Improving accuracy using a dictionary

Rather than just returning the given prediction, we can check whether the word actually exists in our dictionary. If it does, then that is our prediction. If it isn't in the dictionary, we can try and find a word that is similar to it and predict that instead. Note that this strategy relies on our assumption that all CAPTCHA words will be valid English words, and therefore this strategy wouldn't work for a random sequence of characters. This is one reason why some CAPTCHAs don't use words.

There is one issue here—how do we determine the closest word? There are many ways to do this. For instance, we can compare the lengths of words. Two words that have a similar length could be considered more similar. However, we commonly consider words to be similar if they have the same letters in the same positions. This is where the edit distance comes in.

Ranking mechanisms for words

The Levenshtein edit distance is a commonly used method for comparing two short strings...