[Paper Review] Attribute Extraction from Product Titles in eCommerce

Ajinkya More|arXiv (Cornell University)|Aug 15, 2016

Natural Language Processing Techniques12 references20 citations

TL;DR

This paper proposes a sequence labeling approach using Conditional Random Fields (CRF) and Structured Perceptron with a curated normalization scheme to extract product attribute values—particularly 'brand'—from short, unstructured eCommerce titles. The method achieves 92–95% post-normalization precision and significantly improves product discoverability, with a CausalImpact analysis showing over 250,000 additional impressions for products augmented with extracted brand metadata.

ABSTRACT

This paper presents a named entity extraction system for detecting attributes in product titles of eCommerce retailers like Walmart. The absence of syntactic structure in such short pieces of text makes extracting attribute values a challenging problem. We find that combining sequence labeling algorithms such as Conditional Random Fields and Structured Perceptron with a curated normalization scheme produces an effective system for the task of extracting product attribute values from titles. To keep the discussion concrete, we will illustrate the mechanics of the system from the point of view of a particular attribute - brand. We also discuss the importance of an attribute extraction system in the context of retail websites with large product catalogs, compare our approach to other potential approaches to this problem and end the paper with a discussion of the performance of our system for extracting attributes.

Motivation & Objective

• To address the challenge of extracting attribute values from short, unstructured product titles in large eCommerce catalogs.
• To improve product discoverability by enabling faceted search through accurate attribute metadata.
• To develop a scalable, low-resource method for attribute extraction that minimizes manual labeling.
• To evaluate the real-world impact of attribute extraction on key metrics like impressions, clicks, and conversion.
• To extend the approach beyond 'brand' to other attributes such as manufacturer part numbers, character names, and sports teams.

Proposed method

• Model attribute extraction as a sequence labeling problem using CRF and Structured Perceptron algorithms.
• Use distant supervision to generate initial training data from existing product metadata without manual labeling.
• Apply a curated normalization scheme to standardize variations of attribute values (e.g., 'Hewlett Packard' → 'HP').
• Train separate models per attribute due to category-specific applicability and varying data sparsity.
• Leverage a synthetic control group via the CausalImpact R package to estimate causal effects of metadata augmentation.
• Iteratively update the normalization dictionary based on analyst feedback to improve precision.

Experimental results

Research questions

• RQ1 Can sequence labeling models effectively extract attribute values from short, syntactically unstructured product titles?
• RQ2 How does a curated normalization scheme improve precision and handle lexical variations in attribute values?
• RQ3 What is the real-world impact of attribute extraction on product impressions and discoverability?
• RQ4 To what extent can distant supervision reduce the need for manual labeling in training data creation?
• RQ5 Can the same pipeline be generalized across diverse attributes like brand, model number, and character names?

Key findings

• The system achieved 92–95% post-normalization precision for brand attribute extraction, with a 1–3% improvement over raw algorithmic outputs.
• Over 98% of products classified as 'unbranded' were confirmed as true negatives, indicating high specificity.
• The model discovered over 1,000 brand values not present in the existing Walmart brand database.
• CausalImpact analysis showed a predicted increase of over 250,000 impressions for products with extracted brand metadata during the test period.
• Similar positive impacts were observed for other attributes, including increased clicks, add-to-cart rates, and order volume.
• The system was successfully deployed to extract over 20 attributes, with separate models trained per attribute to optimize performance for category-specific needs.