This paper proposed a novel method to extract bilingual translation pairs from the web. Based on the observation that translation pairs tend to appear collectively on the web, a recursive process is used to extract high quality translation pairs from the web. First query the search engine with some seed data and crawl the returned pages. Then identify the Collective Translation Pair Block (CTPB) which contains the collective translation pairs using a heuristic evaluation method. After the CTPB has been identified, a PAT tree is employed to generate the extraction patterns automatically. Then a ranking SVM model is used to re-rank these patterns based on the F measure. The top 10 patterns are adopted to extract the translation pairs with the help of surface pattern. At last in order to get the high quality extraction translation, the extracted translation pairs are verified by a SVM classifier based on the translation relevant between the source and the target language.