Enriching a Fashion Knowledge Graph from Product Textual Descriptions

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.

Introduction

Knowledge Graphs (KG) have become an essential tool for representing and organizing information in various domains. In recent years, they have gained significant attention in the e-commerce sector due to their potential to enhance search and recommendation systems. The Fashion Knowledge Graph (FKG) at Farfetch is a large-scale KG that contains rich metadata about fashion products, brands, and attributes. However, enriching this FKG with product textual descriptions presents unique challenges due to the specific nature of entities in the fashion domain and data scarcity.

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 FKG at Farfetch. They specifically focus on named entity recognition (NER) and entity linking (EL) techniques to extract and link relevant information from product textual descriptions to entities within the FKG.

The Significance of Enriched Fashion Knowledge Graph

The authors highlight the importance of having a complete and enriched FKG as an e-commerce backbone. A well-structured KG can greatly impact downstream applications such as search and recommendations by providing accurate information about products' attributes. This not only improves user experience but also helps businesses make informed decisions based on customer preferences.

Challenges in Enriching a Fashion Knowledge Graph

Unlike more generic KGs like Wikidata or Yago, entities in the fashion domain are highly specific (e.g., product attributes), making it challenging to obtain long textual descriptions for training NER models. Additionally, data scarcity poses another hurdle as labeling datasets for training supervised models is labor-intensive.

To address these challenges, the authors employ transfer learning-based approaches that leverage pre-trained models to train an NER module with limited manually labeled data. This approach reduces the need for large amounts of training data and speeds up the process of model development.

Enriching the Fashion Knowledge Graph

The authors' proposed method consists of two main steps: named entity recognition (NER) and entity linking (EL). The NER module identifies relevant entities within product textual descriptions, while the EL module links these entities to corresponding entities within the FKG.

Named Entity Recognition (NER)

The authors use a transfer learning-based approach to train their NER model due to limited training data availability. They fine-tune a pre-trained BERT model on a small dataset containing manually labeled fashion-related entities. BERT is a state-of-the-art language representation model that has shown promising results in various natural language processing tasks.

After fine-tuning, the NER model achieves 89.75% accuracy in identifying fashion-related entities within product descriptions. This demonstrates that transfer learning can be effective even with limited training data, making it a viable option for enriching KGs in specific domains like fashion.

Entity Linking (EL)

Once relevant entities are identified by the NER module, they are linked to corresponding entities within the FKG using an EL module. The authors employ both rule-based and machine learning approaches for this task, depending on the availability of training data for each entity type.

For example, brand names can be easily linked using simple string matching rules since they are highly specific and have consistent naming conventions. On the other hand, more general attributes like color or material may require machine learning models trained on larger datasets due to their variability.

Experimental Results

To evaluate their proposed method's effectiveness, the authors conduct experiments on a subset of products from Farfetch's FKG. They compare their results with baseline methods, including a rule-based approach and an NER model trained on a larger dataset from a different domain.

The results show that their proposed method outperforms the baseline methods, achieving 89.75% accuracy in NER and 81.25% accuracy in EL. This demonstrates the effectiveness of transfer learning-based approaches for enriching KGs in specific domains with limited training data.

Conclusion

In conclusion, this research paper presents an effective method for enriching a Fashion Knowledge Graph using advanced techniques such as transfer learning-based NER and entity linking. The study showcases how leveraging these techniques can lead to automatic enrichment of product attributes within the graph, improving downstream applications like search and recommendations.

This research was presented at the International Workshop on Knowledge Graph Generation from Text (ESWC 2022), highlighting its relevance in advancing knowledge graph research within the e-commerce sector. With further advancements in natural language processing and machine learning techniques, we can expect to see more efficient methods for enriching KGs in specific domains like fashion, ultimately enhancing user experience and business operations.