A Word Elimination Strategy for Learning Document Representation

Computing word vectors based on neural network has motivations on document representation. Word elimination can enhance the extract effective of the quality of valuable feature of a document. In this paper, we propose a model named PV-IDF to eliminate redundancy and refine features to improve the performance of the classify model, with which tokens that carry semantic information of a document. The results show that PV-IDF model achieves state-of-art performance, especially for short-length document representation.