Introduction
Effective communication relies heavily on "pragmatic intention"—the subtle social cues that convey politeness and respect. However, this intention is often distorted during transmission. This paper addresses two primary hurdles in digital communication:
- Channel Distortion: Automated tools like Machine Translation (MT) frequently strip away politeness markers, potentially turning a polite request into a blunt command.
- Receiver Variance: Politeness is perceived differently across cultures and backgrounds. A message deemed polite by one speaker may be perceived as neutral or rude by another.
While traditional AI "style transfer" treats politeness as a binary state, this research introduces a "fine-grained" paraphrasing method that adjusts messages to match specific contexts and intended perceptions.
Methodology: The Plan-and-Realize Framework
The researchers propose a two-step "Plan-and-Realize" framework to ensure intended politeness is preserved across different communication circumstances.
Fine-Grained Planning
The system treats 18 distinct politeness strategies (such as Subjunctive, Gratitude, and Hedges) as modular components. Using Integer Linear Programming (ILP), the model identifies the optimal combination of these strategies to achieve a target politeness score for a specific receiver or channel. This allows the system to substitute "at-risk" markers—those likely to be lost in translation—with more robust alternatives that carry equivalent pragmatic weight.
Fine-Grained Realization
To generate fluent paraphrases, the authors adapted the Delete-Retrieve-Generate (DRG) paradigm:
- Delete: Markers of strategies slated for replacement are removed from the original text.
- Plan: The ILP output dictates which new markers to incorporate based on the target circumstance.
- Generate: A sequence-to-sequence model (G-GST) is trained on the WikiConv dataset to blend these new markers into the existing context naturally.
Experimental Results
The framework was evaluated using the Stanford Politeness Corpus and WikiConv data. Perception models were trained to estimate the "politeness credit" of various strategies for different senders and receivers.
The ILP-based method demonstrated superior performance over baseline "Greedy" and "Retrieval" approaches in two key scenarios:
- Scenario A (Translation): The model reduced the "perception gap" (the difference between intended and perceived politeness) from 0.43 to 0.21.
- Scenario B (Misaligned Perceptions): In cases of differing cultural perceptions, the model slashed the gap from 1.01 to 0.12.
Human evaluators rated the generated paraphrases with an average naturalness score of 4.2 out of 5, indicating that the strategic adjustments did not compromise linguistic quality.
Discussion and Limitations
A core insight of this work is that pragmatic strategies function like "LEGO bricks"—multiple combinations can achieve the same communicative goal. This sensitivity allows for "compensatory" paraphrasing: if a specific word like "please" is likely to be dropped by a translator, the model might add a "thanks" or a "could you" to maintain the overall tone.
However, challenges remain. Strong negative markers, such as "Swearing," are difficult to offset with positive strategies. Furthermore, the current model assumes an additive relationship between strategies, whereas real-world language often involves complex, non-linear interactions. Future work may explore more diverse stylistic shifts and the integration of these tools as human-in-the-loop writing assistants.
Summary updated