Research On Sentiment Analysis Based On Deep Learning

With the rapid development of network technology,various online social platforms have become the primary choice for people to communicate and obtain information.Compared with traditional communication methods,people can now share their thoughts freely in a short period of time,and this phenomenon has also led to the increasing proliferation of network data.Sentiment analysis of user-generated online reviews can be applied to product performance analysis and product optimization,and has high research value.However,the traditional sentiment research based on sentiment dictionary or machine learning has certain limitations.The former relies on the quality of the dictionary,while the latter relies on highquality data,which cannot satisfy sentiment analysis of massive data.The current deep learning sentiment analysis method can automatically extract relevant features from largescale text data,and extract deep semantic information more easily.Due to the limitations,this paper studies the construction of vector representation and sentiment classification models in sentiment analysis by deep learning methods.The specific work is as follows:Firstly,in terms of word vector transformation of data,in view of the problem that the traditional word vector transformation model lacks the ability to distinguish the weight of text words,this paper uses the TF-IDF algorithm to evaluate the importance of words in the text,and its weight in similar corpora is unbalanced The problem is improved so that it can be combined with the mainstream Word2 vec word vector conversion model to improve the weight of keyword vectors.At the same time,in view of the insufficient emotional information expressed by the vector matrix represented by the traditional single text,this paper proposes a multi-feature fusion method,which integrates part-of-speech features and text features to solve the problem of polysemy in the text,and integrates the Emoji is integrated into it to build a multi-feature vector matrix,which makes the emotional information representation more abundant.Secondly,in the construction of deep learning sentiment analysis model,in view of the lack of judgment of the structure of sentences and different emotional tendencies in the traditional CNN network,this paper uses the characteristics of sentence structure,introduces a transition vocabulary,and proposes a segmented convolution layer,which has different structures according to sentences.Features improve on the pooling layer.At the same time,in view of the problem that the traditional LSTM network is susceptible to the interference of filler characters,this paper proposes a mask matrix to redefine the current neuron state of the input data according to the formula to eliminate interference,and optimize the weight allocation ratio with the attention mechanism that introduces positional features.Then this paper combines the advantages of the improved two models,proposes the Conv Bi LSTM-ATT sentiment analysis model,and confirms the effectiveness of the model in public datasets.Finally,this paper builds a complete emotion analysis system based on vector transformation and emotion analysis technology,designs the overall framework of the system and the core function module based on the algorithm proposed in this paper,and combines the intelligent robot NAO on the basis of page interaction to complete the emotional interaction process of man-machine dialogue,so as to closely combine theory and practice.