Few-shot Setting Research Articles

Knowledge-injected Prompt Learning for Chinese Biomedical Entity Normalization

The Biomedical Entity Normalization (BEN) task aims to align raw, unstructured medical entities to standard entities, thus promoting data coherence and facilitating better downstream medical applications. Recently, prompt learning methods have shown promising results in the natural language processing field. However, existing research falls short in tackling the more complex Chinese BEN task, especially in the few-shot scenario with limited medical data, and the vast potential of the external medical knowledge base has not yet been fully exploited. To address these challenges, this article proposes a novel Knowledge-injected Prompt Learning (PL-Knowledge) method. Specifically, the approach consists of five stages: candidate entity matching, knowledge extraction, knowledge encoding, knowledge injection, and prediction output. By effectively encoding the knowledge items contained in medical entities and incorporating them into tailor-made knowledge-injected templates, the additional knowledge enhances the model’s ability to capture latent relationships between medical entities, thus achieving a better match with the standard entities. Comprehensive experiments are conducted on a benchmark dataset in both few-shot and full-scale settings. This method outperforms existing baselines, with an average accuracy improvement of 12.96 percentage points in few-shot and 0.94 percentage points in full-data cases, showcasing its excellence in the BEN task.

CHA: Conditional Hyper-Adapter method for detecting human–object interaction

Human–object interactions (HOI) detection aims at capturing human–object pairs in images and predicting their actions. It is an essential step for many visual reasoning tasks, such as VQA, image retrieval and surveillance event detection. The challenge of this task is to tackle the compositional learning problem, especially in a few-shot setting. A straightforward approach is designing a group of dedicated models for each specific pair. However, the maintenance of these independent models is unrealistic due to combinatorial explosion. To address the above problems, we propose a new Conditional Hyper-Adapter (CHA) method based on meta-learning. Different from previous works, our approach regards each <verb, object> as an independent sub-task. Meanwhile, we design two kinds of Hyper-Adapter structures to guide the model to learn “how to address the HOI detection”. By combining the different conditions and hypernetwork, the CHA can adaptively generate partial parameters and improve the representation and generalization ability of the model. Finally, our proposed method can be viewed as a plug-and-play module to boost existing HOI detection models on the widely used HOI benchmarks.

Learning to explain is a good biomedical few-shot learner.

Significant progress has been achieved in biomedical text mining using deep learning methods, which rely heavily on large amounts of high-quality data annotated by human experts. However, the reality is that obtaining high-quality annotated data is extremely challenging due to data scarcity (e.g. rare or new diseases), data privacy and security concerns, and the high cost of data annotation. Additionally, nearly all researches focus on predicting labels without providing corresponding explanations. Therefore, in this paper, we investigate a more realistic scenario, biomedical few-shot learning, and explore the impact of interpretability on biomedical few-shot learning. We present LetEx-Learning to explain-a novel multi-task generative approach that leverages reasoning explanations from large language models (LLMs) to enhance the inductive reasoning ability of few-shot learning. Our approach includes (1) collecting high-quality explanations by devising a suite of complete workflow based on LLMs through CoT prompting and self-training strategies, (2) converting various biomedical NLP tasks into a text-to-text generation task in a unified manner, where collected explanations serve as additional supervision between text-label pairs by multi-task training. Experiments are conducted on three few-shot settings across six biomedical benchmark datasets. The results show that learning to explain improves the performances of diverse biomedical NLP tasks in low-resource scenario, outperforming strong baseline models significantly by up to 6.41%. Notably, the proposed method makes the 220M LetEx perform superior reasoning explanation ability against LLMs. Our source code and data are available at https://github.com/cpmss521/LetEx.

Context-Based Meta-Reinforcement Learning With Bayesian Nonparametric Models.

