Sub-event detection for social media is getting increasingly important because social media platforms have played key roles in disseminating vital information to the public. However, conventional methods fail to achieve high-quality analysis of events evolution due to the severe sparseness and noise of tweets data. In this paper, we propose an unsupervised sub-event detection model which learns rich information from hashtags with a Text-CNN model. Furthermore, a two steps training method leveraged by KL divergence is introduced to further reduce the negative influence of incoherent semantics. The experiments show that our method achieves very good performance on datasets of different languages. On the Chinese dataset, in terms of NMI, and BCubed F1 precision, our method has a significant increase of 5.1%, and 11.9%, respectively, over the baseline methods. In most cases, our method significantly improves the performance of sub-event detection compared with state-of-the-art methods.

社交媒体的子事件检测变得越来越重要，因为社交媒体平台在向公众传播重要信息方面发挥了关键作用。然而，由于推文数据的严重稀疏性和噪声，传统方法无法实现对事件演变的高质量分析。在本文中，我们提出了一种无监督的子事件检测模型，该模型使用 Text-CNN 模型从主题标签中学习丰富的信息。此外，引入了利用 KL 散度的两步训练方法，以进一步减少不连贯语义的负面影响。实验表明，我们的方法在不同语言的数据集上取得了非常好的性能。在中文数据集上，在 NMI 和 BCubed F1 精度方面，我们的方法分别显着提高了 5.1% 和 11.9%，超过基线方法。在大多数情况下，与最先进的方法相比，我们的方法显着提高了子事件检测的性能。