Automatic Marketing Theme and Commodity Construction System for E-commerce

Zhiping Wang, Peng Lin, Hainan Zhang, Hongshen Chen, Tianhao Li, Zhuoye Ding, Sulong Xu, Jinghe HuEMNLP 2023

TL;DR

This paper introduces MTCC, an automated system for generating e-commerce marketing themes and curating relevant product collections. By leveraging pre-trained language models and a novel theme-commodity consistency module, MTCC aims to improve user shopping efficiency and online conversion rates, outperforming traditional manual methods.

What Problem Does This Paper Address?

In the booming world of e-commerce, consumers often face an overwhelming number of individual product listings. While traditional recommendation systems present products one by one, this can be inefficient for users with specific needs or interests. For example, someone looking for "warm toddler shoes" would prefer to see a curated collection of such items rather than sifting through unrelated products. Manually creating these themed collections and their associated marketing narratives is a labor-intensive, time-consuming, and often ineffective process, leading to suboptimal online performance. This paper addresses the need for an automated system that can generate compelling marketing themes and assemble relevant product assortments, thereby enhancing the user shopping experience and boosting e-commerce metrics.

Methods

The MTCC system is built around a four-module pipeline:

  • Theme Generation Module: This module utilizes a pre-trained language model (specifically, an e-commerce fine-tuned UniLM) to generate marketing theme titles. It leverages popular keywords from e-commerce search engines to rank and select the most relevant and appealing themes.
  • Commodity Selection Module: This module employs a BERT classification model to identify products that are consistent with a generated marketing theme. It concatenates product information (title, attributes) with the theme and uses a consistency score to select relevant commodities. An enhancement is made by incorporating a pre-trained search engine model to further refine commodity selection by considering attribute conflicts like color and size.
  • Indicator Simulator Module: To predict the online effectiveness of generated themes before they are published, this module uses a BERT classification model trained on online user behavior data (specifically, click-through rates). It simulates the potential CTR of a theme, flagging low-quality themes for further refinement.
  • Theme Rewriter Module: For themes identified as low-quality by the indicator simulator, this module uses the unilm model to rewrite them. It iteratively incorporates information from the selected commodities to generate more effective marketing themes.

Datasets

The paper utilizes data from a publicly available e-commerce platform (JD.com). This includes:

  • Human-created and reviewed commodity-theme pairs for training the theme generation model.
  • Commodity titles, category names, and attributes.
  • Human-written short title-commodity pairs for training the commodity selection module.
  • Online marketing theme log data (user browsing and clicking data) for training the indicator simulator.

The authors do not explicitly mention creating new benchmarks but rather leverage existing data from a large e-commerce platform.

Key Insights and Ideas

  • The Power of Thematic Grouping: The core insight is that users benefit significantly from curated collections of products organized around specific marketing themes, rather than just individual product recommendations. This reduces cognitive load and improves shopping efficiency.
  • Automating Creativity: The paper demonstrates that AI can be effectively used to automate the creative process of generating marketing themes, a task traditionally requiring human expertise.
  • Iterative Refinement: The inclusion of an indicator simulator and a theme rewriter module creates a feedback loop, allowing the system to learn from simulated online performance and continuously improve theme quality.
  • Leveraging Pre-trained Models: The success of MTCC hinges on fine-tuning powerful pre-trained language models (like UniLM and BERT) on domain-specific e-commerce data, enabling them to understand nuanced product attributes and user preferences.

Key Results

  • Theme Generation: The e-commerce fine-tuned UniLM model outperformed baseline models (BART, T5, UniLM-BERT) on metrics like SacreBLEU, ROUGE, and BLEU, producing smoother and more relevant themes.
  • Commodity Selection: The MTCC system, enhanced with a search engine, achieved superior performance in selecting relevant commodities compared to BERT, RoBERTa, and the search engine alone.
  • Indicator Simulation: The e-commerce UniLM model showed better accuracy in predicting CTR compared to BERT and RoBERTa.
  • Theme Rewriting: The e-commerce UniLM model outperformed baselines in generating smoother and more relevant themes.
  • Online A/B Test: Deploying AI-generated marketing themes resulted in an 11% increase in CTR compared to human-produced themes. Furthermore, AI production efficiency was 10 times higher than manual production, significantly saving costs and improving system efficiency.

Limitations and Future Work

  • User Shopping Information: The current system primarily uses instant user interests for theme generation. The authors acknowledge that incorporating actual user shopping history and preferences into the production process would lead to more personalized marketing themes and a deeper personalized experience.
  • Manual Filtering: Despite the automation, marketing themes still require manual filtering before online publication to ensure quality and user experience. Future work could explore further automating this screening process.

Summary updated

BibTeX
@inproceedings{wang2023automatic,
  author = {Zhiping Wang and Peng Lin and Hainan Zhang and Hongshen Chen and Tianhao Li and Zhuoye Ding and Sulong Xu and Jinghe Hu},
  title = {{Automatic Marketing Theme and Commodity Construction System for E-commerce}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
}
Figure