Research On Expression Recommendation Method Based On Sentiment Analysis

As communication increasingly takes the online form,emojis have become a popular tool for Internet users to express themselves and emoji recommendation is now an important part of the core functionality of messaging applications.Currently,emoji recommendation is commonly implemented by searching texts for keywords that match explanatory captions that have been manually appended to emojis in advance,but approaches of this kind appear to have limited effectiveness.A new,improved implementation method that builds upon convolutional neural networks(CNNs)and knowledge graphs is proposed.This method will be able to recommend emojis based on the semantics of the input text and,once the user has made a choice,search the knowledge graph for its underlying meaning and connections to make further recommendation of related emojis.The study subjects in this thesis are Weibo texts and emojis.Using a dataset composed of a collection of Weibo posts that include emojis,a method is devised to make initial emoji recommendations based on the input text and follow-up recommendations that are semantically related to the user's choice in the preceding stage.The thesis consists of two parts: a text sentiment analysis algorithm based on a CNN and an emoji recommendation method based on a knowledge graph.For sentiment analysis,the thesis starts with processing the dataset for noise reduction and converting Weibo texts into a training dataset for supervised learning.Word vector-based representation learning is then carried out on the training dataset in order to transform it into vector data that can be learned by the CNN.Mainstream text sentiment analysis models usually categorize texts into the classes of joy,anger,sorrow and pleasure,which is apparently not sufficiently refined or comprehensive for Weibo texts.Borrowing from theories in psychology,a new categorization is presented composed of the seven emotions of love,sadness,surprise,resignation,anger,joy and ridicule and constructs a CNN to analyze the sentiments of the texts.Such categorization of Weibo texts is achieved by building upon LeNet,the classical CNN structure,to make it suitable for the Weibo text dataset.For emoji recommendation,sentiment entities(mostly words expressing emotions)are extracted from Weibo texts using statistical methods;entity relationships are extracted based on PMI and word vector similarities;entity properties,defined using emojis,are extracted with pointwise mutual information.A sentiment knowledge graph is then created,which,in the final step,is combined with the sentiment categorization model to make intelligent emoji recommendations based on input texts and emoji choices.