TL;DR
This paper tackles the challenge of automatically extracting relevant product attributes and their values from listing titles on e-commerce platforms. They propose a novel Text-to-Text approach, specifically a hierarchical multi-label text classification (HMLTC) model, that leverages buyer search and filtering behavior to prioritize attributes that are most likely to be used by shoppers, significantly improving attribute extraction accuracy compared to existing systems.
What Problem Does This Paper Address?
Online marketplaces rely heavily on structured product attributes (like brand, color, size) to help buyers filter search results and discover relevant items. However, requiring sellers to manually input these attributes is time-consuming and often leads to incomplete or missing information. This, in turn, can cause relevant listings to be missed by buyers using attribute filters. The paper aims to automate this process by predicting the most important attributes and their likely values directly from product titles, thereby reducing seller effort and improving listing discoverability.
Methods
The core of the paper's contribution lies in framing the problem as a "Text-to-Text" task, specifically a hierarchical multi-label text classification (HMLTC) problem that also incorporates ranking. They utilize a Convolutional Neural Network (CNN) Seq2Seq model, trained using the Fairseq framework. This approach allows the model to predict attribute-value pairs, and crucially, to rank their importance.
Key methodological aspects include:
- Leveraging Buyer Behavior: The model is trained on historical buyer data, specifically their search queries and subsequent filtering actions. This allows the model to learn which attributes are most influential in a buyer's journey to find a specific product.
- Hierarchical Multi-Label Text Classification (HMLTC): The model accounts for the hierarchical nature of attributes and their values (e.g., "Material" is an attribute, and "Leather" is a value within that attribute).
- Ranking Output: The model doesn't just predict attributes and values; it also outputs an importance ranking, helping sellers focus on what matters most to buyers.
- Text-to-Text Framework: This framework is chosen for its flexibility, allowing it to identify attribute-value pairs not explicitly present in the title and to handle multi-value attributes.
Datasets
The training and evaluation datasets were derived from two major eBay verticals: "Electronics" and "Fashion." They collected approximately 10 million and 3 million random entities from these verticals, respectively. Each entity included a listing title and a corresponding attribute-value pair that was previously used in a search filtering session. To manage the long-tail distribution of attribute-value pairs, they truncated the data to the top 800 most frequent combinations, which still captured 90% of the buyer filtering activity. A portion of this data was set aside for validation and testing.
Key Insights and Ideas
- Buyer Behavior is Key: The most significant insight is that historical buyer filtering behavior is a powerful signal for identifying important product attributes. This moves beyond simply extracting what's present in the title to understanding what buyers use to refine their searches.
- Text-to-Text for Flexibility: The Text-to-Text approach offers a distinct advantage over traditional methods like Named Entity Recognition (NER) because it can suggest attribute values that aren't explicitly mentioned in the product title. This is crucial for e-commerce where titles are often concise.
- Hierarchical Modeling Matters: Explicitly modeling the hierarchical relationship between attributes and values, combined with ranking, leads to a more nuanced and useful output for both sellers and buyers.
- Short Titles are a Challenge: The relatively short length of product titles (average of 12 tokens) and the unique jargon on platforms like eBay can pose challenges for general-purpose pre-trained language models like BERT and ULMFiT, potentially explaining why the custom Seq2Seq model performed better.
Key Results
- Improved Accuracy: The Seq2Seq-hierarchical model achieved a 33.2% improvement in relevant attribute-value extraction compared to eBay's existing production system (Attribute Extraction Service - AES).
- Outperformance of Baselines: The Seq2Seq models (both single and hierarchical) consistently outperformed BERT and ULMFiT across various metrics (Precision@k and nDCG@k) for both attribute ranking and attribute-value prediction.
- Faster Training: The Seq2Seq model was significantly faster to train (up to 15x faster than BERT and 5x faster than ULMFiT) and did not require pre-trained models.
- Qualitative Success: The model successfully identified relevant attribute-value pairs even when they were not explicitly mentioned in the title, and it could handle multi-value attributes like "size."
Limitations and Future Work
- Reliance on Existing Filters: The model's attribute ranking is inherently influenced by the filtering options that were historically available on the site. It cannot discover attribute pairs that have never been used for filtering.
- Future Work: The authors suggest using training data that includes buyers who were served with random or partially random filtering options to mitigate the feedback loop issue. They also propose further exploring the generalization capabilities of the hierarchical modeling for other downstream tasks.
Summary updated