Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Hierarchical Multi-field Representations for Two-Stage E-commerce Retrieval

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read CHARM uses a block-triangular attention mask to produce hierarchical field-level product embeddings in one BERT forward pass, matching or beating baselines on the Multi-Aspect Amazon Shopping Queries dataset.

desk verdict Block-triangular attention over product fields is a real and useful idea, but the abstract overclaims against MURAL; fix the headline and add error bars. read the letter →

arxiv 2501.18707 v1 pith:W5CD3MUR submitted 2025-01-30 cs.IR

Niklas Freymuth , Dong Liu , Thomas Ricatte , Saab Mansour This is my paper · ORCID
classification cs.IR
keywords ProductRetrievalDenseE-commerceSearchMulti-FieldRepresentationsBlockAttentionHierarchicalTwo-StageField-LevelEmbeddings
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 tries to show that a single transformer encoder can make e-commerce products searchable at several levels of detail at once. It proposes CHARM, which changes the attention mask so that tokens in a later product field can look back at all earlier fields but not forward, producing a separate embedding for each field plus one aggregated embedding. The aggregated embedding finds a shortlist of candidates cheaply, and the field embeddings re-rank that shortlist by the best matching field. On English, Spanish, and Japanese subsets of the Multi-Aspect Amazon Shopping Queries dataset, CHARM matches or outperforms strong bi-encoder and multi-representation baselines while requiring far fewer similarity comparisons than full field-by-field search.

What carries the argument

The load-bearing object is the block-triangular attention mask defined by $M_{i,j}=0$ if $F(i)\ge F(j)$ and $M_{i,j}=-\infty$ otherwise, where $F$ maps a token to its product-field index and fields are ordered from most compressed (color, brand) to most detailed (bullet points, description). This mask lets tokens in a field attend to their own field and every earlier field while blocking later fields, so the special token at each field boundary accumulates a prefix of the product's information. The second piece is the aggregated representation $h_p = \sum_f w_f h_{p,f}$ with $w_f = \mathrm{softmax}(K h_{\mathrm{CLS}})_f$, which is trained to preserve the best field-level matches and supports the first retrieval stage. Together these pieces give the two-stage pipeline its cost profile $O(N(M+k|F|))$ rather than $O(NM|F|)$.

What would settle it

A reproduction that fixes the query encoding three ways—symmetric field-mirroring, plain CLS, and a learned query aggregation—and reports two-stage metrics on the same dataset would settle whether the hierarchy itself, rather than the query encoding, drives the observed gains.

Watch

Extended reading notes

Core claim

The central claim is that the block-triangular attention mask, $M_{i,j}=0$ when the field of token $i$ is at or after the field of token $j$, and $-\infty$ otherwise, turns one BERT forward pass into a cascade of field-level retrieval vectors with progressively finer detail. Each field is prefixed with a special token whose final hidden state is the vector for that field, and the model also learns a softmax-weighted aggregate of those vectors. Training combines three InfoNCE losses: one on the aggregate, one on every field, and one on the best-matching field. At evaluation, the aggregate is used for first-stage candidate selection and the maximum field similarity over the shortlist decides the final ranking. On the three language subsets, CHARM's best-field and two-stage evaluations match or beat the MVR, MURAL, and BiBERT baselines, and ablations show that the triangular mask, the aggregate loss, and the symmetric query encoding each contribute.

Load-bearing premise

The gains depend on a symmetric query encoder that mirrors the product's field-special-token structure, a construction the paper leaves to an appendix and whose removal (CLS only) measurably hurts performance.

Editorial extensions

If this is right

  • A single shared BERT backbone can produce both a cheap aggregated vector and several field-specific vectors, so catalog-scale product retrieval no longer has to choose between speed and field-level precision.
  • Because each field vector carries information only up to that field, the model can report which field matched a query, giving search systems a natural explanation for a product's retrieval.
  • The two-stage cost formula $O(N(M+k|F|))$ means the added accuracy of multi-field matching is available at roughly the same cost as ordinary dense retrieval when $k$ is small relative to the corpus size.
  • On the evaluated datasets, deeper fields such as bullet points and descriptions match longer, more specific queries, so fine-grained product details become reachable by detailed search intents.