Deep reinforcement learning agents usually need to collect a large number of interactions to solve a single task. In contrast, meta-reinforcement learning (meta-RL) aims to quickly adapt to new tasks using a small amount of experience by leveraging the knowledge from training on a set of similar tasks. State-of-the-art context-based meta-RL algorithms use the context to encode the task information and train a policy conditioned on the inferred latent task encoding. However, most recent works are limited to parametric tasks, where a handful of variables control the full variation in the task distribution, and also failed to work in non-stationary environments due to the few-shot adaptation setting. To address those limitations, we propose MEta-reinforcement Learning with Task Self-discovery (MELTS), which adaptively learns qualitatively different nonparametric tasks and adapts to new tasks in a zero-shot manner. We introduce a novel deep clustering framework (DPMM-VAE) based on an infinite mixture of Gaussians, which combines the Dirichlet process mixture model (DPMM) and the variational autoencoder (VAE), to simultaneously learn task representations and cluster the tasks in a self-adaptive way. Integrating DPMM-VAE into MELTS enables it to adaptively discover the multi-modal structure of the nonparametric task distribution, which previous methods using isotropic Gaussian random variables cannot model. In addition, we propose a zero-shot adaptation mechanism and a recurrence-based context encoding strategy to improve the data efficiency and make our algorithm applicable in non-stationary environments. On various continuous control tasks with both parametric and nonparametric variations, our algorithm produces a more structured and self-adaptive task latent space and also achieves superior sample efficiency and asymptotic performance compared with state-of-the-art meta-RL algorithms.

LAWSUIT: a LArge expert-Written SUmmarization dataset of ITalian constitutional court verdicts

AbstractLarge-scale public datasets are vital for driving the progress of abstractive summarization, especially in law, where documents have highly specialized jargon. However, the available resources are English-centered, limiting research advancements in other languages. This paper introduces LAWSUIT, a collection of 14K Italian legal verdicts with expert-authored abstractive maxims drawn from the Constitutional Court of the Italian Republic. LAWSUIT presents an arduous task with lengthy source texts and evenly distributed salient content. We offer extensive experiments with sequence-to-sequence and segmentation-based approaches, revealing that the latter achieve better results in full and few-shot settings. We openly release LAWSUIT to foster the development and automation of real-world legal applications.

Extraction of Substance Use Information From Clinical Notes: Generative Pretrained Transformer-Based Investigation.

Understanding the multifaceted nature of health outcomes requires a comprehensive examination of the social, economic, and environmental determinants that shape individual well-being. Among these determinants, behavioral factors play a crucial role, particularly the consumption patterns of psychoactive substances, which have important implications on public health. The Global Burden of Disease Study shows a growing impact in disability-adjusted life years due to substance use. The successful identification of patients' substance use information equips clinical care teams to address substance-related issues more effectively, enabling targeted support and ultimately improving patient outcomes. Traditional natural language processing methods face limitations in accurately parsing diverse clinical language associated with substance use. Large language models offer promise in overcoming these challenges by adapting to diverse language patterns. This study investigates the application of the generative pretrained transformer (GPT) model in specific GPT-3.5 for extracting tobacco, alcohol, and substance use information from patient discharge summaries in zero-shot and few-shot learning settings. This study contributes to the evolving landscape of health care informatics by showcasing the potential of advanced language models in extracting nuanced information critical for enhancing patient care. The main data source for analysis in this paper is Medical Information Mart for Intensive Care III data set. Among all notes in this data set, we focused on discharge summaries. Prompt engineering was undertaken, involving an iterative exploration of diverse prompts. Leveraging carefully curated examples and refined prompts, we investigate the model's proficiency through zero-shot as well as few-shot prompting strategies. Results show GPT's varying effectiveness in identifying mentions of tobacco, alcohol, and substance use across learning scenarios. Zero-shot learning showed high accuracy in identifying substance use, whereas few-shot learning reduced accuracy but improved in identifying substance use status, enhancing recall and F1-score at the expense of lower precision. Excellence of zero-shot learning in precisely extracting text span mentioning substance use demonstrates its effectiveness in situations in which comprehensive recall is important. Conversely, few-shot learning offers advantages when accurately determining the status of substance use is the primary focus, even if it involves a trade-off in precision. The results contribute to enhancement of early detection and intervention strategies, tailor treatment plans with greater precision, and ultimately, contribute to a holistic understanding of patient health profiles. By integrating these artificial intelligence-driven methods into electronic health record systems, clinicians can gain immediate, comprehensive insights into substance use that results in shaping interventions that are not only timely but also more personalized and effective.

