Controlling Personality-Based Stylistic Variation with Neural Natural Language Generators

  • 2018-05-22 02:07:32
  • Shereen Oraby, Lena Reed, Shubhangi Tandon, T. S. Sharath, Stephanie Lukin, Marilyn Walker
  • 1

Abstract

Natural language generators for task-oriented dialogue must effectivelyrealize system dialogue actions and their associated semantics. In manyapplications, it is also desirable for generators to control the style of anutterance. To date, work on task-oriented neural generation has primarilyfocused on semantic fidelity rather than achieving stylistic goals, while workon style has been done in contexts where it is difficult to measure contentpreservation. Here we present three different sequence-to-sequence models andcarefully test how well they disentangle content and style. We use astatistical generator, Personage, to synthesize a new corpus of over 88,000restaurant domain utterances whose style varies according to models ofpersonality, giving us total control over both the semantic content and thestylistic variation in the training data. We then vary the amount of explicitstylistic supervision given to the three models. We show that our most explicitmodel can simultaneously achieve high fidelity to both semantic and stylisticgoals: this model adds a context vector of 36 stylistic parameters as input tothe hidden state of the encoder at each time step, showing the benefits ofexplicit stylistic supervision, even when the amount of training data is large.

 

Quick Read (beta)

loading the full paper ...