Topic extraction by clustering word embeddings on short online texts

We demonstrate our topic extraction method in which topics are treated as clusters of word embeddings. The OPTICS algorithm is used to find small and arbitrarily-shaped clusters of embeddings, produced by a fastText model. The result is a set of dominant and non-dominant domain-specific topics. The focus of the method is on short online posts which are difficult to analyze with traditional topic extraction approaches because of the word collocation scarcity. The method is tested on dataset of posts from Twitter, LinkedIn and company blogs related to industrial automation. The method significantly outperforms traditional topic extraction approaches by finding relevant and understandable topics with related tokens.