Bayesian filtering applies statistical modeling to messages, by performing a frequency analysis on each word within the message and then evaluating the message as a whole. Therefore, it can ignore a suspicious keyword if the entire message is withinnormal bounds. Heuristic filtering is less effective, since new exception rules may need to be defined when a valid message is labeled as spam. Signature-based filtering is useless against variable-length messages, because the calculated MD5 hash changes all the time. Finally, pattern matching is actually a degraded rule-based technique, where the rules operate at the word level using wildcards, and not at higher levels.