text stringlengths 0 11.8k |
|---|
Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1–22 July 27 - August 1, 2025 ©2025 Association for Computational Linguistics ECOMSCRIPTBENCH: A Multi-task Benchmark for E-commerce Script Planning via Step-wise Intention-Driven Product Association W... |
Introduction In our daily lives, humans commonly plan a se- quence of general prototypical actions, usually in the form of step-by-step instructions, to achieve a specific objective (Abbott et al., 1985). This *Work done during his internship at Amazon.com Inc. †Visiting academic scholar at Amazon.com Inc. Help me plan... |
with friends and family Chat with your AI Shopping Assistant! Type your question here… … … … … … … … … Sure! Here is a plan with some products: 1. Select a convenient date and time for |
your party that works for most guests 2. Make a list of people you want to invite 3. |
4. |
5. Prepare autumn-themed food 6. Buy autumn-themed decorations to |
create a festive atmosphere 7. |
8. |
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 (Bow... |
In e-commerce, there is a growing trend of cus- tomers querying LLM-empowered shopping assis- tants to create scripts tailored to their specific needs or objectives, with each step featuring relevant products. For example, as illustrated in Figure 1, LLM assistant is expected to generate an eight-step script toward the... |
Related Works 2.1 Goal-oriented Script Planning Goal-oriented scripts refer to a coherent and ap- propriate sequence of steps, usually in the form of actions, as instructions for achieving a goal (Reg- neri et al., 2010). They are a common reflection of language planning capabilities, often observed in embodied AI (Gan... |
abilities of large vision-language models. Joshi et al. (2023) designed an interactive text-based gam- ing framework that consists of daily real-world hu- man activities as another benchmark. Nevertheless, none of the prior works have explored script plan- ning in the context of e-commerce, which holds significant pote... |
Problem Definition 3.1 ECOMSCRIPT Task Definitions We first introduce our definition of the proposed e-commerce script planning tasks (ECOMSCRIPT). Since both asking an LLM to generate a script with products associated with each step and evaluating such generations are difficult to accomplish directly, it is challengin... |
The step-product discriminator (T2) can verify the results of each retrieved product associated with each step and remove unnecessary products for sim- ple steps. Finally, the script-products verifier (T3) will check the product-enriched script and ensure that all products can coordinate smoothly within the script to b... |
ECOMSCRIPTBENCH Construction In this section, we introduce our method for synthe- sizing product-enriched scripts to construct an eval- uation benchmark. An overview of the framework is shown in Figure 2. Specifically, our framework consists of four main stages: (1) user objective and script collection, (2) product pur... |
4.1 User Objective Collection <Source User Purchase Review> Objective: Participate in a |
long-range marathon 4.1 Objective-oriented Script Generation Objective: Participate in a |
long-range marathon <A multi-step script> Step 1: Invest in quality |
running shoes Step 2: Join a local running |
group |
... Step 9: Practice hydration |
during runs Step 10: Register for the |
marathon 4.4 Human Annotation & Expert Verification Amazon MTurk Expert (Candidate Data) User Objectives and Scripts Product-enriched |
Scripts 4.2 Purchase Intention Mining Nike Air Zoom Running Shoes <Customers’ Purchase Intentions> Intention 1: Participate in |
professional running group Intention 2: Improve their |
running experience |
... |
... Intention 10: Alleviate foot |
pain caused by inadequate or |
poorly fitting shoes 4.3 Step-Intention Alignment Subtask 1: Script Verification Subtask 2 & 3: Product Discrimination & |