Reading between the lines

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

  • The query encoder's mirroring of product field special tokens is the least specified component: the paper defers details to an appendix, and its ablation shows that dropping this symmetry (using the CLS token instead) hurts results, so the reported gains should be read as contingent on that choice.
  • The block-triangular mask is a special case of a directed attention graph; extending it to arbitrary field subsets, such as letting description attend to category but not to brand, is a natural next step that could make the hierarchy data-driven rather than length-ordered.
  • Because the first retrieval stage is a single aggregated vector, the method should combine with approximate nearest-neighbor indexes for billion-scale catalogs, where the exact k-NN used in the paper would become the bottleneck.
  • The query-field alignment analysis suggests a testable extension: using the matched field as a signal in downstream ranking or as a product-type filter could improve precision beyond the two-stage score.
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

4 major / 5 minor

Summary. The paper introduces CHARM, a dual-encoder retrieval model that replaces BERT's full attention mask with a block-triangular mask over e-commerce product fields. Each field representation attends to its own and earlier fields, producing a cascade of field-level embeddings plus a learned aggregated embedding, and retrieval is performed in two stages: an aggregated-vector shortlist followed by field-level reranking. The authors evaluate on the Multi-Aspect Amazon Shopping Queries dataset and report comparisons against BiBERT, MVR, and MURAL, along with ablations and an explainability analysis of field-query alignment.

Significance. The architectural idea is attractive: a single forward pass generates multiple structured representations with a controllable information hierarchy, and the two-stage pipeline has a clear complexity advantage over multi-representation baselines. The paper also provides a thoughtful ablation design and an analysis of field-query match patterns that supports the interpretability claims. However, the current manuscript overstates its headline comparison, lacks statistical support for small performance differences, and leaves at least one load-bearing evaluation detail underspecified. With revision, the contribution could be a solid addition to product retrieval research.

major comments (4)
  1. [Abstract; Table 1; Section 5.2] The abstract's claim that CHARM 'matches or outperforms state-of-the-art baselines' is not supported by the evidence in Table 1 as presented. MURAL [52] is listed as a baseline in Table 1, where its US R@10 is 63.71 against CHARM's best 34.78; the paper itself notes in Section 4.3 that these results 'are not directly comparable due to the different training setups,' but this caveat is not carried into the abstract or the Section 5.2 narrative. Please restrict the headline claim to the directly comparable baselines (BiBERT, MVR) or provide a matched MURAL comparison.
  2. [Section 5; Tables 1 and 2] All results are reported as single numbers without variance estimates, confidence intervals, or significance tests. The ablations in Table 2 contain differences as small as +0.02 (Added LDiv), and the main-table margins over BiBERT and MVR are often under one point. Without multiple seeds or paired significance tests, the claims that each loss term 'uniquely contributes' and that CHARM 'outperforms' the comparable baselines are not statistically grounded. Please report means and standard deviations over at least three runs, or provide a paired significance test.
  3. [Section 3.2; Appendix A; Table 2] The symmetric query encoding is underspecified. The text says 'We mirror the special tokens and their aggregated representation for the query tokenization and encoder' and defers details to Appendix A, but Appendix A only visualizes the block-triangular attention mask for product fields; it does not state how a query without product fields is segmented, which special tokens are inserted, or how the aggregated query vector is computed. Since the 'Asym. Encoders' ablation in Table 2 shows a performance drop of 0.29-0.40 points when the CLS token is used for the query, the reported results depend on exactly this unspecified construction. Please give the complete query token sequence and aggregation formula.
  4. [Section 5.1; Table 1] The P@10 metric is not reproducible as reported. Section 5.1 states that Precision@{5,10} is evaluated 'by using a classifier model trained to predict if a query-product pair is exact or not,' but gives no architecture, training data, positive/negative sampling, or threshold. Because P@10 appears in the main comparison table and is used in the 'matches or outperforms' narrative, please either describe the classifier in full or remove the metric.
minor comments (5)
  1. [Table 1] The MURAL rows list only two numbers (63.71, 42.28) without column labels; please specify that these are R@10 and NDCG@50 for the US subset, and mark unavailable cells with an em dash.
  2. [Section 3.1, Eq. (1)] The index convention for the attention mask is inconsistent: Eq. (1) uses M_{j,i} while Eq. (3) defines M_{i,j}; please align the notation.
  3. [Appendix A, Figure 7 caption] The caption describes the matrix as 'block-diagonal,' but the method and Eq. (3) define a block-triangular mask; please correct the terminology.
  4. [Section 5.1] The sentence 'Is this is not available' should read 'Since this is not available.'
  5. [Table 1 header] The JP column header 'NCDG@50' should be 'NDCG@50.'

