With the advancement of technology and development of social media, patients discuss medications and other related information including adverse drug reactions (ADRs) with their friends, family or other patients. Although, there are various pros and cons of using social media for automatic ADR monitoring, information on social media provided by patients about drugs are widely considered a valuable resource for post-marketing drug surveillance. In this study, we developed a named entity recognition (NER) system based on conditional random fields to identify ADRs-related information from Twitter data. The representation of words for the input text is one of the crucial steps in supervised learning. Recently, the word vector representation is becoming popular, which uses unlabeled data to provide a generalization for reducing the data sparsity in word representation. This study examines different word representation methods for the ADR recognition task, including token normalization, and two state-of-the-art word embedding methods, namely word2vec and the global vectors (GloVe). The experimental results demonstrate that all of the studied representation scheme can improve the recall rate and overall F-measure with the cost of the reduced precision. The manual analysis of the generated clusters demonstrates that word2vec has stronger cluster trends compared to GloVe.