Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Multi-objective Aligned Bidword Generation Model for E-commerce Search Advertising

T0 review · 4 major / 5 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 One bidword model lifts ad revenue 2.13% in a live e-commerce test

desk verdict Real deployed system and credible online A/B, but the offline retrieval evaluation is likely circular and has a data inconsistency. read the letter →

arxiv 2506.03827 v1 pith:7F7RUW6C submitted 2025-06-04 cs.CL cs.AIcs.IR

Zhenhui Liu , Chunyuan Yuan , Ming Pang , Zheng Fang , Li Yuan , Xue Jiang , Changping Peng , Zhangang Lin
show 2 more authors
Zheng Luo Jingping Shao
This is my paper
classification cs.CLcs.AIcs.IR
keywords queryrewritingbidwordgenerationmulti-objectivepreferencealignmentsponsoredsearche-commerceadvertisinglargelanguagemodeldirectoptimizationlong-tailqueries
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

MoBGM is a bidword-generation model for e-commerce search advertising: it rewrites a shopper's query into bidwords that can actually retrieve ads. The paper's central claim is that one generator can simultaneously optimize three objectives—relevance of the rewrite to the original query, authenticity of the rewrite as a bidword a user would really search, and the advertising revenue it can earn—and that this beats existing query-rewriting methods. Offline, it reports higher Recall@k and NDCG@k than mining-, matching-, and generation-based baselines; in a live A/B test on a shopping platform, it reports +2.13% ad revenue. If the claim holds, query rewriting no longer has to choose between natural rewrites and commercial value.

What carries the argument

The load-bearing mechanism is a discriminator that serves as a multi-head reward model. Its relevance head is a BERT-style classifier over four rewrite relations (synonym, hypernym-to-hyponym, hyponym-to-hypernym, incorrect); its authenticity head is a binary classifier trained on labels derived from search-frequency thresholds; its value head regresses each bidword's CPM. The discriminator produces three reward signals, which are turned into pairwise winner–loser preferences and combined in a MODPO-style alignment loss that updates the generator. This arrangement lets the model optimize relevance, authenticity, and revenue together while avoiding RL training instability.

What would settle it

Compare the authenticity labels used for training against human judgments on a sample of low-frequency queries: if a large share of rewrites that humans call valid, useful bidwords are labeled negative by the frequency threshold, the reward signal is mis-specified and the claimed long-tail benefit should shrink or reverse in a tail-only A/B test.

Watch

Extended reading notes

Core claim

The paper proposes that query rewriting in sponsored search should be treated as multi-objective bidword generation, and that all three objectives can be aligned in a single model without reinforcement learning. A discriminator scores each query–bidword pair on relevance, authenticity, and CPM-based advertising value; those scores become pairwise preference rewards. A multi-objective preference-alignment loss, built on MODPO, then trains the generator to maximize the combined reward, with weights $w_{rel}=0.5$, $w_{au}=0.2$, and $w_{val}=0.3$. The result, MoBGM, outperforms all baselines on the golden set and, when deployed on the shopping platform, lifts ad impressions, CPC, CPM, and ad revenue compared with the previous online model.

Load-bearing premise

The authenticity labels treat 'a bidword a real user would search for' as the same thing as 'search frequency above a threshold', so the model is trained to avoid exactly the rare but valuable long-tail bidwords if that proxy is wrong.

Editorial extensions

If this is right

  • Deploying the model requires no online reinforcement learning: the discriminator's scores become pairwise preferences used in a DPO-style loss, so the large generator can be trained stably and then distilled to a smaller serving model.
  • Because the relevance head covers four rewrite relations, the model can deliberately choose synonym, hypernym, or hyponym rewrites, letting operators control the specificity of the bidword.
  • The caching scheme—pre-generated bidwords for top queries and live generation for the rest—means the same model serves head and tail traffic without adding latency.
  • Online, the gains decompose: fine-tuning drives impressions while multi-objective alignment drives CPC and CPM, so each training stage contributes a different revenue channel.

