PAPER: IntroductionThe usage of deep learning methods has yielded significant improvements in natural language processing (NLP) tasks. These methods are data-hungry and they are nowadays trained on internet-scale databases to achieve good performance on some of the NLP tasks. However, achieving similar performance on other NLP domains, languages or styles becomes difficult due to relevant data unavailability of similar size and quality. The collection and labelling of such data is difficult because of various reasons such as associated cost, unavailability of expert annotators as well as privacy in some domains.Transfer learning (Weiss et al., 2016), self-supervised learning (Goyal, 2022), few-shot/zero-shot learning (Song et al., 2022), meta-learning (Hospedales et al., 2022) are some of the approaches already proposed to solve the data scarcity problems. These methods reduce the number of samples required for training machine learning models to achieve similar performance. However, in practice, it is difficult to decide not just the size of the data but also which data sam-ples should be chosen. With the rising digitization, a large number of data samples are relatively easily available, especially for NLP tasks. However, the methodology for finding a good quality subset of data, that should be labelled and further used for training machine learning models without affecting the performance, is missing.Active learning (AL) is another well studied area in which iteratively data samples are evaluated and added to the training set (Ning et al., 2022;Prakhya et al., 2017;Liu & Huang, 2019;Ash et al., 2019;Gissin & Shalev-Shwartz, 2019). In reality, it is impractical to use because the AL strategy depends on warm-starting the model with information about the task. However, all of these approaches suffer with a cold start as initial samples are chosen randomly or methods with high uncertainty (Yuan et al., 2020). Also, calculating the valuation of each sample is very expensive as it involves training models with added samples in each iteration. Therefore, these methods are rarely used in practical settings.Additionally, the classification methods work in closed-set settings, i. e., they consider only a fixed number of known labels. Such models can incorrectly classify test samples from unknown classes into one of the known classes with high confidence. For example, a classifier trained with a news classification dataset comprising of news with labels politics, science, and business will always wrongly classify sports news in one of those classes. This problem is addressed with Open Set Recognition (OSR) in which the incomplete knowledge of the world is accepted at the training time and the samples, that do not belong to any known category, are classified as unknown category.In this paper, we propose two sampling techniques to solve the above mentioned problems. First sampling method finds a good quality subset of original samples, termed as support set, to accommodate the labelling budget. The second method finds a set of samples from the unknown category, which is termed as amplified set. We propose a training methodology with the combination of a support set and an amplified set to solve the problem of OSR and low labeling budget without affecting performance of the model. Specifically, our contributions are as follow:• a sample sparsification sampling technique to get support set, without the pre-knowledge of labeling set,f X f X f - → XSparsification Amplification that results in drastic reduction in labeling budget with negligible reduction in accuracy,• a sample amplification sampling technique to get amplified set for training a robust classifier that identifies samples from unknown categories,• an application of both, sample sparsification and sample amplification procedures, for text classification task with SOTA results and• the zero-shot multilingual text classification without necessity of task specific target language data.Outline: Section 2 describes the important background work related to the proposed methodologies. The problem formulation is done in Section 3. In Section 4, the sample sparsification and amplification methodologies are presented while their application to the text classification task is illustrated in Section 5. The experimental results to prove the benefits of the proposed methodologies are summarised in Section 6. Finally, the related literature research is mentioned in Section 7 followed by conclusion and future scope in Section 8.
================================================================================

REVIEW
--------------------------------------------------------------------------------
**Summary Of The Paper**  
The paper introduces two complementary sampling techniques—*sample sparsification* and *sample amplification*—to address data scarcity and open-set recognition (OSR) in NLP tasks. Sample sparsification reduces labeling requirements by selecting a representative subset of data using HDBSCAN clustering and adaptive binning (Algorithm 1), while sample amplification generates out-of-distribution (OOD) samples via Voronoi regions (Algorithm 2) to improve OSR. The methods are applied to text classification using SBERT embeddings and evaluated on benchmarks like AGNews, PubMed-RCT, and MLSum, with claims of state-of-the-art (SOTA) results for zero-shot multilingual transfer.  

---

**Strengths**  
1. **Novel Framework for Dual Challenges**: The integration of sample sparsification (for low-label budgets) and amplification (for OSR) within a unified methodology is a compelling contribution to addressing two critical NLP challenges simultaneously.  
2. **Empirical Validation Across Domains**: Experiments span diverse datasets (news, medical, sentiment, etc.), demonstrating broad applicability. Zero-shot multilingual results (Table 3) highlight potential for cross-lingual tasks.  
3. **Methodological Transparency**: Algorithms (Algorithms 1–2) are clearly described, enabling reproducibility. Use of established tools (HDBSCAN, SBERT) ensures accessibility.  
4. **Comparison with Competitors**: Results (Table 2) suggest advantages over SetFit and ALPS, particularly in label efficiency.  

