Octet: Online Catalog Taxonomy Enrichment with Self-Supervision

Yuning Mao, Tong Zhao, Andrey Kan, Chenwei Zhang, Xin Luna Dong, Christos Faloutsos, Jiawei Han

TL;DR

This paper introduces OCTET, a self-supervised framework for automatically enriching online product taxonomies. OCTET tackles the challenge of incomplete and rapidly evolving e-commerce taxonomies by leveraging user queries and item data, significantly improving both term extraction and attachment accuracy compared to existing methods.

What Problem Does This Paper Address?

Online marketplaces like Amazon rely heavily on taxonomies (hierarchical category structures) to organize products, facilitate browsing, and improve search results. However, maintaining these taxonomies is a labor-intensive and costly process for human experts. Existing taxonomies are often incomplete, failing to capture new or emerging product types, and struggle to keep pace with the dynamic nature of e-commerce. This leads to a mismatch between how users search for items and how those items are categorized. The paper addresses the critical research question of how to effectively enrich existing, incomplete taxonomies with new terms and concepts, ensuring consistency with the established structure, without requiring manual annotation.

Methods

OCTET employs a two-stage, self-supervised approach:

  1. Term Extraction: This stage identifies potential new terms that could be added to the taxonomy. OCTET frames this as a sequence labeling task, using a BiLSTM-CRF model. Crucially, it uses distant supervision by leveraging the existing taxonomy terms (V) to label mentions in item titles and user queries. This allows it to learn to extract relevant terms without human-annotated data.
  2. Term Attachment: Once new terms (V') are extracted, this stage determines where they fit within the existing taxonomy (T). OCTET uses a Graph Neural Network (GNN) to capture rich, heterogeneous information. This includes:
  3. Structural Representation: The GNN models the existing taxonomy's structure and, importantly, the interactions between user queries, items, and taxonomy nodes derived from user behavior logs.
  4. Semantic Representation: This captures the meaning of terms, particularly by focusing on head words within multi-word terms and using dependency parsing.
  5. Lexical Representation: This leverages string-level similarities between terms.

These representations are combined to measure the relationship between a new term and existing taxonomy nodes, ultimately predicting the most appropriate parent node. The entire framework is self-supervised, relying only on the existing taxonomy and user interaction data.

Datasets

The primary datasets used are real-world online catalog taxonomies and associated user behavior logs from:

  • Amazon.com: Specifically, the "Grocery & Gourmet Food," "Home & Kitchen," and "Beauty & Personal Care" taxonomies.
  • Yelp.com: The "Business Categories" taxonomy.

The paper also mentions using a large corpus of item titles and user queries from Amazon.com for term extraction. For evaluation, they created a closed-world setup by holding out leaf nodes from existing taxonomies as the "new" terms (V') and an open-world setup where new terms were identified and then manually labeled by experts and crowdsourced workers.

Key Insights and Ideas

  • Self-Supervision is Key: The core innovation is the ability to enrich taxonomies without human labels, relying solely on the existing taxonomy structure and user interaction data. This makes the approach scalable and practical for real-world applications.
  • Heterogeneous Information Fusion: OCTET's strength lies in its ability to combine diverse signals – taxonomy structure, user behavior (query-item-taxonomy interactions), and lexical/semantic term properties – for robust term attachment.
  • Sequence Labeling for Term Extraction: Framing term extraction as a sequence labeling task, rather than just phrase mining, allows OCTET to better capture domain-specific terms and context, leading to higher precision.
  • User Behavior as a Rich Signal: The interactions between users, their queries, and the items they click provide invaluable implicit signals for understanding relationships between concepts, which is often missing in traditional text-based methods.
  • Taxonomy Enrichment is More Than Just Finding Terms: The paper highlights that attaching new terms consistently and correctly to the existing hierarchy is as crucial as extracting them, and OCTET addresses both.

Key Results

  • Term Extraction: OCTET significantly outperforms baseline methods in term extraction, achieving high precision (0.91@K=100) in open-world expert evaluation and superior recall in closed-world evaluation.
  • Term Attachment: OCTET demonstrates superior performance across multiple taxonomies and evaluation settings (closed-world and open-world), outperforming state-of-the-art methods like HiDir and MSejrKu by a significant margin (e.g., achieving 62.5% Edge-F1 and 84.2% Ancestor-F1 on Amazon Grocery & Gourmet Food in closed-world).
  • End-to-End Performance: In an open-world end-to-end evaluation, OCTET successfully enriched an online catalog taxonomy to be twice its original size (2,163 to 4,355 terms) with a high precision of 88.1% for neighbor precision.
  • Feature Importance: The study shows that combining structural, semantic, and lexical representations is crucial for optimal performance, with each component contributing uniquely.

Limitations and Future Work

  • Low Recall in Term Extraction: While precision is high, the recall for term extraction can still be low, especially in open-world crowdsourcing evaluations. This is attributed to the conciseness and noise in item titles, leaving room for improvement.
  • Potential for Taxonomy Errors: The paper acknowledges that OCTET can sometimes detect potential errors in the core taxonomy itself, suggesting that the model's learning can highlight inconsistencies in human-curated structures.
  • Future Work: The authors suggest exploring joint learning of taxonomy enrichment with downstream applications like recommendation and search, which could further enhance the utility of enriched taxonomies. They also mention the possibility of filtering low-confidence terms or allowing top-k predictions to improve precision in term attachment.

Summary updated

Abstract

Taxonomies have found wide applications in various domains, especially online for item categorization, browsing, and search. Despite the prevalent use of online catalog taxonomies, most of them in practice are maintained by humans, which is labor-intensive and difficult to scale. While taxonomy construction from scratch is considerably studied in the literature, how to effectively enrich existing incomplete taxonomies remains an open yet important research question. Taxonomy enrichment not only requires the robustness to deal with emerging terms but also the consistency between existing taxonomy structure and new term attachment. In this paper, we present a self-supervised end-to-end framework, Octet, for Online Catalog Taxonomy EnrichmenT. Octet leverages heterogeneous information unique to online catalog taxonomies such as user queries, items, and their relations to the taxonomy nodes while requiring no other supervision than the existing taxonomies. We propose to distantly train a sequence labeling model for term extraction and employ graph neural networks (GNNs) to capture the taxonomy structure as well as the query-item-taxonomy interactions for term attachment. Extensive experiments in different online domains demonstrate the superiority of Octet over state-of-the-art methods via both automatic and human evaluations. Notably, Octet enriches an online catalog taxonomy in production to 2 times larger in the open-world evaluation.
BibTeX
@inproceedings{mao2020octet,
  author = {Yuning Mao and Tong Zhao and Andrey Kan and Chenwei Zhang and Xin Luna Dong and Christos Faloutsos and Jiawei Han},
  title = {{Octet: Online Catalog Taxonomy Enrichment with Self-Supervision}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2020},
  url = {https://openreview.net/forum?id=CSldvhzIpKw},
}

Loading related content...

Figure