Content Ordering in the Generation of Persuasive Discourse

IJCAI 1997

TL;DR

This paper presents a hierarchical framework for generating persuasive arguments, focusing on the crucial aspect of content ordering. By integrating insights from rhetoric, psychology, and corpus analysis, their system plans argument structure and then refines it to ensure both logical coherence and persuasive impact, even influencing surface textual features like punctuation.

What Problem Does This Paper Address?

Generating persuasive discourse is a complex task in Natural Language Generation (NLG). While many systems focus on the logical structure of arguments (premises leading to conclusions), this paper highlights that the order in which these components are presented significantly impacts an argument's persuasiveness and coherence. Simply having a logically sound argument isn't enough; how it's sequenced matters for convincing an audience. The challenge is to develop a systematic way to determine this optimal ordering, moving beyond just logical validity to consider psychological and rhetorical effectiveness.

Methods

The authors propose a hierarchical planning framework, building upon their previous work. This framework has two main levels:

  • Argument Structure (AS) Level: This level determines the logical form of the argument, including deductive, inductive, and fallacious relationships between premises and conclusions. It uses a set of primitive operators like BEL (belief) and IS_SALIENT (making a proposition salient to the hearer). Crucially, it employs a planner called AbNLP, which uses a concept of encapsulation where abstract operator bodies contain goals rather than operators, allowing for more flexible planning.
  • Eloquence Generation (EG) Level: This level handles stylistic and rhetorical refinements. It doesn't change the AS plan but adds structurally redundant rhetorical statements, marks stylistic preferences, and inserts lexical information.

A key innovation is the explicit modeling of salience as a goal, not just belief. The ordering process itself is divided into two phases:

  1. Goal Fulfillment Ordering: This phase addresses how to satisfy BEL goals, which often require multiple subarguments. It allows for reorderings like pre-order (conclusion first) or post-order (premises first).
  2. Refinement Phase Ordering: This phase resolves ordering between components within an operator's body, which cannot be determined at the higher AS level.

The system uses resource estimation to predict subargument size and other factors to guide ordering decisions, drawing on persuasion-oriented heuristics (e.g., increasing strength, handling skepticism) and coherency-oriented heuristics.

Datasets

The paper mentions that their approach is based on the study of a corpus of natural arguments drawn from scientific papers, advertisements, editorial commentary, and "letters to the editor." No specific dataset is named or introduced.

Key Insights and Ideas

  • Salience is Key: The paper argues that making information salient to the hearer is as important as ensuring they believe it. This allows for enthymematic arguments (where some premises are implicit).
  • Hierarchical Planning for Argument: A multi-level planning approach (AS and EG) provides a structured way to build arguments, separating logical structure from stylistic refinement.
  • Ordering is Crucial for Persuasion: The order of argument components is not arbitrary; it directly influences how persuasive and coherent the discourse is. This is supported by psychological research and rhetorical traditions.
  • Topic Manipulation: Explicitly modeling topic manipulation (using PUSH_TOPIC and POP_TOPIC operators) is vital for generating coherent arguments and is linked to the use of clue words and formatting.
  • Heuristics Drive Ordering: Persuasion-oriented heuristics (based on rhetoric and psychology) and coherency-oriented heuristics are used to guide the content ordering process.
  • Surface Features Emerge from Structure: The ordering process at the AS and EG levels can directly influence surface textual features like punctuation, clue words, and formatting, enhancing comprehension.

Key Results

The paper demonstrates how their framework can generate ordered argument plans. Figure 6 shows an example of an argument plan before and after ordering, illustrating how the structure is refined. Figure 7 shows the result of refinement before and after ordering, highlighting how internal component ordering within subarguments can be adjusted to improve coherence. The authors claim their approach facilitates the subsequent realization of structure into text, particularly the generation of clue words important for argument comprehension.

Limitations and Future Work

  • Resource Estimation: While resource estimation is used to guide ordering, it's noted as a pragmatic notion of universal quantification, suggesting potential for more principled solutions.
  • Parallel Evidence Ordering: The paper dismisses "parallel evidence ordering" as inherently incoherent and requiring significant clue word repair, suggesting this is a constraint handled by the AS level.
  • Implicit Information: While enthymematic arguments are mentioned, the paper doesn't deeply explore the mechanisms for deciding what can be left implicit beyond the IS_SALIENT goal.
  • Future Work: The paper implies that further exploration of how ordering influences surface features and the integration of more sophisticated persuasion and coherency heuristics would be valuable. The authors also suggest that the distinction between intentional and attentional communicative goals could be further explored.

Summary updated

BibTeX
@inproceedings{1997content,
  title = {{Content Ordering in the Generation of Persuasive Discourse}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1997},
}
Figure