Research And Application Of Content-controlled Text Generation

Controlled text generation aims to generate text that meets the user’s expectations to achieve specific goals and tasks.In literary composition and application writing,controlling the generated text content is crucial to the usefulness of text generation systems.Traditional methods control text content by introducing a series of instructions or keywords,but due to the random nature of language models,the generated text is usually still not controllable enough to meet the needs of users in downstream application scenarios.In particular,in narrative text generation,simple keywords cannot control the development of the storyline.To solve this problem,this thesis proposes two improvements to traditional controllable text generation methods.One is the text generation algorithm based on the core event framework,which improves the controllability of text by introducing core events in the text generation control framework.The other is a controlled text generation algorithm based on the "continue,review,rewrite" multi-task fusion model,which integrates the three tasks of continue,review and rewrite to improve the consistency and fluency of the generated text.Both approaches bring significant improvements to controlled text generation to better meet user needs.The main work of this thesis includes the following three aspects:1.To address the controllability problem of text generation,this thesis designs a core event framework-based text generation algorithm.The algorithm process takes the extracted core event framework as the blueprint for text generation and guides the generated text to conform to the domain-specific style and semantics;the core event framework extraction is implemented by various techniques including semantic role annotation and statistical text features.The proposed Plan2Plot2 Write scheme is compared with existing practices for experiments and results are analyzed,and the improvement in ROUGE-L and BLEU metrics is 20.7% and 21.4%,respectively,compared with the baseline scheme.2.To address the consistency problem in controlled text generation,this thesis proposes a controlled text generation algorithm based on a multi-task fusion model of "continue,review,and rewrite",in which the models of multiple tasks are fused and trained jointly.In the process of model-guided text generation,the system will rewrite the generated results when the output of each step does not meet the system’s criteria.The experimental comparison shows that the WRRmt model proposed in this thesis improves47.5% in ROUGE-L,and the Self-BLEU metric decreases 23.7% relative to the preimprovement.3.A content-controlled text generation system is designed and implemented.The system combines the above two controllable text generation algorithms to provide users with a fully functional and process-controllable text generation service.Before text generation,users can use various hints to guide the text generation;during the process,the system will modify the generated results according to the feedback.The content-controlled text generation system in this thesis has good text generation capability.At the same time,the system is easy to deploy and low cost in real text generation scenarios,and the generated text content has high relevance.