Social streams from existing online platforms such as Twitter or X have consistently shown the most up-to-date information about current events. Detecting and categorizing Twitter content connected to events is one of the most challenging tasks. Many researchers have therefore been interested in this issue for many years. In this paper, we propose an approach to identifying events that occurred from tweet text. To accomplish this goal, (1) tweets are first decomposed into sets of representative word-entities using NER technology; (2) a graph representing relationships between co-occurring entities is created; (3) several clustering methods are investigated on the graph; and finally, (4) tweets are assigned to the locally dense subgraphs (i.e., clusters) as identified events. Experiments conducted on two standard Twitter datasets demonstrate that the proposed strategy outperforms state-of-the-art methods. The results also show the achievement of accurately detecting significant events at the top of the ranking results.