Hierarchical Bidirectional Long Short-Term Memory Networks for Chinese Messaging Spam Filtering

Messaging spam filtering is an important research area in the field of natural language processing (NLP). In this paper, we propose a hierarchical bidirectional long short-term memory network based approach for Chinese messaging spam filtering. Considering that a message consists of sentences and a sentence consists of words, we design a hierarchical architecture to generate the representation of a message that aggregates the information of each word in each sentence. Besides, we notice that the errors produced by Chinese segment may affect the performance of our model. So we use the unsegmented characters as input rather than the segmented words like most of the Chinese NLP models. The experimental results demonstrate that our method outperforms most of the state-of-the-art methods on the dataset that is tagged manually by a online medical company. Meanwhile, we also show that the unsegmented character has better performance than segmented word in this task.