UCLA

Companies that can leverage their product descriptions to find meaningful insights have a competitive advantage in reaching and understanding their consumer. In practice, however, this is a challenging task. Product descriptions may be written by a number of different individuals with inconsistent formatting. They may contain typos, inaccurate representations and complex terminology so unique to a single product as to be meaningless to analyze alongside other products.

For this reason, the study of Product Attribute Extraction (PAE) aims to extract meaningful at- tributes from product descriptions with the goal of producing a structured dataset of attribute-value pairs. Prior work has focused largely on conventional natural language processing (NLP) techniques to identify such relationships. In this paper, we combine the architecture of transformers and variational autoencoders (VAE) to merge and format semantically similar product descriptions with the goal of decreasing the complexity of PAE. We demonstrate examples of the feasibility of this approach and assess its potential, shortcomings and application to the field.