EcomScriptBench: A Multi-task Benchmark for E-commerce Script Planning via Step-wise Intention-Driven Product Association

Changlong Yu; Chen Luo; Haiyang Zhang; Hansu Gu; Hui Liu; Jiaxin Bai; Limeng Cui; Qi He; Sheikh Muhammad Sarwar; Shuiwang Ji

REVIEW 3 major objections 4 minor 13 references

EcomScriptBench: A Multi-task Benchmark for E-commerce Script Planning via Step-wise Intention-Driven Product Association

T0 review · 3 major / 4 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 E-commerce script planning is a hard, distinct evaluation target: current LLMs, even fine-tuned, fall short on matching products to plan steps, and injecting purchase intentions is the intervention that most reliably helps.

desk verdict A solid benchmark contribution with an honest limitations section, but the intention-injection gain is likely overstated because the Task 2/3 candidates were generated by the same intention-alignment process being tested. read the letter →

arxiv 2505.15196 v1 pith:HOUW3XIJ submitted 2025-05-21 cs.CL

Weiqi Wang , Limeng Cui , Xin Liu , Sreyashi Nag , Wenju Xu , Chen Luo , Sheikh Muhammad Sarwar , Yang Li

show 9 more authors

Hansu Gu Hui Liu Changlong Yu Jiaxin Bai Yifan Gao Haiyang Zhang Qi He Shuiwang Ji Yangqiu Song

This is my paper · ORCID

classification cs.CL

keywords e-commercescriptplanningbenchmarkpurchaseintentionproductretrievalLLMevaluationsemanticalignmenthumanannotationmulti-task

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 defines e-commerce script planning as three binary subtasks: judging whether a script is plausible for a goal, whether a product genuinely helps a step, and whether all step-product pairs work together. It then builds the first large-scale benchmark for these tasks, with 605,229 product-enriched scripts derived from real e-commerce purchase reviews and 2.4 million products. Testing over twenty language models, the authors find that even the strongest open models and commercial APIs are far from reliable, especially on product-related subtasks. Fine-tuning helps, but the clearest gain comes from first training models on large purchase-intention knowledge bases, which improves accuracy across all three tasks. The paper's intended contribution is a reproducible evaluation target plus evidence that purchase-intention knowledge is the missing link between actions and products.

What carries the argument

The mechanism is step-intention alignment: each product is described by ten purchase intentions distilled from product metadata and reviews, and each script step is matched to products by cosine similarity between SentenceBERT embeddings of the step and the product's intentions, with a similarity threshold to filter noise. This bridges the semantic gap between action-oriented steps and product metadata, and it is the same intentional knowledge that, when injected into LLMs via sequential fine-tuning, produces the reported gains.

What would settle it

Take a fresh set of, say, 1,000 real user shopping plans (actual purchases and step-by-step intentions) collected independently of this benchmark, have experts label them, and measure how well a model fine-tuned on EcomScriptBench transfers; if accuracy drops toward chance or expert agreement on the benchmark's own labels falls far below the reported 96%, the benchmark's synthetic objectives would be shown not to represent real shopping behavior.

Watch

Extended reading notes

Core claim

The central claim is that current (large) language models cannot yet reliably perform e-commerce script planning—simultaneously judging a step-by-step shopping plan, deciding which steps need products, and verifying that recommended products fit the whole plan—and that this capability improves when models are given purchase intentions, the underlying reasons a customer buys a product. The paper establishes this by reporting that the best open-source model reaches only 75%, 68%, and 65% accuracy on the three subtasks, with commercial GPT-4o reaching 78%, 74%, and 72% in its best settings. It also shows that fine-tuning on the benchmark raises accuracy substantially, and that sequentially fine-tuning on two existing intention-knowledge sources before the benchmark yields further gains, notably lifting Mistral-7B from 75.63 to 81.18 on product discrimination. The intended contribution is a benchmark that makes this difficulty measurable, plus a validated direction for improving it.

