Abstract
Cross-lingual sentiment classification (CLSC) at the aspect level presents unique challenges compared to document-level analysis, primarily due to the need for precise alignment of opinionated units across languages. This paper proposes a novel approach using constrained Statistical Machine Translation (SMT) to preserve the boundaries of sentiment-bearing phrases during translation. By bypassing traditional, error-prone word-alignment models, the method allows for direct mapping of sentiment labels between source and target languages. Evaluations on the OpeNER hotel review corpus show that the proposed method maintains high accuracy, with minimal performance loss compared to monolingual baselines.
Introduction: The Problem of Aspect-Level Alignment
While document-level sentiment analysis can often rely on simple label transfer (e.g., a 5-star rating remains valid regardless of language), aspect-level analysis identifies specific opinions about distinct features, such as "the room was clean but the service was slow." Standard machine translation (MT) often reorders or splits these phrases, causing the original sentiment labels to become misaligned or "lost." Traditionally, researchers have employed complex word-alignment models to recover these units, but these models are frequently inaccurate.
Methodology: Constrained SMT and Boundary Preservation
The core of the proposed method is the use of constrained Statistical Machine Translation (SMT) to perform language transfer while strictly preserving the boundaries of opinionated units.
Using the Moses SMT toolkit, the author employs <zone> and <wall> tags to force the translator to treat specific text segments as indivisible blocks. While internal reordering within a zone is permitted, the text inside cannot be mixed with the surrounding context. Furthermore, <x> tags are used to pass metadata—such as unique IDs and polarity labels—directly through the translation pipeline.
The paper investigates two primary scenarios for cross-lingual transfer:
- Setting A: Translating the training data into the test language.
- Setting B: Translating the test data into the training language.
Experimental Setup
The study utilizes the OpeNER opinion corpus, specifically the hotel review component, which includes annotations for both Spanish (ES) and English (EN). The training sets consist of approximately 3,600–3,900 opinionated units, while the test sets contain roughly 500 units. To build the translation system, the author leveraged large-scale parallel corpora (Common Crawl, Europarl v7, News Commentary, and UN data) supplemented by domain-specific data from Booking.com and TripAdvisor.
Key Results
The experimental results indicate that the constrained SMT approach is highly effective. The monolingual baseline achieved accuracies between 81% and 83%. In cross-lingual settings, the performance drop was remarkably small:
- In the English-to-Spanish direction, the accuracy loss was only 2.3% to 3.5% relative to the monolingual baseline.
- The best cross-lingual configurations reached up to 77.6% accuracy.
A key finding was that translating the test set (Setting B) generally outperformed translating the training set (Setting A). While errors in test set translation have a more immediate negative impact, this approach scales better as MT quality improves. Additionally, applying a Language Model (LM) filter to clean the MT training data significantly boosted final classification accuracy.
Conclusion and Future Work
The "boundary" solution elegantly bypasses the alignment problem, creating a simple 1-to-1 mapping between source and target opinion units. Future research could explore the impact of internal reordering within preserved boundaries and the application of these constraints to modern Neural Machine Translation (NMT) systems. Furthermore, incorporating more sophisticated features, such as dependency information or opinion shifters, may further enhance the robustness of the classifier against translation artifacts.
Summary updated