Interactive Latent Knowledge Selection for E-Commerce Product Copywriting Generation

Zeming Wang, Yanyan Zou, Yuejian Fang, Hongshen Chen, Mian Ma, Zhuoye Ding, Bo LongACL 2022

TL;DR

This paper tackles the challenge of generating compelling e-commerce product descriptions by proposing a novel framework that interactively selects relevant "human knowledge" (auxiliary details) from noisy product information. Their method, ILVT, significantly improves both the quality and diversity of generated copywritings, outperforming existing approaches.

What Problem Does This Paper Address?

In the booming e-commerce landscape, product copywriting plays a crucial role in attracting customers and driving sales. While traditional methods rely solely on product titles and attributes, this paper argues that valuable insights are often hidden within auxiliary product details (e.g., shopper reviews, marketer-added descriptions). These details, however, are often noisy and contain irrelevant information. The core problem is how to effectively extract the salient, informative parts of this "human knowledge" and integrate it with existing product information to generate high-quality, diverse, and engaging product descriptions. Existing methods either ignore this rich source of information or simply concatenate it, leading to suboptimal results due to the inherent noise.

Methods

The authors introduce an Interactive Latent Variables model (ILVT) built upon the Transformer architecture. The key innovation lies in its ability to interactively select informative human knowledge from noisy product details.

Here's a breakdown of their approach:

  • Interactive Latent Variables: ILVT employs two types of latent variables:
  • Description Latent Variable ($z_d$): Captures intrinsic product information (title and attributes) and guides the selection of knowledge.
  • Knowledge Latent Variable ($z_k$): Represents the selected knowledge and is influenced by both the description latent variable and the product details.
  • Knowledge Selection: A heuristic matching algorithm is used to filter out noisy fragments from product details. Then, a pseudo-labeling strategy is employed to identify the most relevant knowledge snippet. This selected knowledge is then encoded.
  • Transformer Architecture: Both an encoder and decoder based on the Transformer architecture are used. The encoder processes product titles, attributes, and selected knowledge. The decoder generates the final product description, incorporating the selected knowledge and latent variables.
  • Copy Mechanism: A copy mechanism is integrated into the decoder to allow direct copying of relevant words from the selected knowledge and input product information, enhancing accuracy.

The "interactive" aspect comes from the way $z_d$ and $z_k$ influence each other, allowing for a dynamic selection and generation process.

Datasets

The paper introduces a new dataset called JDK (JD.com Knowledge).

  • JDK Dataset: This dataset comprises 220,000 instances of Chinese e-commerce products. Each instance includes:
  • Product Title
    • Attribute Set
    • Product Details (auxiliary human knowledge)
    • Corresponding Product Description (copywriting)

The dataset covers 30 categories and is split into training, validation, and test sets. The authors provide statistics on the average lengths of these components, highlighting the significant length of product details compared to titles and attributes.

Key Insights and Ideas

  • Human knowledge is valuable but noisy: Auxiliary product details contain rich information for copywriting, but simply including them directly harms generation quality due to noise. Effective selection is crucial.
  • Interactive latent variables are key to diversity and quality: The interplay between description and knowledge latent variables allows the model to not only select relevant information but also to generate more diverse and engaging copywritings.
  • Pseudo-labeling for knowledge selection: In the absence of explicit labels for salient knowledge, a pseudo-labeling strategy based on semantic similarity proves effective for guiding the selection process.
  • Transformer's suitability for this task: The Transformer architecture, with its attention mechanisms, is well-suited for processing and integrating various forms of textual information, including noisy details.

Key Results

  • Automatic Evaluation: ILVT significantly outperforms baseline models (CONVSEQ2SEQ, TRANSFORMER, KOBE, PTRANS) across various metrics, including BLEU (for quality) and Distinct (for diversity).
  • ILVT achieves a BLEU score of 15.24, outperforming the best baseline (PTRANS+PSE, which uses pseudo-labeled knowledge) by a notable margin.
  • ILVT also demonstrates superior diversity, with Distinct-1 of 44.57 and Distinct-2 of 74.80, significantly higher than all baselines.
  • Human Evaluation: Human evaluators rated ILVT's generated copywritings higher than baselines across correctness, diversity, and coherence. ILVT achieved an average score of 4.77 for correctness, 4.62 for diversity, and 4.51 for coherence on a 1-5 scale.
  • Ablation Studies: Removing specific modules like the copy mechanism or the knowledge latent variable led to significant drops in performance, underscoring the importance of each component in ILVT.

Limitations and Future Work

  • Pseudo-labeling limitations: While effective, the pseudo-labeling approach relies on heuristics and might not perfectly capture human judgment of "salient" knowledge.
  • Computational Cost: The interactive nature and Transformer architecture likely contribute to higher computational demands, though this is not explicitly detailed as a limitation.
  • Future Work: The paper suggests that further exploration of how to dynamically model the input product title, attributes, and human-written advertising knowledge to extract salient parts and inject them into generated copywritings could lead to further improvements in diversity and quality. The authors also implicitly suggest that refining the knowledge selection mechanism could be a fruitful area for future research.

Summary updated

BibTeX
@inproceedings{wang2022interactive,
  author = {Zeming Wang and Yanyan Zou and Yuejian Fang and Hongshen Chen and Mian Ma and Zhuoye Ding and Bo Long},
  title = {{Interactive Latent Knowledge Selection for E-Commerce Product Copywriting Generation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
}
Figure