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation chain: CHARM is trained end-to-end on a public benchmark and evaluated on held-out labels; the flagged issues are an overbroad SOTA claim and an underspecified query encoding, which affect correctness and reproducibility rather than circularity.

full rationale

The paper's central chain is not circular. The block-triangular attention mask (Equation 3) defines which tokens each field representation can see; the field-level vectors are read out from special tokens (Equation 4); the aggregated vector is a learned softmax-weighted combination of those field vectors; and all three loss terms (LAgg, LFields, LMax in Equation 6) are trained with InfoNCE against query-product pairs from a public dataset. Evaluation then measures R@10, R@100, NDCG@50, and P@10 on held-out labels, so no metric is constructed from a fitted parameter renamed as a prediction. The two-stage retrieval is an algorithmic combination of the same trained vectors, not a separate fitted quantity. The only load-bearing design choice flagged by the paper itself is the symmetric query encoding, i.e., 'We mirror the special tokens and their aggregated representation for the query tokenization and encoder' (Section 3.2, Evaluation), whose removal degrades performance (Section 5.3: 'Using the processed CLS token for the query representation degrades performance'). That is an empirical design choice with missing implementation detail (deferred to Appendix A, which is not present), but it is not a definitional reduction of the reported results. Similarly, the abstract's 'matches or outperforms state-of-the-art baselines' is overbroad given Table 1, where MURAL scores R@10=63.71 on US versus CHARM's 34.78, and the caveat 'results are not directly comparable due to the different training setups' appears only in the footnote. This is a correctness/overclaim issue, not circularity. No load-bearing argument reduces to a self-citation: MURAL and MADRAL are external prior works with no author overlap with this paper, and the paper does not invoke any uniqueness theorem from its own authors. The analysis sections describe the model's own representations and matching behavior, which is descriptive rather than predictive. Overall, the derivation is self-contained against external benchmarks, so the circularity score is low despite the flagged overclaim and missing appendix.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method adds no new physical or mathematical entities. It relies on learned network parameters and a small set of design choices: the field hierarchy order, shortlist size, loss weights, dataset-specific temperature, and an undescribed P@10 classifier. The central design axioms are that field order by average length reflects information content, that query and product can share a symmetric special-token encoding, and that the Amazon labels and the P@10 classifier are reliable.

free parameters (5)
  • Field hierarchy order = Color, Brand, (Category), Title, Description, Bullet Points
    Ordering by average field length is the paper's approximation of information content (Section 3.1). The ablation 'Other Field Order' in Table 2 shows a different order lowers performance, so this choice is load-bearing and selected by the authors.
  • Shortlist size k = 100
    Used for two-stage retrieval in all reported results (Section 5.1). No sensitivity analysis is given, so the accuracy and cost tradeoff is not characterized.
  • InfoNCE temperature tau = 0.1 (US), 1.0 (ES and JP)
    Set per dataset in Table 7 without explanation. The dataset-specific values are likely chosen by validation or prior practice, making cross-dataset comparisons dependent on this choice.
  • Loss weights lambda_Agg, lambda_Fields, lambda_Max = 1.0 for all three
    Fixed for simplicity in Section 4.2. The ablation in Table 2 shows each loss term contributes, but the weights are not tuned and no sensitivity is reported.
  • P@10 exactness classifier = Not described
    Section 5.1 reports Precision@10 using a classifier trained to predict exactness, but the classifier's architecture, training data, and accuracy are not given. This learned component is a hidden variable in the reported P@10 results.
assumptions (4)
  • ad hoc to paper Product fields ordered by average length form a valid hierarchy from compressed to detailed information.
    Introduced in Section 3.1 as the motivation for the attention mask. No external evidence shows that length implies information content, and the alternative order in Section 5.3 changes results.
  • domain assumption Query and product encoders can share weights and use the same symmetric special-token representation construction.
    Relied on in Section 3.2 Evaluation. The query has no product fields yet is assumed to mirror product special tokens, and the exact construction is only in the missing Appendix A.
  • standard math InfoNCE contrastive loss with one sampled positive and in-batch negatives is a sufficient training objective for retrieval.
    Standard practice in dense retrieval, described in Section 3.1. It assumes the batch negatives are representative enough for learning a good metric space.
  • domain assumption The labeled exact, substitute, complementary, and irrelevant pairs in the Amazon Shopping Queries dataset are correct, and the P@10 classifier is accurate.
    Evaluation in Section 4.1 and Section 5.1 relies on label quality. The P@10 classifier is not described or released.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Multi-field Representations for Two-Stage E-commerce Retrieval." pith.science (2026). https://pith.science/paper/W5CD3MUR

