Research on Real-Time Semantic Understanding and Dynamic Evaluation of Artificial Intelligence Techniques in Multilingual Environments

Abstract

NPMI is used to describe the local co-occurrence properties of words, and the collected text data is constructed into a homogeneous graph composed of words. A word-granularity graph convolutional neural network (VGCN) is used as a graph encoder to achieve parallel training of graph convolutions. The output of the VGCN is used as a kind of embedding vector for BERT, and the VGCN is fused with other embedding using a transformer for feature fusion. The reprojection layer converts the model output to semantic representation vectors, and the post-projection layer converts the semantic representation vectors to vectors for loss computation. NT-Xent is chosen as the loss function of the model, while the real-time semantic understanding effect of the model is evaluated using the Spearman correlation coefficient in statistics. The CBOW model provided by Word2vec is learned from the training set and constantly updated for each parameter to calculate semantic similarity between words. The LDA model is used to extract topic keywords from each document and calculate the probability, which finally completes the dynamic assessment of the credibility of text content based on semantic understanding. The ROUGE-L index and BLEU-4 index of the VGCN-CL model in this paper are 66.47% and 63.86%, respectively, which are 23.20% higher than the baseline model ROUGE and 25.14% higher than BLEU-4. When the content assessment results are true, the accuracy and other metrics of this paper’s model are 90.36%, 89.87%, and 90.09%, respectively, which are the best performance among all models.

Keywords

• npmi
• graph convolutional neural network
• word2vec
• lda model
• real-time semantic understanding
• 68t42