Sentimental Feature Extraction Algorithm For Chinese Restaurant Reviews Based On LSTM And CNN

Aspect-level sentiment analysis aims to mine the subjective sentiment information about a given aspect from a sentence.Feature extraction is a very important part of this task.This paper focuses on feature extraction in the aspect-level sentiment analysis task of Chinese restaurant reviews.The current aspect-level sentiment analysis algorithm has the following shortcomings:First,The current character-word vector training model assigns the same weight to each character vector in a word.In most words,semantic words are distorted,such as "basketball" semantic recognition of "ball".Secondly,attention mechanism can give more weight to semantic-related words in a sentence,but not all words in a sentence are related to a given aspect of semantic;and attention weight calculation usually depends on the aspect word representation.The current attention mechanism calculates the aspect word representation directly by taking the mean value of each word vector in the aspect phrase.In this way,the obtained aspect word representation must be loss.Aiming at the above problems,this paper conducts in-depth research from the extraction of Chinese semantic features and the extraction of emotional features from text data.The main work of this article is summarized as follows:First,this paper proposes a joint model based on CWE.First,the Bi-RNN and Attention mechanisms are used to dynamically learn weights and assign different weights to each character in a word.Then,because the position of Chinese characters in words is usually fixed,this paper uses a two-dimensional CNN to extract N-gram features between character vectors.The related features obtained by the above two methods are added to the original word vector to obtain the final word representation.Secondly,we proposed an aspect-level semantic feature extraction model based on BLSTM-CNN.We have two improvements:(1)Adaptive target word representation considers expressing the target word as a weighted summation of aspect vectors and learning weights through a neural network to achieve the effect of adaptively expressing the target word.(2)CNN was added to the model to enhance the ability of local features of the model,and a proximity strategy was used to solve the problem of CNN cross-target word error extraction of emotional features.