| As the application of deep learning techniques in the field of artificial intelligence is becoming widespread,research in the field of deep learning-based text generation is also becoming popular.In particular,the story ending generation field,as an interesting but challenging task,aims at completing a story ending sentence for an incomplete story.Although some research scholars have made a series of explorations and improvements on this task,and some progress has been made.However,due to the lack of global guidance of content and sentimental information,this often leads to the model unable to control the generated endings.This will result in irrelevant or sentimentally inconsistent endings,or generic endings.These are the two main challenges that remain for this task: the first is how to understand story trends more deeply,especially the key information of the stories.The second is how to generate sentimentally endings that avoid conflicting with the sentimental trends of the stories.To address these issues,this thesis proposes a story ending generation model based on story line and sentiment-aware pre-training.To address the problem of difficulty in understanding story trends,the key information in the story is obtained with the story line extraction module,followed by encoding the story content and story line information using the pre-trained GPT-2 skeleton.To address the difficulty of generating emotionally reasonable endings,the sentiment prediction module is used to perceptually identify the sentiment of the story and predict the sentiment trend of the story.Finally,the sentiment condition generation module is combined with the expected sentiment to generate the endings of the story which are consistent with the context and sentiment of the story.This thesis adopts the combination of automatic and manual evaluation indexes,a large number of comparative experiments,ablation experiments,and case analysis on the ROC-Stories dataset.The experimental results show that the model proposed in this thesis outperforms the previous benchmark model in the vast majority of metrics compared to the benchmark model.This further demonstrates that the introduction of storyline information can effectively enhance the model’s ability to understand the content of the main story line when content consistency is considered.Moreover,the perceptive power of the sentiment prediction module can facilitate the model to better understand the sentiment trends embedded in the story and ultimately generate more logical and consistent story endings.These findings will provide ideas for solving the story ending generation task in the area of sentimental beyond contentuality,which will be conducive to the continued development of the story ending generation task. |
