PaperlayerView exact revision on arXiv

Source-linked AI summary

Persuasive Natural Language Generation -- A Literature Review

Sebastian Duerr, Peter A. Gloor

arXiv:2101.05786v1cs.CLcs.AI

TL;DR

The review asks what research has established about persuasion and natural language generation, an area not comprehensively addressed in the reviewed journals. The authors conduct a literature review and organize findings into business-focused categories, scholarly theories, and relevant tools and datasets. The review synthesized 77 sources into four persuasive-NLG categories and 49 determinants, and identified 6 tools and 17 persuasion or message datasets.

  • Problem

    The review asks what research has established about persuasion and natural language generation, an area not comprehensively addressed in the reviewed journals.

    Source: §1 · p. 1, §4 · p. 11

  • Method

    The authors conduct a literature review and organize findings into business-focused categories, scholarly theories, and relevant tools and datasets.

    Source: §1 · p. 2, §2 · p. 2, §3 · p. 6

  • Results

    The review synthesized 77 sources into four persuasive-NLG categories and 49 determinants, and identified 6 tools and 17 persuasion or message datasets.

    Source: Tools & Datasets · p. 9, §5 · p. 12

  • Takeaways & Limitations

    The review proposes a research agenda that brings previous efforts together and identifies the need to clarify how determinants influence persuasive-message generation.

    Source: §5 · p. 12

    Persuasive NLG may lack crucial information not explicit in text, so the system may be limited to suggesting techniques or linguistic alterations for a human persuader to edit and submit.

    Source: §4 · p. 12

Abstract

from arXiv · show

This literature review focuses on the use of Natural Language Generation (NLG) to automatically detect and generate persuasive texts. Extending previous research on automatic identification of persuasion in text, we concentrate on generative aspects through conceptualizing determinants of persuasion in five business-focused categories: benevolence, linguistic appropriacy, logical argumentation, trustworthiness, tools and datasets. These allow NLG to increase an existing message's persuasiveness. Previous research illustrates key aspects in each of the above mentioned five categories. A research agenda to further study persuasive NLG is developed. The review includes analysis of seventy-seven articles, outlining the existing body of knowledge and showing the steady progress in this research field.

1. Introduction

Persuasive NLG aims to generate messages that persuade a user to accept an argument, while prior research has largely examined persuasion in NLP and psychology. This review asks what research on persuasion and NLG has established and organizes relevant findings into a business-oriented framework.

1. Introduction

  • Persuasive NLG creates communications intended to persuade a user to accept a specific argument, potentially addressing opposing arguments.The review situates persuasion in contexts ranging from compensation requests to health and charitable appeals.Source: §1 · p. 1
  • Prior work studied syntax, persuasion identification, and psychological factors, while Iyer and Sycara (2019) identify subsequent persuadee uptake as an additional step.The review uses this gap to frame its question about the status of research on persuasion and NLG.Source: §1 · p. 1
  • The review examines research concepts and technical intricacies to develop a theory-based framework for persuasive NLG in a business context.It also aims to identify research gaps and survey a substantial part of broader research efforts.Source: §1 · p. 2

2. Method

The review follows a documented literature-review framework, defining a business-oriented scope and conceptualizing persuasion through principles and scholarly theories. It searches selected journals and expands coverage through citation searches.

2. Method

  • The review follows five steps: defining scope, conceptualizing the topic, searching the literature, analyzing and synthesizing it, and developing a research agenda.Its framework emphasizes comprehensive documentation of the literature-search process.Source: §2 · p. 2
  • The review targets applied research on persuasion and NLG, using five categories derived from the business problem of persuading through textual exchange.It describes its coverage as representative, limited to selected journals rather than the full literature.Source: §2 · p. 2
  • The conceptualization relates persuasion to intentional mental-state change and situates persuasive acts within rhetorical facets and business principles.The cited principles include confidence, logical argumentation, appealing to the other party, careful word choice, flattery, and persistence.Source: §2 · p. 3
  • The framework maps benevolence, linguistic appropriacy, logical argumentation, and trustworthiness to cognitive dissonance, language expectancy, probabilistic models, and balance theory.The review also considers data-processing tools and datasets for persuasive NLG.Source: §2 · pp. 3–4
  • The search used selected top journals and topic-related terms, screened retrieved articles for thematic consistency, and added relevant works through backward and forward citation searches.The citation-search stage identified another 48 relevant articles from journals, conferences, and magazines.Source: §2 · pp. 5–6

3. Results

The results synthesize prior research into four persuasive-NLG categories and separately consolidate relevant tools and datasets. The categories are presented alphabetically, not as a ranking of persuasiveness.

3. Results

  • The results section presents four categories of persuasive NLG alongside relevant tools and datasets needed for implementation.The categories are based on the business framework introduced in the topic conceptualization.Source: §3 · p. 6
  • The four categories are ordered alphabetically, so their sequence does not indicate differences in persuasiveness.The section also provides tools and datasets relevant to implementation.Source: §3 · p. 6
  • The section reports the literature analysis and synthesis as the review framework’s fourth step.It follows the framework’s earlier conceptualization step.Source: §3 · p. 6

Benevolence

The benevolence category gathers determinants aimed at creating value for the persuadee. The review identifies eight determinants and relates them to perceived benevolence and cognitive dissonance.

