Sentiment classification method based on adversative sentence semantic block partitioning mechanism

Abstract

The invention discloses a sentiment classification method based on an adversative sentence semantic block partitioning mechanism, which comprises the following steps: 1, representing each sample in atraining set and a testing set as a word vector matrix by utilizing a known word vector dictionary; 2, selecting a proper convolution kernel to carry out convolution on the word vector matrix, and extracting a mapping feature vector to implement dimensionality reduction; 3, constructing an adversative dictionary, and by inquiring positions of adversatives in the samples, carrying out semantic partitioning on extracted mapping features and extracting the most important information in each partitioned block, and forming a final feature space; 4, training a classifier on the basis of the final feature space, and carrying out classification on the samples in the testing set. According to the sentiment classification method, on the basis of the constructed adversative dictionary, partitioning on sentence semantic blocks is implemented, the important semantic information in each segment can be obtained, and meanwhile, position structural features of a sentence are considered, so that correctness of text sentiment classification can be improved.

Description

technical field [0001] The invention belongs to the problem of emotion classification in the field of natural and language processing, and is especially aimed at effectively classifying emotions for emotional expression modes containing multiple semantics, such as transitional sentences in which suppression is first promoted or promotion is suppressed first. Background technique [0002] With the rapid development of the Internet, text information using the network as the medium of communication has attracted more and more attention from enterprises, institutions and individuals. Network information can help government departments understand people's intentions. Enterprises can understand users' opinions on products by developing product reviews to improve products. Product performance, consumers use product reviews to guide consumption behavior. However, a large number of new comments appear on the Internet every day. The comments may be positive at first, but they may turn...

AI Technical Summary

Problems solved by technology

These comment information in real life not only have emotional words that can express the emotional polarity, but also contain transition words to make the comment information have both positive and negative emotions at the same time. This feature makes the text sentiment classification problem more complicated and also makes the Traditional data mining algorithms and existing machine learning methods face severe challenges:

[0003] One of the challenges: the traditional unsupervised classification method based on the sentiment dictionary, which analyzes the emotional polarity of the words in the sentence through the sentiment dictionary, and determines the overall emotional tendency of the sentence by simply summing the polarities of these words. The importance of words It is obviously difficult to get better results without distinction;

[0004] Challenge 2: The text sentiment analysis method based on machine learning (including: k-nearest neighbor, support vector machine SVM, Bayes, etc.) has the following main problems: 1) Using the traditional bag-of-words method for representation, the dimension of the text vector The number is high and the data is relatively sparse, and the training of the model is not used; 2) only the syntactic structure between features is considered and its semantic information is ignored, resulting in a semantic mismatch in the feature mapping results, which cannot well represent the semantics of the document

However, the traditional CNN network ignores the structural features of the sentence when it is used for sentiment analysis. The Max-pooling method extracts a maximum value from the features of the sentence according to the importance, and does not make any distinction on the structure of the sentence.

This feature makes the method ineffective in processing transitional sentences

Images