Sentiment analysis is becoming more and more popular in the field of natural language processing. This specific field of study unites researchers with background in psychology, sociology, anthropology, linguistics, and computer sciences, and as such everyone interested in it has to have an interdisciplinary approach to the subject. The growing interest towards artificial intelligence and deep learning seen in every aspect of computer sciences did not bypass sentiment analysis either. However, tools developed to detect sentiment for Croatian language are mostly developed with traditional and outdated methods. In this paper we present the process of the development of an up-to-date sentiment analysis tool based on neural networks. The first part describes the theoretical framework of both sentiment analysis and deep learning, followed by the description of the making a specialized Twitter corpus about earthquakes. The corpus is annotated by two annotators and as such gives a unique perspective regarding the unusuality of the language used on Twitter. The most common difficulties in the process are briefly described as well. The final chapter of the paper describes the making of a sentiment analysis tool built with recurrent neural networks and trained on the mentioned corpus. The results of the sentiment analysis tool are contextualized both by comparing them to similar tools presented in other papers and the inter- annotator reliability calculated for Twitter corpus.