Controllable natural language generation for audience-centric styles

Moorjani, Samraj

Permalink

https://hdl.handle.net/2142/120261 Copy

Description

Title

Controllable natural language generation for audience-centric styles

Author(s)

Moorjani, Samraj

Issue Date

2023-04-14

Director of Research (if dissertation) or Advisor (if thesis)

Sundaram, Hari

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Natural Langauge Generation
• Persuasiveness
• Audience-centric Styles
• Persuasive Natural Language Generation

Language

eng

Abstract

Adopting contextually appropriate, audience-tailored linguistic styles, namely persuasiveness, is critical to the success of user-centric language generation systems (e.g., chatbots, computer-aided writing, dialog systems). While existing approaches demonstrate textual style transfer with large volumes of data, grounding style on audience-independent factors is innately limiting because many stylistic objectives (e.g., persuasiveness, memorability, empathy) are hard to define without audience feedback. In this thesis, we first propose the novel task of style infusion - infusing the stylistic preferences of audiences in pretrained language generation models. Since humans are better at pairwise comparisons than direct scoring - i.e., is Sample-A more persuasive than SampleB - we leverage limited pairwise human judgments to bootstrap a style analysis model and augment our seed set of judgments. We infuse the learned textual style in a GPT-2 based text generator while balancing fluency and style adoption. With quantitative and novel qualitative assessments, we show that our infusion approach can generate compelling stylized examples with generic text prompts. We then utilize complex linguistic features strongly correlated with persuasiveness to guide the generation of sequence-to-sequence models. We explore modifications of two approaches - an edit-then-prototype model and the style infusion architecture - to exhibit a “tuning-knob”- like control over the speed of text (i.e., how quickly content is covered). We empirically show that our modifications lead to strong controls over generated text and discuss directions to improve the fluency and control of generations further.

Graduation Semester

2023-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/120261

Copyright and License Information

Copyright 2023 Samraj Moorjani

Owning Collections