That's a tough problem when generating word embeddings. Bodies of text will rarely write "walking walk" or "the dancing dance" consequently leading naïve implementations to consider them unrelated.