Abstract
Review summarization aims to generate a condensed summary for a review or multiple reviews. Existing review summarization systems mainly generate summary only based on review content and neglect the authors’ attributes (e.g., gender, age, and occupation). In fact, when summarizing a review, users with different attributes usually pay attention to specific aspects and have their own word-using habits or writing styles. Therefore, we propose an Attribute-aware Sequence Network (ASN) to take the aforementioned users’ characteristics into account, which includes three modules: an attribute encoder encodes the attribute preferences over the words; an attribute-aware review encoder adopts an attribute-based selective mechanism to select the important information of a review; and an attribute-aware summary decoder incorporates attribute embedding and attribute-specific word-using habits into word prediction. To validate our model, we collect a new dataset TripAtt, comprising 495,440 attribute-review-summary triplets with three kinds of attribute information: gender, age, and travel status. Extensive experiments show that ASN achieves state-of-the-art performance on review summarization in both auto-metric ROUGE and human evaluation.
Introduction
Review summarization aims to condense long, rambling user feedback into a punchy headline or a short sentence. Traditional models, such as standard Sequence-to-Sequence or Pointer-Generator Networks, focus entirely on the text of the review. However, summarization is inherently subjective. People with different attributes have varying "aspect preferences" and "word-using habits." For example, a business traveler might prioritize a hotel's location and meeting rooms, while a family might focus on the pool and breakfast.
This paper introduces the Attribute-aware Sequence Network (ASN), a model that incorporates user metadata—like age, gender, and travel status—to generate summaries that are more relevant to specific audience segments. By using a new dataset of nearly 500,000 hotel reviews, the authors demonstrate that "who" is writing or reading a review is just as important as "what" the review says.
Methodology: The ASN Framework
The ASN model enhances the standard encoder-decoder architecture with three specialized modules designed to capture and utilize user characteristics:
- Attribute Encoder: This module maps user attributes (e.g., "Business Traveler," "Female") into dense embeddings. It also maintains an attribute-specific vocabulary memory—a collection of the most representative words for each group, calculated via TF-IDF.
- Attribute-aware Review Encoder: Instead of just reading the text, this encoder uses an Attribute Selection Strategy. It applies a sigmoid gate to the Bi-LSTM hidden states, conditioned on the attribute embedding. This allows the model to "filter" the review, focusing on the sentences that matter most to that specific user type.
- Attribute-aware Summary Decoder: When generating the summary, the decoder incorporates attribute information through attribute prediction, memory-based attention, and a "soft copy" mechanism that allows the model to generate words directly from the attribute-specific memory, even if those words were not in the original review.
The TripAtt Dataset and Attribute Analysis
The researchers introduced a new large-scale dataset called TripAtt, scraped from TripAdvisor, containing 495,440 attribute-review-summary triplets. The dataset explicitly links user demographics—Gender (Male/Female), Age (5 categories), and Travel Status (Solo, Couple, Family, Friends, Business)—to their writing.
Analysis of TripAtt confirms that different demographics exhibit distinct word-using habits. For instance, female users frequently use descriptive terms like "lovely" or "beautiful," while business travelers focus on functional aspects like "meetings" and "conference." The study found that Travel Status was the most influential attribute, likely because it is highly domain-dependent for hotel reviews.
Key Results and Insights
ASN outperformed several strong baselines, including the Pointer-Generator Network (PGN) and SEASS (a selective encoding model). Quantitatively, ASN achieved a ROUGE-L score of 19.74, a significant improvement over standard S2SATT (18.95) and PGN (19.13).
Ablation studies confirmed that removing any of the attribute-based strategies led to a performance drop, proving that the combination of selective encoding and memory-based decoding is necessary. In human evaluations, ASN was rated higher for "aspect-level coverage." For instance, if a reference summary mentioned both "location" and "service," ASN was much more likely to include both compared to baseline models that might miss the service aspect.
Conclusion
The ASN model demonstrates that user attributes act as a powerful prior for attention in summarization. By knowing the user type, the model can "pre-calculate" which information is most important. While currently focused on single-review summarization, future work could adapt this to multi-review scenarios or explore latent user representations when explicit metadata is unavailable.
Summary updated