Improving lexical diversity of domain-specific automatic text generation with out-of-domain training data and style conditioning