Probing Black Box Language Models with Behavioral Testing to Identify Algorithmic Bias