Hybrider Ansatz zur automatisierten Themen-Klassifizierung von Produktrezensionen

Abstract

In e‑commerce, enormous amounts of data are generated through the interaction of customers with Web platforms. Customer feedback in the form of product reviews, for instance, is an example for unstructured data, which processing requires approaches from the fields of computer linguistics and machine learning. As an alternative to the classical approaches of supervised and unsupervised learning, which often reach their limits in the business context and the application domain of product reviews, this article presents a hybrid approach for categorizing product reviews that combines the advantages of machine learning and human expertise. The aim of this paper is to present an approach that allows to automatically extract structured topics and related aspects from product reviews based on practical requirements. Word2Vec is used to train semantic relationships between words that occur in product reviews. In this way, individual words of each review can be compared with in advance defined topic words regarding their similarity and can then be extracted from the reviews. This approach is demonstrated using around five million product reviews of the Amazon online platform. The results are getting compared with those from a common topic modelling technique.

Mehr zum Titel