Load-bearing premise

The gold labels are only as trustworthy as the GPT-generated scripts and intentions, which were human-verified on small samples (200 scripts, 200 intentions, and 200 labels per task); if those synthetic objectives do not reflect how real shoppers actually plan, the measured model performance would not carry over to real e-commerce use.

Editorial extensions

If this is right

E-commerce script planning can be decomposed into three verifiable binary subtasks, enabling a generate-then-filter pipeline for shopping assistants.
Current LLMs, including very large open models and commercial APIs, are not yet reliable at product-step matching, leaving room for targeted improvement.
Fine-tuning on the benchmark improves all tested models, so the annotated data itself is a usable training signal.
Injecting purchase-intention knowledge from large intention bases improves product discrimination and whole-script verification beyond fine-tuning alone.
Simple prompting tricks like chain-of-thought do not help these tasks, whereas product-related external signals do.

Reading between the lines

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

If intention injection generalizes, a practical design for shopping assistants would separate script generation from product grounding: plan with an LLM, then retrieve and verify products through intention-based matching rather than raw title queries.
The benchmark's three-task decomposition could be reused for other domains where plans need to be paired with artifacts, such as travel itineraries with bookings or DIY projects with parts.
A testable extension is to replace the fixed similarity threshold with a learned ranker over intention-step pairs; the reported fraction of intentions matching objectives suggests the current filter may discard useful matches.
Because the dataset is not public, a harder open benchmark built from public reviews would be needed to confirm that the difficulty is intrinsic to the task and not an artifact of the specific generated scripts.

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.

Desk Editor's Note

read the letter

The paper does something real: it formalizes e-commerce script planning as three sequential discriminative subtasks, builds a large dataset (605K scripts, 2.4M products, 24M intentions) from real Amazon data, and runs a broad model evaluation. The human annotation pipeline is thorough — worker qualification, five votes per entry, 78% pairwise agreement, expert verification at 96% acceptance. The central finding that current LLMs plateau at 65–75% accuracy on these tasks holds up and is not definitionally circular. The paper is also upfront about its limitations, including the use of GPT-4o-mini and the non-public data.

The soft spots are real. First, the dataset is not released. For a benchmark paper that blocks independent verification and reuse. Privacy is a constraint, but without data or detailed evaluation artifacts the community cannot build on this work. That alone justifies a conditional verdict.

The bigger issue is the causal claim about intentions. Task 2 and Task 3 candidates are pre-selected by the same step-intention alignment pipeline that the FolkScope/MIND fine-tuning is meant to improve: GPT-4o-mini writes purchase intentions, SentenceBERT ranks products by similarity to those intentions, and humans label only the top matches. A model fine-tuned on LLM-distilled intentions will naturally agree more with this construction. The large gains for Mistral-7B (75.63 to 81.18 on Task 2, 73.18 to 78.94 on Task 3) could reflect learning the label-generation distribution rather than an independently-validated ability. Expert verification checks label consistency, not selection bias. I would not call the claim false, but it is not established. A held-out set with differently-sampled candidates, or an analysis of performance on low-similarity pairs, would be needed to support it.

Minor issues: the similarity threshold tau is 0.45 in the main text and 0.4 in Appendix A.1.3; several tables lack error bars despite three-seed fine-tuning; zero-shot PTLM numbers are weak enough that the assertion-based evaluation format deserves more discussion.

This paper is for e-commerce NLP researchers and LLM evaluation folks. It deserves serious refereeing and likely acceptance after revision. I would ask for the threshold fix, error bars, and a re-analysis of Table 3 that separates the alignment effect from genuine product-step fit. If the authors can release anonymized evaluation data, this becomes a much stronger community resource.