Product-Script Verification Step 2: Join a local running group Nike Air Zoom Running Shoes Intention: Participate in |
professional running group Women's Casual Long Sleeve |
Button Down Loose Striped Dress Intention: Use as a beach cover-up |
in vacations. High Similarity Determine the plausibility and feasibility of the |
script: <A multi-step script> The script looks good! Low Similarity Determine whether the |
product can help achieving |
the step: <An E-commerce Product> <A Step from a Script> This product can be helpful! Does these products work for |
the script below? <A Product-enriched Script> These products can serve |
the script well! Figure 2: An overview of our benchmark curation and evaluation pipeline for ECOMSCRIPTBENCH. quality, we discard reviews that are too short or contain excessive punctuation or hashtags. With these objectives, we further instruct the LLM to generate goal-oriented scripts based on them. We similarly modi... |
Type #Data (Unlabeled) #Token Expert. Scripts 605,229 71.5 94.0% Steps 5,928,271 7.48 94.0% w. products 3,018,276 6.98 - |
w.o. products 2,909,995 7.98 - |
Products 2,401,087 19.31 - |
Intentions 24,010,870 10.27 98.5% Task 1 5,000 (592,729) - |
95.5% Task 2 5,000 (5,919,278) - |
96.5% Task 3 5,000 (589,801) - |
97.0% Table 1: Statistics of the ECOMSCRIPTBENCH bench- mark. #Token refers to average number of tokens used. Expert. refers to expert acceptance rate. BERT (Reimers and Gurevych, 2019) to calculate the average embedding similarity between each step and the purchase intentions of each product. For each step, we rank al... |
100K 200K 300K 400K 500K 600K Number of Scripts No Product 1 Product 2 Products 3 Products Figure 3: Distribution of the number of retrieved prod- ucts at each step in ECOMSCRIPTBENCH. Kappa (Fleiss, 1971) is 0.53, indicating sufficient agreement. More details are in Appendix B. Expert Verification: To verify the quali... |
Experiments and Analyses 5.1 ECOMSCRIPTBENCH Statistics We first present the statistics of ECOMSCRIPT- BENCH in Table 1. In total, we collect 605K scripts with 5.9 million steps. Of a sample of 200 scripts, 94% were annotated as plausibly correct by ex- pert annotators. Among them, approximately 3 million steps are dee... |
Methods Backbone Script Verification Product Discrimination Product-Script Veri. Acc AUC Ma-F1 Acc AUC Ma-F1 Acc AUC Ma-F1 Random N/A 50.00 - |
50.00 50.00 - |
50.00 50.00 - |
50.00 Majority N/A 60.98 - |
60.05 57.67 - |
57.10 56.46 - |
56.24 PTLM (Zero-shot) RoBERTa-Large 340M 52.04 51.79 51.21 50.80 50.74 50.68 51.39 51.37 51.32 DeBERTa-Large 435M 51.98 52.06 51.82 52.00 51.96 51.23 52.34 52.59 51.81 CAR 435M 52.77 52.75 51.95 51.98 52.10 51.88 53.06 53.25 52.90 CANDLE 435M 53.76 53.61 53.20 52.89 53.10 52.28 52.40 52.37 51.91 VERA-xl 3B 53.63 53.50... |
69.98 64.83 - |
64.36 61.16 - |
60.22 Meta-Llama-3-70B 71.74 - |
71.52 66.02 - |
65.05 62.00 - |
61.33 Meta-Llama-3.1-8B 71.45 - |
71.30 65.74 - |
65.69 61.63 - |
60.96 Meta-Llama-3.1-70B 72.65 - |
72.42 66.15 - |
65.54 62.50 - |
62.22 Meta-Llama-3.1-405B 75.26 - |
74.97 68.16 - |
67.33 65.66 - |
65.65 Gemma-2-2B 66.82 - |
66.80 60.56 - |
60.22 58.95 - |
58.10 Gemma-2-9B 71.27 - |
70.98 65.14 - |
64.15 61.07 - |
60.40 Gemma-2-27B 71.77 - |
71.27 66.86 - |
66.20 63.15 - |
62.70 Phi-3.5-mini 4B 68.18 - |
68.05 61.92 - |
61.15 60.36 - |
59.79 Falcon2 11B 71.73 - |
71.68 65.70 - |
65.12 61.89 - |
61.65 Mistral-7B-v0.3 72.38 - |
71.49 66.42 - |
65.77 62.18 - |
61.47 Mistral-Nemo 12B 73.18 - |
72.51 66.98 - |
66.78 62.95 - |
62.71 Mixtral-8x7B-v0.1 75.06 - |
74.25 66.39 - |
65.59 63.64 - |
62.84 PTLM & LLM (Fine-tuned) RoBERTa-Large 340M 79.18 79.27 78.86 72.26 72.32 71.74 70.26 70.38 69.83 DeBERTa-v3-Large 435M 81.10 80.76 81.03 74.26 74.56 73.78 72.00 71.93 71.99 Meta-LLaMa-3-8B 83.48 83.38 82.64 75.75 75.52 75.73 73.06 73.33 72.84 Meta-LLaMa-3.1-8B 85.24 85.07 84.64 76.44 76.51 75.53 74.48 74.44 74.38... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.