Introduction
Lexical framing—the choice of specific words to describe an issue—can significantly influence public perception without altering underlying facts. For instance, "assistance to the poor" often garners more support than "welfare," despite referring to the same concept. While large language models (LLMs) excel at general style transfer, reframing political or argumentative text remains difficult due to data scarcity and the risk of "meaning flips," where a model accidentally changes the factual content of a sentence while attempting to adjust its tone.
The ENTRUST Framework
To address these challenges, the authors developed ENTRUST (ArgumENT Reframing with langUage modelS and enTailment). This framework employs a three-stage pipeline designed to surgically alter the emotional connotation of a sentence while strictly preserving its denotation.
- Synthetic Data Generation: To overcome the lack of parallel reframing datasets, the researchers used a Connotation-guided Masked Language Model (MLM). By masking words with specific emotional weights (e.g., "fear") and prompting a RoBERTa model to suggest contextually appropriate alternatives with different connotations (e.g., "trust"), they created a large-scale training set from Reddit's Change My View community.
- Controllable Generation: A BART model was fine-tuned on this synthetic data. The model uses control codes and demarcator tokens (
[SEP]) to identify and rewrite specific spans of text according to a target emotion. - Entailment Filtering: To prevent hallucinations, the system generates multiple candidates and uses a Natural Language Inference (NLI) model to select the version that most strongly entails the original input, ensuring factual consistency.
Key Insights
The core innovation of ENTRUST lies in its treatment of reframing as a surgical operation on word "sense" rather than "reference." The authors found that standard generative models often fail to maintain meaning when changing tone (e.g., incorrectly swapping "soft power" for "military strength"). The NLI "safety filter" acts as a critical guardrail, significantly improving the reliability of the output. Furthermore, the model demonstrates context-awareness, choosing different reframings for the same phrase depending on the surrounding sentence structure.
Experimental Results
ENTRUST demonstrates state-of-the-art performance in both automated and human evaluations. It achieved a semantic similarity score of 92.9 (SBERT), nearly matching human-level performance (93.9).
In human trials, ENTRUST was rated as significantly more trustworthy than original partisan inputs. Notably, in "Appeal to Fear" tasks, the system reduced fear ratings from 3.36 to 2.39 on a 5-point scale, actually outperforming human experts in its ability to neutralize fearful language while maintaining the original argument's meaning.
Conclusion and Future Directions
While ENTRUST provides a robust method for reducing polarization in text, future work aims to automate the detection of biased spans and incorporate "stance" to handle more complex rhetorical devices like concessions. The authors also emphasize the ethical importance of ensuring such technology is used to foster constructive dialogue rather than for subtle manipulation.
Summary updated