In this paper, we investigate a range ofudstrategies for combining multiple machine learningudtechniques for recognizing Arabic characters, where weudare faced with imperfect and dimensionally variable inputudcharacters. Experimental results show that combinedudconfidence-based backoff strategies can produce moreudaccurate results than each technique produces by itselfudand even the ones exhibited by the majority votingudcombination.
展开▼