Title: Sequential text-term selection in vector space models
Authors: Jingyuan Liu - Xiamen University (China)
Hansheng Wang - Peking University (China)
Feifei Wang - School of Statistics, Renmin University of China (China) [presenting]
Abstract: Text mining has attracted more and more attention with the accumulation of text documents in all fields. We focus on the use of textual information to explain continuous variables in the framework of linear regressions. To handle the unstructured texts, one common practice is to structuralize the text documents via vector space models. However, using words or phrases as the basic analysis terms in vector space models is in high debate. In addition, vector space models often lead to an extremely large term set and suffer from the curse of dimensionality, which makes term selection important and necessary. Toward this end, we propose a novel term screening method for vector space models under a linear regression setup. We first split the whole term space into different subspaces according to the length of terms, and then conduct term screening in a sequential manner. We prove the screening consistency of the method, and assess the empirical performance of the proposed method with simulations based on a dataset of online consumer reviews for cellphones. Then we analyze the associated real data. Results show that the sequential term selection technique can effectively detect the relevant terms by a few steps.