Using Text Reviews for Product Entity Completion

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 983–991, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Using Text Reviews for Product Entity Completion Mrinmaya Sachan Tanveer Faruquie L. V. Subramaniam Mukesh K. Mohania IBM Research India New Delhi, India {mrsachan, ftanveer, lvsubram, mkmukesh}@in.ibm.com Abstract In this paper we address the problem of obtaining structured information about products in the form of attribute-value pairs by leveraging a combination of enter- prise internal product descriptions and ex- ternal data. Product descriptions are short text strings used internally within enter- prises to describe a product. These strings usually comprise of the Brand name, name of the product, and its attributes like size, color, etc. Existing product data quality solutions provide us the capability to stan- dardize and segment these descriptions into their composing attributes using do- main specific rulesets. We provide tech- niques that can leverage the supervision provided by these existing rulesets for ex- tracting missing values from other exter- nal text data sources accurately. We use a large real life data collection to demon- strate the effectiveness of our approach. 1 Introduction Enterprises usually store information of its prod- ucts in the form of unstructured text strings. Such product descriptions contain the name of the prod- uct and its specific attributes. These product de- scriptions are usually written by multiple people and could contain overlapping information or even the same information written differently. For ex- ample, a superstore may source the same product from many different vendors and each vendor may give varying descriptions of the same product. In- formation could be scattered through various de- partments and held by certain employees or sys- tems instead of being available centrally. This re- sults into varying standards and vocabulary. Consider the following product descriptions ob- tained from an enterprise selling cameras. They are provided by different vendors supplying the cameras to the enterprise: Nikon D90 4288×2848 703 g Digicam F/1.8 Nikon Digital 90 Cam 12.3MP 1.8 F-Len(1.55lb) Nikon D-90 Camera with Nikkor 50mm 1.8D Expert knowledge specific to the domain (that D90, D-90 and Digital 90, 4288 × 2848 and 12.3 MP, F-number 1.8 and 1.8 D focal length, and 703 g and 1.55lbs are same) is required to conclude that the entries above are the same product. Cou- pled with data entry errors, the problem of identi- fying a standard representation of the product be- comes even harder. The problem of obtaining a structured represen- tation of such product descriptions is similar to the ‘Attribute-Value pair’ mining problem. Attribute represents an aspect of the product. It could be anything from a manufacturing detail like model number to information like color, size and weight. Due to its practical applications, the problem has drawn interest from the research community as well as the industry. There are many products which provide solutions for standardizing, match- ing, merging and validating such descriptions. Popular ones include Oracle middleware, Silver- creek, Ethoscontent product-copy, Trilliumsoft- ware and IBM Data Stage-Quality Stage. …

Content truncated to 500 words