Text steganalysis faces severe challenges from the ever-improving quality of text information hiding. Text steganalysis of social networks is more challenging because of the sparsity of steganographic text in massive network texts. Detecting sparse steganographic text mainly relies on accurate modeling of steganographic features, which is still a challenging problem for existing methods. In this paper, we propose a text steganalysis method that exploits language styles. We designed a global and local fusion language style feature extraction module based on RoBertA and TF-IDF, and used a bidirectional recurrent neural network (BiRNN) to analyze the sequence changes of language style features to achieve effective detection of sparse text steganography. Experimental results demonstrate the effectiveness of the proposed method.