Abstarct: Extracting emotions from text consists of textual analysis for the existence of any form of conveyed emotion from the author. The practicality of such an endeavor could be to enhance the sale of a product by review analysis, to prevent suicide by analyzing individuals’ writings for certain emotions, or categorizing books baxsed on the emotions they entice in the reader in order to upgrade an existing recommender system. In extracting emotions, deep learning language models such as BERT have created a new wave. However, not only are such models dependent on the target text, but also do they require to be trained or at least fine-tuned, which is both time and energy consuming. On the other hand, rule-baxsed emotion extraction methods that rely on keyword detection are computationally light and mostly independent of the target text although they are privy to problems such as lexical mis-matching. To address such issues, we proposed a new approach for extracting emotions from ISEAR memories where emotions are modeled by emotionally charged word selected via a deliberate procedure and where each ISEAR memory’s semantic similarity to a modeled emotion is calculated via an off-the-shelf SBERT language model. To select words that properly model emotions we deployed to methods: 1) In a random fashion, for each category of emotion we selected a certain number of words from the NRC emotion dictionary, 2) We cross-referenced NRC keywords against the words from another text with the highest TF-IDF numbers and only selected those NRC keywords for emotion modeling that also have a high TF-IDF score in another text. Then, the cosine similarity of each ISEAR memory and each modeled emotion was calculated by the same SBERT model that was used to create word embeddings. The calculated similarities are recognized as emotional vectors for each memory. Machine learning classifiers and a ranking method are separately used to classify each emotion vector into one of the target emotions. Our results indicate a definite superiority to rule-baxsed methods, and they challenge BERT metrics regarding extracting emotions from the same dataset. From the practical perspective, we endeavored to predict genres of books reviews by creating emotional vectors. We show that in doing so, our proposed method works expectedly better that rule-baxsed methods.