Introduction
Sentiment transfer, a specific application of style transfer, aims to alter the sentiment of a sentence while preserving its core semantic content. A significant hurdle in this domain is the scarcity of parallel corpora—pairs of sentences identical in meaning but differing in sentiment. Historically, researchers have circumvented this by employing complex unsupervised architectures like Generative Adversarial Networks (GANs). However, this study investigates whether mining "pseudo-parallel" data from existing disjoint sets can enable the use of superior supervised Seq2Seq models.
Methodology: Mining Parallel Data
The researchers developed a pipeline to synthesize a parallel corpus from non-parallel positive and negative review sets ($X^1$ and $X^2$):
- Semantic Embedding: Sentences are mapped into high-dimensional space using Universal Sentence Encoders (USE).
- Similarity Matching: For each sentence in one set, the most semantically similar sentence in the opposite set is identified using Euclidean distance.
- Training: A standard Transformer Seq2Seq model is trained on these pairs. To handle bidirectional transfer, special tokens (e.g.,
<from1to2>) are prepended to the input.
Dataset Characteristics and Difficulty
The approach was tested on the Yelp and Amazon review datasets. A key finding was the variance in "matchability" between datasets, which directly impacts the quality of the synthetic parallel data.
As shown in the CDF above, the Amazon dataset exhibits lower overall similarity scores compared to Yelp. This suggests that Amazon reviews are more context-dependent, making it harder to find truly parallel pairs through simple semantic matching. This data distribution explains why the Amazon dataset proved significantly more challenging for the model.
Key Results and Performance
The proposed Seq2Seq_all model was compared against state-of-the-art unsupervised baselines, including CrossAligned and DeleteAndRetrieve, across three metrics: Style Transfer Accuracy (STAcc), BLEU (content preservation), and Perplexity (fluency).
Fluency and Content Preservation
The Seq2Seq approach demonstrated a massive leap in fluency. On the Yelp dataset, it achieved a perplexity of 15.1, drastically lower than the 68.7 achieved by the best unsupervised baseline. This indicates that training on a natural language distribution, even if the pairs are noisy, prevents the "Frankenstein" sentences common in unsupervised word-swapping methods.
Comparative Ranking
Average ranking positions for Yelp and Amazon across Polarity, Meaning, and Grammar. Source, p. 8
The boxplots illustrate that the Seq2Seq model (green) consistently ranks highly across all evaluation pillars. It is particularly dominant in Grammar (fluency) and remains highly competitive in Polarity (accuracy) and Meaning (content preservation). The model's robustness to noise is evident, as using all matched pairs often outperformed filtered sets.
Conclusion
This research demonstrates that modern semantic similarity metrics are sufficiently powerful to create effective synthetic parallel data. By shifting the problem from unsupervised learning to a "pseudo-supervised" Seq2Seq task, the researchers achieved superior fluency and balanced performance, suggesting that data volume and model architecture can often compensate for noise in training pairs.
Summary updated