Aiming at the problem that blog texts are the streaming data captured by different acquisition modality, each kind of which has its particular quality evaluation mode, this paper proposes a text quality evaluation (TQA) model based on 3D CNN correlated with blog text data. In order to achieve accurate TQA value, the model adopted a Bi-LSTM-based architecture to process video-related blog text as auxiliary part to provide additional information for our TQA architecture. First, the auxiliary part constructs feature vector for each video-related text by the model originating from Bi-LSTM and Seq2Seq. Then, the feature vector was feed to a well-trained decoder to reconstruct the original input data. Then, the feature vector complied with the blog textual data are inputted into end-to-end TQA modal based on the 3D CNN straightly. Comprehensive experimental results on the blog text/video dataset from the well-known truism website “http://www.mafengwo.cn/” have shown that the proposed model reflects the subjective quality of online texts more accurately, and has better overall blog TQA assessment performance than the other state-of-the-art non-reference methods.

针对博客文本是通过不同的获取方式捕获的流数据的问题，每种获取模式都有其特定的质量评估模式，本文提出了一种基于3D CNN的文本质量评估（TQA）模型，该模型与博客文本数据相关联。为了获得准确的TQA值，该模型采用基于Bi-LSTM的体系结构来处理与视频相关的博客文本，作为辅助部分，以为我们的TQA体系结构提供其他信息。首先，辅助部分通过源自Bi-LSTM和Seq2Seq的模型为每个视频相关文本构建特征向量。然后，将特征向量输入到训练有素的解码器中，以重建原始输入数据。然后，将符合博客文本数据的特征向量直接输入到基于3D CNN的端到端TQA模态中。