Abstract: Text used for classification often suffers from semantic ambiguity and sparse features, and the meaning of certain words in the sentence may not be consistent with the semantics represented by the actual label of the text, which can lead to classification errors. To address the above issues, a multi-task text classification model combining feature projection and negative supervision is proposed. The main task uses feature projection networks to extract purified vectors with obvious class features and perform classification. The auxiliary task gives the model negative supervision to expand the differences between different categories of text vectors and eliminate the negative impact of individual words. In addition, RoBERTa and BiLSTM are used to simultaneously extract features from positive and negative samples to capture rich semantic information. The model was tested on the THUCNews title classification and micro-loan semantic similarity analysis dataset, and the results show that the model has better performance than existing models.

Key words: text classification, feature projection, negative supervision, multi-task model, RoBERTa, BiLSTM