Referee Report

3 major / 4 minor

Summary. The paper defines E-commerce Script Planning (EcomScript) as three sequential binary tasks: script verification (Task 1), step-product discrimination (Task 2), and script-product verification (Task 3). It then proposes an automatic pipeline in which GPT-4o-mini generates user objectives, scripts, purchase-necessity judgments, product keywords, and product purchase intentions; SentenceBERT is used to associate up to three products with each step via step-intention alignment; and human annotators label a sampled subset of 5,000 entries per task. The resulting EcomScriptBench contains 605,229 scripts, 2.4 million products, and 24 million purchase intentions. Experiments with a wide range of PTLMs, open LLMs, and API LLMs show that zero-shot models perform only moderately, that fine-tuning improves accuracy, and that sequentially fine-tuning on FolkScope and MIND intention knowledge further improves performance. The paper concludes that e-commerce script planning is a difficult task for current LLMs and that injecting purchase-intention knowledge is a promising intervention.

Significance. If the central claims were fully supported, the paper would fill a real gap: it provides a multi-task evaluation framework for e-commerce script planning, a large-scale product-enriched script knowledge base, a detailed human annotation protocol with expert verification (96.33% acceptance on a 200-entry sample per task, Fleiss kappa 0.53), and a broad model comparison. These are genuine strengths. However, the significance is substantially conditional on two issues. First, the dataset is not released, so the central artifact cannot be inspected or used. Second, the Task 2 and Task 3 test labels are constructed by the same intention-alignment mechanism whose benefit the paper claims to demonstrate, which threatens the comparative claim about intention injection. The claim that models perform only moderately on this particular, synthetically generated, intention-filtered distribution is plausible, but the broader conclusion about e-commerce script planning ability is not yet established.

major comments (3)

[§4.4] The claim that injecting purchase-intention knowledge improves EcomScript performance is not yet established, because the Task 2 and Task 3 gold labels are produced by the very intention-alignment mechanism whose benefit is being tested. In §4.3, positive candidates are constructed by (i) having GPT-4o-mini decide purchase necessity and emit product keywords, (ii) ranking filtered products by SentenceBERT cosine similarity between the step and GPT-4o-mini-generated purchase intentions, and (iii) taking the top-3 matches; annotators then label only these pre-selected candidates, and the expert verification of 200 entries per task checks only the same pre-selected set. Consequently, the large gains in Table 3 (e.g., Mistral-7B Task 2 75.63→81.18 and Task 3 73.18→78.94) may reflect fine-tuning on FolkScope/MIND, which are the same type of LLM-distilled intention knowledge used to construct the labels, matching the label-generation distribution rather than improving ability on independently defined product-step compatibility. Please add an evaluation setting in which Task 2/3 labels are obtained without intention-aligned candidate selection, and include a control condition fine-tuned on product metadata alone to separate the effect of intention knowledge from additional in-domain pretraining.
[Ethics Statement] The manuscript states, 'Due to data privacy issues, our dataset will not be made public.' This prevents the central contribution—a benchmark—from being inspected, used, or reproduced by the community, and no public leaderboard or evaluation API is offered. For a benchmark paper, please provide a public evaluation interface or release an anonymized subset; otherwise the benchmark's difficulty and the reported model rankings cannot be independently verified.
[§4.1 and §4.4] Because all test scripts are generated by GPT-4o-mini and only scripts that pass human plausibility annotation are retained as candidates, the benchmark measures model performance on a specific distribution of LLM-generated scripts and pre-filtered product associations. The conclusion that 'current (L)LMs face significant challenges with EcomScript tasks' is therefore too broad; it should be qualified as applying to this synthetic, intention-filtered distribution unless the authors also include a sample of independently collected user-provided scripts and product-step associations.

minor comments (4)

