Zero-shot Classification Research Articles

WC-SBERT: Zero-Shot Topic Classification Using SBERT and Light Self-Training on Wikipedia Categories

In natural language processing (NLP), zero-shot topic classification requires machines to understand the contextual meanings of texts in a downstream task without using the corresponding labeled texts for training, which is highly desirable for various applications. In this article, we propose a novel approach to construct a zero-shot task-specific model called WC-SBERT with satisfactory performance. The proposed approach is highly efficient since it uses light self-training requiring target labels (target class names of downstream tasks) only, which is distinct from other research that uses both the target labels and the unlabeled texts for training. In particular, during the pre-training stage, WC-SBERT uses contrastive learning with multiple negative ranking losses to construct the pre-trained model based on the similarity between Wiki categories. For the self-training stage, online contrastive loss is utilized to reduce the distance between a target label and Wiki categories of similar Wiki pages to the label. Experimental results indicate that compared to existing self-training models, WC-SBERT achieves rapid inference on approximately 6.45 million Wiki text entries by utilizing pre-stored Wikipedia text embeddings, significantly reducing inference time per sample by a factor of 2,746 to 16,746. During the fine-tuning step, the time required for each sample is reduced by a factor of 23–67. Overall, the total training time shows a maximum reduction of 27.5 times across different datasets. Most importantly, our model has achieved state-of-the-art (SOTA) accuracy on two of the three commonly used datasets for evaluating zero-shot classification, namely the AG News (0.84) and Yahoo! Answers (0.64) datasets. The code for WC-SBERT is publicly available on GitHub, 1 and the dataset can also be accessed on Hugging Face. 2

At First Sight! Zero-shot Classification of Astronomical Images with Large Multimodal Models

Abstract Vision-Language multimodal Models (VLMs) offer the possibility for zero-shot classification in astronomy: i.e., classification via natural language prompts, with no training. We investigate two models, GPT-4o and LLaVA-NeXT, for zero-shot classification of low-surface brightness galaxies and artifacts, as well as morphological classification of galaxies. We show that with natural language prompts these models achieved significant accuracy (above 80% typically) without additional training/fine tuning. We discuss areas that require improvement, especially for LLaVA-NeXT, which is an open source model. Our findings aim to motivate the astronomical community to consider VLMs as a powerful tool for both research and pedagogy, with the prospect that future custom-built or fine-tuned models could perform better.

Open-Pose 3D zero-shot learning: Benchmark and challenges

With the explosive 3D data growth, the urgency of utilizing zero-shot learning to facilitate data labeling becomes evident. Recently, methods transferring language or language-image pre-training models like Contrastive Language-Image Pre-training (CLIP) to 3D vision have made significant progress in the 3D zero-shot classification task. These methods primarily focus on 3D object classification with an aligned pose; such a setting is, however, rather restrictive, which overlooks the recognition of 3D objects with open poses typically encountered in real-world scenarios, such as an overturned chair or a lying teddy bear. To this end, we propose a more realistic and challenging scenario named open-pose 3D zero-shot classification, focusing on the recognition of 3D objects regardless of their orientation. First, we revisit the current research on 3D zero-shot classification and propose two benchmark datasets specifically designed for the open-pose setting. We empirically validate many of the most popular methods in the proposed open-pose benchmark. Our investigations reveal that most current 3D zero-shot classification models suffer from poor performance, indicating a substantial exploration room towards the new direction. Furthermore, we study a concise pipeline with an iterative angle refinement mechanism that automatically optimizes one ideal angle to classify these open-pose 3D objects. In particular, to make validation more compelling and not just limited to existing CLIP-based methods, we also pioneer the exploration of knowledge transfer based on Diffusion models. While the proposed solutions can serve as a new benchmark for open-pose 3D zero-shot classification, we discuss the complexities and challenges of this scenario that remain for further research development. The code is available publicly at https://github.com/weiguangzhao/Diff-OP3D

Significantly improving zero-shot X-ray pathology classification via fine-tuning pre-trained image-text encoders