Benevolence

  • Eight benevolence determinants concern altering the persuadee’s perceived benevolence through dissonant or consonant measures.The review says implementation in persuasive NLG can be facilitated by identifying determinants’ absence or impact.Source: Benevolence · p. 6, Table 3 · p. 6
  • The category’s determinants are ordered alphabetically, not to imply a particular sequence.Table 3 presents each determinant, a concise definition, examples, and corresponding citations.Source: Benevolence · p. 6, Table 3 · p. 6

Linguistic Appropriacy

The review organizes persuasive NLG around linguistic appropriacy, logical argumentation, and trustworthiness. It identifies language matching, consistent reasoning, and trust-related determinants as ways to shape persuasive messages.

Linguistic Appropriacy

  • Fourteen linguistic-appropriacy determinants aim to match an individual’s stylome, aligning generated language with persuadee expectations.The review says this congruence can generate persuasiveness.Source: Linguistic Appropriacy · p. 7

Logical Argumentation

  • Consistent logical arguments increase persuasiveness, and the review identifies fourteen determinants with implementations such as first-order logic and semantic argumentation graphs.The determinants are presented alphabetically.Source: Logical Argumentation · p. 8

Trustworthiness

  • Fifteen trustworthiness determinants address increasing or decreasing trust, including psychological profiling to influence a persuasive act’s persuasiveness.The review links trust to the persuadee’s assessment of the persuader’s reliability and honesty.Source: Trustworthiness · pp. 8–9

Tools & Datasets

The review consolidates tools and datasets used in technical persuasion research as resources relevant to persuasive NLG. It reports six tools and seventeen persuasion or message datasets.

Tools & Datasets

  • The analyzed studies use varied tools and datasets to computationally process data for technical analyses of persuasion in NLP or NLG.The review consolidates these resources in an overview table.Source: Tools & Datasets · p. 9, Table 6 · p. 10
  • The review identifies six tools and seventeen persuasion or message datasets.The resources were selected for use in persuasion, textual exchange or debate, and NLP.Source: Tools & Datasets · p. 9
  • The overview classifies resources as tools or datasets and provides synopses, applicable links, and citations.The entries are sorted alphabetically.Source: Tools & Datasets · pp. 9–10

4. Discussion

The discussion proposes a framework and research agenda for persuasive NLG, identifying determinants and resources for future study. It also cautions that generated suggestions may need human input because systems can lack crucial information.

4. Discussion

  • The agenda calls for clearer understanding of how forty-nine determinants affect persuasive-message generation and identifies twenty-one tools and datasets for training persuasive NLG systems.Figure 3 presents the proposed future-research agenda.Source: §4 · p. 11, Figure 3 · p. 11
  • The framework combines four categories with academic theories, but the review says these prerequisites have not been comprehensively considered in the journals examined.The authors propose theorizing successful persuasiveness first, before using the framework to investigate generation approaches.Source: §4 · p. 11
  • Future work should empirically implement benevolence and examine how linguistic analysis can derive language appropriate to a persuadee.The discussion gives hostage negotiation as an example that may transfer to business situations.Source: §4 · p. 11
  • The discussion presents consistent argumentative logic as a way to organize persuasive acts and says a highly trusted persuader is more likely to be chosen as an interaction partner.Connectives, hedges, or logical operators are named as possible logic determinants.Source: §4 · p. 12
  • The proposed research asks how to implement and integrate determinants of language appropriacy, trustworthiness, benevolence, and logical argumentation, which depend on environmental aspects.The discussion also asks which tools and datasets best contribute to deep-learning training for persuasive NLG.Source: §4 · p. 12
  • Persuasive NLG may lack crucial information not explicit in text, so the persuader may need to edit and submit its suggested message and remains responsible.The system may instead serve as an assistant that suggests techniques or linguistic alterations.Source: §4 · p. 12

5. Conclusion

The review synthesizes 77 sources into 49 determinants across four categories and proposes a research agenda for persuasive NLG. It cautions that the framework may be incomplete and that determinants depend on context.

5. Conclusion

  • The review integrates 77 sources into 49 determinants across four categories, summarized in Table 7.The authors use this synthesis to address the status of research on persuasion and natural language generation.Source: §5 · p. 12, Table 7 · p. 12
  • Research has moved beyond identifying persuasion toward generation, but the reviewed work remains fragmented and lacks sophisticated integration.The review also identifies tools and datasets as critical technical resources for AI success.Source: §5 · p. 13
  • The review is mainly limited to work from 2000–2020 and selected top journals, and the authors cannot guarantee the framework is complete.Backward and forward searching may also introduce less sophisticated sources.Source: §5 · p. 13
  • No single determinant suffices: persuasion depends on multiple interactions in context, including factors such as the persuadee’s mental state and environment.The authors reject a one-size-fits-all approach and suggest continuous checks and updates to models, data, and tools.Source: §5 · p. 13
Loading 2101.05786v1…