Maximizing Stylistic Control and Semantic Accuracy in NLG: Personality Variation and Discourse Contrast

Neural generation methods for task-oriented dialogue typically generate from\na meaning representation that is populated using a database of domain\ninformation, such as a table of data describing a restaurant. While earlier\nwork focused solely on the semantic fidelity of outputs, recent work has\nstarted to explore methods for controlling the style of the generated text\nwhile simultaneously achieving semantic accuracy. Here we experiment with two\nstylistic benchmark tasks, generating language that exhibits variation in\npersonality, and generating discourse contrast. We report a huge performance\nimprovement in both stylistic control and semantic accuracy over the state of\nthe art on both of these benchmarks. We test several different models and show\nthat putting stylistic conditioning in the decoder and eliminating the semantic\nre-ranker used in earlier models results in more than 15 points higher BLEU for\nPersonality, with a reduction of semantic error to near zero. We also report an\nimprovement from .75 to .81 in controlling contrast and a reduction in semantic\nerror from 16% to 2%.\n