Sentiment and emotion classification over noisy labels

摘要

With the rapid development of social media, online users are allowed to share their opinions conveniently. However, the ground truth for sentiments and emotions in social media is often constructed through surveys, hashtags or emoticons, where the labels may contain many errors. There are also amateurs and malicious users expressing offensive opinions or spreading fraudulent reviews, which has been identified as a growing threat to the trustworthiness of online comments. Thus, it is valuable for us to reconcile this noise in the ground truth when training sentiment and emotion classifiers. In this paper, we propose a hidden de-noising classification model (HDCM) that does not need any outsourcing systems or lexicons to estimate the actual sentimental or emotional category of each instance from corpora with noisy labels. The simplicity of assigning the category to a document by users under any contexts, and the authority of a user in assigning categories to documents with various domains are modeled as the unobserved hidden constraints in HDCM. Extensive evaluations using datasets with different scales of noisy labels validate the effectiveness of the proposed model for both sentiment and emotion classification tasks.