Reading between the lines

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

  • If the authenticity proxy mislabels genuinely useful rare bidwords as fake, the model may systematically avoid the long tail it claims to rescue; a human-labeled validation of the frequency threshold would reveal how much of the long-tail benefit is real.
  • Optimizing toward CPM as the value signal rewards bidwords with high cost per mille, which may push generation toward competitive head phrases and away from cheaper niche queries, especially with the chosen value weight of 0.3.
  • The same discriminator-reward pattern could be reused for other constrained generation tasks in e-commerce—product titles, ad creatives, or category tags—wherever a scalar business metric can be regressed and paired with pairwise preference data.
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 proposes MoBGM, a query-bidword generation model for e-commerce sponsored search. The system consists of an LLM generator (Llama-3-Chinese-8B-Instruct) that is post-pretrained on query-title data, fine-tuned on mined query-bidword pairs, and then aligned through a MODPO-style multi-objective preference loss. A BERT-based discriminator provides three reward signals: query-bidword relevance, bidword authenticity (defined by search frequency), and CPM-based advertising value. Offline experiments on JD click-log data compare MoBGM with mining-, matching-, and generation-based baselines using Recall@k, NDCG@k, and human-evaluated relevance/authenticity. An online A/B test on 5% of traffic reports a statistically significant +2.13% ad-revenue improvement, and the model is described as deployed in production.

Significance. If the offline claims are valid, the main contribution is a practical, deployable framework that unifies relevance, authenticity, and revenue objectives in one generation pipeline, with strong commercial evidence from the online A/B test. The online experiment is an independent validation of revenue/CPM gains and is the strongest part of the paper. The paper does not provide code or a reproducible evaluation harness, and the offline evaluation is weakened by an undisclosed golden-set heuristic and missing statistical reporting; these issues must be addressed for the offline superiority claim to be convincing.

major comments (4)
  1. [Section 4.1] The 'heuristic policy' used to identify the 10 most valuable bidwords from click logs is never disclosed. Since the CPM head in Eq. (5) is trained on click-log CPM and the authenticity head in Section 3.4.2 is trained on search-frequency thresholds, the golden labels may be constructed from exactly the signals that MoBGM is optimized to predict. In that case, Recall@k and NDCG@k measure how well the model reproduces a CPM/frequency heuristic rather than independent quality, and baselines that do not optimize CPM are systematically disadvantaged. The ablation 'w/o. CPM' in Table 3 shows Recall@3 dropping from 15.18 to 14.24, which is consistent with this concern. Please disclose the heuristic and, if it uses CPM or search frequency, re-evaluate MoBGM on labels constructed from independent sources.
  2. [Section 3.4.2] The authenticity label is defined by whether a query 'can match a bidword and the search frequency exceeds a threshold,' so rare but genuine long-tail bidwords are labeled negative by construction. This conflicts with the paper's stated motivation of handling long-tail queries in Section 1, and it means the authenticity reward may suppress the very long-tail bidwords the model is supposed to generate. Please report the threshold, analyze the long-tail distribution of positive and negative labels, and include a long-tail-stratified evaluation.
  3. [Tables 2 and 3] The same model MoBGM is reported with Relevance 97.49 in Table 2 and 92.49 in Table 3, with no explanation for the discrepancy. In addition, Section 4.5.1 asserts that MoBGM 'significantly outperforms' the baselines, but no significance tests, confidence intervals, or error bars are provided for any offline metric. This makes the offline superiority claim hard to verify and needs to be fixed.
  4. [Section 4.6.2] The online A/B test is a valuable independent validation, but it compares MoBGM only with the previous online model and with variants of MoBGM, not with the offline baselines such as EEQR and BEQUE. The paper should therefore not present the online results as direct confirmation that MoBGM outperforms those baselines on retrieval quality; the online evidence supports revenue/CPM gains but does not by itself validate the recall/NDCG superiority claim.
minor comments (5)
  1. [Section 4.1 and Table 1] The Preference Alignment Dataset is described as 990,000 queries (360,000 sampled from each of head, middle, and tail), but Table 1 reports 984,271 training and 49,384 validation queries, totaling 1,033,655. Please reconcile these numbers.
  2. [Equation (9)] The placement of beta_w and beta_l as coefficients on the log-ratio terms divided by w_rel is confusing; please clarify whether these are separate KL coefficients for winner and loser and how the formula follows from MODPO.
  3. [Section 3.4.2] Please specify whether the frequency threshold is applied to the query or to the bidword, and how the threshold value was chosen.
  4. [Section 4.3] The human-evaluated Relevance and Authenticity metrics do not report the number of annotators, the number of annotated samples, or inter-annotator agreement; please add these details.
  5. [Throughout] There are minor typos and wording issues, such as 'post pre-train' in Section 1 and 'advertises buys' in Section 3.1.

