| Personalised news headline generation technology is used to generate news headlines that match users’ interests and reading habits,given their historical reading behaviour data and candidate news content.For users,this technology can effectively reduce information overload and meet their personalised needs.For news platforms,the technology can accelerate the speed of news going live and increase user click-through rates.Therefore,personalised news headline generation technology is of great significance and is becoming a new research hotspot.In the past few years,personalised news headline generation techniques have made great progress.Inspired by the field of news recommendation,there is a new research trend to personalise news headlines by injecting users’ interests into the headline generator.Models based on this technology have been very successful.However,there are still problems to be solved in the two key tasks of user interest representation and user interest injection.Firstly,the interest of users in news headlines is first and foremost in the physical words in the headlines,and existing user interest representations lack modelling of the physical words in the news being viewed.Furthermore,only a small proportion of user interests are relevant to the candidate news,and embedding user interests directly into the headline generator may lead to a large amount of information redundancy,which may adversely affect the text summary.In view of the above issues,the following research work was conducted in this paper:(1)Entity-enhanced personalised news headline generation(NREE): this paper proposes a new approach to modelling user interests,using news subjects and entities to model news content,ensuring that news modelling both highlights important entities and contains rich context to understand user preferences.An attention mechanism is used to aggregate news vectors into interests,highlighting the extent to which different news items contribute to interests.Experiments on the MIND dataset and the PENS recommendation dataset demonstrate the effectiveness of NREE’s use of entities in the user interest modelling process.Experiments on the PENS dataset demonstrate the effectiveness of this paper’s interest modelling in enhancing the quality and level of personalisation of headline generation in a personalised news headline generation task.(2)Enhanced user interest-aware personalised headline generation(EUI-PENS): This paper further proposes a user interest filtering module.The user interest filtering module of EUI-PENS is essentially an interest-aware attention mechanism that provides information related to user interests and candidate news contexts to the embedding module.The validity of the model was verified on the PENS dataset and ablation experiments were conducted for each module in the model.The experiments demonstrated that the user interest filtering module is a good enhancement for different user interest representations. |
PDF Full Text Request