The goal of price comparison platform is to offer the best product price for its customers. Product descriptions are obtained in two ways: directly from the partner online shops or by scraping online shop websites. In most cases product descriptions are not well structured nor clean, thus it becomes difficult to match them and compare their prices. To overcome this problem companies often employ whole teams to compare product descriptions manually. This use case describes implementation of automatic online product descriptions matching system for undisclosed client.

Different descriptions of the same product occur often. It is clear that is not trivial to match products with such unstructured descriptions

Our matching system first transforms unstructured descriptions into structured product descriptions using semantic parsing. The next step is extracting relevant attributes and forming feature vectors based on similarity measures. These feature vectors are then used to train matching function. The goal of matching function is to determine a probabilistic score of match between two products based on constructed feature vectors.