Relation Extraction in Underexplored Biomedical Domains: A Diversity-optimised Sampling and Synthetic Data Generation Approach

Abstract The sparsity of labeled data is an obstacle to the development of Relation Extraction (RE) models and the completion of databases in various biomedical areas. While being of high interest in drug-discovery, the literature on natural products, reporting the identification of potential bioactive compounds from organisms, is a concrete example of such an overlooked topic. To mark the start of this new task, we created the first curated evaluation dataset and extracted literature items from the LOTUS database to build training sets. To this end, we developed a new sampler, inspired by diversity metrics in ecology, named Greedy Maximum Entropy sampler (https://github.com/idiap/gme-sampler). The strategic optimization of both balance and diversity of the selected items in the evaluation set is important given the resource-intensive nature of manual curation. After quantifying the noise in the training set, in the form of discrepancies between the text of input abstracts and the expected output labels, we explored different strategies accordingly. Framing the task as an end-to-end Relation Extraction, we evaluated the performance of standard fine-tuning (BioGPT, GPT-2, and Seq2rel) and few-shot learning with open Large Language Models (LLMs) (LLaMA 7B-65B). In addition to their evaluation in few-shot settings, we explore the potential of open LLMs as synthetic data generators and propose a new workflow for this purpose. All evaluated models exhibited substantial improvements when fine-tuned on synthetic abstracts rather than the original noisy data.We provide our best performing (F1-score = 59:0) BioGPT-Large model for end-to-end RE of natural products relationships along with all the training and evaluation datasets. See more details at https://github.com/idiap/abroad-re.

Few-Shot Steel Defect Detection Based on a Fine-Tuned Network with Serial Multi-Scale Attention

Detecting defects on a steel surface is crucial for the quality enhancement of steel, but its effectiveness is impeded by the limited number of high-quality samples, diverse defect types, and the presence of interference factors such as dirt spots. Therefore, this article proposes a fine-tuned deep learning approach to overcome these obstacles in unstructured few-shot settings. Initially, to address steel surface defect complexities, we integrated a serial multi-scale attention mechanism, concatenating attention and spatial modules, to generate feature maps that contain both channel information and spatial information. Further, a pseudo-label semi-supervised learning algorithm (SSL) based on a variant of the locally linear embedding (LLE) algorithm was proposed, enhancing the generalization capability of the model through information from unlabeled data. Afterwards, the refined model was merged into a fine-tuned few-shot object detection network, which applied extensive base class samples for initial training and sparsed new class samples for fine-tuning. Finally, specialized datasets considering defect diversity and pixel scales were constructed and tested. Compared with conventional methods, our approach improved accuracy by 5.93% in 7-shot detection tasks, markedly reducing manual workload and signifying a leap forward for practical applications in steel defect detection.