Deep neural networks are increasingly used in medical imaging for tasks such as pathological classification, but they face challenges due to the scarcity of high-quality, expert-labeled training data. Recent efforts have utilized pre-trained contrastive image-text models like CLIP, adapting them for medical use by fine-tuning the model with chest X-ray images and corresponding reports for zero-shot pathology classification, thus eliminating the need for pathology-specific annotations. However, most studies continue to use the same contrastive learning objectives as in the general domain, overlooking the multi-labeled nature of medical image-report pairs. In this paper, we propose a new fine-tuning strategy that includes positive-pair loss relaxation and random sentence sampling. We aim to improve the performance of zero-shot pathology classification without relying on external knowledge. Our method can be applied to any pre-trained contrastive image-text encoder and easily transferred to out-of-domain datasets without further training, as it does not use external data. Our approach consistently improves overall zero-shot pathology classification across four chest X-ray datasets and three pre-trained models, with an average macro AUROC increase of 4.3%. Additionally, our method outperforms the state-of-the-art and marginally surpasses board-certified radiologists in zero-shot classification for the five competition pathologies in the CheXpert dataset.

Analysis of LLMs for educational question classification and generation

Large language models (LLMs) like ChatGPT have shown promise in generating educational content, including questions. This study evaluates the effectiveness of LLMs in classifying and generating educational-type questions. We assessed ChatGPT's performance using a dataset of 4,959 user-generated questions labeled into ten categories, employing various prompting techniques and aggregating results with a voting method to enhance robustness. Additionally, we evaluated ChatGPT's accuracy in generating type-specific questions from 100 reading sections sourced from five online textbooks, which were manually reviewed by human evaluators. We also generated questions based on learning objectives and compared their quality to those crafted by human experts, with evaluations by experts and crowdsourced participants.Our findings reveal that ChatGPT achieved a macro-average F1-score of 0.57 in zero-shot classification, improving to 0.70 when combined with a Random Forest classifier using embeddings. The most effective prompting technique was zero-shot with added definitions, while few-shot and few-shot + Chain of Thought approaches underperformed. The voting method enhanced robustness in classification. In generating type-specific questions, ChatGPT's accuracy was lower than anticipated. However, quality differences between ChatGPT-generated and human-generated questions were not statistically significant, indicating ChatGPT's potential for educational content creation. This study underscores the transformative potential of LLMs in educational practices. By effectively classifying and generating high-quality educational questions, LLMs can reduce the workload on educators and enable personalized learning experiences.

Evaluating large language models for user stance detection on X (Twitter)

Current stance detection methods employ topic-aligned data, resulting in many unexplored topics due to insufficient training samples. Large Language Models (LLMs) pre-trained on a vast amount of web data offer a viable solution when training data is unavailable. This work introduces Tweets2Stance - T2S, an unsupervised stance detection framework based on zero-shot classification, i.e. leveraging an LLM pre-trained on Natural Language Inference tasks. T2S detects a five-valued user’s stance on social-political statements by analyzing their X (Twitter) timeline. The Ground Truth of a user’s stance is obtained from Voting Advice Applications (VAAs). Through comprehensive experiments, a T2S’s optimal setting was identified for each election. Linguistic limitations related to the language model are further addressed by integrating state-of-the-art LLMs like GPT-4 and Mixtral into the T2S framework. The T2S framework’s generalization potential is demonstrated by measuring its performance (F1 and MAE scores) across nine datasets. These datasets were built by collecting tweets from competing parties’ Twitter accounts in nine political elections held in different countries from 2019 to 2021. The results, in terms of F1 and MAE scores, outperformed all baselines and approached the best scores for each election. This showcases the ability of T2S, particularly when combined with state-of-the-art LLMs, to generalize across different cultural-political contexts.

Benchmarking PathCLIP for Pathology Image Analysis.

Accurate image classification and retrieval are of importance for clinical diagnosis and treatment decision-making. The recent contrastive language-image pre-training (CLIP) model has shown remarkable proficiency in understanding natural images. Drawing inspiration from CLIP, pathology-dedicated CLIP (PathCLIP) has been developed, utilizing over 200,000 image and text pairs in training. While the performance the PathCLIP is impressive, its robustness under a wide range of image corruptions remains unknown. Therefore, we conduct an extensive evaluation to analyze the performance of PathCLIP on various corrupted images from the datasets of osteosarcoma and WSSS4LUAD. In our experiments, we introduce eleven corruption types including brightness, contrast, defocus, resolution, saturation, hue, markup, deformation, incompleteness, rotation, and flipping at various settings. Through experiments, we find that PathCLIP surpasses OpenAI-CLIP and the pathology language-image pre-training (PLIP) model in zero-shot classification. It is relatively robust to image corruptions including contrast, saturation, incompleteness, and orientation factors. Among the eleven corruptions, hue, markup, deformation, defocus, and resolution can cause relatively severe performance fluctuation of the PathCLIP. This indicates that ensuring the quality of images is crucial before conducting a clinical test. Additionally, we assess the robustness of PathCLIP in the task of image-to-image retrieval, revealing that PathCLIP performs less effectively than PLIP on osteosarcoma but performs better on WSSS4LUAD under diverse corruptions. Overall, PathCLIP presents impressive zero-shot classification and retrieval performance for pathology images, but appropriate care needs to be taken when using it.

