Abstract
Aspect-Based Sentiment Analysis (ABSA) is a granular sentiment analysis task that identifies sentiment towards specific features of an entity. While English has abundant labeled data for ABSA, many languages like Hindi and French suffer from significant data sparsity. This paper proposes a method to leverage resource-rich English data to improve ABSA in these low-resource languages. By utilizing bilingual word embeddings, an Out-of-Vocabulary (OOV) translation strategy, and a dual-LSTM architecture (A3) that separately processes word embeddings and hand-crafted features, the authors demonstrate significant performance improvements in both multi-lingual and cross-lingual settings.
Introduction
Sentiment analysis has evolved from simple document-level classification to Aspect-Based Sentiment Analysis (ABSA), which identifies specific features—such as the "battery life" of a phone or the "service" at a restaurant—and assigns sentiment to each. However, a major hurdle in applying ABSA to languages like Hindi and French is data sparsity. Supervised machine learning thrives on massive datasets, but high-quality labeled corpora for niche tasks like ABSA are often unavailable for these languages.
Furthermore, models frequently encounter Out-of-Vocabulary (OOV) words—terms seen during testing that were absent from the training set. In resource-poor scenarios, these OOV words act as "blind spots." This research explores using English as a bridge to fill these gaps, leveraging its status as a resource-rich language to boost performance in Hindi and French.
Methodology
The proposed pipeline combines deep learning with traditional linguistic features through several key components:
- Bilingual Word Embeddings: Skip-gram models were trained on parallel corpora (English-Hindi and English-French) to ensure that semantically similar words in both languages occupy the same neighborhood in a shared vector space.
- OOV Handling Strategy: If a word in the target language is missing from the embedding model, the system translates it to English and utilizes the corresponding English word's vector. This "safety net" preserves semantic context even when specific vocabulary is missing.
- Hand-crafted Features: Sentiment scores are extracted from established English resources like SentiWordNet, Bing Liu’s lexicon, and MPQA by translating target language words into English.
- Neural Architectures: The researchers evaluated three neural network variants. The most effective, Architecture A3, utilizes two separate Bidirectional LSTMs to process word embeddings and hand-crafted features independently before merging them in the final fully-connected layers.
The Insight of Feature Overshadowing
A critical finding was the superiority of Architecture A3. Word embeddings typically have high dimensionality (e.g., 300), while hand-crafted features are low-dimensional. In a simple concatenation (Architecture A1), the high-dimensional embeddings can "overshadow" the features, preventing the network from effectively learning from the lexicons. Separate LSTMs allow the model to extract meaningful patterns from both sources independently.
Experimental Setup
The study utilized several key benchmarks:
- Hindi: A dataset containing 5,417 review sentences and 4,509 aspect terms.
- French: The SemEval-2016 shared task dataset for restaurant reviews (3,482 aspect terms).
- English (Cross-lingual Training): The SemEval-2014 ABSA dataset covering laptop and restaurant domains.
- Parallel Corpora: 7.2 million sentences from Amazon product reviews were used to train the bilingual embeddings.
Results and Discussion
The proposed approach demonstrated dramatic improvements over monolingual baselines across all scenarios. The "OOV strategy" alone accounted for a performance increase of over 2 percentage points in several tests.
As shown in the results tree, the A3 architecture combined with bilingual embeddings and OOV handling achieved the highest accuracy:
- Hindi Multi-lingual: Reached 76.29% accuracy, up from a 63.64% baseline.
- Hindi Cross-lingual: Improved from a baseline of 16.29% to 60.39%.
- French Multi-lingual: Achieved 71.72% accuracy.
- French Cross-lingual: Achieved 69.49% accuracy.
These results suggest that while deep learning is powerful, the qualitative information found in traditional sentiment lexicons remains a vital "driving force" for accurate sentiment prediction, especially when processed through an architecture that respects the different scales of the input features.
Conclusion
This research successfully demonstrates that cross-lingual and multi-lingual strategies can effectively mitigate data sparsity in ABSA. By bridging low-resource languages with English through shared vector spaces and specialized neural architectures, models can achieve state-of-the-art performance. Future work may explore these methods for related tasks such as aspect term extraction and implicit sentiment analysis, where sentiment is implied by context rather than explicit polar words.
Summary updated