Abstract
Generating high-quality, long-form product reviews is a significant challenge in Natural Language Generation (NLG). While standard models often produce repetitive or aimless text, this paper introduces a "coarse-to-fine" decoding framework that mimics the human writing process. By decomposing generation into aspect planning, syntactic sketching, and word-level realization, the model produces reviews that are more informative, coherent, and fluent than traditional word-level generators.
Introduction
The generation of long-form text is often considered a "boss level" in NLG. Standard Recurrent Neural Networks (RNNs) and Generative Adversarial Networks (GANs) excel at local word prediction but frequently lack global coherence. This leads to reviews that drift between topics or become trapped in repetitive loops. The core issue is that existing models attempt to generate text in a vacuum, whereas human writers follow a hierarchy: deciding what to talk about (aspects), how to structure the sentence (syntax), and finally which specific words to use.
Methodology: Coarse-to-Fine Decoding
The proposed framework decomposes review generation into three distinct stages to ensure topical and structural control.
- Aspect Sequence Generation: Based on the user, product, and rating, an attention-based GRU decoder predicts a high-level "content flow" (e.g., Sound → Service → Price).
- Aspect-Aware Sketch Generation: For each predicted aspect, the model generates a "sketch"—a sentence skeleton containing a mix of words and semantic placeholders like Part-of-Speech (POS) tags (e.g., "this NN sounds RB great").
- Sketch-based Review Generation: A final decoder fills the sketch's blanks with specific words (e.g., "this microphone sounds surprisingly great"), guided by aspect-specific word distributions.
To facilitate this, the authors utilize unsupervised topic modeling (Twitter-LDA) to extract aspects and align them with sentences during training.
Experimental Setup
The model was evaluated on three large-scale benchmarks:
- AMAZON (Electronics): ~681k reviews.
- YELP (Restaurants): ~1.06M reviews.
- RATEBEER (Beer): ~2.48M reviews.
The datasets were pre-processed to focus on reviews under 100 tokens and filtered for active users and products to maintain high data quality.
Results and Discussion
The coarse-to-fine model consistently outperformed strong baselines, including context-aware models (gC2S, Attr2Seq) and GAN-based models (SeqGAN, LeakGAN).
As shown in the results table, the model achieved a Perplexity of 26.55 and a BLEU-1 score of 28.22 on the Amazon dataset, significantly surpassing the state-of-the-art ExpansionNet. Beyond automated metrics, human evaluation confirmed the model's superiority, with fluency scores (3.54/5.0) approaching the human "Gold Standard" (4.01). The model also demonstrated better aspect coverage, generating more substantive content than GAN-based alternatives.
Conclusion
By breaking down the complex task of review generation into hierarchical sub-tasks, the coarse-to-fine framework provides a controllable and effective method for producing high-quality text. While the current "sketches" are relatively simple, future work could integrate dependency parsing and more diverse syntactic templates to further enhance the complexity and variety of generated reviews.
Summary updated