Circularity Check

1 steps flagged · score 6.0 of 10

Offline Recall/NDCG evaluation is partially circular: the golden 'most valuable bidwords' are chosen by an undisclosed click-log heuristic while MoBGM is trained to predict CPM and search-frequency from the same click logs; online and human metrics remain independent.

  1. self definitional [Section 4.1 (Golden Dataset) vs Section 3.4.3 (Eq. 5) and Section 3.4.2 (Eq. 4)]
    "For the sampled 90000 queries distinct from the Preference Alignment Dataset, we use a heuristic policy to identify the 10 most valuable bidwords from the click logs. This set will be used as the golden standard to evaluate performance. ... We use the CPM ... as an optimization objective. We collect bidword data from the online click logs and compute the CPM value for every bidword. ... When a query can match a bidword and the search frequency exceeds a threshold, the query is considered a positive sample; otherwise, it is marked as a negative sample."

    The offline metrics Recall@k and NDCG@k measure agreement between generated bidwords and a 'golden standard' constructed by an undisclosed heuristic over click logs. The model's discriminator is explicitly trained to fit CPM values from those same click logs (Eq. 5) and to classify authenticity by search-frequency thresholds from search logs (Eq. 4); these discriminator scores then define the preference-alignment rewards used to train MoBGM. Unless the golden-set heuristic uses independent human judgments, the evaluation target is effectively the same log-derived value/frequency signal the model is optimized to predict, so top-k accuracy reduces to reproducing the heuristic.

full rationale

The paper's central offline claim that MoBGM significantly outperforms state-of-the-art methods rests on Recall@k and NDCG@k against a Golden Dataset whose '10 most valuable bidwords' are selected by an undisclosed heuristic from click logs. The same click logs supply the CPM regression targets in Eq. 5 and the search-frequency-based authenticity labels in Eq. 4, which drive the multi-objective preference rewards. Thus the offline evaluation is at least partially self-referential: it may reward the model for reproducing the click-log value/frequency heuristic rather than for independently better rewriting. I do not find load-bearing self-citation: the self-citations in the paper ([27], [39], [40]) are background references, not justifications of the method's uniqueness or validity. The human-expert relevance/authenticity metrics and the online A/B test (+2.13% ad revenue, p<0.05) provide independent evidence of commercial value, which keeps the circularity partial rather than total. An additional data inconsistency—MoBGM relevance is 97.49 in Table 2 but 92.49 in Table 3 for the same model—further weakens confidence in the offline numbers but is not itself a circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on three hand-set hyperparameter groups (preference weights, KL coefficients, authenticity threshold), plus domain assumptions about the validity of the discriminator rewards and the golden dataset heuristic. No new physical or conceptual entities are introduced.

free parameters (3)
  • Preference alignment weights (w_rel, w_au, w_val) = 0.5, 0.2, 0.3
    Chosen based on validation performance (Section 4.4); these weights control the trade-off among relevance, authenticity, and CPM in the multi-objective alignment loss (Eq. 9).
  • KL penalty coefficients (beta_w, beta_l) = 1.0, 0.25
    Set in Section 4.4; control the KL divergence penalty for winner and loser in the preference alignment loss.
  • Authenticity frequency threshold = Not specified
    In Section 3.4.2, a query is a positive authenticity sample if it can match a bidword and its search frequency exceeds an unnamed threshold; the threshold is a hand-set parameter that shapes the authenticity discriminator and reward.
assumptions (4)
  • domain assumption The BERT-based discriminators provide valid estimates of relevance, authenticity, and CPM for unseen bidwords.
    The entire preference alignment reward (Eqs. 6-9) treats discriminator outputs as accurate rewards; if the discriminators are biased, the alignment will optimize the wrong objective.
  • ad hoc to paper Search frequency exceeding a threshold is a valid proxy for whether a user would genuinely search a bidword.
    Section 3.4.2 defines authenticity this way; no evidence is given that frequency-thresholding matches human notions of authentic bidwords, yet it generates the training labels for the authenticity discriminator.
  • standard math The multi-objective DPO-style loss in Eq. 9 is a valid extension of MODPO and correctly balances three rewards.
    The paper extends MODPO [41] without derivation; it assumes the MODPO objective remains valid when the reward function is a weighted margin of three discriminator scores.
  • domain assumption The heuristic policy that selects 'most valuable bidwords' for the golden set is a reasonable ground truth for retrieval quality.
    Section 4.1 uses an unstated heuristic to build the golden standard for Recall@k and NDCG@k; the merit of the offline results depends on this heuristic.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-objective Aligned Bidword Generation Model for E-commerce Search Advertising." pith.science (2026). https://pith.science/paper/7F7RUW6C

