Jul 2026· IEEE Transactions on Neural Networks and Learning Systems· Vol PP, pp. 1-15· 0 citations
Medicine
TL;DR
A novel weakly supervised (WS) learning MLTC framework consisting of a novel category word selection method, namely category word selection with significance ranking and crowd-sourcing (Cws-src), and a generic WS learning MLTC method, namely WS multilabel text classification with correlation-aware label propagation (Wmltc-clp), which estimates accurate pseudolabels by propagating them over a text correlation graph.
Abstract
Multilabel text classification (MLTC) methods require enormous labeled training samples to ensure the model's performance, which involves significant manual labor costs. An alternative to conducting MLTC is to only employ predefined representative words of classes, namely category words, as the weak supervision. In this article, we propose a novel weakly supervised (WS) learning MLTC framework consisting of two parts. First, we propose a novel category word selection method, namely category word selection with significance ranking and crowd-sourcing (Cws-src), which generates confident category words by manually selecting from the topically reranked words using a new TW-ITF weighting scheme, thereby effectively mitigating the noises in pseudolabels by filtering repetitive and less significant terms for each class, leading to improved classification performance. Subsequently, we propose a generic WS learning MLTC method, namely WS multilabel text classification with correlation-aware label propagation (Wmltc-clp), which estimates accurate pseudolabels by propagating them over a text correlation graph. To evaluate the proposed framework, we conduct extensive experiments on nine benchmark datasets, including five sentiment analysis datasets and four prevalent MLTC datasets. The results demonstrate that Cws-src can generate more confident category words and Wmltc-clp can achieve significant improvements over the WS learning baselines. The maximum performance gains of Wmltc-clp over the best WS learning baseline methods reach 0.096, 0.081, 0.075, and 0.02 on Micro- $F1$ , Macro- $F1$ , average precision (AP), and ranking loss (RL) across all benchmark datasets.
Multi-label classification of financial news is frequently affected by incomplete and noisy annotations, while obtaining expert-curated labels at scale is prohibitively expensive. This study proposes a weakly supervised classification framework that combines large language model (LLM) zero-shot annotation with a serial label-correction strategy. The framework first uses an LLM to generate initial weak labels and then refines them through a two-stage Correct→Clean procedure that recovers missing labels via centrality-weighted graph propagation before suppressing label noise. Systematic experiments on a financial subset of Reuters-21578 show that, under an extreme mixed-corruption setting with 80% missing labels and 15% noise labels, Correct→Clean increases the Micro-F1 from 0 to 0.6748. In an end-to-end evaluation, the proposed framework achieves a Micro-F1 of 0.8882 with reduced-dimensional features, recovering 88.69% of the performance gap to fully supervised learning. Additional experiments on the RCV1 Topics and AAPD datasets confirm that the advantage of Correct→Clean is consistently reproduced across domains and dataset sizes. These findings demonstrate that coupling LLM-generated annotations with ordered label correction offers an effective means of addressing the joint effects of missing and noisy labels, providing a promising approach to financial text classification when expert annotations are scarce.
Large language models (LLMs) struggle to classify text into taxonomies with many semantically similar labels, as the distinctions are domain-specific and not captured by pre-training. To handle large label spaces, a common approach retrieves top-$K$ candidate labels by embedding similarity and prompt the LLM to choose among them. However, top-$K$ retrieval reduces the number of candidates but does not help the model tell similar ones apart. When two similar labels both appear as candidates, the model lacks the signal to choose correctly between them. We propose a framework that (1) identifies which label pairs the model struggles to distinguish, (2) expands the candidate set to include confusable labels, and (3) generates targeted rules to differentiate between similar candidates. The framework requires no fine-tuning, and the generated rules transfer to smaller, cheaper models. On three benchmarks (WOS, Flipkart, LEDGAR), our approach improves Macro F1 by up to 10.0pp over retrieval baselines, with smaller models (2B--20B) gaining up to 11.5pp via cross-model transfer.
Large language models (LLMs) can generate synthetic training data for text classification, but the quality of generated samples is heterogeneous: some fall in correct class regions of the embedding space while others land in peripheral or cross-class zones. We propose a geometric filtering framework that evaluates each LLM-generated sample by its Euclidean distance to real class examples in a sentence embedding space, selecting only geometrically consistent candidates. A soft weighting mechanism transforms filter scores into sample weights for classifier training. Evaluated across 13 datasets, 5 classifiers, 10 augmentation methods, and over 6,700 configurations, our method achieves +2.61 percentage points (pp) over SMOTE ($p<0.0001$, Cohen's $d=0.95$, 88.9% win rate). The approach generalizes to named entity recognition (+9.26pp, 100% win rate) without filter modification, and is robust across 5 LLMs from 4 providers. A key finding is that the simplest distance-based filter consistently outperforms complex multi-criteria alternatives.
This work proposes distilling the model into a probabilistic classifier, enabling lightweight deployment without repeated LLM calls, and demonstrates that LSR improves macro-F1 scores by an average of 7.0% compared to standard zero-shot classification baselines.
Nathan Vandemoortele, Bram Steenwinckel, F. Ongenae et al.· Discover Computing· 0 citations
This paper proposes a novel Partial label-based Self-training framework (PaSta) that leverages partial label learning technique to overcome the limitations of existing methods and designs a partial label-based classification model with two well-crafted loss functions to guide the model learning at both label and representation spaces.
Yujing Liu, Yixin Liu, Yu Zheng et al.· 0 citations
Semi-supervised semantic segmentation has long turned on one question, which pseudo-labels to trust, and a generation of selection rules, dynamic thresholds, per-class curricula, soft confidence weights, answered it for the noisy, under-confident ResNet teachers of their day. Self-supervised foundation encoders change the regime: with a DINOv2 teacher, confidence saturates, so the filtering that helped a weak teacher can hurt a strong one. We propose CW-BASS v2, a saturation-aware pseudo-label selection method that reads the teacher's confidence regime rather than committing to one rule. It pairs held-out calibration, an unbiased per-class noise estimate, with a self-adaptive confidence floor that provably bounds retention away from 1, and combines them in a one-pass gate: measure the reliability of the teacher's confident set, pi_kept = Pr[correct | c>= tau], on a held-out slice, and filter strictly when it meets the confidence demanded (pi_kept>= tau), falling back to the adaptive floor otherwise. The boundary is the pre-existing operating threshold, not a value tuned to mIoU, and across six DINOv2 teachers it makes the correct strict-vs-floor call blind. CW-BASS v2 thus recovers the UniMatch V2 operating point on the saturated benchmarks by selecting strict (Pascal VOC 1/8 87.4 against its reported 87.9; Cityscapes within 0.5), and improves on it where the confident set is unreliable (pi_kept ~ 89%, ADE20K), where the floor edges ahead (+1.5 mIoU, single seed). The gate is principled because the failure it avoids is measured, not assumed: on a reliable, saturated teacher the confidence distribution's dynamic range collapses (98% of Pascal pixels>= 0.95), so an adaptive cutoff floods the retention mask and self-training decays into confirmation bias.
Ebenezer Tarubinga· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
