the 45K query-passage pairs. The paper acknowledges this in Section 8, but the abstract and conclusion present the results without this caveat. Since all main tables and the tokenization analysis depend on this benchmark, the authors should provide a quantitative validation of the labels (e.g., multi-annotator agreement on a sample) or explicitly state in the abstract that the results are preliminary and may not generalize to real Amharic search. The supplementary 2AIRTC evaluation, which shows only comparable performance, reinforces the need for this qualification.
minor comments (7)
  1. [Table 6 caption] Table 6 caption: '2AIRC' should be '2AIRTC'.
  2. [Figure 1] Figure 1: the label 'BERT-Mediu-nAmharic-embed' contains a typo; it should read 'BERT-Medium-Amharic-embed'.
  3. [Appendix A.4] Appendix A.4: 'across across' should be 'across'.
  4. [Throughout] Model names are spelled inconsistently (e.g., 'RoBERTa-Base-Amharic-Embed' vs. 'RoBERTa-Base-Amharic-embed'); please standardize throughout.
  5. [Section 5.1] Section 5.1: report the exact number of query-passage pairs and train/test split sizes after deduplication rather than 'approximately 45K'.
  6. [Section 6.1] Section 6.1, third bullet: the claim that Amharic-specific models achieve superior performance with significantly fewer parameters should be qualified, since the fine-tuned multilingual model in Section 6.5 achieves higher performance with more parameters; the efficiency advantage holds only in the zero-shot comparison.
  7. [Table 2] In Table 2, the caption states significant improvements 'over the strongest baseline' but it is unclear whether the paired t-test is against BM25 or against the bi-encoder; please clarify the comparison and the number of queries used for significance testing.

Circularity Check

1 steps flagged · score 6.0 of 10

The central 'language-specific beats multilingual scaling' claim is forced by comparing fine-tuned Amharic models to zero-shot multilingual baselines; the paper's own fine-tuned Arctic control outperforms the proposed model.

  1. fitted input called prediction [Abstract; Sections 6.1 and 6.5, Tables 1 and 3]
    "Abstract: 'Our proposed RoBERTa-Base-Amharic-Embed model (110M parameters) achieves a 17.6% relative improvement in MRR@10 and a 9.86% gain in Recall@10 over the strongest multilingual baseline, Arctic Embed 2.0 (568M parameters).' Section 6.5: 'While our primary comparison focuses on zero-shot multilingual models, we also investigate the impact of retrieval-specific supervised fine-tuning. To this end, we fine-tune the strongest multilingual baseline, Snowflake-Arctic-Embed (568M parameters), using Amharic query–passage pairs."

    The headline gain is measured against a zero-shot Arctic Embed 2.0 (Table 1), while RoBERTa-Base-Amharic-Embed received four epochs of contrastive fine-tuning on the same 45K Amharic query-passage pairs used to build the evaluation set. Section 6.5 then fine-tunes Arctic Embed 2.0 on the identical pairs and obtains MRR@10 0.827 vs. 0.775 and Recall@10 0.942 vs. 0.913, i.e., the multilingual model is stronger when supervision is matched. The paper's central conclusion that language-specific adaptation, not scaling, drives gains is therefore an artifact of comparing fine-tuned models to zero-shot baselines: the 'language-specific' advantage is, by construction, the supervised-fine-tuning advantage. The paper's own control removes the claimed effect.

full rationale

The only load-bearing circular pattern is the comparative claim in the abstract and conclusions: the 'strongest multilingual baseline' is selected under a zero-shot protocol while the proposed models are fine-tuned, so the reported superiority is built into the comparison rather than demonstrated. The paper is otherwise self-contained: the Amharic Passage Retrieval Dataset is constructed from public AMNEWS data with explicit heuristic labels, training and evaluation are separated into held-out splits, and supplementary evaluation on the external 2AIRTC collection provides independent, though annotation-limited, evidence. No self-citation chain, uniqueness import, or ansatz-by-citation is used. The fine-tuned multilingual control in Table 3 is an honest internal check, but it undermines rather than supports the headline conclusion. Score 6 because the central language-specific-versus-scaling claim reduces to the fine-tuning asymmetry; it is not 8-10 because the retrieval measurements themselves are genuine empirical outputs and the external benchmark gives some independent content.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper's central claim rests on the validity of the self-constructed heuristic benchmark and on the assumption that improvements transfer beyond that benchmark. These are domain assumptions, not standard mathematical axioms. No new physical or conceptual entities are introduced.