@misc{pith2026250118707,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Multi-field Representations for Two-Stage E-commerce Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W5CD3MUR}},
  note         = {Machine review of arXiv:2501.18707}
}
read the original abstract

Dense retrieval methods typically target unstructured text data represented as flat strings. However, e-commerce catalogs often include structured information across multiple fields, such as brand, title, and description, which contain important information potential for retrieval systems. We present Cascading Hierarchical Attention Retrieval Model (CHARM), a novel framework designed to encode structured product data into hierarchical field-level representations with progressively finer detail. Utilizing a novel block-triangular attention mechanism, our method captures the interdependencies between product fields in a specified hierarchy, yielding field-level representations and aggregated vectors suitable for fast and efficient retrieval. Combining both representations enables a two-stage retrieval pipeline, in which the aggregated vectors support initial candidate selection, while more expressive field-level representations facilitate precise fine-tuning for downstream ranking. Experiments on publicly available large-scale e-commerce datasets demonstrate that CHARM matches or outperforms state-of-the-art baselines. Our analysis highlights the framework's ability to align different queries with appropriate product fields, enhancing retrieval accuracy and explainability.

Figures

Figures reproduced from arXiv: 2501.18707 by the authors.

Figure 1
Figure 1. Retrieval with CHARM. Given a product with mul [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Schematic overview of CHARM. Our model receives a tokenized product, including a special token per product [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Average length of queries matching a product field [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Average entropy of product type distributions across [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Preservation of ‘best‘ matches in two-stage retrieval [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Exemplary block-diagonal attention matrix. Each [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Modeling shopper interest broadness with entropy-driven dialogue policy in the context of arbitrarily large product catalogs

    cs.IR 2025-09 conditional novelty 5.0 of 10

    Retrieval-score entropy, normalized over the top 50 candidates, routes a conversational recommender between direct recommendations and clarifying questions, and the production AB test shows longer conversations.

Reference graph

Works this paper leans on

64 extracted references · 39 canonical work pages · cited by 1 Pith paper

  1. [52]

    Xiaojie Sun, Keping Bi, Jiafeng Guo, Sihui Yang, Qishen Zhang, Zhongyi Liu, Guannan Zhang, and Xueqi Cheng. 2024. A Multi-Granularity-Aware Aspect Learning Model for Multi-Aspect Dense Retrieval. InProceedings of the 17th ACM International Conference on Web Search and Data Mining . 674–682

    work page 2024

  2. [1]

    Amin Abolghasemi, Suzan Verberne, and Leif Azzopardi. 2022. Improving BERT- based query-by-document retrieval with multi-task optimization. In European Conference on Information Retrieval . Springer, 3–12

    work page 2022

  3. [2]

    Ricardo Baeza-Yates, Berthier Ribeiro-Neto, et al . 1999. Modern information retrieval. Vol. 463

    work page 1999

  4. [3]

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural Machine Translation by Jointly Learning to Align and Translate. InInternational Conference on Learning Representations (ICLR)

    work page 2015

  5. [4]

    Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah

  6. [5]

    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 American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), Jill Burstein, Christy...

    doi:10.18653/v1/n 2019

  7. [6]

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. The Faiss library. (2024). arXiv:2401.08281 [cs.LG]

    arXiv 2024

  8. [7]

    Yixing Fan, Xiaohui Xie, Yinqiong Cai, Jia Chen, Xinyu Ma, Xiangsheng Li, Ruqing Zhang, Jiafeng Guo, et al . 2022. Pre-training methods in information retrieval. Foundations and Trends® in Information Retrieval 16, 3 (2022), 178–317

    work page 2022

Show all 64 references
  1. [8]

