基于文本增强的共注意机制的多模态标签推荐

针对新型社交平台用户发布帖子时通常会使用标签来标记帖子的关键词或话题来提高自己在社交媒体中参与度的问题,使用了层级结构,从单词、短语和句子三个层级来提取文本特征。并且提出文本内容的汇总注意机制,将每个层级的语义内容总结为一个特征向量,然后提出一个文本增强的共注意模型,将每个层级的语义分别与图像模态进行语义融合。同时,考虑到不同用户使用标签的偏好习惯等各不相同,引入一个外部存储单元来记录每个用户的历史标签习惯,计算当前待推荐帖子与历史帖子之间的相似度影响向量,建立用户的个性化模块。在真实数据集上的实验结果表明,文中基于多模态帖子内容理解和个性化模块分析模型相比与其他模型,在精确率、召回率和F₁分数上都有很大提升;提出的两个关于多模态内容理解的注意力机制和用户的个性化建模都对整体推荐效果有显著的贡献。     

A Metadata Model Based on Coupling Testing Information to Increase Testability of Component

A software component must be tested every time it is reused in order to assure quality of component itself and system in which it is to be integrated. So how to increase testability of component has become a key technology in the software engineering community. Here a method is introduced to increase component testability. And meanings of component testability and relative effective ways to increase testability are summarized. Then definitions of component coupling testing criterion, DU-I （Definition-Use Information） and OP-Vs （ Observation-Point Values ） are given. Base on these, a definition-use table is introduced, which includes DU-A and OP-Vs item, to help component testers to understand and observe interior details about component under test better. Then a framework of testable component based on above DU-table is given. These facilities provide ways to detect errors, observe state variables by observation-points based monitor mechanism. Moreover, above methods are applied to our application developed by ourselves before, and some test cases are generated. Then our method is compared with Orso method and Kan method using the same example, presenting the comparison results. The results illustrate the validity of our method, effectively generating test cases and killing more mutants.     

基于迁移学习的小数据集命名实体识别研究

针对军事重要目标实体自动获取的问题,提出一种将基于转换器的轻量级双向编码表征(a lite BERT,ALBERT)、双向门控循环单元(Bi-Gated recurrent unit,BiGRU)、条件随机场(conditional random field,CRF)相结合的小样本数据集命名实体识别方法.考虑到军事重要...