assumptions (3)
  • domain assumption Headline-article pairs from AMNEWS form valid query-passage relevance judgments.
    Section 5.1 constructs the Amharic Passage Retrieval Dataset on this heuristic; no human relevance labels exist. All non-paired passages are treated as negatives during training and evaluation.
  • domain assumption The 10% held-out split of AMNEWS is an unbiased test bed for retrieval effectiveness.
    Section 5.1: split stratified by category, but duplicate or near-duplicate articles may not be fully removed beyond MD5 hashing.
  • domain assumption 2AIRTC relevance judgments, despite acknowledged incompleteness, are reliable enough for supplementary conclusions.
    Appendix A uses 2AIRTC zero-shot evaluation while stating that incomplete judgments penalize dense models; conclusions are treated as indicative.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimized Text Embedding Models and Benchmarks for Amharic Passage Retrieval." pith.science (2026). https://pith.science/paper/KZEWYUAQ

@misc{pith2026250519356,
  author       = {Pith},
  title        = {Pith review of: Optimized Text Embedding Models and Benchmarks for Amharic Passage Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KZEWYUAQ}},
  note         = {Machine review of arXiv:2505.19356}
}
read the original abstract

Neural retrieval methods using transformer-based pre-trained language models have advanced multilingual and cross-lingual retrieval. However, their effectiveness for low-resource, morphologically rich languages such as Amharic remains underexplored due to data scarcity and suboptimal tokenization. We address this gap by introducing Amharic-specific dense retrieval models based on pre-trained Amharic BERT and RoBERTa backbones. Our proposed RoBERTa-Base-Amharic-Embed model (110M parameters) achieves a 17.6% relative improvement in MRR@10 and a 9.86% gain in Recall@10 over the strongest multilingual baseline, Arctic Embed 2.0 (568M parameters). More compact variants, such as RoBERTa-Medium-Amharic-Embed (42M), remain competitive while being over 13x smaller. Additionally, we train a ColBERT-based late interaction retrieval model that achieves the highest MRR@10 score (0.843) among all evaluated models. We benchmark our proposed models against both sparse and dense retrieval baselines to systematically assess retrieval effectiveness in Amharic. Our analysis highlights key challenges in low-resource settings and underscores the importance of language-specific adaptation. To foster future research in low-resource IR, we publicly release our dataset, codebase, and trained models at https://github.com/kidist-amde/amharic-ir-benchmarks.

Figures

Figures reproduced from arXiv: 2505.19356 by the authors.

Figure 1
Figure 1. Average subword fertility across embedding [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Subword tokenization comparison for a repre [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Effect of base model size on ColBERT perfor [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Negation failure case: The model retrieves the same top passage for both a positive (Query 1) and a [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: MRR@10 scores with 5 training epochs. The [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: NDCG@10 scores with 5 training epochs. Peak score (0.774) occurs at 5e-5 learning rate and batch size 256. Larger batch sizes generally benefit from more aggressive learning. 2e-5 5e-5 Learning Rate 64 128 256 Batch Size 0.873 0.885 0.874 0.884 0.868 0.887 Recall@10 (E…
Figure 9
Figure 9. Figure 9: NDCG@10 with 3 training epochs. Perfor￾mance is highest at 5e-5/128, and all batch sizes benefit from higher learning rates. 2e-5 5e-5 Learning Rate 64 128 256 Batch Size 0.853 0.871 0.858 0.880 0.848 0.875 Recall@10 (Epochs=3) 0.850 0.855 0.860 0.865 0.870 0.875 0.880…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 21 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Ife Adebara, AbdelRahim Elmadany, Muhammad Abdul-Mageed, and Alcides Alcoba Inciarte. 2023. https://doi.org/10.18653/v1/2023.findings-acl.97 SERENGETI : Massively multilingual language models for A frica . In Findings of the Association for Computational Linguistics: ACL 2023, pages 1498--1537, Toronto, Canada. Association for Computational Linguistics

    doi:10.18653/v1/2023.findings-acl.97 2023

  4. [4]

