An Improved Multi-Channel CNN-LSTM Model For Twitter Text Sentiment Analysis

In recent years,with the large-scale commercialization of 4G technologies and a rapid development of the mobile Internet.In order to analyze people's opinions and emotional tendencies in a timely manner,requiring a text sentiment analysis of the opinions expressed by users for timely response and improvement measures.The traditional sentiment analysis method have some shortcomings,such as severe dependence on emotional dictionary and manual maintenance of text features.At present,the deep learning model represented by CNN-LSTM makes up for the shortcomings of traditional methods,but it has different emphases in different emotional granularity units,but ignores the overall emotional judgment.This thesis proposed a Multi-Channel CNN-LSTM model for English text sentiment analysis.The specific innovations are as follows:1.At present,the most widely used single-channel CNN-LSTM performs poorly in special contexts because the information contained in the words is insufficient,and the information contained in some characters could refer more information.Change the single input channel of CNN-LSTM model into multi input channel,which including a word channel and a word channel,so that it can accept both word embedding and character embedding for sentiment analysis.Concatenate the two channel output vectors and then feed into the fully connected layer,after that use Softmax layer for classification.This solves the problem of relying on the quality of word segmentation and the inability to process sequence data efficiently.This improved model could increase the Accuracy from 0.81 to 0.82 and increase the Recall from 0.55 to 0.82 on the Sentiment140 dataset.2.In order to solve the problems of large input matrix,poor ability of discrimination and easy over-fitting after addition of multi-channel structure,the thesis added Attention mechanism to the two channels,and calculated the Attention score via learning different weights to solve the inaccurate problem.The lack of discrimination due to the huge size of the input features.The experimental results show that the model achieves an accuracy of 0.89 and 0.99 on the Sentiment140 and IMDB datasets respectively,whichis 32% and 1% higher than the accuracy of the CNN-LSTM model of 0.67 and 0.98,Precision increased by1.2% and 6.2%,and F1-score increased by 0.1% and 3.7%,respectively.3.Using the proposed model to design a platform for social media sentiment analyzation,users could upload corpus,text information acquisition,training model,analyzation and prediction of sentiment,result display and export results,furthermore help users analyze public opinions on the network in time.