News Generation

Introduction

This project is a part of the Teknofest 2024 Türkçe Doğal Dil İşleme competition. The aim of the project is to generate news title and content from a given image.

Dataset

The dataset is collected from the news website. The dataset consist of news titles, news content and images. The dataset is in Turkish Language.

Data-Preprocessing

• Sample Data:

  ---

    title = "Balıkesir’de tarihi bina yangında küle döndü"
    word_index = {'Balıkesir’de': 9, 'tarihi': 5, 'bina': 3, 'yangında': 7, 'küle': 5, 'döndü': 6 }
    tokens: [start_token, 9, 5, 3, 7, 5, 6, end_token]

  Input	Output
  Image + start_token	9
  Image + start_token + 9	5
  Image + start_token + 9 + 5	3
  Image + start_token + 9 + 5 + 3	7
  Image + start_token + 9 + 5 + 3 + 7	5
  Image + start_token + 9 + 5 + 3 + 7 + 5	6
  Image + start_token + 9 + 5 + 3 + 7 + 5 + 6	end_token

Model

The model is a combination of CNN and LSTM, where the image is fed to the Encoder(CNN) and the output of the CNN is fed to the Decoder(LSTM) along with the input text.