---

**Weaknesses**  
1. **High-Dimensional Voronoi Limitations (Severe)**:  
   - Algorithm 2 constructs Voronoi regions directly from raw SBERT embeddings (768D), which are computationally infeasible in high dimensions. The paper projects embeddings to 2D via t-SNE (Fig. 3) for Voronoi computation, but this risks distorting geometric relationships, undermining OOD sample validity. No analysis of distortion effects is provided.  

2. **Ambiguous Hyperparameter Choices (Major)**:  
   - Sparsity/increment factors (e.g., 1, 5, 10) are heuristically chosen without theoretical justification (Sec. 4.1). Sensitivity analyses (Appendix A.1–A.2) are vague; e.g., why fix bins at 10 across all experiments? No ablation studies confirm robustness to these choices.  

3. **Overreliance on SVM (Moderate)**:  
   - All experiments use RBF SVM (Sec. 5), despite modern NLP relying on deep learning models (e.g., transformers). This limits generalizability and raises questions about the methodology’s compatibility with complex models.  

4. **Insufficient OSR Validation (Major)**:  
   - Amplified sets (Algorithm 2) are claimed to improve OSR, but no ground-truth OOD splits are used for validation. Instead, artificial OOD samples are created by holding out classes (Table 5), which lacks realism. Comparison to existing OSR methods (e.g., cbsSVM [Fei & Liu, 2016]) is absent.  

5. **Label-Agnostic Claim Contradicted by Results (Major)**:  
   - While Sec. 4.1 claims sparsification requires no pre-knowledge of labels, Table 4 shows performance degrades significantly with smaller label sets (e.g., 0.9867 ± 0.002 vs. 0.9429 ± 0.003 for 240k vs. 2.6k samples). This contradicts the “label-agnostic” premise.  

---

**Questions For The Authors**  
1. **High-Dimensional Voronoi Validity**: How is the t-SNE projection (Fig. 3) validated as preserving sufficient geometric properties of the original SBERT embeddings for Voronoi region construction? What metrics quantify distortion?  
2. **Hyperparameter Robustness**: What ablation studies confirm the sensitivity of results to sparsity/increment factors (e.g., bins=10, increment=2)? Why not vary these systematically?  
3. **Deep Learning Compatibility**: Why is the methodology restricted to SVMs? Can it be extended to transformer-based classifiers (e.g., BERT, RoBERTa)?  
4. **OSR Ground Truth**: How are amplified sets validated as true OOD samples? Are they compared to gold-standard OOD splits in benchmark datasets (e.g., ImageNet-C for vision)?  
5. **Label Set Dependency**: How is the contradiction between the “label-agnostic” claim (Sec. 4.1) and performance degradation in Table 4 resolved? Does sparsification implicitly assume partial label knowledge?  

---

**Limitations Not Addressed By The Authors**  
1. **Scalability Concerns**: The paper does not discuss computational feasibility for large-scale datasets or high-dimensional embeddings.  
2. **Theoretical Foundations**: The binning heuristic (increasing sparsity outward) lacks theoretical grounding.  
3. **Cross-Modality Generalization**: The reliance on SBERT embeddings suggests the methodology is tied to modalities with strong pretrained encoders (text/images), excluding others (audio, video).  
4. **Ethical and Practical Implications**: No discussion of ethical risks (e.g., biases in amplified OOD samples) or practical deployment barriers (e.g., annotation cost savings vs. computational overhead).  

---

**Soundness**: 3/4 (Methods are plausible but lack rigorous validation for critical components like Voronoi regions and hyperparameters.)  
**Contribution**: 3/4 (Novel framework for dual challenges, but overlaps with prior work on clustering-based sampling and OSR.)  
**Confidence**: 4/5 (Well-motivated methods with empirical support, though key limitations remain unresolved.)  
**Rating**: 6/10 (Borderline Accept – promising direction but insufficient validation and theoretical depth to fully convince.)  

**Brief Justification For Rating**:  
While the paper proposes an innovative framework combining sparsification and amplification for NLP, critical flaws undermine its credibility. The high-dimensional Voronoi approach relies on flawed dimensionality reduction, and the lack of rigorous statistical validation (e.g., p-values, ablation studies) weakens empirical claims. Without addressing these gaps, the contribution remains underdeveloped for top-tier acceptance.

================================================================================
