# Enriching a Fashion Knowledge Graph from Product Textual Descriptions

**Authors:** João Barroca, Abhishek Shivkumar, Beatriz Quintino Ferreira, Evgeny Sherkhonov, João Faria  
**arXiv:** [2206.01087](https://arxiv.org/abs/2206.01087)  
**Date:** 2022-06-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.01087v1/

## Key points

- Authors: João Barroca, Abhishek Shivkumar, Beatriz Quintino Ferreira, Evgeny Sherkhonov, and João Faria
- Focus on enriching the large-scale Fashion Knowledge Graph (FKG) at Farfetch using named entity recognition (NER) and entity linking (EL) techniques from product textual descriptions
- Importance of a complete and enriched FKG for e-commerce applications such as search and recommendations
- Challenges in fashion domain knowledge graph enrichment include specificity of entities and data scarcity for training supervised models
- Approach includes transfer learning-based NER module training with limited labeled data and EL module linking named entities to relevant entities in FKG
- Experiments show promising results with 89.75% accuracy in NER using pre-trained model
- Research presented at the International Workshop on Knowledge Graph Generation from Text (ESWC 2022), highlighting relevance in advancing e-commerce knowledge graph research

## Summary

In their paper titled "Enriching a Fashion Knowledge Graph from Product Textual Descriptions," authors João Barroca, Abhishek Shivkumar, Beatriz Quintino Ferreira, Evgeny Sherkhonov, and João Faria delve into the application of techniques for enriching the large-scale Fashion Knowledge Graph (FKG) at Farfetch. <br>
Knowledge Graphs are recognized for their utility in representing information and have become integral to various e-commerce applications. The authors specifically focus on named entity recognition (NER) and entity linking (EL) techniques to extract and link rich metadata from product textual descriptions to entities within the FKG. <br>
The significance of having a complete and enriched FKG as an e-commerce backbone is highlighted, as it can greatly impact downstream applications such as search and recommendations. However, enriching a Knowledge Graph in the fashion domain presents unique challenges. Unlike more generic KGs like Wikidata and Yago, entities in the fashion domain are highly specific (e.g., product attributes), and obtaining long textual descriptions is not always feasible. Additionally, data scarcity poses a hurdle as labeling datasets for training supervised models is labor-intensive. <br>
To address these challenges, the authors employ a transfer learning-based approach to train an NER module with limited manually labeled data. Subsequently, an EL module links identified named entities to relevant entities within the FKG. Despite relying on rule-based or machine learning models due to limited training data availability, experiments with a pre-trained model demonstrate promising results with 89.75% accuracy in NER.<br>
Overall, this research showcases how leveraging advanced techniques for enriching a Fashion Knowledge Graph can lead to automatic enrichment of product attributes within the graph. The study was presented at the International Workshop on Knowledge Graph Generation from Text (ESWC 2022), underscoring its relevance in advancing knowledge graph research within the e-commerce sector.