Like the Name Blender, the Word Blender enumerates all combinations of the words supplied. Each combination is then scored against a language score with additional edit distance component. Language models are statistical models that can determine how likely a string of characters is to occur in a given language. The language model helps the Worde Blender identify the more natural combinations of the starter words that are likely easier to pronounce. The edit distance component rewards combinations that more closely match each of words supplied. Combinations below a certain threshold are discarded. The end result is a new, but plausible name that is both easy to read and pronounce.
The technical details: The Word Blender scores names by combining a language score with an edit distance score. The language score is created by using a language model with Kneser-Ney smoothing to model the probability character sequences (n-grams) of a string. n-gram frequencies are determined by analyzing the word frequencies from the Project Gutenberg frequent word list.