Weakly Supervised Short Text Categorization Using World Knowledge

摘要

Short text categorization is an important task in many NLP applications, such as sentiment analysis, news feed categorization, etc. Due to the sparsity and shortness of the text, many traditional classification models perform poorly if they are directly applied to short text. Moreover, supervised approaches require large amounts of manually labeled data, which is a costly, labor intensive, and time-consuming task. This paper proposes a weakly supervised short text categorization approach, which does not require any manually labeled data. The proposed model consists of two main modules: (1) a data labeling module, which leverages an external Knowledge Base (KB) to compute probabilistic labels for a given unlabeled training data set, and (2) a classification model based on a Wide & Deep learning approach. The effectiveness of the proposed method is validated via evaluation on multiple datasets. The experimental results show that the proposed approach outperforms unsupervised state-of-the-art classification approaches and achieves comparable performance to supervised approaches.