Lot or Not: Identifying Multi-Quantity Offerings in E-Commerce

Gal Lavee, Ido GuyACL 2022

TL;DR

This paper tackles the challenge of automatically identifying "lot" offerings (multiple identical items sold together) in e-commerce, a crucial task for improving buyer experience and seller efficiency. They propose a novel approach using natural language processing on product titles, achieving high accuracy in both classifying lot offerings and predicting their exact quantity.

What Problem Does This Paper Address?

In the vast world of e-commerce, sellers often list items in "lots" – groups of identical products sold as a single unit. This is common for items like multi-packs of batteries or boxes of pens. While these lot offerings are valuable for buyers seeking bulk discounts and for sellers managing inventory, e-commerce platforms struggle to automatically identify them. This is because sellers rarely explicitly flag their listings as "lots" or specify the quantity. This lack of identification hinders crucial e-commerce tasks like matching products to catalogs, ranking search results, and providing accurate pricing guidance. The paper aims to bridge this gap by developing methods to accurately detect lot offerings and their sizes using only the product title.

Methods

The core of this research lies in applying advanced natural language processing (NLP) techniques to product titles. Instead of relying on explicit "lot" tags, which are rarely used, the authors focus on the rich linguistic cues present in titles. They frame the problem in two ways:

  1. Binary Classification: Determining if a title represents a lot offering or a single item.
  2. Lot Size Prediction: Predicting the exact number of items in a lot.

For lot size prediction, they move beyond traditional regression, treating it as a sequence labeling task. This means the model predicts a "lot size token" for each word in the title, allowing for more nuanced prediction. They experiment with various model architectures, including:

  • Feature-based models: Using regular expressions and n-grams.
  • Recurrent Neural Networks (RNNs): Specifically LSTMs, which are good at processing sequential data.
  • Transformer-based models: Leveraging powerful architectures like BERT, which excel at capturing contextual information.

A key innovation is their custom tokenization scheme, designed to better handle the unique language and numerical patterns found in e-commerce titles, especially for identifying lot-specific tokens. They also explore the benefits of pre-training models on large text corpora and fine-tuning them on e-commerce data.

Datasets

The researchers created a new dataset by manually labeling nearly 20,000 offering titles from eBay across multiple categories. This dataset includes labels for both lot status and lot size. They also created a "Heterogeneous" dataset combining examples from various categories, which was particularly challenging due to the underrepresentation of lot offerings.

Key Insights and Ideas

  • Titles are Rich with Information: Even without explicit "lot" tags, product titles contain strong linguistic signals (e.g., "pack of 6," "3-pack," "lot of 10") that can be leveraged to identify lot offerings.
  • Sequence Labeling for Lot Size: Treating lot size prediction as a sequence labeling problem, rather than simple regression, significantly improves accuracy by focusing on identifying specific "lot size tokens" within the title.
  • Custom Tokenization Matters: Standard NLP tokenization methods can be suboptimal for e-commerce titles. A custom approach that intelligently separates punctuation and identifies numerical quantities is crucial for better performance.
  • Pre-trained Models are Powerful: Leveraging pre-trained language models (like BERT) and fine-tuning them on e-commerce data yields substantial performance gains, especially when labeled data is scarce.
  • Transformer Architectures Shine: Transformer-based models, particularly those with attention mechanisms, demonstrate superior performance in both lot classification and size prediction tasks.

Key Results

The study achieved impressive results:

  • High Binary Accuracy: The best-performing model, a transformer encoder (TRANS_ENC_SZ), achieved binary accuracy of up to 0.945 on the Heterogeneous dataset.
  • Accurate Lot Size Prediction: The same transformer model achieved lot size accuracy of up to 0.922 on the Heterogeneous dataset.
  • Lot Size Prediction is as Accurate as Classification: Surprisingly, the models were almost as good at predicting the exact lot size as they were at simply classifying an offering as a lot or not.
  • Transformer vs. RNN: Transformer encoder architectures generally outperformed RNN-based models, suggesting that local word structure is more critical than global word ordering for this task.
  • Custom Tokenization Advantage: Their custom tokenization scheme outperformed standard BPE tokenization, especially for the lot classification task.

Limitations and Future Work

The authors acknowledge that while their models achieve high accuracy using only titles, incorporating other signals like images, descriptions, and price could further improve performance. They also note that false negative errors (misclassifying a lot as a non-lot) were more common than false positives.

Future work could explore:

  • Multi-modal approaches: Integrating information from images and descriptions.
  • Price analysis: Using price as an additional signal for lot identification.
  • Cross-lingual generalization: Adapting these methods to other languages.
  • Real-time deployment: Investigating the complexity-accuracy trade-off for practical production environments, suggesting simpler models like LSTM might be preferable for efficiency.

Summary updated

BibTeX
@inproceedings{lavee2022lot,
  author = {Gal Lavee and Ido Guy},
  title = {{Lot or Not: Identifying Multi-Quantity Offerings in E-Commerce}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
}
Figure