Research On Segmentation And Recognition Of Criminal Suspects’ Dialogue Topics Based On Deep Learning

The communication service application in the suspect ’s mobile phone will generate a large amount of dialogue text,and some key dialogue paragraphs can be used as evidence for the police to solve the case.However,in the face of massive dialogue texts,it is difficult to use conventional methods to efficiently find dialogue paragraphs that can be used as evidence from the suspect ’s dialogue texts.Aiming at this problem,a method of topic segmentation and recognition of criminal suspects ’ dialogue based on deep learning is proposed to assist the police to quickly find the target paragraph from the massive dialogue text.The method is divided into two stages.The first stage : the dialogue topic segmentation is performed on the dialogue text containing multiple topics,and each topic paragraph after segmentation contains only one topic.The second stage : the dialogue topic recognition is performed on the segmented topic paragraph.The police can quickly find relevant paragraphs according to the identified topic categories to improve the search efficiency.Based on the above two stages,this article conducts research.The main work contents and results are as follows :(1)Dialogue text processing of criminal suspects: In order to solve the problem of lack of training data for dialogue topic segmentation and dialogue topic recognition of criminal suspects,1694 dialogue texts are obtained from the mobile phone of criminal suspects.Through the cleaning,correction and screening of dialogue texts,1103 high-quality dialogue texts of criminal suspects.(2)Dialogue topic segmentation : Aiming at the problem that the Text Tiling text segmentation algorithm is not suitable for processing dialogue text,it is proposed to enhance the Text Tiling algorithm through the next sentence prediction mechanism of the BERT model and apply it to the criminal suspect dialogue topic segmentation.The next sentence is used to predict the probability of whether the two sentences input are continuous as the similarity of the two sentences,so that it is suitable for the suspect dialogue text.The experimental results show that the proposed segmentation method improves the F1 value by 2.4% compared with the Text Tiling algorithm on the test data.(3)Dialogue topic recognition: BERT model has a good effect on text feature extraction by using multi-attention mechanism.In order to make full use of the semantic features extracted by the BERT model,the BERT-Bi LSTM model is proposed.The Bi LSTM model is used to learn the vectorized representation features output by the BERT model in a way different from the multi-head attention mechanism,that is,learning features in a cyclic accumulation manner.By comparing the experimental results of BERT-Bi LSTM model and BERT model on test data,F1 is increased by 1.2 %,which proves that the new hybrid model can improve the accuracy of dialogue topic recognition.After completing the training of the two-stage model,the topic paragraphs obtained in the first stage are also used as the test data in the second stage.The experimental result F1 is 60.5%,which proves that the topic segmentation and recognition method of criminal suspect dialogue based on deep learning proposed in this article is feasible.