1. Predictive Publication Characteristics System (Views, Likes, Clicks, Comments, etc.)

  • Category: NLP
  • Client: Commercial Client
  • Project date: 2021

Utilizing historical data from diverse sources and the interactions associated with publications, we have developed a predictive system that analyzes article content to forecast various characteristics of publications. These characteristics include views, link clicks, likes, etc. By leveraging this system, organizations can optimize their content strategies and make data-driven decisions to maximize engagement and interaction with their publications.

Application of predictive models

Used Approaches

Header vectorization*

  • TF-IDF
  • BERT

Regression model**

  • FFN
  • Catboost

*Vectorization is the process of extracting features from data (text) that are used to build predictive models

** Regression models - predictive function of the independent variable and parameters with the addition of a random variable

Metrics for assessing the quality of predictive models

  • Mean Squared Error (MSE)
  • RMSE
  • Mean Absolute Error (MAE) ~ 2500 views

2. Domain-Specific Search and Q&A Systems

  • Category: NLP
  • Client: Commercial Clients
  • Project date: 2019-2023

Before the emergence of ChatGPT, there was a need for systems that could swiftly find answers in specific, narrow areas like technical documentation, instructions, and user guides. In such instances, ChatGPT may not provide the correct solution. Our NLP team has been instrumental in creating such systems, specifically:

Our NLP team has been instrumental in creating such systems, including:

  • Information search systems able to locate specific articles within a vast knowledge base by using text descriptions
  • Voice assistants that find required answers in technical documentation based on ambiguously formed questions (as users often lack technical terminology)
  • Automated systems for recognizing and classifying breakdowns and malfunctions of industrial equipment. In the event of production equipment failure, these systems recognize specific malfunction terminology based on voice calls from production machine operators and promptly generate equipment repair requests.