Abstract
Sentence-level sentiment analysis (SLSA) aims to identify the overall sentiment polarity conveyed in a given sentence. While deep learning models currently define the state-of-the-art, their reliance on the independent and identically distributed (i.i.d.) assumption often leads to performance degradation in real-world scenarios where training and target data distributions differ. This paper proposes a supervised solution based on the non-i.i.d. paradigm of gradual machine learning (GML). The approach begins with labeled observations and gradually labels target instances in order of increasing difficulty through iterative knowledge conveyance. By leveraging deep neural networks (DNNs) for feature extraction and constructing a factor graph to model relations, the proposed GML framework outperforms pure DNN solutions across multiple benchmarks.
Introduction
Most modern sentiment analysis models rely on the i.i.d. assumption—the premise that training data and real-world target data share the same distribution. However, language is inherently messy and context-dependent. A model trained on movie reviews may struggle with the specific nuances of product feedback or social media posts. Even sophisticated models like BERT can fail on complex sentence structures where local cues (like the word "not") mislead the global sentiment prediction. The authors argue that pure DNN solutions are limited by this distribution misalignment and a lack of explicit "knowledge conveyance" between labeled and unlabeled data.
The SLSA-GML Framework
To address these limitations, the authors introduce a supervised GML solution. Unlike standard classifiers that perform a single-pass prediction, GML operates iteratively, moving from "easy" to "hard" instances.
- Feature Extraction via DNNs: The framework uses two specialized DNN architectures to identify relationships between sentences. First, an Affective-aware Polarity Classifier (based on the Entailment as Few-Shot Learner model) creates a polarity-sensitive embedding space. This allows for the identification of K-Nearest Neighbor (KNN) relations. Second, a Semantic Deep Network (built on RoBERTa-Large) is trained to detect implicit sentiment relations (similar or opposite) between arbitrary sentence pairs.
- Factor Graph Construction: These relations are modeled as binary factors in a factor graph. KNN factors capture local similarity, while semantic factors capture global or implicit logic.
- Gradual Inference: The system identifies unlabeled instances with the highest "evidential certainty" (lowest entropy), labels them, and uses these new labels as evidence for subsequent iterations. This "divide and conquer" strategy allows the model to adapt to the target data distribution on the fly.
Experimental Results
The researchers evaluated their approach on four major sentiment benchmarks: MR (movie reviews), CR (customer reviews), SST (Stanford Sentiment Treebank), and Twitter2013. The GML approach consistently achieved state-of-the-art (SOTA) performance, outperforming models like EFL, SentiLARE, and RoBERTa-Large.
Key accuracy gains over the previous best performer (EFL) included:
- CR: +1.68%
- MR: +0.89%
- SST: +1.76%
- Twitter2013: +1.58%
Ablation studies confirmed that the synergy between KNN relations and semantic relations is vital. While KNN captures straightforward similarities, the semantic network is crucial for identifying opposite polarities in complex reasoning tasks.
Conclusion
This work demonstrates that while DNNs are excellent feature extractors, their performance can be significantly enhanced when integrated into a non-i.i.d. GML framework. By acting as a bridge between labeled training data and unlabeled target data, GML effectively mitigates distribution misalignment. Future work will explore extending this framework to document-level analysis and multi-label classification, as well as testing its efficacy in weakly supervised settings with minimal labeled data.
Summary updated