Semi-supervised Category-specific Review Tagging on Indonesian E-Commerce Product Reviews

Meng Sun, Marie Stephen Leo, Eram Munawwar, Paul C. Condylis, Sheng-yi Kong, Seong Per Lee, Albert Hidayat, Muhamad Danang KeriantoACL 2020

TL;DR

This paper tackles the challenge of automatically generating descriptive tags for product reviews on Indonesian e-commerce platforms. They introduce a semi-supervised pipeline that effectively extracts "category-specific" topics, which are crucial for buyers but difficult to tag using traditional supervised methods due to their domain-specific nature. This approach enables more efficient filtering of reviews, improving the user experience.

What Problem Does This Paper Address?

In e-commerce, product reviews are a goldmine of customer feedback, influencing purchasing decisions. However, manually sifting through millions of reviews to find relevant information is a daunting task for buyers. The authors identify two types of topics within reviews: "generic" topics (like customer service or delivery) and "category-specific" topics (attributes unique to a product type, e.g., "cable hole" for phone cases). While generic topics can be tagged with supervised learning, category-specific topics are too diverse and numerous to be effectively labeled this way. This paper aims to bridge this gap by developing a scalable, semi-supervised method to automatically generate category-specific tags for Indonesian e-commerce product reviews.

Methods

The proposed pipeline consists of several key stages:

  • Phrase Extraction: They leverage Stanford NLP's dependency parser to extract noun-based phrases from review text. Specifically, they focus on Universal Dependency Relations (UDP) like amod (adjectival modifier), nsubj (nominal subject), and compound (compound noun).
  • Topic Generation: For each product category, they employ the Unsupervised Aspect Extraction (UAE) model (He et al., 2017). This neural network model uses an attention mechanism to group extracted phrases into latent topics based on their embeddings. The UAE model has an embedding layer, an attention layer, and an auto-encoder layer.
  • Category-specific Topic Filtering: To distinguish between generic and category-specific topics, they use a supervised filtering approach. A general word list is created from frequent words in generic phrases. Topics are then classified as category-specific if a low percentage of their phrases are considered "general" (i.e., contain words from the general word list).
  • Category-specific Tag Generation: Finally, for each product, they group review phrases to their corresponding filtered category-specific topics. The most frequent phrase within each topic for that product is selected as the category-specific tag.

Datasets

The authors utilize a large dataset of 89.5 million products across 18 product categories from Tokopedia, an Indonesian e-commerce platform. They do not introduce a new benchmark dataset but rather apply their method to real-world, large-scale e-commerce review data.

Key Insights and Ideas

  • The Value of Category-Specific Tags: The paper strongly emphasizes that category-specific topics are crucial for buyers looking to filter reviews effectively, as they highlight product attributes that matter most within a particular domain.
  • Semi-Supervised Approach for Scalability: Recognizing the impossibility of manual labeling for the vast number of category-specific topics, the authors' semi-supervised approach is key to achieving scalability on a large e-commerce platform.
  • Leveraging Neural Attention for Topic Modeling: The use of the UAE model with its attention mechanism is a significant methodological choice, allowing for more nuanced phrase grouping into topics compared to traditional bag-of-words models like LDA.
  • The Importance of Phrase-Level Analysis: By focusing on extracted phrases rather than individual words, the method captures more meaningful semantic units relevant to product attributes.

Key Results

  • The proposed pipeline was successfully implemented and scaled on Tokopedia's extensive product review data.
  • Manual evaluation by an internal product team showed that the generated category-specific topics are coherent and relevant.
  • The method efficiently generates category-specific product tags, aiding in review filtering.
  • The average time to generate category-specific topics per category was approximately 2 hours for around 0.5 million reviews.
  • On average, 5 category-specific topics were generated per category.

Limitations and Future Work

  • Language Nuances: The paper acknowledges the challenges of working with informal Indonesian language, which includes slang, abbreviations, and variations in sentence structure, though their method is shown to be robust.
  • Manual Labeling for Filtering: While the semi-supervised approach reduces the manual labeling burden, some manual labeling is still required for the filtering stage.
  • Future Work: The paper implies that the generated tags are uploaded to a data warehouse, suggesting potential for further integration into the e-commerce platform's search and filtering functionalities. The success of the current pipeline also opens avenues for applying similar techniques to other languages or e-commerce platforms.

Summary updated

BibTeX
@inproceedings{sun2020semi,
  author = {Meng Sun and Marie Stephen Leo and Eram Munawwar and Paul C. Condylis and Sheng-yi Kong and Seong Per Lee and Albert Hidayat and Muhamad Danang Kerianto},
  title = {{Semi-supervised Category-specific Review Tagging on Indonesian E-Commerce Product Reviews}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
}
Figure