Success Story - en

Extraction of structured product descriptions from customer reviews

Motivation for launching the project by the client: usually, customer reviews reflect certain characteristics of the goods. It was necessary to automate the process of highlighting keywords that are responsible for a specific characteristic. In addition, the task was to predict the product at the request of customers based on previously left reviews.


What we had initially: MVideo did not have such functionality for processing reviews, there was a need to add new functionality.


Project goals: select a "descriptive" characteristic for each review and build a graph based on this data. 


MIL Team's solution: application of a semi-automatic method for extracting terms from the text of a product review. Building a knowledge graph, including matching terms with given technical characteristics, and training vector representations of graph elements to predict a product by the recall. 


Tools for building the model:
  • Adaptive Text Rank based on technical characteristics and a set of sentiment words for highlighting terms;
  • SOTA BERT model for matching terms and specifications;
  • TransE method for training vector representations of graph elements;
  • ABAE method for highlighting "important" characteristics for products based on a set of reviews.

The model results: Sets of terms for various categories of goods have been obtained, graphs have been built and a model for highlighting "important" characteristics has been pretrained.


Client: Mvideo
Technological stack: Python, Tensorflow
Research Group NLP Engineering Division
Made on
Tilda