Identifying relevant descriptors for tweet sets

Twitter is a media where information flows in vast volumes. Even considering a particular topic, the collected tweets come in a wide variety of forms. This makes the data description problem complex. In this paper, given a set of tweets, we consider the problem of producing a set of relevant descriptors to characterize the tweets. Messages and words are considered in an embedding space learned with Doc2Vec, a model well suited for short documents. We propose to leverage this model to identify text units that may span over several words. We then propose to measure the impact of the units on their document representation through ablation. The most important units are selected as descriptors. Experiments have been conducted in the context of a tweet cluster description problem on two datasets. One is about the storm Alex, which struck France in October 2020, and the other about the beginning of the Russia-Ukraine war in February 2022. The results show the interest of our method compared to existing work.