Abstract
Aspect-Based Sentiment Analysis (ABSA) is a critical task for fine-grained opinion mining. Traditional approaches treat ABSA as a collection of discriminative classification problems, often ignoring the semantic richness of labels and requiring task-specific architectures. This paper introduces GAS (Generative Aspect-based Sentiment analysis), a unified framework that reformulates ABSA as a sequence-to-sequence generation task. By leveraging the pre-trained T5 model and incorporating natural language label semantics, GAS achieves significant performance improvements across multiple benchmarks, establishing a new state-of-the-art for several complex sentiment analysis tasks.
Introduction
Aspect-Based Sentiment Analysis (ABSA) aims to extract detailed sentiment information from text, identifying aspect terms (e.g., "pizza"), opinion terms (e.g., "delicious"), aspect categories (e.g., "food quality"), and sentiment polarities (e.g., "positive"). Historically, these sub-tasks were addressed using discriminative models that predict class indices. However, this paradigm suffers from two primary limitations:
- Neglect of Label Semantics: Classification models treat labels as arbitrary indices, failing to utilize the linguistic meaning of terms like "service" or "positive."
- Architectural Rigidity: Models designed for specific tasks (e.g., aspect-sentiment pair extraction) often require complete redesigns to handle more complex structures like triplets.
GAS addresses these issues by treating ABSA as a text-to-text problem, allowing a single architecture to generalize across various tasks while leveraging the semantic knowledge embedded in natural language labels.
Methodology
The GAS framework utilizes the T5 (Text-to-Text Transfer Transformer) model to transform input sentences into structured sentiment descriptions. The authors propose two primary modeling paradigms:
- Annotation-style Modeling: The model reconstructs the input sentence while embedding labels within brackets (e.g., The [pizza | food quality | positive] was [delicious | positive].).
- Extraction-style Modeling: The model generates a structured list of sentiment tuples directly (e.g., (pizza, food quality, positive); (server, service, positive).).
Prediction Normalization
Generative models occasionally "hallucinate" or produce minor lexical variations (e.g., pluralizing a term found in the input). To ensure validity, GAS employs Prediction Normalization. This process uses Levenshtein distance to map generated tokens back to the most likely valid candidates from the original input or the predefined label set.
Experimental Results
The framework was evaluated on four major ABSA tasks—Aspect Opinion Pair Extraction (AOPE), Unified ABSA (UABSA), Aspect Sentiment Triplet Extraction (ASTE), and Target Aspect Sentiment Detection (TASD)—using standard SemEval benchmarks (Laptop14, Rest14, Rest15, Rest16).
Performance Gains
GAS demonstrated substantial improvements over previous state-of-the-art (SOTA) discriminative models. Notably, in the ASTE task, GAS achieved an average gain of 7.6 F1 points across the four datasets, as shown in the comparison below.
Analysis of Paradigms
The results indicate that while Annotation-style modeling is effective for simpler tasks, Extraction-style modeling is superior for complex tasks like ASTE and TASD. This is attributed to the fact that Annotation-style forces the model to reconstruct the entire input sentence, introducing unnecessary complexity and potential for error in long sequences.
Conclusion
The GAS framework proves that a unified generative approach can outperform specialized discriminative models in fine-grained sentiment analysis. By treating labels as data and utilizing a flexible Seq2Seq architecture, GAS achieves new SOTA results while maintaining high generality. Future research will focus on improving the handling of implicit aspects and refining normalization techniques for semantically correct but lexically divergent predictions.
Summary updated