Introduction
Generating long-form, coherent product reviews from structured data is a significant challenge in natural language generation. Unlike short snippets or objective reports, product reviews are subjective and require high internal consistency across multiple sentences. Standard Recurrent Neural Networks (RNNs) often struggle with long-range coherence, leading to repetitive or contradictory text that "loses the thread" of the initial input constraints. Researchers from Peking University have addressed this by developing a hierarchical model designed to generate multi-sentence reviews from a simple list of aspect-sentiment scores.
The HRGM-a Architecture
The core of this research is the Hierarchical Review Generation Model with aligned attention (HRGM-a). This model treats a review as a structured composition rather than a simple sequence of words. It employs a two-layer hierarchical decoder to mimic the top-down planning process humans use when writing:
- Paragraph-level LSTM ($LSTM^P$): This layer manages the high-level flow, tracking which aspect (e.g., fuel economy, interior) is being discussed and passing that context to the next level.
- Sentence-level LSTM ($LSTM^S$): This layer generates the actual words for a specific sentence based on instructions from the paragraph layer.
A key innovation is the aligned attention mechanism. Instead of standard attention, it uses a gate vector to combine the paragraph-level state with a specific aspect-oriented vector. This ensures that the sentence-level decoder remains strictly focused on the intended data point, preventing "aspect drift."
Experimental Setup
The model was trained and evaluated on a newly crawled dataset of 43,060 Chinese car reviews from AutoHome. Each review covers exactly eight aspects: Space, Power, Control, Fuel Consumption, Comfort, Appearance, Interior, and Price. Ratings (1–5) were normalized to a range of [-1.0, 1.0], and text was segmented using the Jieba library. The dataset was split 75/25 for training and testing, ensuring no overlap in input vectors between sets.
Results and Evaluation
The HRGM-a model significantly outperformed baseline models in both quantitative and qualitative evaluations. Quantitatively, it achieved a BLEU-4 score of 0.0412, surpassing the best sequential model (0.0369) and retrieval-based baselines (0.0133).
Qualitatively, three human judges rated the outputs on a 5-point scale for readability, accuracy, and usefulness. As shown in the results table, HRGM-a achieved readability scores nearly on par with human-written "Gold" standards and outperformed other methods in accuracy and usefulness.
The results demonstrate that the hierarchical approach maintains better internal consistency than traditional retrieval methods, which often produce contradictory statements (e.g., describing a car as both "narrow" and "spacious").
Conclusion
By mimicking the top-down planning process of human writing, the HRGM-a model successfully bridges the gap between structured scores and natural, multi-sentence reviews. While challenges in content richness and occasional logic errors remain, this hierarchical framework provides a robust foundation for future work in complex data-to-text generation tasks, with potential applications in other domains beyond product reviews.
Summary updated