    Luyu Gao and Jamie Callan. 2021. Condenser: a pre-training architecture for dense retrieval. arXiv preprint arXiv:2104.08253 (2021)

    2021 arXiv

  2. [9]

    Lin, and Jamie Callan

    Luyu Gao, Xueguang Ma, Jimmy J. Lin, and Jamie Callan. 2022. Tevatron: An Efficient and Flexible Toolkit for Dense Retrieval. ArXiv abs/2203.05765 (2022)

    2022 arXiv

  3. [10]

    Luyu Gao, Yunyi Zhang, Jiawei Han, and Jamie Callan. 2021. Scaling Deep Contrastive Learning Batch Size under Memory Limited Setup. In6th Workshop on Representation Learning for NLP, RepL4NLP 2021 . Association for Computational Linguistics (ACL), 316–321

    2021

  4. [12]

    Jiafeng Guo, Yinqiong Cai, Yixing Fan, Fei Sun, Ruqing Zhang, and Xueqi Cheng

  5. [13]

    Raia Hadsell, Sumit Chopra, and Yann LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06) , Vol. 2. IEEE, 1735–1742

    2006

  6. [14]

    Kailash A Hambarde and Hugo Proenca. 2023. Information retrieval: recent advances and beyond. IEEE Access (2023)

    2023

  7. [15]

    ACM Transactions on Information Systems (TOIS) 40, 4 (2022), 1–42

    Semantic models for the first-stage retrieval: A comprehensive review. ACM Transactions on Information Systems (TOIS) 40, 4 (2022), 1–42

    2022

  8. [16]

    Sebastian Hofstätter, Sheng-Chieh Lin, Jheng-Hong Yang, Jimmy Lin, and Allan Hanbury. 2021. Efficiently teaching an effective dense retriever with balanced topic aware sampling. In Proceedings of the 44th International ACM SIGIR Confer- ence on Research and Development in Info...

    2021

  9. [17]

    Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning deep structured semantic models for web search using clickthrough data. In Proceedings of the 22nd ACM international conference on Information & Knowledge Management . 2333–2338

    2013

  10. [18]

    Yunzhong He, Yuxin Tian, Mengjiao Wang, Feier Chen, Licheng Yu, Mao- long Tang, Congcong Chen, Ning Zhang, Bin Kuang, and Arul Prakash. 2023. Que2engage: Embedding-based retrieval for relevant and engaging products at facebook marketplace. In Companion Proceedings of the ACM W...

    2023

  11. [19]

    Nan Jiang, Dhivya Eswaran, Choon Hui Teo, Yexiang Xue, Yesh Dattatreya, Sujay Sanghavi, and Vishy Vishwanathan. 2022. On the Value of Behavioral Representations for Dense Retrieval. arXiv preprint arXiv:2208.05663 (2022)

    2022 arXiv

  12. [20]

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data 7, 3 (2019), 535–547

    2019

  13. [21]

    Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Baner- jee, and Fillia Makedon. 2020. A survey on contrastive self-supervised learning. Technologies 9, 1 (2020), 2

    2020

  14. [22]

    Akshay Kekuda, Yuyang Zhang, and Arun Udayashankar. 2024. Embedding based retrieval for long tail search queries in ecommerce. In Proceedings of the 18th ACM Conference on Recommender Systems (Bari, Italy) (RecSys ’24). Association for Computing Machinery, New York, NY, USA, 7...

    2024

  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. 39–48

    2020

  16. [24]

    Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open- domain question answering. arXiv preprint arXiv:2004.04906 (2020)

    2020 arXiv

  17. [25]

    Weize Kong, Swaraj Khadanga, Cheng Li, Shaleen Kumar Gupta, Mingyang Zhang, Wensong Xu, and Michael Bendersky. 2022. Multi-aspect dense retrieval. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3178–3186

    2022

  18. [26]

    Vihan Lakshman, Choon Hui Teo, Xiaowen Chu, Priyanka Nigam, Abhinandan Patni, Pooja Maknikar, and SVN Vishwanathan. 2021. Embracing structure in data for billion-scale semantic product search. arXiv preprint arXiv:2110.06125 (2021)

    2021 arXiv

  19. [27]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Opti- mization. International Conference on Learning Representations (ICLR) (2015)

    2015

  20. [28]

    Millicent Li, Tongfei Chen, Benjamin Van Durme, and Patrick Xia. 2024. Multi- Field Adaptive Retrieval. arXiv preprint arXiv:2410.20056 (2024)

    2024 arXiv

  21. [29]

    Rui Li, Yunjiang Jiang, Wenyun Yang, Guoyu Tang, Songlin Wang, Chaoyi Ma, Wei He, Xi Xiong, Yun Xiao, and Eric Yihong Zhao. 2019. From Semantic Retrieval to Pairwise Ranking: Applying Deep Learning in E-commerce Search (SIGIR’19). Association for Computing Machinery, New York,...

