HOW to construct a news headline that captures reader curiosity without entering the domain of click-bait?
Personalized News Headline
A generic framework for personalized news headline generation based on PENS
Background
HOW to construct a news headline that captures reader curiosity without entering the domain of click-bait?
Personalized News Headline
Model Framework
How to learn personalization?
Personalized news recommendation is the matching between candidate news and user interests, where accurate news representations and user preferences are learned.
How to inject the user preference into a news headline generator？
Our proposed headline generator is a variant of transformer encoder and pointer network decoder, with three different manners to inject user preferences learned to the generator.
It can be considered as a sequential decision-making process. The reward is estimated by personalization, fluency and factualness.
Experiments
Firstly, we can see that every personalized news headline generation method can outperform non-personalized methods like PG. It might be that our proposed framework can generate personalized news headlines by incorporating user interests.
Secondly, we can find that user modeling makes a difference in generating personalized headlines.
Thirdly, it is reported that the second way of injecting user interests gets the best performance on most of the user modeling methods.
However, it still remains a large room for explorations on better injecting user representations into the generation process since the second way seems to be defective at some time.
Case Study
We believe the reason is that PG is trained via supervised learning to maximize the log-likelihood of ground-truth news headlines. While our framework is trained via RL technique where coverage score is considered as an indicator to encourage the generation to be more complete. In addition, the exhibited cases show that our framework can produce user-specific news headlines in accordance with their individual interests reflected by historical click behaviors.
How to make your own website - Find more