ENTRUST: Argument Reframing with Language Models and Entailment

Tuhin Chakrabarty, Christopher Hidey, Smaranda Muresan

TL;DR

This paper introduces ENTRUST, a novel method for argument reframing that aims to make arguments more trustworthy by adjusting their connotative meaning (e.g., reducing partisanship or fear-mongering) while preserving their original denotative meaning. ENTRUST combines controllable text generation with an entailment component, demonstrating significant improvements over existing baselines in human evaluations.

What Problem Does This Paper Address?

Framing, the way an issue is presented, can significantly influence public opinion. This paper focuses on lexical framing, where subtle word choices can alter the perceived meaning of an argument without changing its core factual content (denotation). For instance, "assistance to the poor" elicits a more positive response than "welfare." The challenge is to automatically reframe arguments to be more trustworthy (e.g., less partisan, less fear-inducing) while ensuring the reframed argument still means the same thing as the original. Existing methods struggle to achieve this balance, often altering meaning or failing to control the reframing process effectively.

Methods

The ENTRUST system has two main components:

  1. Connotation-guided Masked Language Model (MLM) for Data Creation: The authors leverage a lexical resource that annotates words with emotional connotations (like fear, trust, joy). They use this to identify words in existing arguments that can be replaced with synonyms carrying different connotations. A masked language model (like RoBERTa) is then used to generate plausible substitutions that maintain the original sentence's denotative meaning. To guide this process, they prepend "control codes" representing the desired emotional associations (e.g., "trust") and use special tokens [SEP] to signal which words to focus on for replacement.
  2. Controllable Text Generation with Entailment: A BART language model is fine-tuned on the generated parallel data. This model is trained to reframe arguments based on a target emotion (e.g., "trust"). Crucially, a post-decoding step uses a Natural Language Inference (NLI) model (fine-tuned on MNLI) to score multiple generated reframings. The system selects the reframing that best preserves the original argument's denotative meaning (i.e., is entailed by the original argument).

Datasets

  • Training Data: The authors create a parallel corpus of naturally occurring arguments and their reframed versions. This corpus is derived from the "Change My View" subreddit, filtered to include premises (justifications for arguments).
  • Evaluation Data:
    • Arguments with high partisan skew collocations (from Webson et al., 2020).
  • Arguments containing "appeal to fear or prejudice" fallacies (from Da San Martino et al., 2019).
  • Human-generated reframings for comparison.

Key Insights and Ideas

  • Connotation vs. Denotation is Key: The paper clearly distinguishes between the factual meaning (denotation) and the emotional/associative meaning (connotation) of words, highlighting how manipulating the latter can influence perception without altering the former.
  • Controlled Generation is Crucial: Simply using a powerful language model like BART isn't enough. Explicit control signals (like [SEP] tokens) and a post-hoc entailment check are vital for ensuring the reframed output is both semantically similar and achieves the desired connotative shift.
  • Trust as a Reframing Goal: Focusing on "trust" as a target connotation is a practical and impactful application, aiming to make arguments more persuasive and less manipulative.
  • Human Evaluation is Paramount: While automatic metrics are useful, human judgment is essential for truly assessing the quality of reframed arguments, especially regarding nuanced aspects like trustworthiness and fear reduction.

Key Results

  • Automatic Evaluation: ENTRUST maintains semantic similarity to input arguments better than most baselines, performing comparably to humans on partisan collocation tasks and outperforming all but one baseline on appeal to fear tasks.
  • Human Evaluation: ENTRUST significantly outperforms all baselines in fluency, meaning preservation, trustworthiness (for partisan tasks), and reduction of fear (for appeal to fear tasks). It achieves scores close to human performance.
  • Entailment's Impact: The post-decoding entailment step demonstrably helps maintain denotative meaning, as shown by comparing BART variants with and without this component.

Limitations and Future Work

  • Limitations:
    • The automatic metrics, while useful, can sometimes penalize human reframings that deviate more significantly from the original structure, even if they are semantically sound.
  • While ENTRUST aims for trustworthiness, the potential for dual-use in generating persuasive arguments is acknowledged.
  • Future Work:
    • Directly incorporating the role of "stance" in framing (for arguments and counter-arguments).
  • Expanding to generating "concessions" (points of agreement before disagreement).
  • Further exploring the ethical implications and ensuring responsible use of the technology.

Summary updated

Abstract

Framing involves the positive or negative presentation of an argument or issue depending on the audience and goal of the speaker (Entman 1983). Differences in lexical framing, the focus of our work, can have large effects on peoples' opinions and beliefs. To make progress towards reframing arguments for positive effects, we create a dataset and method for this task. We use a lexical resource for "connotations" to create a parallel corpus and propose a method for argument reframing that combines controllable text generation (positive connotation) with a post-decoding entailment component (same denotation). Our results show that our method is effective compared to strong baselines along the dimensions of fluency, meaning, and trustworthiness/reduction of fear.
BibTeX
@inproceedings{chakrabarty2021entrust,
  author = {Tuhin Chakrabarty and Christopher Hidey and Smaranda Muresan},
  title = {{ENTRUST: Argument Reframing with Language Models and Entailment}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2021},
  doi = {10.18653/v1/2021.naacl-main.394},
}
Figure