@misc{pith2026250603827,
  author       = {Pith},
  title        = {Pith review of: Multi-objective Aligned Bidword Generation Model for E-commerce Search Advertising},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7F7RUW6C}},
  note         = {Machine review of arXiv:2506.03827}
}
read the original abstract

Retrieval systems primarily address the challenge of matching user queries with the most relevant advertisements, playing a crucial role in e-commerce search advertising. The diversity of user needs and expressions often produces massive long-tail queries that cannot be matched with merchant bidwords or product titles, which results in some advertisements not being recalled, ultimately harming user experience and search efficiency. Existing query rewriting research focuses on various methods such as query log mining, query-bidword vector matching, or generation-based rewriting. However, these methods often fail to simultaneously optimize the relevance and authenticity of the user's original query and rewrite and maximize the revenue potential of recalled ads. In this paper, we propose a Multi-objective aligned Bidword Generation Model (MoBGM), which is composed of a discriminator, generator, and preference alignment module, to address these challenges. To simultaneously improve the relevance and authenticity of the query and rewrite and maximize the platform revenue, we design a discriminator to optimize these key objectives. Using the feedback signal of the discriminator, we train a multi-objective aligned bidword generator that aims to maximize the combined effect of the three objectives. Extensive offline and online experiments show that our proposed algorithm significantly outperforms the state of the art. After deployment, the algorithm has created huge commercial value for the platform, further verifying its feasibility and robustness.

Figures

Figures reproduced from arXiv: 2506.03827 by the authors.

Figure 1
Figure 1. The architecture of the multi-objective aligned bidword generation model. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The deployment of MoBGM and the role of the [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 20 canonical work pages

  1. [1]

    Sanjay Agrawal, Srujana Merugu, and Vivek Sembium. 2023. Enhancing e- commerce product search through reinforcement learning-powered query refor- mulation. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. 4488–4494

    work page 2023

  2. [2]

    Ioannis Antonellis, Hector Garcia-Molina, and Chi-Chao Chang. 2008. Simrank++ query rewriting through link analysis of the clickgraph. InProceedings of the 17th international conference on World Wide Web. 1177–1178

    work page 2008

  3. [3]

    Mohammad Gheshlaghi Azar, Mark Rowland, Bilal Piot, Daniel Guo, Daniele Calandriello, Michal Valko, and Rémi Munos. 2023. A General Theoretical Para- digm to Understand Learning from Human Preferences. arXiv:2310.12036 [cs.AI] https://arxiv.org/abs/2310.12036

    arXiv 2023

  4. [4]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng X...

    arXiv 2023

  5. [5]

    Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, Jérémy Scheurer, Javier Rando, Rachel Freedman, Tomasz Korbak, David Lindner, Pedro Freire, et al. 2023. Open problems and fundamental limitations of reinforcement learning from human feedback. arXiv preprint arXiv:2307.15217 (2023)

    arXiv 2023

  6. [6]

    Xiuying Chen, Daorui Xiao, Shen Gao, Guojun Liu, Wei Lin, Bo Zheng, Dongyan Zhao, and Rui Yan. 2020. RPM-Oriented Query Rewriting Framework for E- commerce Keyword-Based Sponsored Search (Student Abstract). In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 34. 13769–13770

    work page 2020

  7. [7]

    Hang Cui, Ji-Rong Wen, Jian-Yun Nie, and Wei-Ying Ma. 2002. Probabilistic query expansion using query logs. In Proceedings of the 11th international conference on World Wide Web. 325–332

    2002

  8. [8]

    Yiming Cui, Ziqing Yang, and Xin Yao. 2023. Efficient and Effective Text Encoding for Chinese LLaMA and Alpaca. arXiv preprint arXiv:2304.08177 (2023). https: //arxiv.org/abs/2304.08177

    arXiv 2023

Show all 45 references
  1. [9]

    Aijun Dai, Zhenyu Zhu, Haiqing Hu, Guoyu Tang, Lin Liu, and Sulong Xu. 2024. Enhancing E-Commerce Query Rewriting: A Large Language Model Approach with Domain-Specific Pre-Training and Reinforcement Learning. InProceedings of the 33rd ACM International Conference on Informatio...

