| Sociality has become one of the core features of the mobile Internet,which occupies 99.6%of the Internet.Therefore,social networks have great political,commercial,and scientific value.Question generation is an important task in the field of natural language processing,which aims to enable computers to automatically generate questions described in natural language.Answer-agnostic question generation is a subfield of question generation,which aims to generate meaningful and fluent questions only through contextual input,so as to help humans better understand and analyze relevant information.In view of the characteristics of massive unlabeled texts in social networks,answer-agnostic question generation for social networks helps to achieve cyberspace security from the perspective of application content security.This approach can transform text into interactive and disseminated questions,encouraging user engagement and knowledge sharing,and promoting effective management of cyberspace.However,answer-agnostic question generation for social networks currently faces three main challenges:(1)The inherent difficulty in answer-agnostic question generation,as generating appropriate questions with less input information than answer-aware question generation is more challenging.In social networks,there are more colloquial and fragmented texts,which also increase the difficulty of text processing;(2)The difficulty in selecting candidate answers,as the task is decomposed into two subtasks:candidate answer extraction and answer-aware question generation,which is effective but brings challenges in candidate answer selection.In social networks,there may be some internet jargons that are rare in daily life,which could increase the difficulty of selecting candidate answers;(3)The difficulty in evaluating the quality of generated questions,as this task lacks an answer to guide the generation,makes it difficult to determine if the generated question is answerable or valuable,especially when compared to answer-aware question generation.In social networks,the lack of manually annotated information may render traditional metrics inadequate for comprehensive and accurate evaluation.This dissertation proposes an exploration for addressing these issues and challenges,which includes the following specific tasks:(1)To address the inherent difficulty of generating answer-agnostic questions,a two-stage question generation model based on a statistic-based candidate answer extraction and a pre-trained language model is proposed,named RAKE+BART.The model decomposes the answer-agnostic question generation task into two stages:candidate answer extraction and question generation.In the first stage,the model utilizes the RAKE algorithm in statistical learning to calculate the relative weight of keywords in the text,thereby extracting the key information of the text.In the second stage,the keywords extracted by the RAKE algorithm are used as part of the input of the pre-trained language model BART to guide the generation of questions.The experimental results show that the candidate answers extracted by the RAKE algorithm help the model to better understand the text and improve the quality of the generation of unanswered questions,which is manifested in the improvement of BLEU,ROUGE,and METEOR indicators.This method can handle the problem of fragmented text in social networks because even if the text is segmented,it usually does not omit important vocabulary.Therefore,statistical methods can be used to capture these underlying key information.(2)To address the challenge of selecting candidate answers,an explicit candidate answer extraction and question generation model based on a pre-trained language model,named QGAE,is proposed to optimize the two-stage process of answer-agnostic question generation tasks.QGAE is a model based on the BART pre-trained language model designed for the field of answer-agnostic question generation.It realizes multi-task endto-end question-answer pair generation by combining answer extraction and question generation modules in parallel.The experimental results show that compared with the two-stage model RAKE+BART,the end-to-end model QGAE significantly improves multiple indicators of candidate answer extraction and question generation,and achieves the SOTA in the field of current answer-agnostic question generation.This method can effectively solve the problem of rare slang being ignored in social networks,because it can understand the text and identify the potential key information in it.(3)To address the difficulty in evaluating the quality of generated questions,a question generation evaluation metric based on question-answer consistency,named QAC,is proposed to evaluate the answerability and quality of questions generated by the answer-agnostic question generation model.This metric evaluates the answerability of the question by adding a question-answering module after the answer-agnostic question generation model,and using the quality of the answers obtained by question answering to indirectly reflect the quality of the question.The experimental results show that the QAC metric has a high consistency with the human evaluation results on real social network data.Compared with the RAKE+BART model,the QGAE model has better answerability and higher quality on the QAC metric.In practical social network scenarios,this metric can be used for self-evaluation of the model. |
