| News is an important form for people to obtain real-time information.The rapid development of the Internet has enabled people to read news conveniently through channels such as portal websites and news apps.However,in the era of information explosion,the quality of news has not been improved simultaneously with the quantity,which undoubtedly greatly increases the cost of filtering invalid information.In the financial field,in order to effectively track and evaluate the development status of enterprises or industries,data analysts need to quickly analyze the daily news text data and extract major events in the news.The entire process includes text filtering,text deduplication,and Multiple steps,such as information condensing,consume a lot of labor costs;at the same time,in order to accurately obtain event information from long-length financial news,data analysts often need to possess strong financial background knowledge and information positioning capabilities.The difference in abilities of analysts has caused the instability of the quality of the acquisition event,and the manual processing method is becoming more and more difficult to deal with the contradiction between the increasing number of texts and the increasing efficiency requirements.Therefore,how to use the model to automatically filter texts and extract event sentences from massive financial news,and realize the process of refining long texts to short texts,has important research value and application prospects.This paper first constructed an event sentence extraction dataset Fin News composed of financial news for model training and performance verification.This paper defines the event sentence extraction task as a sentence-level sequence labeling problem,and uses the framework of hierarchical encoders and decoders to construct the event sentence extraction model Fin Extracor.Through comparative experiments,the effectiveness of the model on the Fin News dataset is verified.Aiming at the problem of the inconsistency between the training objective of the event sentence extraction model and the evaluation index,this paper proposes a strategy reinforcement learning training method based on the Rouge score,and adopts a two-stage training method to improve the model convergence speed.The experimental results show that the reinforcement learning method significantly improves the Rouge index of abstract extraction.When the model encodes news chapters,it is common to face a more serious long-dependency problem due to the long text.In response to this problem,this paper designs an attention mechanism based on news title information.Through this attention mechanism,the topic information in the title is aligned with each sentence to determine the relevance of each sentence to the full text topic.The experimental results show that the attention mechanism based on news titles alleviates the problem of information loss in chapter topics in the decoding process,and has a significant improvement in multiple indicators.Experiments have proved that compared with other extraction models,the model designed in this paper has achieved 73.62% F1 on the Fin News data set,and has also achieved the current best results on the Rouge indicator. |
PDF Full Text Request