Incorporating target-aware knowledge into prompt-tuning for few-shot stance detection

Stance detection, a fundamental task in natural language processing, identifies user stances in texts towards specific targets. The diverse targets and ever-changing expressions make it challenging to attain comprehensive knowledge from limited data. Existing methods focus on incorporating supplementary knowledge, neglecting fusion consistency during training, which is critical for preserving the rationality of the inference. In this paper, we introduce TAP, a novel approach for few-shot stance detection. TAP extends the verbalizer hierarchically, a mapping function in prompt-tuning. Constructed using a log-odds ratio of topics and targets, the verbalizer refines candidates with prior knowledge, forming the foundation for subsequent hierarchical distillation. The hierarchical distillation, a technique based on pilot experiments on the hierarchical verbalizer, ensures the fusion of diverse knowledge during prompt-tuning, maintaining consistency throughout the training process. Notably, TAP constructs verbalizers without external knowledge augmentation. The hierarchical distillation involves a joint loss function, contributing to the model’s robustness and training consistency. Extensive experiments are conducted on SemEval2016t6 and ArgMin datasets with 13 different targets. The proposed method is evaluated on various few-shot and full-data settings with F1-Macro and F1-Micro scores. On average, TAP achieves overall improvements of 4.71% and 3.76% over state-of-the-art baselines on Semeval2016t6 and ArgMin datasets, respectively, in few-shot scenarios.

TGIN: Translation-Based Graph Inference Network for Few-Shot Relational Triplet Extraction.

Extracting relational triplets aims at detecting entity pairs and their semantic relations. Compared with pipeline models, joint models can reduce error propagation and achieve better performance. However, all of these models require large amounts of training data, therefore performing poorly on many long-tail relations in reality with insufficient data. In this article, we propose a novel end-to-end model, called TGIN, for few-shot triplet extraction. The core of TGIN is a multilayer heterogeneous graph with two types of nodes (entity node and relation node) and three types of edges (relation-entity edge, entity-entity edge, and relation-relation edge). On the one hand, this heterogeneous graph with entities and relations as nodes can intuitively extract relational triplets jointly, thereby reducing error propagation. On the other hand, it enables the triplet information of limited labeled data to interact better, thus maximizing the advantage of this information for few-shot triplet extraction. Moreover, we devise a graph aggregation and update method that utilizes translation algebraic operations to mine semantic features while retaining structure features between entities and relations, thereby improving the robustness of the TGIN in a few-shot setting. After updating the node and edge features through layers, TGIN propagates the label information from a few labeled examples to unlabeled examples, thus inferring triplets from these unlabeled examples. Extensive experiments on three reconstructed datasets demonstrate that TGIN can significantly improve the accuracy of triplet extraction by 2.34%~10.74% compared with the state-of-the-art baselines. To the best of our knowledge, we are the first to introduce a heterogeneous graph for few-shot relational triplet extraction.

Towards Open Vocabulary Learning: A Survey.

In the field of visual scene understanding, deep neural networks have made impressive advancements in various core tasks like segmentation, tracking, and detection. However, most approaches operate on the close-set assumption, meaning that the model can only identify pre-defined categories that are present in the training set. Recently, open vocabulary settings were proposed due to the rapid progress of vision language pre-training. These new approaches seek to locate and recognize categories beyond the annotated label space. The open vocabulary approach is more general, practical, and effective than weakly supervised and zero-shot settings. This paper thoroughly reviews open vocabulary learning, summarizing and analyzing recent developments in the field. In particular, we begin by juxtaposing open vocabulary learning with analogous concepts such as zero-shot learning, open-set recognition, and out-of-distribution detection. Subsequently, we examine several pertinent tasks within the realms of segmentation and detection, encompassing long-tail problems, few-shot, and zero-shot settings. As a foundation for our method survey, we first elucidate the fundamental principles of detection and segmentation in close-set scenarios. Next, we examine various contexts where open vocabulary learning is employed, pinpointing recurring design elements and central themes. This is followed by a comparative analysis of recent detection and segmentation methodologies in commonly used datasets and benchmarks. Our review culminates with a synthesis of insights, challenges, and discourse on prospective research trajectories. To our knowledge, this constitutes the inaugural exhaustive literature review on open vocabulary learning.

