Motivation for the customer to launch a project:
customer reviews often reflect certain characteristics of products. It was necessary to automate the process of identifying keywords responsible for specific characteristics. In addition, the task was set to predict products based on customer requests based on previously left reviews.
Description of the initial situation:
MVideo did not have such functionality for processing reviews; there was a need to add new functionality.
Project goals:
MIL Team solution:
application of a semi-automatic method for extracting terms from the text of a product review. Construction of a knowledge graph, including the comparison of terms with given technical characteristics, and training of vector representations of graph elements to predict a product based on a review.
To build the model we used:
Simulation results:
Sets of terms for various product categories were obtained, graphs were constructed, and a model for identifying “important” characteristics was pretrained.
Customer: Mvideo
Technology stack: Python, Tensorflow
customer reviews often reflect certain characteristics of products. It was necessary to automate the process of identifying keywords responsible for specific characteristics. In addition, the task was set to predict products based on customer requests based on previously left reviews.
Description of the initial situation:
MVideo did not have such functionality for processing reviews; there was a need to add new functionality.
Project goals:
- highlight a “descriptive characteristic” for each review and build a graph based on this data.
MIL Team solution:
application of a semi-automatic method for extracting terms from the text of a product review. Construction of a knowledge graph, including the comparison of terms with given technical characteristics, and training of vector representations of graph elements to predict a product based on a review.
To build the model we used:
- Adaptive Text Rank based on technical characteristics and a set of sentiment words to highlight terms;
- SOTA BERT model for matching terms and technical characteristics;
- TransE method for training vector representations of graph elements;
- ABAE method for identifying “important” characteristics for products from a set of reviews.
Simulation results:
Sets of terms for various product categories were obtained, graphs were constructed, and a model for identifying “important” characteristics was pretrained.
Customer: Mvideo
Technology stack: Python, Tensorflow