Zero-shot classification of small target on sea bottom using model-agnostic meta-learning.

A model-agnostic meta-learning (MAML)-based active target classifier to identify small targets (e.g., mines) on the sea bottom in different ocean environments from those present in the training data is proposed. To better classify the targets deviating from those in the training set, MAML is applied to the out-of-distribution samples. Frequency-domain target and clutter scattering signals from various tasks with varying bottom types (silt/clay) and incident angles (low/moderate/high) are utilized as training data samples. MAML significantly outperforms conventional neural networks during the test. The improved generalization of MAML is explained using loss landscape in the form of a smooth convex curve.

Googling Politics? Comparing Five Computational Methods to Identify Political and News-related Searches from Web Browser Histories

ABSTRACT Search engines play a crucial role in today’s information environment. Yet, political and news-related (PNR) search engine use remains understudied, mainly due to the lack of suitable measurement methods to identify PNR searches. Existing research focuses on specific events, topics, or news articles, neglecting the broader scope of PNR search. Furthermore, self-reporting issues have led researchers to use browsing history data, but scalable methods for analyzing such data are limited. This paper addresses these gaps by comparing five computational methods to identify PNR searches in browsing data, including browsing sequences, context-enhanced dictionary, Traditional Supervised Machine Learning (SML), Transformer-based SML, and zero-shot classification. Using Dutch Google searches as a test case, we use Dutch browsing history data obtained via data donations in May 2022 linked to surveys (N users = 315; N records = 9,868,209; N searches = 697,359), along with 35.5k manually annotated search terms. The findings highlight substantial variation in accuracy, with some methods being more suited for narrower topics. We recommend a two-step approach, applying zero-shot classification followed by human evaluation. This methodology can inform future empirical research on PNR search engine use.

A comparative study of large language model-based zero-shot inference and task-specific supervised classification of breast cancer pathology reports.

Although supervised machine learning is popular for information extraction from clinical notes, creating large annotated datasets requires extensive domain expertise and is time-consuming. Meanwhile, large language models (LLMs) have demonstrated promising transfer learning capability. In this study, we explored whether recent LLMs could reduce the need for large-scale data annotations. We curated a dataset of 769 breast cancer pathology reports, manually labeled with 12 categories, to compare zero-shot classification capability of the following LLMs: GPT-4, GPT-3.5, Starling, and ClinicalCamel, with task-specific supervised classification performance of 3 models: random forests, long short-term memory networks with attention (LSTM-Att), and the UCSF-BERT model. Across all 12 tasks, the GPT-4 model performed either significantly better than or as well as the best supervised model, LSTM-Att (average macro F1-score of 0.86 vs 0.75), with advantage on tasks with high label imbalance. Other LLMs demonstrated poor performance. Frequent GPT-4 error categories included incorrect inferences from multiple samples and from history, and complex task design, and several LSTM-Att errors were related to poor generalization to the test set. On tasks where large annotated datasets cannot be easily collected, LLMs can reduce the burden of data labeling. However, if the use of LLMs is prohibitive, the use of simpler models with large annotated datasets can provide comparable results. GPT-4 demonstrated the potential to speed up the execution of clinical NLP studies by reducing the need for large annotated datasets. This may increase the utilization of NLP-based variables and outcomes in clinical studies.

