诗歌生成中的韵律规范和主题一致性一直以来都是自然语言生成领域的研究热点。为提升诗歌生成中的韵律规范,提出了基于Transformer结合韵律特征的诗歌生成模型(Transformer and prosodic features poetry generation model,TPPG)。根据...诗歌生成中的韵律规范和主题一致性一直以来都是自然语言生成领域的研究热点。为提升诗歌生成中的韵律规范,提出了基于Transformer结合韵律特征的诗歌生成模型(Transformer and prosodic features poetry generation model,TPPG)。根据韵律特征建立平仄韵律词库和平声韵脚词库,在Transformer编码器中引入平仄韵律编码,模型训练过程中可以捕获更多平仄韵律特征的信息,学习到多种诗歌韵律;最终根据建立的平声韵脚词库规范诗歌生成韵脚,运用极大后验概率对于候选的诗歌选择当前赋有韵律特征规范的最优诗句,整体提升诗歌规范性和流畅性。实验结果表明TPPG模型生成的诗歌能够很好地符合韵律,在人工评价和机器评价中均有提高。展开更多
面向推荐系统的协同知识感知注意网络(Collaborative Knowledge-Aware Attentive Network for Recommender Systems,简称CKAN)没有优化与项目直接关联实体的嵌入表示,且没有考虑不同传播层次的重要性以及用户点击物品间的关联信息,使推...面向推荐系统的协同知识感知注意网络(Collaborative Knowledge-Aware Attentive Network for Recommender Systems,简称CKAN)没有优化与项目直接关联实体的嵌入表示,且没有考虑不同传播层次的重要性以及用户点击物品间的关联信息,使推荐实体不够准确。因此,提出一种融合项目直接关联实体传播的推荐算法(Items are Directly Associated with Entity Propagation - Collaborative Knowledge-aware Attentive Network for Recommender Systems,简称IEP-CKAN),在IEP-CKAN中将优化与项目直接关联实体的嵌入表示,并融入Ripple Net模型思想在利用波纹集信息时给予靠近中心节点的层次信息更大或更小的权重,再通过无序编码器挖掘用户点击物品间的关联信息,最终计算出推荐结果。实验结果表明,与CKAN相比该方法在Last. FM数据集上将AUC和F1的值提高了1.1%和1.3%,在Book-Crossing数据集上将AUC和F1的值提高了0.9%和1.0%,证明该方法能够进一步提高推荐模型的性能。The Collaborative Knowledge-aware Attentive Network for Recommender Systems (CKAN) does not optimize the embedded representation of entities directly associated with the project. Moreover, the importance of different communication levels and the correlation information between items clicked by users are not considered, so the recommended entity is not accurate enough. Therefore, This paper proposes a recommendation algorithm (Items are Directly Associated with Entity Propagation - Collaborative Knowledge-aware Attentive) Network for Recommender Systems (IEP-CKAN for short) puts the embedded representation of entities directly related to the optimization project in IEP-CKAN, and incorporates the idea of Ripple Net model to give a greater or smaller weight to the hierarchical information near the central node when making use of the information of the ripple set. Then through the disordered encoder mining the user clicks between the items of the association information, the final calculation of the recommendation result. The experimental results show that compared with CKAN, the proposed method increases the values of AUC and F1 by 1.1% and 1.3 % in Last. FM dataset and 0.9% and 1.0% in Book-Crossing dataset, which proves that the proposed method can further improve the performance of the recommended model.展开更多
文摘诗歌生成中的韵律规范和主题一致性一直以来都是自然语言生成领域的研究热点。为提升诗歌生成中的韵律规范,提出了基于Transformer结合韵律特征的诗歌生成模型(Transformer and prosodic features poetry generation model,TPPG)。根据韵律特征建立平仄韵律词库和平声韵脚词库,在Transformer编码器中引入平仄韵律编码,模型训练过程中可以捕获更多平仄韵律特征的信息,学习到多种诗歌韵律;最终根据建立的平声韵脚词库规范诗歌生成韵脚,运用极大后验概率对于候选的诗歌选择当前赋有韵律特征规范的最优诗句,整体提升诗歌规范性和流畅性。实验结果表明TPPG模型生成的诗歌能够很好地符合韵律,在人工评价和机器评价中均有提高。
文摘面向推荐系统的协同知识感知注意网络(Collaborative Knowledge-Aware Attentive Network for Recommender Systems,简称CKAN)没有优化与项目直接关联实体的嵌入表示,且没有考虑不同传播层次的重要性以及用户点击物品间的关联信息,使推荐实体不够准确。因此,提出一种融合项目直接关联实体传播的推荐算法(Items are Directly Associated with Entity Propagation - Collaborative Knowledge-aware Attentive Network for Recommender Systems,简称IEP-CKAN),在IEP-CKAN中将优化与项目直接关联实体的嵌入表示,并融入Ripple Net模型思想在利用波纹集信息时给予靠近中心节点的层次信息更大或更小的权重,再通过无序编码器挖掘用户点击物品间的关联信息,最终计算出推荐结果。实验结果表明,与CKAN相比该方法在Last. FM数据集上将AUC和F1的值提高了1.1%和1.3%,在Book-Crossing数据集上将AUC和F1的值提高了0.9%和1.0%,证明该方法能够进一步提高推荐模型的性能。The Collaborative Knowledge-aware Attentive Network for Recommender Systems (CKAN) does not optimize the embedded representation of entities directly associated with the project. Moreover, the importance of different communication levels and the correlation information between items clicked by users are not considered, so the recommended entity is not accurate enough. Therefore, This paper proposes a recommendation algorithm (Items are Directly Associated with Entity Propagation - Collaborative Knowledge-aware Attentive) Network for Recommender Systems (IEP-CKAN for short) puts the embedded representation of entities directly related to the optimization project in IEP-CKAN, and incorporates the idea of Ripple Net model to give a greater or smaller weight to the hierarchical information near the central node when making use of the information of the ripple set. Then through the disordered encoder mining the user clicks between the items of the association information, the final calculation of the recommendation result. The experimental results show that compared with CKAN, the proposed method increases the values of AUC and F1 by 1.1% and 1.3 % in Last. FM dataset and 0.9% and 1.0% in Book-Crossing dataset, which proves that the proposed method can further improve the performance of the recommended model.
