Semi-supervised Sentiment Analysis
Semi-supervised Sentiment Analysis (Label Propagation and Self-Training for Opinion Mining) · Also known as: SSSA, semi-supervised opinion mining, label-propagation sentiment classification, self-training sentiment analysis
Semi-supervised sentiment analysis combines a small set of manually labeled text samples with a large pool of unlabeled text to train opinion classifiers. By propagating sentiment signals from labeled seeds to unlabeled data through self-training, label propagation, or consistency regularization, the approach achieves competitive accuracy without the cost of labeling large corpora.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use semi-supervised sentiment analysis when you have abundant unlabeled text but annotation is expensive or slow — common in domain-specific corpora (medical notes, legal documents, low-resource languages). It is the right choice when fully supervised fine-tuning is infeasible due to label scarcity. Avoid it when labeled data is already plentiful (several thousand examples), as full supervision will outperform. Also avoid it when unlabeled text differs substantially in domain or style from the labeled seed, since confident pseudo-labels may encode systematic domain shift rather than true sentiment signal, causing error propagation.
Strengths & limitations
- Dramatically reduces annotation cost by leveraging large pools of free unlabeled text.
- Compatible with any base classifier, from logistic regression to large pre-trained transformers.
- Self-training and label propagation are well-understood and easy to implement with standard libraries.
- Can be combined with data augmentation or consistency regularization for further gains.
- Particularly effective when a pre-trained language model provides strong initial representations.
- Error propagation: confident but wrong pseudo-labels reinforce themselves across iterations, degrading accuracy.
- Convergence is sensitive to the confidence threshold; too high leaves most unlabeled data unused, too low introduces noise.
- Performance heavily depends on domain match between labeled seeds and unlabeled pool.
- Evaluation is complicated — without a large labeled test set it is hard to distinguish genuine improvement from overfitting to pseudo-labels.
Frequently asked
How large does the labeled seed set need to be?
With strong pre-trained encoders like BERT, as few as 50–200 labeled examples per class can bootstrap effective semi-supervised training. Without pre-trained representations, at least 500–1000 labeled examples per class are advisable to prevent runaway error propagation.
What is the difference between self-training and label propagation?
Self-training iteratively re-trains the model by adding high-confidence predictions on unlabeled data as pseudo-labels. Label propagation instead builds a graph of example similarity and spreads label information along graph edges without iterative re-training, making it less susceptible to error accumulation but more memory-intensive.
Does semi-supervised sentiment analysis work for domain-specific text?
Yes, but the labeled seed and unlabeled pool should share the same domain. A model seeded with movie-review labels and applied to clinical notes will propagate domain-shifted pseudo-labels and perform poorly.
How do I evaluate the method fairly?
Hold out a separate labeled test set that is never used for pseudo-label generation or threshold tuning. Report F1 per class and macro-F1 to surface class-imbalance effects, and compare against a fully supervised baseline trained on the same seed set.
Can I combine this with active learning?
Yes. A common hybrid queries a human annotator for the most uncertain unlabeled examples rather than assigning pseudo-labels, then uses the newly labeled examples to re-train. This combines the cost savings of semi-supervision with the label quality of active annotation.
Sources
- Zhu, X. (2005). Semi-Supervised Learning Literature Survey. Technical Report 1530, Computer Sciences, University of Wisconsin-Madison. link ↗
- Pang, B., & Lee, L. (2008). Opinion Mining and Sentiment Analysis. Foundations and Trends in Information Retrieval, 2(1–2), 1–135. DOI: 10.1561/1500000011 ↗
How to cite this page
ScholarGate. (2026, June 3). Semi-supervised Sentiment Analysis (Label Propagation and Self-Training for Opinion Mining). ScholarGate. https://scholargate.app/en/deep-learning/semi-supervised-sentiment-analysis
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- BERT-based ClassificationDeep learning↔ compare
- LDA Topic ModelDeep learning↔ compare
- Self-supervised Sentiment AnalysisDeep learning↔ compare
- Semi-supervised BERT-based ClassificationDeep learning↔ compare