Semi-supervised Formality Style Transfer using Language Model Discriminator and Mutual Information Maximization

Kunal Chawla, Diyi Yang

Abstract: Kunal Chawla, Diyi Yang. Findings of the Association for Computational Linguistics: EMNLP 2020.

2020.

Introduction

Formality Style Transfer (FST) involves converting informal text into a grammatically correct, formal equivalent while preserving the original meaning. For example, transforming "I flippin' LOVE that movie!" into "I truly enjoy that movie." Despite its utility, FST faces two significant challenges:

  1. Data Scarcity: Creating high-quality parallel datasets is labor-intensive and expensive.
  2. The "Safe Response" Trap: Standard models often produce generic, repetitive outputs that lose the original sentence's nuance.

While Generative Adversarial Networks (GANs) are common in style transfer, they often rely on unstable binary classifiers. This paper introduces a semi-supervised framework that leverages pre-trained language models and mutual information maximization to overcome these hurdles.

Methodology

The proposed model is built upon the BART architecture and incorporates three key components to enhance training stability and output quality:

Overview of the proposed semi-supervised framework featuring LM Discriminator, MMI, and cyclic loss.
  • Language Model (LM) Discriminator: Instead of a binary classifier, the authors use two pre-trained language models (one formal, one informal). This provides a "denser" training signal by offering feedback on the probability of every token belonging to the target style, leading to more fluent generation.
  • Maximum Mutual Information (MMI) Objective: Traditional models maximize $P(y|x)$. By maximizing the mutual information—which also considers $P(x|y)$—the model is forced to generate outputs uniquely tied to the source content, reducing generic responses.
  • Bi-directional Cyclic Training: The model is trained to translate in both directions (Informal $\leftrightarrow$ Formal). This enables the use of unlabeled data via a cyclic reconstruction loss: a sentence translated to the target style and back should ideally match the original input.

Datasets

The model was evaluated on several benchmarks:

  • GYAFC (Grammarly Yahoo Answers Corpus): The primary benchmark, featuring "Entertainment & Music" (E&M) and "Family & Relationships" (F&R) domains.
  • Twitter & BookCorpus: Served as large-scale unlabeled data for informal and formal styles, respectively.
  • Yelp & Amazon Sentiment Transfer: Used to test the model's generalization to sentiment-based style transfer.

Key Results

The proposed model (referred to as Ours*) achieved state-of-the-art results across multiple benchmarks.

Performance comparison on GYAFC datasets (E&M and F&R) across multiple metrics.
  • Automated Metrics: On the GYAFC dataset, the model reached BLEU scores of 76.52 (E&M) and 80.29 (F&R), outperforming previous state-of-the-art models.
  • Human Evaluation: Human annotators rated the outputs as more fluent and formal than baselines, with high content preservation.
  • Generalization: The LM-discriminator approach proved robust, outperforming specialized unsupervised models on sentiment transfer tasks in the Yelp and Amazon datasets.

Discussion and Limitations

While the framework significantly advances FST, certain challenges remain:

  • Sarcasm and Rhetorical Questions: The model often fails to preserve sarcastic tones, a common issue in NLP.
  • Context and Slang: Heavy slang or implied context can occasionally lead to incorrect formal mappings.
  • Named Entities: Ensuring proper nouns (e.g., "50 Cent") are not incorrectly "translated" remains a technical hurdle.

Future research may focus on integrating external knowledge bases to better handle context-heavy and entity-specific nuances.

Summary updated

BibTeX
@inproceedings{chawla2020semi,
  author = {Kunal Chawla and Diyi Yang},
  title = {{Semi-supervised Formality Style Transfer using Language Model Discriminator and Mutual Information Maximization}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  url = {https://openreview.net/forum?id=NysgX90g51},
}
Figure