Semantic similarity is a core technique of many NLP research fields. However, state-of-the-art metrics for semantic similarity computation often operate at different levels, e.g., words or sentences. In this paper, semantic similarity computation metrics are firstly introduced and the quality is measured in order to determine their advantages and limitations; then a new semantic similarity metric based on multi-features fusion is proposed. Distributed representations of words are used for alignment-based disambiguation operation, and Wikipedia tags are used to enhance the performance of our approach. The proposed metric is unsupervised, and can be applied at different levels e.g., single words or entire documents. The metric is evaluated on both English and Chinese datasets, it is shown that the precision and recall scores are higher than metrics which simply using knowledge base or distributed representation of words.