Comparison of Short-Text Embeddings for Unsupervised Event Detection in a Stream of Tweets

In this article, we apply recent short-text embeddings techniques to the automatic detection of events in a stream of tweets. We model this task as a dynamic clustering problem. Our experiments are conducted on a publicly available English corpus of tweets and on a French similar dataset annotated by our team. We show that recent techniques based on deep neural networks (ELMo, Universal Sentence Encoder, BERT, SBERT), although promising on many applications, are not well adapted for this task, even on the English corpus. We also experiment with different types of fine-tuning in order to improve the results of these models on French data. We propose a detailed analysis of the results obtained, showing the superiority of traditional tf-idf approaches for this type of task and corpus.