On the Opportunities and Challenges of Foundation Models for GeoAI (Vision Paper)

Large pre-trained models, also known as foundation models (FMs), are trained in a task-agnostic manner on large-scale data and can be adapted to a wide range of downstream tasks by fine-tuning, few-shot, or even zero-shot learning. Despite their successes in language and vision tasks, we have not yet seen an attempt to develop foundation models for geospatial artificial intelligence (GeoAI). In this work, we explore the promises and challenges of developing multimodal foundation models for GeoAI. We first investigate the potential of many existing FMs by testing their performances on seven tasks across multiple geospatial domains, including Geospatial Semantics, Health Geography, Urban Geography, and Remote Sensing. Our results indicate that on several geospatial tasks that only involve text modality, such as toponym recognition, location description recognition, and US state-level/county-level dementia time series forecasting, the task-agnostic large learning models (LLMs) can outperform task-specific fully supervised models in a zero-shot or few-shot learning setting. However, on other geospatial tasks, especially tasks that involve multiple data modalities (e.g., POI-based urban function classification, street view image–based urban noise intensity classification, and remote sensing image scene classification), existing FMs still underperform task-specific models. Based on these observations, we propose that one of the major challenges of developing an FM for GeoAI is to address the multimodal nature of geospatial tasks. After discussing the distinct challenges of each geospatial data modality, we suggest the possibility of a multimodal FM that can reason over various types of geospatial data through geospatial alignments. We conclude this article by discussing the unique risks and challenges to developing such a model for GeoAI.

Knowledge-injected prompt learning for actionable information extraction from crisis-related tweets

In crisis situations, emergency response organizations can derive significant benefits from actionable information available on social media. Such information provides effective guidance and support for rescue operations. Unlike traditional sources of information, social media reflects the latest developments during a crisis, as well as actionable suggestions and the needs of affected individuals. However, existing methods that leverage pre-trained language models for this task require a large amount of annotated data for fine-tuning. This data acquisition process is costly and impractical in some data-scarce scenarios. To address this challenge, we propose a Knowledge-injected Actionable Information Extraction Model (KAIEM) that can extract actionable information from crisis-related tweets in few-shot and zero-shot settings. Based on prompt learning, KAIEM automatically constructs the label word space of verbalizers using word embeddings, eliminating the need for manual design or external knowledge bases. This mechanism enables our model to make accurate predictions without relying on labeled data for fine-tuning. Additionally, KAIEM employs knowledge probing to inject conceptual knowledge into prompt templates, thereby enhancing the semantics of crisis-related tweets. We conducted a series of experiments on a real-world crisis-related dataset. The experimental results demonstrate the effectiveness of KAIEM in few-shot and zero-shot scenarios.

TPKE-QA: A gapless few-shot extractive question answering approach via task-aware post-training and knowledge enhancement

Few-shot extractive question answering (EQA) is a challenging task in natural language processing, whose current methods are mainly based on pretrained language models (PLMs). Data augmentation is often employed to improve the answer predictions of EQA models in few-shot settings. However, due to the differences between pretraining objectives and the EQA task, as well as embedding space alignment bottlenecks, the performance of few-shot EQA models must be improved. We propose TPKE-QA, a few-shot extractive Question Answering approach via Task-aware Post-training and Knowledge Enhancement, with entity-noun-oriented span selection in post-training, which can automatically generate EQA-style examples from a large-scale unlabeled corpus. By post-training based on generated examples, the gap between PLMs and the EQA task is effectively filled. To avoid embedding space alignment issues, a knowledge-enhanced sequence generation and knowledge injection approach for the EQA task enables gapless knowledge enhancement and fine-tuning on the post-trained model. In experiments, TPKE-QA achieved state-of-the-art results in most few-shot settings on the MRQA 2019 benchmark.

