<p>Road accidents are common and often result in the loss of many lives throughout the world. Analyzing different types of road accidents helps in planning and managing road networks. Social media, like Twitter, is vital for exchanging information globally and is a cost-effective alternative to detecting such accidents. Many challenges remain in the literature, with the primary focus on detecting the category, cause, and severity of the road accident tweet; eventually, this becomes a multi-task learning problem. Also, the language pre-trained model, called Bidirectional Encoder Representations from Transformers (BERT), suffers from out-of-vocabulary (OOV) words, which can be found in tweets for the three downstream tasks, namely category, cause, and severity of road accidents. Motivated by this, in this work, we propose a framework that combines strategies such as multi-task learning, the context-dependent BERT model, and context-free static word embedding called fastText for fine-tuning in detecting three downstream tasks from the road accident tweet. For this, we crawled Twitter to collect the tweets to leverage multi-task learning. Our framework shows that the information across Twitter datasets for related accident tasks can be helpful in understanding task-specific features. We show that our framework outperforms the baselines for the multi-task classification of road accident tweets. The findings of this work may offer crucial direction to road safety organizations, particularly in any given study area, for the proactive deployment of effective remedies so as to ensure vehicle safety.</p>