Introduction
Traditional sentiment analysis models often struggle with domain adaptation. A classifier trained on Amazon electronics reviews frequently fails when applied to movie critiques or hotel feedback because it relies heavily on domain-specific keywords like "battery life" or "cinematography." This paper introduces sentiment flow—the sequential progression of local sentiments (positive, negative, neutral) within a text—as a robust, domain-independent structural model for web review argumentation.
By abstracting away from specific content and focusing on the rhetorical "flow" of opinions, the researchers demonstrate that humans structure arguments consistently across disparate domains. Whether discussing a faulty hard drive or a disappointing stay, reviewers often follow similar patterns: starting with a compliment, transitioning to specific complaints, and concluding with a summary judgment.
Methodology: Modeling the Flow
The core of the proposed approach is representing a review as a sequence of local sentiments. To ensure this model is generalizable, the authors introduced three abstraction transformations designed to "clean" the raw flow:
- Change: Removes repeating consecutive sentiments (e.g.,
[pos, pos, neg]becomes[pos, neg]), focusing on transitions rather than duration. - NoLoops: Eliminates repeating sequences (e.g.,
[pos, neg, pos, neg]becomes[pos, neg]), merging repetitive points. - 2Class: Removes neutral sentiments to highlight the primary conflict between positive and negative stances.
To compare these flows, the researchers utilized a normalized minimum edit distance (a variant of Levenshtein distance). This metric calculates the similarity between a new review's flow and known patterns in training data, regardless of length. A Random Forest classifier then maps these structural patterns to a global sentiment rating.
Experimental Setup
The model was evaluated using three distinct English-language corpora:
- Product Domain: Amazon reviews for various consumer goods.
- Hotel Domain: TripAdvisor reviews from the ArguAna corpus.
- Movie Domain: Rotten Tomatoes reviews from the Cornell Movie Review Data.
These datasets present significant challenges for cross-domain transfer. Movie reviews, for instance, are typically longer and contain more neutral "plot description" than the concise, highly subjective reviews found in the hotel domain.
Key Insights and Results
The study yielded several critical findings regarding the nature of online argumentation:
- Structural Consistency: While raw sentiment flows rarely repeat across domains, the abstracted flows (Change and NoLoops) revealed universal structural patterns common to all review types.
- Neutrality as Noise: In many contexts, neutral sentences represent objective descriptions rather than argumentative moves. Removing them (2Class) often clarified the underlying structure and improved predictive power.
- Domain Robustness: While traditional Bag-of-Words (BoW) models excel in-domain, they falter during cross-domain transitions. In contrast, sentiment flow features maintained steady performance. For example, when training on Hotels and testing on Movies, the Sentiment Flow model achieved 60.9% accuracy, vastly outperforming the BoW baseline of 39.8%.
- Generality: In the movie domain, abstraction increased the "aggregate recall" of known patterns from 0% to nearly 100%, indicating that the model successfully captured the universal "language" of review structure.
Conclusion and Future Directions
Sentiment flow offers a powerful lens for understanding the structural skeleton of web reviews. While it currently ignores sentiment intensity and specific aspects (e.g., price vs. quality), its strength lies in domain independence.
Future work suggests using sentiment flow as a "pivot feature" to align domain-specific models. Additionally, exploring weighted flows—where certain parts of a review, such as the conclusion, carry more weight—could further refine the model's accuracy without sacrificing its generalizability.
Summary updated