Multi-Relation Extraction for Cybersecurity Based on Ontology Rule-Enhanced Prompt Learning

In the domain of cybersecurity, available annotated data are often scarce, especially for Chinese cybersecurity datasets, often necessitating the manual construction of datasets. The scarcity of samples is one of the challenges in researching cybersecurity, especially for the “no-relation” class. Since the annotation process typically focuses only on known relation classes, there are usually no training samples for the “no-relation” class. This poses a zero-shot classification problem, where during the classification process, there is a tendency to classify into a class with a relationship. Zero-shot classification tasks are particularly challenging in this context. Moreover, most relation classification models currently need to traverse all relations to calculate the class with the highest probability. Therefore, the problem of “computational redundancy” is another challenge faced. Thus, how to accurately and efficiently acquire cyberspace knowledge from heterogeneous data sources and address the challenges such as sample scarcity, zero-shot recognition, and computational redundancy is the main focus of this chapter. To address these problems, this chapter designs a multi-relation extraction model based on ontology rule-enhanced prompt learning, which is a parameter-sharing-based multi-task model. By introducing prompt learning, which has shown significant effectiveness in the few-shot domain, this chapter designs prompt templates combining discrete and continuous tokens and uses rule injection in prompt learning to solve the difficulties in zero-shot recognition of “no-relation” and computational redundancy issues, achieving efficient and accurate multi-relation extraction. Specifically, by constructing sub-prompts to achieve an efficient combination of templates, a parameter-sharing structure is used to implement knowledge extraction step by step: The first step constructs entity prompt templates combining discrete and continuous tokens, identifying the classes of two entities based on prompt learning. The second step involves rule injection, identifying whether it belongs to the “no-relation” class based on the combination of sub-prompts; if there is no connection between the classes of two entities, it is classified as “no relation”; if a connection exists, the candidate relation set is filtered out. The third step uses the pre-trained model and vectors from the first step, utilizing prompt learning and rule judgment to determine the relation class from the candidate relation set. Finally, the effectiveness of our model is validated on the general datasets TACRED, ReTACRED, and the cybersecurity dataset constructed in this paper.

Fusing differentiable rendering and language–image contrastive learning for superior zero-shot point cloud classification

Zero-shot point cloud classification involves recognizing categories not encountered during training. Current models often exhibit reduced accuracy on unseen categories without 3D pre-training, emphasizing the need for improved precision and interoperability. We propose a novel approach integrating differentiable rendering with contrastive language–image pre-training. Initially, differentiable rendering autonomously learns representative viewpoints from the data, enabling the transformation of point clouds into multi-view images while preserving key visual information. This transformation facilitates optimized viewpoint selection during training, refining the final feature representation. Features are extracted from the multi-view images and integrated into a global multi-view feature using a cross-attention mechanism. On the textual side, a large language model (LLM) is provided with 3D heuristic prompts to generate 3D-specific text reflecting category-specific traits, from which textual features are derived. The LLM’s extensive pre-trained knowledge enables it to capture abstract notions and categorical features relevant to distinct point cloud categories. Visual and textual features are aligned in a unified embedding space, enabling zero-shot classification. Throughout training, the Structural Similarity Index (SSIM) is integrated into the loss function to encourage the model to discern more distinctive viewpoints, reduce redundancy in multi-view imagery, and enhance computational efficiency. Experimental results on the ModelNet10, ModelNet40, and ScanObjectNN datasets demonstrate classification accuracies of 75.68%, 66.42%, and 52.03%, respectively, surpassing prevailing methods in zero-shot point cloud classification accuracy.

Prompt-guided DETR with RoI-pruned masked attention for open-vocabulary object detection

Prompt-OVD is an efficient and effective DETR-based framework for open-vocabulary object detection that utilizes class embeddings from CLIP as prompts, guiding the Transformer decoder to detect objects in base and novel classes. Additionally, our RoI-pruned masked attention helps leverage the zero-shot classification ability of the Vision Transformer-based CLIP, resulting in improved detection performance at a minimal computational cost. Our experiments on the OV-COCO and OV-LVIS datasets demonstrate that Prompt-OVD achieves an impressive 21.2 times faster inference speed than the first end-to-end open-vocabulary detection method (OV-DETR), while also achieving higher APs than four two-stage methods operating within similar inference time ranges. We release the code at https://github.com/DISL-Lab/Prompt-OVD.

