Currently, supervised learning based methods and techniques such as deep learning methods, have achieved great success in the text mining area. When researchers develop these giant models, they usually assume the availability of massive annotated training data. However, the real-world usefulness of these models will be impaired because in the real world, readily available annotated data are scarce. Unlabeled training data can also be rare in some circumstances such as personalization modeling. Fortunately, there is useful information existing in many inexpensive and readily available resources. In this thesis, we show how to utilize auxiliary information from text and graphs to alleviate the scarcity of training data in the sentiment analysis, text classification, and personalized word...[ Read more ]

Currently, supervised learning based methods and techniques such as deep learning methods, have achieved great success in the text mining area. When researchers develop these giant models, they usually assume the availability of massive annotated training data. However, the real-world usefulness of these models will be impaired because in the real world, readily available annotated data are scarce. Unlabeled training data can also be rare in some circumstances such as personalization modeling. Fortunately, there is useful information existing in many inexpensive and readily available resources. In this thesis, we show how to utilize auxiliary information from text and graphs to alleviate the scarcity of training data in the sentiment analysis, text classification, and personalized word embeddings task. To alleviate the scarcity of annotated training data, we utilize auxiliary information to obtain supervision signals, so that models are trained using these signals rather than annotations. To alleviate the scarcity of unlabeled training data, we utilize auxiliary information to design heuristics to enrich training data. For example, for sentiment analysis, supervision can be opinion words in the text. For text classification, supervision can be generated words from a masked language model via querying about the topic of a document. For personalized word embeddings learning, auxiliary information can be a social graph. To utilize supervision signals from text data, we propose a variational weakly-supervised framework for the sentiment analysis and text classification task. To utilize auxiliary information from text data, we use a social network as a regularization that encourages users to learn from friends’ corpus, thus gaining more training data.