Research On Machine-Generated Text Detection Based On Layer-wise Relevance Propagation

With the popularity of the Internet,the network is filled with a large amount of interactive information.The form of the information is complicated and the source is obscure.It is impossible to determine whether it was written by humans or generated by machines.Current state-of-the-art text generation models have been able to generate text that approaches the style of human language.Such generated texts may be misused for fake news generation,fake product reviews generation,spamming/phishing and other false information,which makes it extremely difficult for people to get credible information from the Internet.Generated text detection,that can distinguish the generated texts from the real texts,play a vital role in mitigating such misuse of the generated texts.Most existing generated text detection methods employed learning algorithms that incorporated a wide variety of features to detect generated texts,which included text content,text structure,generation technology and other aspects to discover the generated texts.However,with the improvement of the generated text quality,the existing methods can only gain the features frequently appear in the sentences,but the relevant features with their semantic correlation at the deep semantic level.It is difficult to ensure the accuracy.In addition,the existing detection methods usually can only detect the text generated from their own or similar models.The general detection mechanism of the generated texts has become the difficulty of current research.To solve the above problems,this paper studies the recognition and detection of machine-generated text and proposes the detection scheme of machine-generated text based on Layer-wise Relevance Propagation.We use Layer-wise Relevance Propagation method,take the currently popular Generative Adversarial Networks model as an example of text generation model,and find the relevant representation of the machine-generated text at the deep semantic level.Moreover,we use the attention mechanism to adjust the weight of the relevance features to achieve the accurate detection of machine-generated text.The main research work is as follows:Firstly,the interpretable method of deep neural network based on deep learning is used to explain the Generative Adversarial Networks.In detail,expanding LRP method for the generator and the discriminator of Generative Adversarial Networks respectively,and building Layer-wise Relevance Propagation models for the generator and discriminator.And the relevance propagation rules are used for different network layers or neural links.Two sets of the feature sets can be obtained,which are named as document vectors.Secondly,summing up the document vectors for the generator and discriminator to obtain the relevant features of the generated texts.Thirdly,the attention mechanism is introduced to adaptively update the feature weights,and the preprocessed relevant features are used as the input of the Softmax layer in the detection model to achieve machine-generated text detection.Finally,experiments based on real-world datasets demonstrate that the proposed model achieves more accurate machine-generated text detection and has sufficient generalization ability on the datasets with sentences of different length.