    Mehdi Ali, Michael Fromm, Klaudia Thellmann, Richard Rutmann, Max L \"u bbering, Johannes Leveling, Katrin Klug, Jan Ebert, Niclas Doll, Jasper Buschhoff, Charvi Jain, Alexander Weber, Lena Jurkschat, Hammam Abdelwahab, Chelsea John, Pedro Ortiz Suarez, Malte Ostendorff, Samuel Weinbach, Rafet Sifa, Stefan Kesselheim, and Nicolas Flores-Herr. 2024. https:...

    doi:10.18653/v1/2024.findings-naacl.247 2024

  5. [5]

    Israel Abebe Azime, Mitiku Yohannes Fuge, Atnafu Lambebo Tonja, Tadesse Destaw Belay, Aman Kassahun Wassie, Eyasu Shiferaw Jada, Yonas Chanie, Walelign Tewabe Sewunetie, and Seid Muhie Yimam. 2024 a . Enhancing Amharic-LLaMA : Integrating task specific and generative datasets. arXiv preprint arXiv:2402.08015

    work page Pith review arXiv 2024

  6. [6]

    Israel Abebe Azime and Nebil Mohammed. 2021. An Amharic news text classification dataset. arXiv preprint arXiv:2103.05639

    work page Pith review arXiv 2021

  7. [7]

    Israel Abebe Azime, Atnafu Lambebo Tonja, Tadesse Destaw Belay, Mitiku Yohannes Fuge, Aman Kassahun Wassie, Eyasu Shiferaw Jada, Yonas Chanie, Walelign Tewabe Sewunetie, and Seid Muhie Yimam. 2024 b . https://doi.org/10.18653/v1/2024.findings-emnlp.25 Walia- LLM : Enhancing A mharic- LL a MA by integrating task-specific and generative datasets . In Findin...

    doi:10.18653/v1/2024.findings-emnlp.25 2024

  8. [8]

    Tadesse Destaw Belay, Abinew Ayele, and Seid Muhie Yimam. 2021. https://aclanthology.org/2021.winlp-1.5/ The development of pre-processing tools and pre-trained embedding models for A mharic . In Proceedings of the Fifth Workshop on Widening Natural Language Processing, pages 25--28, Punta Cana, Dominican Republic. Association for Computational Linguistics

    work page 2021

Show all 49 references
  1. [9]

    Daniel Fernando Campos, Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, Li Deng, and Bhaskar Mitra. 2016. MS MARCO : A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268

    2016 arXiv

  2. [10]

    Antoine Chaffin and Raphaël Sourty. 2024. https://github.com/lightonai/pylate Pylate: Flexible training and retrieval for late interaction models

    2024

  3. [11]

    Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading W ikipedia to answer open-domain questions. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1870--1879, Vancouver, Canada. Associat...

    2017

  4. [12]

    Jiangui Chen, Ruqing Zhang, Jiafeng Guo, Maarten de Rijke, Yiqun Liu, Yixing Fan, and Xueqi Cheng. 2023. A unified generative retriever for knowledge-intensive language tasks via prompt learning. In SIGIR. ACM

    2023

  5. [13]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm \'a n, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. https://doi.org/10.18653/v1/2020.acl-main.747 Unsupervised cross-lingual representation learning ...

    2020 doi

  6. [14]

    Zhuyun Dai and Jamie Callan. 2020. Context-aware term weighting for first stage passage retrieval. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 1533--1536

    2020

  7. [15]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT : Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Hum...

