A memory-based frequent-pattern discovering algorithm for large-scale corpus was presented.First

the origi-nal corpus was partitioned into several parts using appropriate dividing policy.Then each partition was processed inde-pendently to produce a temporary result

and the union of all temporary results is the final frequent-pattern set.The algo-rithm prunes a subtree once it is sure that none of the corresponding pattern will be frequent.Experiment shows that it takes no more than 1.6 gigabytes of memory to discover all patterns appearing more than 100 times for a 3.6 gigabytes news corpus

the average speed is 3.28 magabytes per second.