Abstract
Aspect-Based Sentiment Analysis (ABSA) is a granular task that identifies specific entities (aspects) and their associated sentiments within a text. While most high-performing systems are heavily tuned for specific languages, this paper proposes a language-agnostic deep learning framework. By employing a dual-stream Bi-LSTM architecture that independently processes word embeddings and universal hand-crafted features, the model achieves state-of-the-art results across six diverse languages, including English, Spanish, German, and Hindi.
Introduction
Standard sentiment analysis often fails to capture the nuances of complex reviews. For instance, in the sentence "The food was great, but the service was slow," a global sentiment score might be neutral, obscuring the positive feedback on food and the negative feedback on service. ABSA addresses this by identifying aspect terms and classifying their respective sentiments.
A major hurdle in ABSA is "language divergence." Most existing models rely on language-specific tools like dependency parsers, making them difficult to port. This research seeks a robust, single architecture capable of performing across different language families without extensive linguistic engineering.
Methodology
The researchers investigated four neural network architectures based on Bidirectional Long Short-Term Memory (Bi-LSTM) units to determine the optimal way to integrate raw word embeddings with hand-crafted features.
The architectures tested were:
- A1: A baseline using only word embeddings.
- A2: Concatenation of embeddings and features at the input layer.
- A3: Merging features with sentence embeddings learned by a Bi-LSTM.
- A4 (Dual-Stream): Two separate Bi-LSTM streams—one for word embeddings and one for features—whose outputs are concatenated before the final layers.
Universal Features
To maintain language agnosticism, the model uses:
- Distributional Thesaurus (DT): Helps identify similar words based on context, improving extraction of unseen terms.
- Sentiment Lexicons: Lists of positive and negative words available for over 130 languages, used for classification.
Experimental Results
The model was evaluated on benchmarks including SemEval-2016 (English, Spanish, French, Dutch), GermEval-2017 (German), and a multi-domain Hindi dataset.
The A4 architecture emerged as the most robust, contributing to state-of-the-art (SOTA) performance in 9 out of 12 language/task pairs. Notable improvements included a +12.5 point F1-score increase for aspect extraction in Hindi and a +12.3 point accuracy boost for sentiment classification in French.
The success of A4 is attributed to its ability to capture the sequence of hand-crafted features (e.g., recognizing that a negation followed by a negative word implies positivity) without the high-dimensional word embeddings "drowning out" the feature signal.
Discussion and Limitations
While the model demonstrates significant cross-lingual robustness, certain challenges remain:
- Multi-token Aspects: The system occasionally fails to group tokens into single units (e.g., identifying "customer" and "service" separately instead of "customer service").
- Linguistic Nuance: Sarcasm and metaphors are still difficult to classify correctly.
- Context Windows: The fixed ±5 word window for sentiment classification can miss relevant clues if they are positioned further away in the sentence.
Conclusion
This work demonstrates that a dual-stream Bi-LSTM architecture can effectively bridge the gap between raw data and hand-crafted linguistic features in a language-agnostic manner. By avoiding complex, language-specific pre-processing, the model provides a scalable and portable solution for global ABSA tasks.
Summary updated