Unsupervised Extraction of Attributes and Their Values from Product Description

International Joint Conference on Natural Language Processing, pages 1339–1347, Nagoya, Japan, 14-18 October 2013. Unsupervised Extraction of Attributes and Their Values from Product Description Keiji Shinzato Satoshi Sekine Rakuten Institute of Technology {keiji.shinzato, satoshi.b.sekine}@mail.rakuten.com Abstract This paper describes an unsupervised method for extracting product attributes and their values from an e-commerce product page. Previously, distant supervi- sion has been applied for this task, but it is not applicable in domains where no reli- able knowledge base (KB) is available. In- stead, the proposed method automatically creates a KB from tables and itemizations embedded in the product’s pages. This KB is applied to annotate the pages automati- cally and the annotated corpus is used to train a model for the extraction. Because of the incompleteness of the KB, the anno- tated corpus is not as accurate as a manu- ally annotated one. Our method tries to fil- ter out sentences that are likely to include problematic annotations based on statisti- cal measures and morpheme patterns in- duced from the entries in the KB. The experimental results show that the perfor- mance of our method achieves an average F score of approximately 58.2 points and that filters can improve the performance. 1 Introduction E-commerce enables consumers to purchase a large number of products from a variety of cat- egories such as books, electronics, clothing and foods. In recent years, this market has grown rapidly around the world. Online shopping is re- garded as a very important part of our daily lives. Structured product data are most important for online shopping malls. In particular, product at- tributes and their values (PAVs) are crucial for many applications such as faceted navigation and recommendation. However, since structured infor- mation is not always provided by the merchants, it is important to build technologies to create this structured information (such as PAVs) from un- structured data (such as a product description). Because of the wide variety of product types, such technology should not rely on a manually anno- tated corpus. One of the promising methods for information extraction (IE) without a manually annotated corpus is distant supervision (Mintz et al., 2009), which leverages an existing knowledge base (KB) such as Wikipedia or Freebase to an- notate texts using the KB instances. These popu- lar KBs, however, are not very helpful for distant supervision in an e-commerce domain for the fol- lowing reasons. (1) An infobox in a Wikipedia ar- ticle is not always tailored towards e-commerce. For instance, as of May 2013, the infobox at- tributes of wine in English Wikipedia included energy, carbohydrates, fat, protein and alcohol1. These are not particularly useful for users seek- ing their favorite wines through online shopping. Instead, the grape variety, production area, and vintage of the wine would be of greater interest. (2) On the other hand, Freebase contains PAVs for limited types of products such as digital cam- eras2. However, since Freebase is currently only available in English, we cannot use Freebase in a distant supervision method for other …

Content truncated to 500 words