Currently,the amount of sports news is increasing,given the number of sports available.As a result,manually writing sports news requires high labor costs to achieve the intended efficiency.Therefore,it is necessary to...Currently,the amount of sports news is increasing,given the number of sports available.As a result,manually writing sports news requires high labor costs to achieve the intended efficiency.Therefore,it is necessary to develop the automatic generation of sports news.Most available news gen-eration methods mainly rely on real-time commentary sentences,which have the following limitations:(1)unable to select suitable commentary sentences for news generation,and(2)the generated sports news could not accurately describe game events.Therefore,this study proposes a sports news generation with text-editing model(SNG-TE)is proposed to generate sports news,which includes selector and rewriter modules.Within the study context,a weight adjustment mechanism in the selector module is designed to improve the hit rate of important sentences.Furthermore,the text-editing model is introduced in the rewriter module to ensure that the generated news sentences can cor-rectly describe the game events.The annotation and generation experiments are designed to evaluate the developed model.The study results have shown that in the annotation experiment,the accuracy of the sentence annotated by the selector increased by about 8%compared with other methods.Moreover,in the generation experiment,the sports news generated by the rewriter achieved a 49.66 ROUGE-1 score and 21.47 ROUGE-2,both of which are better than the available models.Additionally,the proposed model saved about 15 times the consumption of time.Hence,the proposed model provides better performance in both accuracy and efficiency,which is very suitable for the automatic generation of sports news.展开更多
为解决现有图像仿真中动漫风格迁移网络存在图像失真和风格单一等问题,提出了适用于动漫人脸风格迁移和编辑的TGFE-TrebleStyleGAN(text-guided facial editing with TrebleStyleGAN)网络框架。利用潜在空间的向量引导生成人脸图像,并在...为解决现有图像仿真中动漫风格迁移网络存在图像失真和风格单一等问题,提出了适用于动漫人脸风格迁移和编辑的TGFE-TrebleStyleGAN(text-guided facial editing with TrebleStyleGAN)网络框架。利用潜在空间的向量引导生成人脸图像,并在TrebleStyleGAN中设计了细节控制模块和特征控制模块来约束生成图像的外观。迁移网络生成的图像不仅用作风格控制信号,还用作约束细粒度分割后的编辑区域。引入文本生成图像技术,捕捉风格迁移图像和语义信息的关联性。通过在开源数据集和自建配对标签的动漫人脸数据集上的实验表明:相较于基线模型DualStyleGAN,该模型的FID降低了2.819,SSIM与NIMA分别提升了0.028和0.074。集成风格迁移与编辑的方法能够确保在生成过程中既保留原有动漫人脸细节风格,又具备灵活的编辑能力,减少了图像的失真问题,在生成图像特征的一致性和动漫人脸图像风格相似性中表现更优。展开更多
基金funded by the Research Project of Natural Science at Anhui Universities in 2021,Research on relation extraction of emergency plan knowledge graph based on deep embedding clustering(No.KJ2021A0994).
文摘Currently,the amount of sports news is increasing,given the number of sports available.As a result,manually writing sports news requires high labor costs to achieve the intended efficiency.Therefore,it is necessary to develop the automatic generation of sports news.Most available news gen-eration methods mainly rely on real-time commentary sentences,which have the following limitations:(1)unable to select suitable commentary sentences for news generation,and(2)the generated sports news could not accurately describe game events.Therefore,this study proposes a sports news generation with text-editing model(SNG-TE)is proposed to generate sports news,which includes selector and rewriter modules.Within the study context,a weight adjustment mechanism in the selector module is designed to improve the hit rate of important sentences.Furthermore,the text-editing model is introduced in the rewriter module to ensure that the generated news sentences can cor-rectly describe the game events.The annotation and generation experiments are designed to evaluate the developed model.The study results have shown that in the annotation experiment,the accuracy of the sentence annotated by the selector increased by about 8%compared with other methods.Moreover,in the generation experiment,the sports news generated by the rewriter achieved a 49.66 ROUGE-1 score and 21.47 ROUGE-2,both of which are better than the available models.Additionally,the proposed model saved about 15 times the consumption of time.Hence,the proposed model provides better performance in both accuracy and efficiency,which is very suitable for the automatic generation of sports news.
文摘为解决现有图像仿真中动漫风格迁移网络存在图像失真和风格单一等问题,提出了适用于动漫人脸风格迁移和编辑的TGFE-TrebleStyleGAN(text-guided facial editing with TrebleStyleGAN)网络框架。利用潜在空间的向量引导生成人脸图像,并在TrebleStyleGAN中设计了细节控制模块和特征控制模块来约束生成图像的外观。迁移网络生成的图像不仅用作风格控制信号,还用作约束细粒度分割后的编辑区域。引入文本生成图像技术,捕捉风格迁移图像和语义信息的关联性。通过在开源数据集和自建配对标签的动漫人脸数据集上的实验表明:相较于基线模型DualStyleGAN,该模型的FID降低了2.819,SSIM与NIMA分别提升了0.028和0.074。集成风格迁移与编辑的方法能够确保在生成过程中既保留原有动漫人脸细节风格,又具备灵活的编辑能力,减少了图像的失真问题,在生成图像特征的一致性和动漫人脸图像风格相似性中表现更优。
