This paper focuses on the automatic acquisition of semantic relationships from Chinese corpus, motivated by improving the performances of our QA systems named NL-WAS. Linguistic patterns designed for Chinese sentences are applied to a collection of texts to extract synonymy relationship, hyponymy relationship, and meronymy relationship. Patterns are broken down into unambiguous and ambiguous, and different strategies are adopted to refine the candidates extracted using this two kinds of patterns. Compared to other previous works, we apply not only strict unambiguous patterns but also loose unambiguous patterns to extract relationships and proposed efficient approach to refine the outputs of these patterns for the sake of high recall and high precision. The experimental result shows that the proposed method can delete most noisy pairs of terms and improve accuracy and efficiency of NL-WAS. At the same time, our method is complementary to statistically based approaches that find semantic relationships between terms.