    2019

  22. [30]

    Hang Li and Jun Xu. 2014. Semantic Matching in Search.Foundations and Trends® in Information Retrieval 7, 5 (2014), 343–469. https://doi.org/10.1561/1500000035

    2014 doi

  23. [31]

    Yizhi Li, Zhenghao Liu, Chenyan Xiong, and Zhiyuan Liu. 2021. More robust dense retrieval with contrastive dual learning. In Proceedings of the 2021 ACM SIGIR International Conference on Theory of Information Retrieval . 287–296

    2021

  24. [32]

    Jimmy Lin, Rodrigo Nogueira, and Andrew Yates. 2022. Pretrained transformers for text ranking: Bert and beyond . Springer Nature

    2022

  25. [33]

    Sen Li, Fuyu Lv, Ruqing Zhang, Dan Ou, Zhixuan Zhang, and Maarten de Rijke

  26. [34]

    Xinyu Ma, Jiafeng Guo, Ruqing Zhang, Yixing Fan, and Xueqi Cheng. 2022. Pre-train a discriminative text encoder for dense retrieval via contrastive span prediction. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retriev...

    2022

  27. [35]

    Alessandro Magnani, Feng Liu, Min Xie, and Somnath Banerjee. 2019. Neural product retrieval at walmart. com. In Companion Proceedings of The 2019 World Wide Web Conference. 367–372

    2019

  28. [36]

    YA Malkov and DA Yashunin. 2018. Efficient and Robust Approximate Near- est Neighbor Search Using Hierarchical Navigable Small World Graphs. IEEE Transactions on Pattern Analysis and Machine Intelligence 42, 4 (2018), 824–836

    2018

  29. [37]

    Chen Luo, Xianfeng Tang, Hanqing Lu, Yaochen Xie, Hui Liu, Zhenwei Dai, Limeng Cui, Ashutosh Joshi, Sreyashi Nag, Yang Li, et al. 2024. Exploring Query Understanding for Amazon Product Search. In 2024 IEEE International Conference on Big Data (BigData) . IEEE, 2343–2348

    2024

  30. [38]

    Franco Maria Nardini, Cosimo Rulli, and Rossano Venturini. 2024. Efficient Multi- vector Dense Retrieval with Bit Vectors. In European Conference on Information Retrieval. Springer, 3–17

    2024

  31. [39]

    Priyanka Nigam, Yiwei Song, Vijai Mohan, Vihan Lakshman, Weitian (Allen) Ding, Ankit Shingavi, Choon Hui Teo, Hao Gu, and Bing Yin. 2019. Semantic Product Search. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (Anchorage, AK...

    2019

  32. [40]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    2018 arXiv

  33. [41]

    Aashiq Muhamed, Sriram Srinivasan, Choon-Hui Teo, Qingjun Cui, Belinda Zeng, Trishul Chilimbi, and SVN Vishwanathan. 2023. Web-scale semantic product search with large language models. In Pacific-Asia Conference on Knowledge Discovery and Data Mining . Springer, 73–85. Under R...

    2023

  34. [42]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research 21, 140 (2020), 1–67

    2020

  35. [43]

    Chandan K Reddy, Lluís Màrquez, Fran Valero, Nikhil Rao, Hugo Zaragoza, Sambaran Bandyopadhyay, Arnab Biswas, Anlu Xing, and Karthik Subbian. 2022. Shopping queries dataset: A large-scale ESCI benchmark for improving product search. arXiv preprint arXiv:2206.06588 (2022)

    2022 arXiv

  36. [44]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In Conference on Empirical Methods in Natural Language Processing. https://api.semanticscholar.org/CorpusID:201646309

    2019

  37. [45]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing system...

