This paper describes our system used in the ACL 2015 Workshop on Noisy User-generated Text Shared Task for Named Entity Recognition (NER) in Twitter. Our system uses Conditional Random Fields to train two separate classifiers for the two evaluations: predicting 10 fine-grained types, and segmenting named entities. We focus our efforts on generating word representations from large amount of unlabeled newswire data and tweets. Our experiment results show that cluster features derived from word representations significantly improve Twitter NER performances. Our system is ranked 2nd for both evaluations.