ArcheType: A Novel Framework for Open-Source Column Type Annotation Using Large Language Models

Existing deep-learning approaches to semantic column type annotation (CTA) have important shortcomings: they rely on semantic types which are fixed at training time; require a large number of training samples per type; incur high run-time inference costs; and their performance can degrade when evaluated on novel datasets, even when types remain constant. Large language models have exhibited strong zero-shot classification performance on a wide range of tasks and in this paper we explore their use for CTA. We introduce ArcheType, a simple, practical method for context sampling, prompt serialization, model querying, and label remapping, which enables large language models to solve CTA problems in a fully zero-shot manner. We ablate each component of our method separately, and establish that improvements to context sampling and label remapping provide the most consistent gains. ArcheType establishes a new state-of-the-art performance on zero-shot CTA benchmarks (including three new domain-specific benchmarks which we release along with this paper), and when used in conjunction with classical CTA techniques, it outperforms a SOTA DoDuo model on the fine-tuned SOTAB benchmark.

Использование машинного обучения для тематической разметки текстовых материалов корпуса устной речи

The research aims to determine the effectiveness of the thesaurus method for forming a list of topic classes when using machine learning for the topic classification of text materials of sociolinguistic interviews. The paper considers the potential of using machine learning in the topic annotation of linguistic corpus materials. The polytopical nature of the analyzed material is due to its genre belonging to dialogical speech. The hierarchical structure of the topics, identified as a result of a preliminary introspective analysis of the texts, can be described using a thesaurus. The results of using the unsupervised machine learning method are discussed involving two sets of topic class names: a list of topics used in manual text annotation and an extended list of micro-topics whose names were selected from a Russian language thesaurus. The paper is novel in that it is the first to propose the thesaurus method for selecting topic labels for the zero-shot classification of weakly structured Russian texts. The research findings show that using a more detailed lexical description for topic classes improves the classification result.

SKZC: self-distillation and k-nearest neighbor-based zero-shot classification

Zero-shot learning represents a formidable paradigm in machine learning, wherein the crux lies in distilling and generalizing knowledge from observed classes to novel ones. The objective is to identify unfamiliar objects that were not included in the model’s training, leveraging learned patterns and knowledge from previously encountered categories. As a crucial subtask of open-world object detection, zero-shot classification can also provide insights and solutions for this field. Despite its potential, current zero-shot classification models often suffer from a performance gap due to limited transfer ability and discriminative capability of learned representations. In pursuit of advancing the subpar state of zero-shot object classification, this paper introduces a novel model for image classification which can be applied to object detection, namely, self-distillation and k-nearest neighbor-based zero-shot classification method. First, we employ a diffusion detector to identify potential objects in images. Then, self-distillation and distance-based classifiers are used for distinguishing unseen objects from seen classes. The k-nearest neighbor-based cluster heads are designed to cluster the unseen objects. Extensive experiments and visualizations were conducted on publicly available datasets on the efficacy of the proposed approach. Precisely, our model demonstrates performance improvement of over 20% compared to contrastive clustering. Moreover, it achieves a precision of 0.910 and a recall of 0.842 on CIFAR-10 datasets, a precision of 0.737, and a recall of 0.688 on CIFAR-100 datasets for the macro average. Compared to a more recent model (SGFR), our model realized improvements of 10.9%, 13.3%, and 7.8% in Sacc, Uacc, and H metrics, respectively. This study aims to introduce fresh ideas into the domain of zero-shot image classification, and it can be applied to open-world object detection tasks. Our code is available at https://www.github.com/CmosWolf1/Code_implementation_for_paper_SKZC.

A systematic evaluation of text mining methods for short texts: Mapping individuals’ internal states from online posts

Short texts generated by individuals in online environments can provide social and behavioral scientists with rich insights into these individuals’ internal states. Trained manual coders can reliably interpret expressions of such internal states in text. However, manual coding imposes restrictions on the number of texts that can be analyzed, limiting our ability to extract insights from large-scale textual data. We evaluate the performance of several automatic text analysis methods in approximating trained human coders’ evaluations across four coding tasks encompassing expressions of motives, norms, emotions, and stances. Our findings suggest that commonly used dictionaries, although performing well in identifying infrequent categories, generate false positives too frequently compared to other methods. We show that large language models trained on manually coded data yield the highest performance across all case studies. However, there are also instances where simpler methods show almost equal performance. Additionally, we evaluate the effectiveness of cutting-edge generative language models like GPT-4 in coding texts for internal states with the help of short instructions (so-called zero-shot classification). While promising, these models fall short of the performance of models trained on manually analyzed data. We discuss the strengths and weaknesses of various models and explore the trade-offs between model complexity and performance in different applications. Our work informs social and behavioral scientists of the challenges associated with text mining of large textual datasets, while providing best-practice recommendations.