    2019

  8. [16]

    Thibault Formal, Carlos Lassance, Benjamin Piwowarski, and St \'e phane Clinchant. 2021 a . Splade v2: Sparse lexical and expansion model for information retrieval. arXiv preprint arXiv:2109.10086

    2021 arXiv

  9. [17]

    Thibault Formal, Benjamin Piwowarski, and St \'e phane Clinchant. 2021 b . Splade: Sparse lexical and expansion model for first stage ranking. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2288--2292

    2021

  10. [18]

    Andargachew Mekonnen Gezmu, Binyam Ephrem Seyoum, Michael Gasser, and Andreas N \"u rnberger. 2018. https://aclanthology.org/W18-3809/ Contemporary A mharic corpus: Automatically morpho-syntactically tagged A mharic corpus . In Proceedings of the First Workshop on Linguistic R...

    2018

  11. [19]

    Samuel Humeau, Kurt Shuster, Marie - Anne Lachaux, and Jason Weston. 2020. Poly-encoders: Architectures and pre-training strategies for fast and accurate multi-sentence scoring. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, Apri...

    2020

  12. [20]

    Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2022. Unsupervised dense information retrieval with contrastive learning. In TMLR

    2022

  13. [21]

    Jeff Johnson, Matthijs Douze, and Herv \'e J \'e gou. 2019. Billion-scale similarity search with GPUs . IEEE Transactions on Big Data, 7(3):535--547

    2019

  14. [22]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)...

    2020

  15. [23]

    Omar Khattab and Matei Zaharia. 2020. ColBERT : Efficient and effective passage search via contextualized late interaction over BERT . In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, page 39–48, New York, NY, ...

    2020

  16. [24]

    Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav...

    2019 doi

  17. [25]

    Donald Metzler, Yi Tay, and Dara Bahri. 2021. Rethinking search. ACM SIGIR Forum, 55:1 -- 27

    2021

  18. [26]

    Hellina Hailu Nigatu and Inioluwa Deborah Raji. 2024. https://doi.org/10.1145/3630106.3658546 ``i searched for a religious song in amharic and got sexual content instead'': Investigating online harm in low-resourced languages on youtube . In Proceedings of the 2024 ACM Confere...

    2024

  19. [27]

    Rodrigo Nogueira and Kyunghyun Cho. 2019. Passage re-ranking with BERT . arXiv preprint arXiv:1901.04085

    2019 arXiv

  20. [28]

    Rodrigo Nogueira, Wei Yang, Kyunghyun Cho, and Jimmy Lin. 2019. Multi-stage document ranking with BERT . arXiv preprint arXiv:1910.14424

    2019 arXiv

  21. [29]

    Kelechi Ogueji, Yuxin Zhu, and Jimmy Lin. 2021. https://doi.org/10.18653/v1/2021.mrl-1.11 Small data? No problem! Exploring the viability of pretrained multilingual language models for low-resourced languages . In Proceedings of the 1st Workshop on Multilingual Representation ...

    2021 doi

  22. [30]

    Stephen Della Pietra, Mark Epstein, Salim Roukos, and Todd Ward. 1997. https://doi.org/10.3115/976909.979639 Fertility models for statistical natural language understanding . In Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics and Eighth ...

    1997

  23. [31]

    Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. https://doi.org/10.18653/v1/P19-1493 How multilingual is multilingual BERT ? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4996--5001, Florence, Italy. Association for Compu...

    2019 doi

  24. [32]

    Ronak Pradeep, Kai Hui, Jai Gupta, Adam Lelkes, Honglei Zhuang, Jimmy Lin, Donald Metzler, and Vinh Tran. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.83 How does generative retrieval scale to millions of passages? In Proceedings of the 2023 Conference on Empirical Method...

    2023 doi

  25. [33]

    Nils Reimers and Iryna Gurevych. 2019. https://arxiv.org/abs/1908.10084 Sentence-bert: Sentence embeddings using siamese bert-networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics

    2019 arXiv

  26. [34]

    Robertson and Steve Walker

    Stephen E. Robertson and Steve Walker. 1997. On relevance weights with little relevance information. In Proceedings of the 20th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, page 16–24, New York, NY, USA. Association for Comput...

