محمد نظري

Nowadays, the volume of data is increasing. Searching, filtering, and finding the content of interest to users in this huge space is difficult and challenging. Categorization can be very useful in this field by narrowing the search space and categorizing topics. Especially with recent advances in natural language processing, many researchers are now interested in developing programs that use text classification methods. Various methods have been introduced so far for document classification, which can be mentioned as traditional methods and methods based on neural networks. In traditional methods, due to the high dimensions and the sparsity of the document representation vectors, the cost of calculations is high and their accuracy is low. In addition, in traditional methods, the semantic relationship between words is ignored. In methods based on neural networks, which are known as word embedding, each word is represented in fixed dimensions. Methods based on words and sentence embedding, when the documents are large, it is difficult to recognize the difference between documents using similarity measures due to the fact that the vectors are close to each other. On the other hand, in these methods, words are looked at locally and the global connection between words is ignored. Therefore, in this research, we introduce a method for document classification using a combination of LDA and Word2vec in order to consider both local and global features of words in the text. Then we model the data in the form of a graph and then classify the data using a graph autoencoder. In this research, we used the plot synopsis of movies to classify them according to their genres. The results of the classification and eva‎luation graph show the superiority of the introduced model compared to the previous methods. In general, it can be said that the classification results show a 7% increase in the f-score compared to previous works. We also solved the problem of cold start by using the model introduced in movie recommender systems.

متن‌كاوي , رده‌بندي اسناد , گراف شباهت , استخراج ويژگي از متن , مدل موضوعي آگاه از زمينه

Text mining , Text Classification , Text feature extraction , Context-aware Topic Model

Mohammad Nazari

Hossein Rahmani