GlyReShot: A glyph-aware model with label refinement for few-shot Chinese agricultural named entity recognition

Chinese agricultural named entity recognition (NER) has been studied with supervised learning for many years. However, considering the scarcity of public datasets in the agricultural domain, exploring this task in the few-shot scenario is more practical for real-world demands. In this paper, we propose a novel model named GlyReShot, integrating the knowledge of Chinese character glyph into few-shot NER models. Although the utilization of glyph has been proven successful in supervised models, two challenges still persist in the few-shot setting, i.e., how to obtain glyph representations and when to integrate them into the few-shot model. GlyReShot handles the two challenges by introducing a lightweight glyph representation obtaining module and a training-free label refinement strategy. Specifically, the glyph representations are generated based on the descriptive sentences by filling the predefined template. As most steps come before training, this module aligns well with the few-shot setting. Furthermore, by computing the confidence values for draft predictions, the refinement strategy selectively utilizes the glyph information only when the confidence values are relatively low, thus mitigating the influence of noise. Finally, we annotate a new agricultural NER dataset and the experimental results demonstrate effectiveness of GlyReShot for few-shot Chinese agricultural NER.

Knowledge-Enriched Prompt for Low-Resource Named Entity Recognition

Named Entity Recognition (NER) in low-resource settings aims to identify and categorize entities in a sentence with limited labeled data. Although prompt-based methods have succeeded in low-resource perspectives, challenges persist in effectively harnessing information and optimizing computational efficiency. In this work, we present a novel prompt-based method to enhance low-resource NER without exhaustive template tuning. First, we construct knowledge-enriched prompts by integrating representative entities and background information to provide informative supervision tailored to each entity type. Then, we introduce an efficient reverse generative framework inspired by question answering (QA), which avoids redundant computations. Finally, we reduce costs by generating entities from their types while retaining model reasoning capacity. Experiment results demonstrate that our method outperforms other baselines on three datasets under few-shot settings.

Evaluating GPT-V4 (GPT-4 with Vision) on Detection of Radiologic Findings on Chest Radiographs.

Background Generating radiologic findings from chest radiographs is pivotal in medical image analysis. The emergence of OpenAI's generative pretrained transformer, GPT-4 with vision (GPT-4V), has opened new perspectives on the potential for automated image-text pair generation. However, the application of GPT-4V to real-world chest radiography is yet to be thoroughly examined. Purpose To investigate the capability of GPT-4V to generate radiologic findings from real-world chest radiographs. Materials and Methods In this retrospective study, 100 chest radiographs with free-text radiology reports were annotated by a cohort of radiologists, two attending physicians and three residents, to establish a reference standard. Of 100 chest radiographs, 50 were randomly selected from the National Institutes of Health (NIH) chest radiographic data set, and 50 were randomly selected from the Medical Imaging and Data Resource Center (MIDRC). The performance of GPT-4V at detecting imaging findings from each chest radiograph was assessed in the zero-shot setting (where it operates without prior examples) and few-shot setting (where it operates with two examples). Its outcomes were compared with the reference standard with regards to clinical conditions and their corresponding codes in the International Statistical Classification of Diseases, Tenth Revision (ICD-10), including the anatomic location (hereafter, laterality). Results In the zero-shot setting, in the task of detecting ICD-10 codes alone, GPT-4V attained an average positive predictive value (PPV) of 12.3%, average true-positive rate (TPR) of 5.8%, and average F1 score of 7.3% on the NIH data set, and an average PPV of 25.0%, average TPR of 16.8%, and average F1 score of 18.2% on the MIDRC data set. When both the ICD-10 codes and their corresponding laterality were considered, GPT-4V produced an average PPV of 7.8%, average TPR of 3.5%, and average F1 score of 4.5% on the NIH data set, and an average PPV of 10.9%, average TPR of 4.9%, and average F1 score of 6.4% on the MIDRC data set. With few-shot learning, GPT-4V showed improved performance on both data sets. When contrasting zero-shot and few-shot learning, there were improved average TPRs and F1 scores in the few-shot setting, but there was not a substantial increase in the average PPV. Conclusion Although GPT-4V has shown promise in understanding natural images, it had limited effectiveness in interpreting real-world chest radiographs. © RSNA, 2024 Supplemental material is available for this article.

