Time-sync comments denoising via graph convolutional and contextual encoding

Abstract

Time-Sync Comments (TSC), which is a new kind of textual comments on online video websites, has showed its great potential in fine-grain video analysis. However, as a crowd-sourced resource, there are many low quality comments in TSC data and this is an impediment to make full use of TSC. Thus a denoising method is necessary when we are dealing with these comments. In this study, we propose GCCED, a graph convolutional and contextual encoding denoising model for TSC semantic denoising problem. A TSC graph is built on the whole corpus and semantic embedding of words are learned through graph convolution. Moreover, we exploit the relations between TSC and its context and design an embedding method based on the word graph. Experiments on real world dataset are conducted and the result demonstrate the proposed model outperforming other baselines in almost all classification metrics.