Method and device for doubly judging bad short messages by pre-training model and short message address

Abstract

The invention discloses a method and device for doubly judging bad short messages through a pre-training model and a short message address and relates to the technical field of information. The deviceis composed of a classified short message sample set, a pre-training model module, a short message collector, a text processor, an address extractor, a web crawler and a bad short message judger. According to the invention, the problem that traditional machine learning depends on features in bad short message recognition is solved; compared with deep learning, a large number of training sets arenot needed, judgment can be conducted through url short links in the short messages, and the short messages with sparse semantic information can be well recognized; meanwhile, the property of the short message judged by combining the text information and the short message address has better interpretability and more intuitive analysis effect than the property of the short message judged only according to the ip of the short message address.

Description

technical field [0001] The present invention relates to the field of information technology, especially the field of information security technology. Background technique [0002] At present, the information security of mobile phone text messages has been paid attention to by the whole society. The research on the identification of bad text messages is mainly based on the analysis method based on text classification and the analysis method based on the url in the text message. [0003] In the analysis of bad information based on text, it is mainly based on traditional machine learning algorithms and methods based on deep learning. Traditional machine learning, such as the invention patent number CN110147448A, selects features through the construction of dual feature engineering. The text features of bad text messages are ever-changing, and the text feature extraction method in traditional machine learning cannot be fully applied to the actual situation of bad text message cl...

AI Technical Summary

Problems solved by technology

However, since a part of the mask is required for input, BERT ignores the dependency between the masked positions, so there is a difference between pretraining and fine-tuning effects (pretrain-finetuned discrepancy)

Images