Font transformer for few-shot font generation

Automatic font generation is of great benefit to improving the efficiency of font designers. Few-shot font generation aims to generate new fonts from a few reference samples, and has recently attracted a lot of attention from researchers. This is valuable but challenging, especially for ideograms with high diversity and complex structures. Existing models based on convolutional neural networks (CNNs) struggle to generate glyphs with accurate font style and stroke details in the few-shot setting. This paper proposes the TransFont, exploiting the long-range dependency modeling ability of the Vision Transformer (ViT) for few-shot font generation. For the first time, we empirically show that the ViT is better at glyph image generation than CNNs. Furthermore, based on the observation of the high redundancy in the glyph feature map, we introduce the glyph self-attention module for mitigating the quadratic computational and memory complexity of the pixel-level glyph image generation, along with several new techniques, i.e., multi-head multiple sampling, yz axis convolution, and approximate relative position bias. Extensive experiments on two Chinese font libraries show the superiority of our method over existing CNN-based font generation models, the proposed TransFont generates glyph images with more accurate font style and stroke details.

Few-shot defect classification via feature aggregation based on graph neural network

The effectiveness of deep learning models is greatly dependent on the availability of a vast amount of labeled data. However, in the realm of surface defect classification, acquiring and annotating defect samples proves to be quite challenging. Consequently, accurately predicting defect types with only a limited number of labeled samples has emerged as a prominent research focus in recent years. Few-shot learning, which leverages a restricted sample set in the support set, can effectively predict the categories of unlabeled samples in the query set. This approach is particularly well-suited for defect classification scenarios. In this article, we propose a transductive few-shot surface defect classification method, which using both the instance-level relations and distribution-level relations in each few-shot learning task. Furthermore, we calculate class center features in transductive manner and incorporate them into the feature aggregation operation to rectify the positioning of edge samples in the mapping space. This adjustment aims to minimize the distance between samples of the same category, thereby mitigating the influence of unlabeled samples at category boundary on classification accuracy. Experimental results on the public dataset show the outstanding performance of our proposed approach compared to the state-of-the-art methods in the few-shot learning settings. Our code is available at https://github.com/Harry10459/CIDnet.

Prompt-based learning framework for zero-shot cross-lingual text classification

Cross-lingual text classification is a challenging task that aims to train classifiers with data in one language, known as the source language, and apply the acquired knowledge to data in another language, referred to as the target language. Recent advancements in multilingual pre-trained language models (PLMs) have made significant progress in addressing cross-lingual issues, and the application of prompt-based learning has further improved task performance. However, these models still face challenges such as the gap between cross-lingual classification tasks and pre-training tasks of PLMs, as well as issues related to scarce resources and data noise, which hinder the full exploitation of the implicit knowledge in PLMs. In this paper, we propose a Prompt-based Cross-lingual Learning (PCL) framework that combines language-agnostic continuous prompt learning with self-learning process. Specifically, PCL framework leverages language-agnostic prompts and PLMs to achieve semantic transfer between source and target languages. To enhance the semantic relationship between prompts and category labels, a label attention module is introduced. Additionally, a set of self-training rules is proposed, which includes a scoring function. In a few-shot setting, noisy data is dynamically filtered through scoring and ranking of the data. During each training iteration, both the model and scoring function weights are updated, further improving the discrimination capability of the model. In summary, the proposed PCL framework builds upon cross-lingual prompt learning, effectively removing noisy data and applying it to zero-shot cross-lingual text classification, which is beneficial for engineering applications. The findings of this study have implications for prompt learning method. The PCL framework achieves state-of-the-art performance in cross-lingual text classification task, with a 14% performance improvement compared to basic soft prompt learning. This demonstrates its potential in addressing classification problems in resource-limited scenarios.