    2019

  38. [46]

    Stephen Robertson, Hugo Zaragoza, et al . 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends® in Information Retrieval 3, 4 (2009), 333–389

    2009

  39. [47]

    Nicholas Rossi, Juexin Lin, Feng Liu, Zhen Yang, Tony Lee, Alessandro Magnani, and Ciya Liao. 2024. Relevance Filtering for Embedding-based Retrieval. In Pro- ceedings of the 33rd ACM International Conference on Information and Knowledge Management (Boise, ID, USA) (CIKM ’24)....

    2024

  40. [48]

    Gerard Salton and Christopher Buckley. 1988. Term-weighting approaches in automatic text retrieval. Information processing & management 24, 5 (1988), 513–523

    1988

  41. [49]

    Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, Qiaoqiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021. RocketQAv2: A Joint Training Method for Dense Passage Retrieval and Passage Re-ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language P...

    2021

  42. [50]

    Kihyuk Sohn. 2016. Improved deep metric learning with multi-class n-pair loss objective. Advances in neural information processing systems 29 (2016)

    2016

  43. [51]

    Xiaojie Sun, Keping Bi, Jiafeng Guo, Xinyu Ma, Yixing Fan, Hongyu Shan, Qishen Zhang, and Zhongyi Liu. 2023. Pre-training with Aspect-Content Text Mutual Prediction for Multi-Aspect Dense Retrieval. In Proceedings of the 32nd ACM In- ternational Conference on Information and K...

    2023

  44. [53]

    Sivic and Zisserman. 2003. Video Google: A text retrieval approach to object matching in videos. InProceedings ninth IEEE international conference on computer vision. IEEE, 1470–1477

    2003

  45. [54]

    Deepak Vishwakarma and Suresh Kumar. 2025. Fine-Tuned BERT Algorithm- Based Automatic Query Expansion for Enhancing Document Retrieval System. Cognitive Computation 17, 1 (2025), 1–16

    2025

  46. [55]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2020. Transformers: State-of-the-Art Natural Language Processing. In Proceedings of the 2020 Conference on Empirical Methods...

    2020

  47. [56]

    Zhibo Xiao, Luwei Yang, Wen Jiang, Yi Wei, Yi Hu, and Hao Wang. 2020. Deep Multi-Interest Network for Click-through Rate Prediction. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (Virtual Event, Ireland) (CIKM ’20). Association f...

    2020

  48. [57]

    Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N

    Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All You Need. In Advances in Neural Information Processing Systems . https://arxiv.org/ abs/1706.03762

    2017 arXiv

  49. [58]

    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 , Greg K...

    2021 doi

  50. [59]

    Shunyu Zhang, Yaobo Liang, Ming Gong, Daxin Jiang, and Nan Duan. 2022. Multi- View Document Representation Learning for Open-Domain Dense Retrieval. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Smaranda Mur...

    2022 doi

  51. [60]

    Wayne Xin Zhao, Jing Liu, Ruiyang Ren, and Ji-Rong Wen. 2024. Dense text retrieval based on pretrained language models: A survey. ACM Transactions on Information Systems 42, 4 (2024), 1–60

    2024

  52. [61]

    Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate Nearest Neighbor Neg- ative Contrastive Learning for Dense Text Retrieval. In International Conference on Learning Representations

    2021

  53. [65]

    (B)rand",

    Jianghong Zhou, Bo Liu, Jhalak Nilesh Acharya, Yao Hong, Kuang-chih Lee, and Musen Wen. 2023. Leveraging Large Language Models for Enhanced Product Descriptions in eCommerce. In The 2023 Conference on Empirical Methods in Natural Language Processing. 88. Hierarchical Multi-fie...

    2023

  54. [1993]

    siamese

    Signature verification using a" siamese" time delay neural network. Ad- vances in neural information processing systems 6 (1993)

    1993

  55. [2022]

    ACM Trans

    Semantic Models for the First-Stage Retrieval: A Comprehensive Review. ACM Trans. Inf. Syst. 40, 4, Article 66 (March 2022), 42 pages. https://doi.org/10. 1145/3486250

    2022

  56. [2024]

    In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Barcelona, Spain) (KDD ’24)

    Text Matching Indexers in Taobao Search. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Barcelona, Spain) (KDD ’24). Association for Computing Machinery, New York, NY, USA, 5339–5350. https://doi.org/10.1145/3637528.3671654

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.