Extraction of Water Bodies from High-Resolution Aerial and Satellite Images Using Visual Foundation Models

Water, indispensable for life and central to ecosystems, human activities, and climate dynamics, requires rapid and accurate monitoring. This is vital for sustaining ecosystems, enhancing human welfare, and effectively managing land, water, and biodiversity on both the local and global level. In the rapidly evolving domain of remote sensing and deep learning, this study focuses on water body extraction and classification through the use of recent deep learning models of visual foundation models (VFMs). Specifically, the Segment Anything Model (SAM) and Contrastive Language-Image Pre-training (CLIP) models have shown promise in semantic segmentation, dataset creation, change detection, and instance segmentation tasks. A novel two-step approach involving segmenting images via the Automatic Mask Generator method of the SAM and the zero-shot classification of segments using CLIP is proposed, and its effectiveness is tested on water body extraction problems. The proposed methodology was applied to both remote sensing imagery acquired from LANDSAT 8 OLI and very high-resolution aerial imagery. Results revealed that the proposed methodology accurately delineated water bodies across complex environmental conditions, achieving a mean intersection over union (IoU) of 94.41% and an F1 score of 96.97% for satellite imagery. Similarly, for the aerial imagery dataset, the proposed methodology achieved a mean IoU of 90.83% and an F1 score exceeding 94.56%. The high accuracy achieved in selecting segments predominantly classified as water highlights the effectiveness of the proposed model in intricate environmental image analysis.

S3A: Towards Realistic Zero-Shot Classification via Self Structural Semantic Alignment

Large-scale pre-trained Vision Language Models (VLMs) have proven effective for zero-shot classification. Despite the success, most traditional VLMs-based methods are restricted by the assumption of partial source supervision or ideal target vocabularies, which rarely satisfy the open-world scenario. In this paper, we aim at a more challenging setting, Realistic Zero-Shot Classification, which assumes no annotation but instead a broad vocabulary. To address the new problem, we propose the Self Structural Semantic Alignment (S3A) framework, which extracts the structural semantic information from unlabeled data while simultaneously self-learning. Our S3A framework adopts a unique Cluster-Vote-Prompt-Realign (CVPR) algorithm, which iteratively groups unlabeled data to derive structural semantics for pseudo-supervision. Our CVPR algorithm includes iterative clustering on images, voting within each cluster to identify initial class candidates from the vocabulary, generating discriminative prompts with large language models to discern confusing candidates, and realigning images and the vocabulary as structural semantic alignment. Finally, we propose to self-train the CLIP image encoder with both individual and structural semantic alignment through a teacher-student learning strategy. Our comprehensive experiments across various generic and fine-grained benchmarks demonstrate that the S3A method substantially improves over existing VLMs-based approaches, achieving a more than 15% accuracy improvement over CLIP on average. Our codes, models, and prompts are publicly released at https://github.com/sheng-eatamath/S3A.

Data-Free Generalized Zero-Shot Learning

Deep learning models have the ability to extract rich knowledge from large-scale datasets. However, the sharing of data has become increasingly challenging due to concerns regarding data copyright and privacy. Consequently, this hampers the effective transfer of knowledge from existing data to novel downstream tasks and concepts. Zero-shot learning (ZSL) approaches aim to recognize new classes by transferring semantic knowledge learned from base classes. However, traditional generative ZSL methods often require access to real images from base classes and rely on manually annotated attributes, which presents challenges in terms of data restrictions and model scalability. To this end, this paper tackles a challenging and practical problem dubbed as data-free zero-shot learning (DFZSL), where only the CLIP-based base classes data pre-trained classifier is available for zero-shot classification. Specifically, we propose a generic framework for DFZSL, which consists of three main components. Firstly, to recover the virtual features of the base data, we model the CLIP features of base class images as samples from a von Mises-Fisher (vMF) distribution based on the pre-trained classifier. Secondly, we leverage the text features of CLIP as low-cost semantic information and propose a feature-language prompt tuning (FLPT) method to further align the virtual image features and textual features. Thirdly, we train a conditional generative model using the well-aligned virtual image features and corresponding semantic text features, enabling the generation of new classes features and achieve better zero-shot generalization. Our framework has been evaluated on five commonly used benchmarks for generalized ZSL, as well as 11 benchmarks for the base-to-new ZSL. The results demonstrate the superiority and effectiveness of our approach. Our code is available in https://github.com/ylong4/DFZSL.