[§4.3 vs. Appendix A.1.3] The similarity threshold is inconsistently reported: §4.3 states τ = 0.45, while Appendix A.1.3 states τ = 0.4 and reports a 95% filtering rate. Please reconcile these numbers and specify the threshold actually used for the released statistics.
[§3.3] The pilot study reports that roughly 68% of step-as-query searches returned only a limited assortment, but 'limited assortment' is not defined and no variance or error bar is given; please specify the measurement protocol so the claim is checkable.
[Table 1] The column header '#Data (Unlabeled)' is confusing because the 5,000 rows per task are the labeled benchmark entries; please rename the column to make clear that the parenthetical refers to the larger unlabeled candidate pool.
[Abstract and Conclusion] The paper calls EcomScriptBench the 'very first' e-commerce script planning benchmark; since the dataset is not public, please clarify whether the claim refers to the construction framework, the knowledge base, or the evaluation subset, and qualify it accordingly.

Circularity Check

1 steps flagged · score 5.0 of 10

Partial construction-level circularity: the intention-injection improvement is evaluated on a test set whose Task 2/3 candidates were pre-selected by the same step-intention alignment mechanism whose benefit the paper claims; human labels do not remove this selection bias.

other [Section 4.3 (Step-Intention Alignment) and Section 5.3 (The Effect of Injecting Intentions), Table 3]
"For each step, we rank all filtered products to select the top three that best align with the actionable step in the script ... Each step and its selected products form the candidate data for the second task ... we select two intention knowledge bases based on products from Amazon, FolkScope (Yu et al., 2023), and MIND (Xu et al., 2024) as sources of intentions."

The Task 2/3 evaluation items are not sampled from the full product space: they are the top-3 products ranked by SentenceBERT cosine similarity between each step and GPT-4o-mini-generated purchase intentions, subject to a threshold. The intervention whose benefit is claimed in Table 3 fine-tunes models on FolkScope/MIND, which are also LLM-distilled purchase-intention KBs. The test distribution is therefore generated by the same type of intentional knowledge whose injection is the proposed remedy. A model that learns to reproduce intention-similarity ranking will look better on Task 2/3 by construction, because the candidate set was pre-filtered to contain intention-aligned products.

full rationale

The benchmark's basic finding that current (L)LMs struggle on EcomScript is not circular: all three subtasks have human-annotated gold labels (AMT majority vote plus expert verification), so the difficulty result is externally grounded. The circularity is confined to the comparative claim in Section 5.3. There, the products that appear in Task 2 and Task 3 test items were selected by the paper's own step-intention alignment strategy, i.e., by ranking products according to similarity between steps and LLM-generated purchase intentions. The intervention fine-tunes on FolkScope and MIND, two LLM-distilled purchase-intention resources from the same author group. Consequently, the measured improvement from injecting intentions partly reduces to matching the distribution used to build the evaluation set. The human labels break full definitional circularity, because annotators could have rejected many intention-aligned candidates; indeed the Task 2 majority accuracy of 57.67% shows labels are not identical to the selection rule. For this reason I score this as partial (5), not full (8-10) circularity. The self-citations to FolkScope/MIND are numerous but not independently load-bearing: the transfer experiment is an empirical comparison, and intention quality is also supported by the paper's own expert acceptance rates. The unmentioned limitation is that the evaluation design conflates 'purchase-intention knowledge helps e-commerce script planning' with 'purchase-intention knowledge helps predict the output of a purchase-intention-based candidate generator.'

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

The central claim rests on the validity of the automatically constructed benchmark. The key free parameter is the similarity threshold tau used in product association; it is chosen from the data distribution and reported differently in two sections. The main domain assumptions are that purchase intentions bridge steps and products, that the three subtasks capture script planning, and that GPT-4o-mini produces realistic goals and scripts. No new physical, chemical, or mathematical entities are postulated; the 'product-enriched script' and the three tasks are novel organizational constructs, not independent entities with falsifiable predictions.

free parameters (3)

similarity threshold tau = 0.45 in Section 4.3, 0.4 in Appendix A.1.3
Hand-chosen lower bound on Sentence-BERT cosine similarity between steps and purchase intentions to filter product associations; determined by observing the similarity distribution, and the two sections of the paper report different values.
maximum products per step = 3
Design choice to limit dataset size and overlap; not fitted but influences task difficulty.
number of intentions per product = 10
Design choice for coverage; not fitted but affects the matching signal.