    2024

  2. [10]

    Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. KTO: Model Alignment as Prospect Theoretic Optimization. arXiv:2402.01306 [cs.LG] https://arxiv.org/abs/2402.01306

    2024 arXiv

  3. [11]

    Jianfeng Gao, Shasha Xie, Xiaodong He, and Alnur Ali. 2012. Learning lexicon models from search logs for query expansion. In Proceedings of EMNLP

    2012

  4. [12]

    Mihajlo Grbovic, Nemanja Djuric, Vladan Radosavljevic, Fabrizio Silvestri, and Narayan Bhamidipati. 2015. Context-and content-aware embeddings for query rewriting in sponsored search. In Proceedings of the 38th international ACM SIGIR conference on research and development in ...

    2015

  5. [13]

    Yunlong He, Jiliang Tang, Hua Ouyang, Changsung Kang, Dawei Yin, and Yi Chang. 2016. Learning to rewrite queries. In Proceedings of the 25th ACM Inter- national on Conference on Information and Knowledge Management . 1443–1452

    2016

  6. [14]

    Rosie Jones, Benjamin Rey, Omid Madani, and Wiley Greiner. 2006. Generating query substitutions. In Proceedings of the 15th international conference on World Wide Web. 387–396

    2006

  7. [15]

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)

    2020 arXiv

  8. [16]

    Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, Vol. 1. Minneapolis, Minnesota, 2

    2019

  9. [17]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Gener- ation, Translation, and Comprehension. arXiv:1910.13461 [cs.CL] ht...

    2019 arXiv

  10. [18]

    Kaiwen Li, Tao Zhang, and Rui Wang. 2021. Deep Reinforcement Learning for Multiobjective Optimization. IEEE Transactions on Cybernetics 51, 6 (June 2021), 3103–3114. https://doi.org/10.1109/tcyb.2020.2977661

    2021

  11. [19]

    Sen Li, Fuyu Lv, Taiwei Jin, Guiyang Li, Yukun Zheng, Tao Zhuang, Qingwen Liu, Xiaoyi Zeng, James Kwok, and Qianli Ma. 2022. Query Rewriting in TaoBao Search. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. 3262–3271

    2022

  12. [20]

    Yijiang Lian, Zhijie Chen, Jinlong Hu, Kefeng Zhang, Chunwei Yan, Muchenxuan Tong, Wenying Han, Hanju Guan, Ying Li, Ying Cao, et al. 2019. An end-to-end Generative Retrieval Method for Sponsored Search Engine–Decoding Efficiently into a Closed Target Domain. arXiv preprint ar...

    2019 arXiv

  13. [21]

    Y Liu, M Ott, N Goyal, J Du, M Joshi, D Chen, O Levy, M Lewis, L Zettlemoyer, and V Stoyanov. 1907. RoBERTa: A robustly optimized BERT pretraining approach. arXiv [Preprint](2019). arXiv preprint arXiv:1907.11692 (1907)

    2019 arXiv

  14. [22]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. arXiv:1711.05101 [cs.LG] https://arxiv.org/abs/1711.05101

    2019 arXiv

  15. [23]

    Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. 2023. Query rewriting for retrieval-augmented large language models. arXiv preprint arXiv:2305.14283 (2023)

    2023 arXiv

  16. [24]

    Yu Meng, Mengzhou Xia, and Danqi Chen. 2024. SimPO: Simple Preference Optimization with a Reference-Free Reward. arXiv:2405.14734 [cs.CL] https: //arxiv.org/abs/2405.14734

    2024 arXiv

  17. [25]

    Akash Kumar Mohankumar, Nikit Begwani, and Amit Singh. 2021. Diversity driven query rewriting in search advertising. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining . 3423–3431

    2021

  18. [26]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schul- man, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, a...

