Research On Sentiment Analysis Of Lyrics Based On Deep Learning

In recent years,the field of natural language processing has developed rapidly,and neural networks have been widely used.At the same time,in the context of the vigorous development of artificial intelligence,5G and other technologies,in order to achieve functional diversity and enrich user experience in major mobile music software,intelligent search and recommendation functions for songs have gradually emerged and become popular.The sentiment analysis of lyrics is The key technology to realize these functions.In a song,the song's melody and the lyrics complement each other,and they jointly convey the sentiment and connotation of the song.Among them,the lyrics not only contain semantic information,but also can be used as the carrier of the melody,structure and rhythm characteristics of the audio signal.Therefore,the lyrics express the sentiment connotation of a song to a large extent,and the sentiment analysis of the lyrics is very important.Value.However,the research of lyrics sentiment analysis has problems such as the lack of open source data sets and the difficulty of extracting the semantics of lyrics due to the unique grammatical structure.In order to solve the above problems,this article has launched the research on lyrics sentiment analysis for different types of lyrics text data.The main work and innovations are as follows:1.In response to the lack of data on lyrics sentiment analysis on the Internet,a data set that can be used for lyrics sentiment analysis was constructed.The data set extracts lyrics from 15,000 mainstream Chinese songs and classifies them according to "Positive" and "Negative".Each category is divided into three levels: strong,medium and weak.Each sentiment category contains about 2000 lyrics,and the number of words in each lyrics is between 30 and 100.2.Aiming at the problem that the traditional LSTM architecture will have certain deviations when processing long sequence information,this paper proposes a sentiment analysis model based on the Bi-DLSTM network.This variant LSTM network includes an expanded jump connection layer that can be used in the information Parallel computing is realized when the transmission span becomes larger,which is more conducive to the retention of long sequence information.The comparison experiments between this model and a variety of baseline sentiment analysis models in NLPCC2013,NLPCC2014 sentiment data sets,Weibo sentiment corpus and self-built lyrics data sets show that the model is in the evaluation index accuracy rate,precision rate,recall rate and F1 value Both are better than the comparative model,which reflects the effectiveness of the model.3.Aiming at the problem that potential dependencies between categories are usually ignored in text sentiment analysis tasks,this paper proposes a sentiment analysis model with a multi-module attention mechanism.In this model,Bi-LSTM is used to model the text features of lyrics.Based on the word vector and relative position vector as the input of the network layer,the lyrics-emotion category matching module is introduced into the attention mechanism to more accurately Capture the sentiment type of the text.At the same time,the category dependency module is introduced to extract the dependencies between different sentiment levels,and gradually capture the association between each sentiment level from top to bottom,so as to achieve the result of improving the classification accuracy.This model and several baseline methods have been compared experiments on self-built lyrics dataset and educational dataset,and a detailed analysis has been carried out.Experimental results show that the model is better than other text classification methods in classification accuracy and other indicators.