assumptions (4)

domain assumption Purchase intention is a reliable semantic bridge between script steps and products.
The whole dataset construction uses similarity between step text and product purchase intentions to select product associations (Section 4.3). If intention semantics do not align with step action semantics, the benchmark's product associations are noisy.
domain assumption The three subtasks adequately decompose e-commerce script planning into testable binary decisions.
The paper defines the task as three sequential subtasks (Section 3.1). If these subtasks do not cover the true capability, the benchmark measures something narrower than e-commerce script planning.
domain assumption GPT-4o-mini can generate realistic user objectives, scripts, and purchase intentions from Amazon reviews and product metadata.
The dataset is built by few-shot prompting GPT-4o-mini (Sections 4.1 and 4.2), and quality is checked only on small expert samples. If the generations are systematically unrealistic, the benchmark loses validity.
standard math Sentence-BERT cosine similarity scores reflect task-relevant compatibility between steps and products.
The product retrieval uses Sentence-BERT embeddings with a cosine threshold (Section 4.3). This is an unproved modeling choice about semantic similarity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of EcomScriptBench: A Multi-task Benchmark for E-commerce Script Planning via Step-wise Intention-Driven Product Association." pith.science (2026). https://pith.science/paper/HOUW3XIJ

@misc{pith2026250515196,
  author       = {Pith},
  title        = {Pith review of: EcomScriptBench: A Multi-task Benchmark for E-commerce Script Planning via Step-wise Intention-Driven Product Association},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HOUW3XIJ}},
  note         = {Machine review of arXiv:2505.15196}
}

read the original abstract

Goal-oriented script planning, or the ability to devise coherent sequences of actions toward specific goals, is commonly employed by humans to plan for typical activities. In e-commerce, customers increasingly seek LLM-based assistants to generate scripts and recommend products at each step, thereby facilitating convenient and efficient shopping experiences. However, this capability remains underexplored due to several challenges, including the inability of LLMs to simultaneously conduct script planning and product retrieval, difficulties in matching products caused by semantic discrepancies between planned actions and search queries, and a lack of methods and benchmark data for evaluation. In this paper, we step forward by formally defining the task of E-commerce Script Planning (EcomScript) as three sequential subtasks. We propose a novel framework that enables the scalable generation of product-enriched scripts by associating products with each step based on the semantic similarity between the actions and their purchase intentions. By applying our framework to real-world e-commerce data, we construct the very first large-scale EcomScript dataset, EcomScriptBench, which includes 605,229 scripts sourced from 2.4 million products. Human annotations are then conducted to provide gold labels for a sampled subset, forming an evaluation benchmark. Extensive experiments reveal that current (L)LMs face significant challenges with EcomScript tasks, even after fine-tuning, while injecting product purchase intentions improves their performance.

Figures

Figures reproduced from arXiv: 2505.15196 by the authors.

**Figure 1.** An example of a product-enriched script for planning the objective of plan an autumn-themed party with friends and family, with relevant products associated with some steps. Note that for simpler steps, such as the first two, no products are needed. capability, also known as goal-oriented script planning (Bower et al., 1979; Schank and Abelson, 1975), forms the foundation of situationally grounded planning for comp… view at source ↗

**Figure 2.** An overview of our benchmark curation and evaluation pipeline for E [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Distribution of the number of retrieved prod [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 5 canonical work pages

[7]

Falcon2-11b technical report.CoRR, abs/2407.14885. Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Aakanksha Chowdh- ery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti,...

arXiv
[8]

Jianmo Ni, Jiacheng Li, and Julian J

Gemma: Open models based on gemini re- search and technology.CoRR, abs/2403.08295. Jianmo Ni, Jiacheng Li, and Julian J. McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. InProceedings of the 2019 Conference on Empirical Methods in Natu- ral Language Processing and the 9th International Joint Conference on...