    2024

  19. [27]

    Ming Pang, Chunyuan Yuan, Xiaoyu He, Zheng Fang, Donghao Xie, Fanyi Qu, Xue Jiang, Changping Peng, Zhangang Lin, Zheng Luo, et al. 2025. Generative Retrieval and Alignment Model: A New Paradigm for E-commerce Retrieval. arXiv preprint arXiv:2504.01403 (2025)

    2025 arXiv

  20. [28]

    Wenjun Peng, Guiyang Li, Yue Jiang, Zilong Wang, Dan Ou, Xiaoyi Zeng, Derong Xu, Tong Xu, and Enhong Chen. 2024. Large language model based long-tail query rewriting in taobao search. In Companion Proceedings of the ACM on Web Conference 2024. 20–28

    2024

  21. [29]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems 36 (2024)

    2024

  22. [30]

    Stefan Riezler and Yi Liu. 2010. Query rewriting using monolingual statistical machine translation. Computational Linguistics 36, 3 (2010), 569–582

    2010

  23. [31]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  24. [32]

    Zhenqiao Song, Jiaze Chen, Hao Zhou, and Lei Li. 2021. Triangular Bidword Generation for Sponsored Search Auction. In Proceedings of the 14th ACM Inter- national Conference on Web Search and Data Mining . 707–715

    2021

  25. [33]

    Alessandro Sordoni, Yoshua Bengio, Hossein Vahabi, Christina Lioma, Jakob Grue Simonsen, and Jian-Yun Nie. 2015. A hierarchical recurrent encoder-decoder for generative context-aware query suggestion. In proceedings of the 24th ACM international on conference on information an...

    2015

  26. [34]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, and Amjad Alma- hairi. 2023. Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv:2307.09288 [cs.CL] https://arxiv.org/abs/2307.09288

    2023 arXiv

  27. [35]

    Svitlana Vakulenko, Shayne Longpre, Zhucheng Tu, and Raviteja Anantha. 2021. Question rewriting for conversational question answering. In Proceedings of the 14th ACM international conference on web search and data mining . 355–363

    2021

  28. [36]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems (2017)

    2017

  29. [37]

    Yaxuan Wang, Hanqing Lu, Yunwen Xu, Rahul Goutam, Yiwei Song, and Bing Yin. 2021. QUEEN: Neural query rewriting in e-commerce. InThe Web Conference

    2021

  30. [38]

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al

  31. [39]

    Chunyuan Yuan, Ming Pang, Zheng Fang, Xue Jiang, Changping Peng, and Zhangang Lin. 2024. A Semi-supervised Multi-channel Graph Convolutional Network for Query Classification in E-commerce. In Companion Proceedings of the ACM Web Conference 2024. 56–64

    2024

  32. [40]

    Chunyuan Yuan, Yiming Qiu, Mingming Li, Haiqing Hu, Songlin Wang, and Sulong Xu. 2023. A multi-granularity matching attention network for query intent classification in e-commerce retrieval. In Companion Proceedings of the ACM Web Conference 2023. 416–420

    2023

  33. [41]

    Zhanhui Zhou, Jie Liu, Jing Shao, Xiangyu Yue, Chao Yang, Wanli Ouyang, and Yu Qiao. 2024. Beyond One-Preference-Fits-All Alignment: Multi-Objective Direct Preference Optimization. arXiv:2310.03708 [cs.LG] https://arxiv.org/abs/2310. 03708

    2024 arXiv

  34. [42]

    Simiao Zuo, Qingyu Yin, Haoming Jiang, Shaohui Xi, Bing Yin, Chao Zhang, and Tuo Zhao. 2023. Context-Aware Query Rewriting for Improving Users’ Search Experience on E-commerce Websites. In The 61st Annual Meeting Of The Association For Computational Linguistics

    2023

  35. [2017]

    arXiv:1707.06347 [cs.LG] https://arxiv.org/abs/1707.06347

    Proximal Policy Optimization Algorithms. arXiv:1707.06347 [cs.LG] https://arxiv.org/abs/1707.06347

    arXiv

  36. [2021]

    https://www.amazon.science/publications/queen-neural-query-rewriting- in-e-commerce

  37. [2022]

    Emergent abilities of large language models.arXiv preprint arXiv:2206.07682 (2022)

    2022 arXiv

Pith tools

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