چكيده به لاتين
Human verbal communication involves emotional messages that are conveyed through the use of emotional words. Today, both individuals and organizations need to use relevant "opinions, feelings, and public opinions" in making decisions about their work processes. The purpose of sentiment analysis and opinion mining is to identify a person's attitudes and feelings about specific topics such as movies, products, etc. The problem of integrating advanced neural language models with emotional information remains an area ripe for exploration. Therefore, this thesis presents a model for generating emotional text and recognizing conversational emotional words based on the Affect-LM model. For this purpose, LSTM neural network was used to identify, recognize and generate emotional sentences. In order to improve the performance of the LSTM network, the additional energy term has been used in word prediction, which includes an input vector that includes the emotional group information obtained from the words in the text during training, and the output of the network that works on emotional words. which differentiates the emotional information conveyed by each word. In order to improve the accuracy of detecting and generating emotional sentences, the Affect-LM model has been used to extract the content context and the content context vector of the input data before generating emotional sentences and to train the neural network. In order to test and evaluate the proposed model, three different emotional databases including Twitter, DAIC and SEMAINE databases were used. The results of this research show that using the proposed network to detect emotions and generate emotional text in all target groups, which includes positive, negative, angry, anxiety, and worry emotional sentences, has a good performance and most of the emotional sentences In the Twitter database, it includes the sentences that have anxiety and the least amount of negative emotion sentences. The DAIC database contains the most negative and positive emotional sentences, and in the SEMAINE database, negative emotion sentences have the highest amount, followed by positive emotion sentences and sad sentences. be In general, by using the proposed method and improving the Affect-LM model by determining the content context and the content context vector of the sentences, emotions can be well recognized in different databases and the generation of emotional sentences in different emotional groups with definition labels has been created.