arXiv 2019
[11]

InCompanion of the 2024 Inter- national Conference on Management of Data, SIG- MOD/PODS 2024, Santiago AA, Chile, June 9-15, 2024, pages 148–160

COSMO: A large-scale e-commerce com- mon sense knowledge generation and serving sys- tem at amazon. InCompanion of the 2024 Inter- national Conference on Management of Data, SIG- MOD/PODS 2024, Santiago AA, Chile, June 9-15, 2024, pages 148–160. ACM. Changlong Yu, Weiqi Wang, Xin Liu, Jiaxin Bai, Yangqiu Song, Zheng Li, Yifan Gao, Tianyu Cao, and Bing Yin...

work page 2024
[1384]

like the product

ACM. Jiatao Zhang, Lanling Tang, Yufan Song, Qiwei Meng, Haofu Qian, Jun Shao, Wei Song, Shiqiang Zhu, and Jason Gu. 2024a. FLTRNN: faithful long-horizon task planning for robotics with large language models. InIEEE International Conference on Robotics and Automation, ICRA 2024, Yokohama, Japan, May 13- 17, 2024, pages 6680–6686. IEEE. Liyu Zhang, Weiqi W...

arXiv 2024
[1979]

Scripts in memory for text.Cognitive psychol- ogy, 11(2):177–220. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemen...

work page
[1985]

Marah I Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harki- rat S

The representation of scripts in memory.Jour- nal of memory and language, 24(2):179–199. Marah I Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harki- rat S. Behl, Alon Benhaim, Misha Bilenko, Jo- han Bjorck, Sébastien Bubeck, Martin Cai, Caio César Teodoro Mendes, Weizhu Ch...

arXiv 2024
[2016]

InProceedings of the 25th Interna- tional Conference on World Wide Web, WWW 2016, Montreal, Canada, April 11 - 15, 2016, pages 1373–

Mining user intentions from medical queries: A neural network based heterogeneous jointly mod- eling approach. InProceedings of the 25th Interna- tional Conference on World Wide Web, WWW 2016, Montreal, Canada, April 11 - 15, 2016, pages 1373–

work page 2016
[2020]

InAd- vances in Neural Information Processing Systems 33: Annual Conference on Neural Information Process- ing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual

Language models are few-shot learners. InAd- vances in Neural Information Processing Systems 33: Annual Conference on Neural Information Process- ing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual. Chunkit Chan, Cheng Jiayang, Weiqi Wang, Yuxin Jiang, Tianqing Fang, Xin Liu, and Yangqiu Song. 2024a. Exploring the potential of chatgpt on sen- ten...

work page 2020

Show all 13 references

[2022]

In2022 International Conference on Robotics and Automa- tion, ICRA 2022, Philadelphia, PA, USA, May 23-27, 2022, pages 8847–8854

The threedworld transport challenge: A vi- sually guided task-and-motion planning benchmark towards physically realistic embodied AI. In2022 International Conference on Robotics and Automa- tion, ICRA 2022, Philadelphia, PA, USA, May 23-27, 2022, pages 8847–8854. IEEE. Shibo H...

2022 arXiv
[2023]

InProceedings of the 17th Conference of the European Chapter of the Asso- ciation for Computational Linguistics, EACL 2023, Dubrovnik, Croatia, May 2-6, 2023, pages 3018–

Incorporating task-specific concept knowl- edge into script learning. InProceedings of the 17th Conference of the European Chapter of the Asso- ciation for Computational Linguistics, EACL 2023, Dubrovnik, Croatia, May 2-6, 2023, pages 3018–

2023
[2024]

Yangning Li, Shirong Ma, Xiaobin Wang, Shen Huang, Chengyue Jiang, Haitao Zheng, Pengjun Xie, Fei Huang, and Yong Jiang

Ecomedit: An automated e-commerce knowledge editing framework for enhanced prod- uct and purchase intention understanding.CoRR, abs/2410.14276. Yangning Li, Shirong Ma, Xiaobin Wang, Shen Huang, Chengyue Jiang, Haitao Zheng, Pengjun Xie, Fei Huang, and Yong Jiang. 2024. Ecomgp...

arXiv
[2266]

Association for Computational Linguistics. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Ko...

