Emotion Detection with n-stage Latent Dirichlet Allocation for Turkish Tweets

Özet

Understanding the reason behind the emotions placed in the social media plays a key role to learn mood characterization of any written texts that are not seen before. Knowing how to classify the mood characterization leads this technology to be useful in a variety of fields. The Latent Dirichlet Allocation (LDA), a topic modeling algorithm, was used to determine which emotions the tweets on Twitter had in the study. The dataset consists of 4000 tweets that are categorized into 5 different emotions that are anger, fear, happiness, sadness, and surprise. Zemberek, Snowball, and first 5 letters root extraction methods are used to create models. The generated models were tested by using the proposed n-stage LDA method. With the proposed method, we aimed to increase model’s success rate by decreasing the number of words in the dictionary. By using the multi-stages LDA, we were able to perform better (2-stages:70.5%, 3-stages:76.4%) than the state of the art result (60.4%) which was achieved using the plain LDA for 5 classes. Understanding the reason behindthe emotions placed in the social media plays a key role to learn moodcharacterization of any written texts that are not seen before. Knowing how toclassify the mood characterization leads this technology to be useful in avariety of fields. The Latent Dirichlet Allocation (LDA), a topic modelingalgorithm, was used to determine which emotions the tweets on Twitter had inthe study. The dataset consists of 4000 tweets that are categorized into 5different emotions that are anger, fear, happiness, sadness, and surprise. Zemberek,Snowball, and first 5 letters root extraction methods are used to createmodels. The generated models were tested by using the proposed n-stage LDAmethod. With the proposed method, we aimed to increase model’s success rate bydecreasing the number of words in the dictionary. Using the multi-stage LDA(2-stages:70.5%, 3-stages:76.375%) method, the success rate was increasedcompared to normal LDA (60.375%) for 5 class.