IR methods are increasingly being applied over microblogs to extract real-time information, such as during disaster events. In such sites, most of the user-generated content is written informally – the same word is often spelled differently by different users, and words are shortened arbitrarily due to the length limitations on microblogs. Stemming is a common step for improving retrieval performance by unifying different morphological variants of a word. In this study, we show that rule-based stemming meant for formal text often cannot capture the arbitrary variations of words in microblogs. We propose a context-specific stemming algorithm, based on word embeddings, which can capture many more variations of words than what can be detected by conventional stemmers. Experiments on a large set of English microblogs posted during a recent disaster event shows that, the proposed stemming gives considerably better retrieval performance compared to Porter stemming.