Applying UDA Semi‑Supervised Learning to Financial Text Classification: Experiments and Insights

The article introduces the challenge of scarce labeled data in vertical domains such as finance and proposes semi‑supervised learning, specifically the UDA (Unsupervised Data Augmentation) algorithm, as a promising solution.

It outlines three main contributions: (1) evaluating UDA on a financial text classification case study, (2) testing UDA on lightweight models, and (3) extending the original UDA study with experiments on out‑of‑domain data and mislabeled samples.

Background sections explain why financial NLP tasks suffer from limited annotations and high labeling costs, motivating the use of a small amount of expert‑labeled data combined with large amounts of unlabeled data.

The UDA technique is described: a supervised cross‑entropy loss on labeled data plus an unsupervised consistency loss that enforces similar predictions for augmented versions of the same unlabeled example, using KL divergence. The consistency assumption and the algorithm’s model‑agnostic nature are highlighted.

Experiments on public benchmarks (IMDb, Yelp, Amazon, DBpedia) show that UDA can approach or surpass SOTA performance with very few labeled examples, though gains diminish as more labeled data become available.

Real‑world experiments at Entropy‑Simple Technology focus on a six‑class financial short‑text classification task. Two models—BERT_base (pre‑trained on 10M finance documents) and TextCNN—are trained with and without UDA under varying amounts of labeled data, data‑augmentation strategies (EDA, back‑translation, TF‑IDF word replacement), and noisy‑label conditions.

Key findings include: UDA effectively leverages unlabeled data to achieve performance comparable to doubling the labeled set; the benefit is larger with fewer labels; lightweight TextCNN benefits almost as much as BERT; data‑augmentation choice strongly influences results, with EDA outperforming other methods; and UDA mitigates the impact of moderate label noise.

The article concludes that while UDA adds valuable signal from unlabeled data—especially in low‑resource settings—it does not fully replace the need for more labeled data or complementary techniques such as transfer learning, and that improving augmentation methods remains an important research direction.