[Project in collaboration with Telenor-NTNU AILab]
Deep Learning for NLP and Robo-Journalism
1 - Deep Generative models for title generation
Having a truthful title that reflects the content of an article and at the same time drives traffic is a challenging problem. For a given article, many possible titles exist. This project revolves around generating good suggestions for titles.
There are two major directions this project can take: One is generating qualitatively good titles based on text only. The other revolves around using feedback directly from the system in the following way: titles that drive engagement are considered good, titles that relatively speaking are generating little engagement are bad.
2 - Qualitative features for good titling
Our journalists produce approximately 1500 articles per day, which are then exposed on the newspapers’ web pages. Having a good title for each article is important for 1) getting people to click on the article and 2) getting people to read the article. Too much emphasis on 1. results in clickbait, while too much focus on 2. may result in suboptimal unique users. The question is: what is a good balance, and how do journalists create good titles to balance criterion 1. and 2. This project’s goal is how to use traffic data to analyze what are qualitative features that journalists should pay attention to when creating a title for an article.
This project is in collaboration with Amedia (http://www.amedia.no/).
Tags: Deep Learning, Generative Models, NLP
- Abstractive text summarization using sequence-to-sequence rnns and beyond 
- Sequence to Sequence Learning with Neural Networks, Ilya Sutskever Oriol Vinyals Quoc V. Le
- Controllable Text Generation - Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, Eric P. Xing
- Learning to Generate Reviews and Discovering Sentiment, Alec Radford, Rafal Jozefowicz, Ilya Sutskever
- Generating Sentences from a Continuous Space, Samuel R. Bowman, Luke Vilnis, Oriol Vinyals, Andrew M. Dai, Rafal Jozefowicz, Samy Bengio
A minimal background in Machine Learning is requested (at least one of two courses)