Spell Checking is an integral part of modern word-processing applications. Current spellcheckers can only detect and correct non-word errors. They cannot effectively deal with real-word errors; misspelled words that result in valid English words. Current techniques for detecting real-word errors require huge volume of training corpus and the learned knowledge is represented by opaque set of features that are not apparent. This paper proposes a new method for dealing with real-word errors using selectional preferences of predicates for arguments in a case slot. The method requires very little in terms of resources and can use existing lexicons slightly modified to suit the above task.