    1997

  27. [35]

    Robertson and Hugo Zaragoza

    Stephen E. Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Framework: BM25 and Beyond . Foundations and Trends in Information Retrieval. NOW Publishers

    2009

  28. [36]

    Phillip Rust, Jonas Pfeiffer, Ivan Vuli \'c , Sebastian Ruder, and Iryna Gurevych. 2021. https://doi.org/10.18653/v1/2021.acl-long.243 How good is your tokenizer? on the monolingual performance of multilingual language models . In Proceedings of the 59th Annual Meeting of the ...

    2021 doi

  29. [37]

    Walelign Sewunetie, Atnafu Tonja, Tadesse Belay, Hellina Hailu Nigatu, Gashaw Gebremeskel, Zewdie Mossie, Hussien Seid, and Seid Yimam. 2024. https://aclanthology.org/2024.gitt-1.1/ Gender bias evaluation in machine translation for A mharic, T igrigna, and Afaan Oromoo . In Pr...

    2024

  30. [38]

    Lingfeng Shen, Weiting Tan, Sihao Chen, Yunmo Chen, Jingyu Zhang, Haoran Xu, Boyuan Zheng, Philipp Koehn, and Daniel Khashabi. 2024. https://doi.org/10.18653/v1/2024.findings-acl.156 The language barrier: Dissecting safety challenges of LLM s in multilingual contexts . In Find...

    2024 doi

  31. [39]

    Cohen, and Donald Metzler

    Yi Tay, Vinh Quang Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer memory as a differentiable search index. In NeurIPS

    2022

  32. [40]

    James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. Fever: a large-scale dataset for fact extraction and verification. arXiv preprint arXiv:1803.05355

    2018 arXiv

  33. [41]

    Cagri Toraman, Eyup Halit Yilmaz, Furkan S ahin\" u c, and Oguzhan Ozcelik. 2023. https://doi.org/10.1145/3578707 Impact of tokenization on language models: An analysis for Turkish . ACM Trans. Asian Low-Resour. Lang. Inf. Process., 22(4)

    2023 doi

  34. [42]

    Ahmet \"U st \"u n, Gosse Bouma, and Gertjan van Noord. 2019. https://doi.org/10.26615/978-954-452-056-4_140 Cross-lingual word embeddings for morphologically rich languages . In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RAN...

    2019 doi

  35. [43]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual E5 text embeddings: A technical report. arXiv preprint arXiv:2402.05672

    2024 arXiv

  36. [44]

    Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In ICLR

    2021

  37. [45]

    Andrew Yates, Rodrigo Nogueira, and Jimmy Lin. 2021. Pretrained transformers for text ranking: BERT and beyond. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Tutorials, pages 1...

    2021

  38. [46]

    Tilahun Yeshambel, Josiane Mothe, and Yaregal Assabie. 2020. https://doi.org/10.1007/978-3-030-58219-7_5 2AIRTC : The Amharic adhoc information retrieval test collection . In Experimental IR Meets Multilinguality, Multimodality, and Interaction: 11th International Conference o...

    2020 doi

  39. [47]

    Tilahun Yeshambel, Josiane Mothe, and Yaregal Assabie. 2021. Morphologically annotated Amharic text corpora. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2349--2355

    2021

  40. [48]

    Puxuan Yu, Luke Merrick, Gaurav Nuti, and Daniel Campos. 2024. Arctic-Embed 2.0 : Multilingual retrieval without compromise. arXiv preprint arXiv:2412.04506

    2024 arXiv

  41. [49]

    Qingcheng Zeng, Lucas Garay, Peilin Zhou, Dading Chong, Yining Hua, Jiageng Wu, Yikang Pan, Han Zhou, Rob Voigt, and Jie Yang. 2023. https://doi.org/10.24963/ijcai.2023/698 GreenPLM : cross-lingual transfer of monolingual pre-trained language models at almost no cost . In Proc...

    2023 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.