For the underlying data, see the spreadsheet of features with Mann-Whitney U-test results.
In the chart immediately above, features to the right are more characteristic of the girls' corpus (high 'rho'), and those to the left are characteristic of the boy's corpus. Some terms may be difficult to read; the same order could be found by sorting the spreadsheet by girlboy_rho. The distributions observed for features closer to 0 on the Y axis (low p-values) would be more improbable if the corpora were random samples of the same distribution; higher p-values suggest that differences we are seeing could more easily be the product of random variations. P-values do not tell us anything about the size of the difference itself. Also, as corpus linguist Adam Kilgariff memorably pointed out, Language is never, ever, ever, random. But since we know already the corpora are different, it is interesting to see which features tend to be more characteristic of one or another.
Ted Underwood usefully discussed his experimentation with Mann-Whitney calculations in a blog post in 2011.
The above is the result of work in Python in a Jupyter notebook (viewable as HTML or in a version that can be run locally with Jupyter/IPython).