OpenBrand: Open Brand Value Extraction from Product Descriptions

Kassem Sabeh, Mouna Kacimi, Johann GamperACL 2022

TL;DR

This paper introduces OpenBrand, a novel approach for automatically extracting brand names from product descriptions. By incorporating character-level embeddings alongside traditional word embeddings, OpenBrand significantly improves the ability to discover new and unseen brand names, outperforming existing state-of-the-art methods.

What Problem Does This Paper Address?

In e-commerce, understanding product attributes is crucial for customer purchasing decisions. Brands play a significant role, often influencing perceptions of quality, features, and style. However, accurately extracting brand names from unstructured product descriptions is challenging, especially when dealing with new or emerging brands that haven't been seen before. Existing methods often rely on pre-defined dictionaries or hand-crafted rules, which fail to adapt to the dynamic nature of brand landscapes. More recent deep learning approaches, while powerful, often limit their representations to word embeddings, hindering their ability to generalize to unseen brand names due to variations in spelling, sub-branding, or brand fragmentation.

Methods

The core of OpenBrand is a BiLSTM-CRF-Attention model enhanced with character-level embeddings. Here's a breakdown of the key components:

  • Embedding Layer: This layer combines two types of embeddings:
  • Word Embeddings: Pre-trained GloVe embeddings are used to capture semantic meaning of words.
  • Character-Level Embeddings: This is the novel contribution. Two architectures are explored for generating character embeddings:
  • CNN-based: Convolutional Neural Networks are used to capture position-invariant morphological features, which are effective for learning task-specific representations.
  • LSTM-based: Recurrent Neural Networks (specifically BiLSTMs) are used to capture sequential and position-dependent character features. These character embeddings are concatenated with word embeddings to create a richer representation for each token.

  • Contextual Layer: A Bidirectional LSTM (BiLSTM) layer processes the combined embeddings to capture contextual information. This is followed by a self-attention mechanism, which allows the model to focus on the most relevant parts of the input sequence for making tagging decisions.

  • CRF Layer: A Conditional Random Field (CRF) layer is used as the output layer. CRFs are effective for sequence labeling tasks as they model dependencies between output tags, ensuring more coherent and accurate predictions (e.g., if a token is tagged as "inside a brand," the next token is more likely to also be "inside a brand" rather than "beginning of a brand").

The model uses the {B, I, O} (Beginning, Inside, Outside) tagging scheme to identify brand name entities.

Datasets

  • AZ-base Dataset: A newly introduced dataset created from a subset of the Amazon Review Dataset. It contains over 250k product titles across five categories (Grocery & Gourmet Food, Toys & Games, Sports & Outdoors, Electronics, Automotive). The dataset is split into 70% training, 10% validation, and 20% testing.
  • AZ-zero-shot Dataset: A subset of AZ-base specifically designed to evaluate zero-shot extraction. This dataset has no overlapping brand values between the training and test sets, containing over 8k new and unique brand values.
  • AZ-new-cat Dataset: Another subset of products from Amazon, but with three new categories (Clothing, Shoes & Jewelry; Pet Supplies; Cell Phones & Accessories). This dataset tests the model's ability to generalize to unseen product domains.

Key Insights and Ideas

  • Character Embeddings are Key for Generalization: The most significant finding is that character-level embeddings are crucial for discovering unseen brand names. They help the model learn robust representations that are less sensitive to exact word forms, enabling better generalization.
  • CNNs Outperform LSTMs for Character Embeddings: While both CNN and LSTM-based character encoders improved performance, CNNs offered a better balance of performance and computational efficiency, with significantly lower training times.
  • Attention and CRF Layers Enhance Sequence Tagging: The combination of self-attention and CRF layers in the model architecture proved effective in capturing complex dependencies and improving the overall accuracy of brand name extraction.
  • Zero-Shot Performance is Challenging but Achievable: The AZ-zero-shot dataset highlights the difficulty of extracting completely new brand names. However, OpenBrand still shows superior performance compared to baselines, demonstrating its ability to leverage sub-sequence similarities.
  • Brand Entity Complexity Matters: Detecting brand names composed of multiple entities (e.g., "Gentle Monster") is more challenging. OpenBrand shows a more graceful degradation in performance on multi-entity brands compared to OpenTag.

Key Results

  • Overall Performance: OpenBrand consistently outperformed state-of-the-art baselines (BiLSTM, BiLSTM-CRF, OpenTag) across all categories in the AZ-base dataset, achieving up to a 6.1% improvement in F1 score over OpenTag.
  • Impact of Character Embeddings: Adding character-level embeddings to baseline models improved their F1 scores significantly. For instance, adding a CNN representation to a BiLSTM-CRF model improved performance by 1.15%, while an attention layer only improved it by 0.14%.
  • Zero-Shot Performance: On the AZ-zero-shot dataset, OpenBrand-CNN achieved an F1 score of 43.44%, outperforming OpenTag (41.53%).
  • Cross-Category Generalization: On the AZ-new-cat dataset, OpenBrand achieved gains of up to 2.7% in F1 score over OpenTag, demonstrating good generalization capabilities.
  • Computational Efficiency: OpenBrand-CNN had a 22% increase in average training time per epoch compared to the baseline BiLSTM-CRF-Att, while OpenBrand-LSTM had a 59% increase, highlighting the efficiency of the CNN approach.

Limitations and Future Work

  • Ambiguity in Product Descriptions: The paper notes that some product descriptions can be ambiguous, containing values of other brands or compatibility information that can confuse the model. This is particularly observed in the Automotive category.
  • Compound Brand Values: While OpenBrand handles compound brand values better than baselines, it still faces challenges with them.
  • Disambiguation: A natural extension is to address the problem of disambiguation, which requires more training data to understand patterns better.
  • Ambiguous Product Descriptions: Extending the tagging model to capture ambiguous product descriptions is suggested, especially when extracting brand values from sources other than concise product descriptions.
  • Computational Cost: While CNNs are more efficient than LSTMs for character embeddings, the overall model still involves complex deep learning architectures, which can be computationally intensive.

Summary updated

BibTeX
@inproceedings{sabeh2022openbrand,
  author = {Kassem Sabeh and Mouna Kacimi and Johann Gamper},
  title = {{OpenBrand: Open Brand Value Extraction from Product Descriptions}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
}
Figure