En | Ru

News headline generator for Match TV

Neural Networks, AI

Main information

01/

Task

To launch sports channels in Telegram and Twitter, which will be led by artificial intelligence. The project was timed to coincide with the 2020 European Football Championship

Match TV is a federal sports TV channel, as well as a website with news, interviews, sports statistics and match broadcasts

Client

Neural network service for real-time news headline generation

Solution

How it works

02/

Test the service

Headlines with summaries of the news are published on Match TV channels

The editor looks through the options and selects the most interesting ones

A neural network analyzes news headlines and creates its own variations

Quote

«The neural network was trained on Match TV news headlines and the yellow press. That's why the choices it offered were quite provocative. The best headlines were selected and published manually»

Project Developer

Sergey Verentsov

Service features

03/

The neural network answered questions. The editor could ask them through the bot

The service could finish the headline - the editor just needed to make up the beginning

The neural network was packed into a Telegram chat bot for convenience

Quote

«To train the neural network, we collected all of Match TV's headlines since the site's launch in 2003. As you know, the more data that is used to train a service, the better the results»

Project Developer

Sergey Verentsov

The team and timeline

04/

Development time

2 weeks

May 2020

Sergey Verentsov

Project Developer

Stages of development

05/

preparation of the dataset for training the neural network

training on Match TV and other sports media headlines

testing the service together with the client

integration of a neural network into a Telegram chatbot

complete service launch

Technologies

06/

EORA's own chatbot builder

zDialog

(Generative Pre-trained Transformer 2) is an open-source AI developed by the American company OpenAI

GPT-2

Go to zDialog

Project results

07/

The neural network successfully worked for 1.5 months from June to July 2021

Project Developer

Sergey Verentsov

Now we have created a pipeline for generating any text - not only news titles, but also, for example, product names, product descriptions. The only thing we will need is data on which the neural network can learn