This paper presents a method which improves the performance of Vector Space Model (VSM) when applying it to Chinese Frequently Asked Questions (FAQ). This method combines unigram and bigram models in determining the similarity of document vectors. The performance is further improved by applying shallow lexical semantics and the document length information. Experiments showed that the proposed methods outperform baselines (segmentation and bigram) across different datasets which include FAQs from restricted domains and open domains.