2024 arXiv
[3032]

Association for Computational Linguistics. Hugo Touvron, Louis Martin, Kevin Stone, Peter Al- bert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti 14 Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton- Ferrer, Moya Chen, Guillem Cuc...

2023 arXiv

Pith tools

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

[1] [7]

Falcon2-11b technical report.CoRR, abs/2407.14885. Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Aakanksha Chowdh- ery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti,...

arXiv

[2] [8]

Jianmo Ni, Jiacheng Li, and Julian J

Gemma: Open models based on gemini re- search and technology.CoRR, abs/2403.08295. Jianmo Ni, Jiacheng Li, and Julian J. McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. InProceedings of the 2019 Conference on Empirical Methods in Natu- ral Language Processing and the 9th International Joint Conference on...

arXiv 2019

[3] [11]

InCompanion of the 2024 Inter- national Conference on Management of Data, SIG- MOD/PODS 2024, Santiago AA, Chile, June 9-15, 2024, pages 148–160

COSMO: A large-scale e-commerce com- mon sense knowledge generation and serving sys- tem at amazon. InCompanion of the 2024 Inter- national Conference on Management of Data, SIG- MOD/PODS 2024, Santiago AA, Chile, June 9-15, 2024, pages 148–160. ACM. Changlong Yu, Weiqi Wang, Xin Liu, Jiaxin Bai, Yangqiu Song, Zheng Li, Yifan Gao, Tianyu Cao, and Bing Yin...

work page 2024

[4] [1384]

like the product

ACM. Jiatao Zhang, Lanling Tang, Yufan Song, Qiwei Meng, Haofu Qian, Jun Shao, Wei Song, Shiqiang Zhu, and Jason Gu. 2024a. FLTRNN: faithful long-horizon task planning for robotics with large language models. InIEEE International Conference on Robotics and Automation, ICRA 2024, Yokohama, Japan, May 13- 17, 2024, pages 6680–6686. IEEE. Liyu Zhang, Weiqi W...

arXiv 2024

[5] [1979]

Scripts in memory for text.Cognitive psychol- ogy, 11(2):177–220. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemen...

work page

[6] [1985]

Marah I Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harki- rat S

The representation of scripts in memory.Jour- nal of memory and language, 24(2):179–199. Marah I Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harki- rat S. Behl, Alon Benhaim, Misha Bilenko, Jo- han Bjorck, Sébastien Bubeck, Martin Cai, Caio César Teodoro Mendes, Weizhu Ch...

arXiv 2024

[7] [2016]

InProceedings of the 25th Interna- tional Conference on World Wide Web, WWW 2016, Montreal, Canada, April 11 - 15, 2016, pages 1373–

Mining user intentions from medical queries: A neural network based heterogeneous jointly mod- eling approach. InProceedings of the 25th Interna- tional Conference on World Wide Web, WWW 2016, Montreal, Canada, April 11 - 15, 2016, pages 1373–

work page 2016

[8] [2020]

InAd- vances in Neural Information Processing Systems 33: Annual Conference on Neural Information Process- ing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual

Language models are few-shot learners. InAd- vances in Neural Information Processing Systems 33: Annual Conference on Neural Information Process- ing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual. Chunkit Chan, Cheng Jiayang, Weiqi Wang, Yuxin Jiang, Tianqing Fang, Xin Liu, and Yangqiu Song. 2024a. Exploring the potential of chatgpt on sen- ten...

work page 2020