Jiabin Zheng

Jiabin Zheng

Auto-generated profile, based on publicly available information

Jiabin Zheng

School of Computer Science, Peking University, Beijing, China

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
12 April 2025 in Neural Networks
Verified
Journal Article
Closed access
MGMeng Gao
Meng Gao
YXYutao Xie
Yutao Xie
WCWei Chen
Wei Chen
FZFeng Zhang
Feng Zhang
FDFei Ding
Fei Ding
KWKam-Fai Wong
Kam-Fai Wong

29 July 2024 in IEEE Access
Verified
Journal Article
Open access
JZJiabin Zheng
Jiabin Zheng
HWHanlin Wang
Hanlin Wang
YJYao Jiahui
Yao Jiahui
Large language models (LLMs) have demonstrated their vast potential and value in natural language processing tasks and beyond. However, when these models are applied to develop consultation systems for industrial domains, such as e-government, intelligent diagnosis, and legal consultancy, they encounter many unresolved technical issues. These include the vast scale of the models, lack of tight fusion with existing industry knowledge, along with occurrences of model hallucinations and inadequate explainability. Unlike general-purpose dialogue systems, building a consultation system for a specific industrial domain requires not only the integration of extensive external knowledge from the Internet but also the incorporation of precise, specialized knowledge from specific industries, making the challenges even more complex. In response to these challenges, we propose the Inter-External Knowledge Fusion Contrastive Learning Technique. This technique facilitates the integration of internal industry knowledge with widely accessible external knowledge from the Internet and provides a universal framework for building lightweight, domain-specific consultation systems. Utilizing this technique enables the straightforward creation of precise, professional, and constantly updated domain-specific consultation systems applicable across various industries. Overcoming the inherent limitations associated with LLMs, this technique achieves a performance level comparable to LLMs. To validate the effectiveness of our proposed technique, we conducted extensive experiments in the development of real-world industry consulting systems. By testing on seven real datasets covering diverse tasks, we demonstrate the system’s exceptional performance: our lightweight consultation system utilizes only 4% of the parameters of an LLM but achieves over 90% of its performance level.

21 October 2024
Verified
Conference Paper
Open access
FZFeng Zhang
Feng Zhang
WCWei Chen
Wei Chen
FDFei Ding
Fei Ding
TWTengjiao Wang
Tengjiao Wang
DLDawei Lu
Dawei Lu
JZJiabin Zheng
Jiabin Zheng
Multi-label few-shot image recognition aims to identify multiple unseen objects using only a handful of examples. Recent methods typically tune pre-trained vision-language models with shared or class-specific prompts. However, they still have drawbacks. Tuning a shared prompt is insufficient for all samples especially when the tasks are complex and tuning specific prompts for each class is inevitable to lose generalization ability, thus failing to capture diverse visual knowledge. To address these issues, we propose to meta-tune a generalized prompt pool, enabling each prompt to act as an expert for multi-label few-shot image recognition. Specifically, we first construct a diverse prompt pool to handle complex samples and tasks effectively. Then, the meta-tuning strategy is designed to learn meta-knowledge and transfer it from source tasks to target tasks, enhancing the generalization of prompts. Extensive experimental results on two widely used multi-label image recognition datasets demonstrate the effectiveness of our method.

30 April 2024 in IEEE Access
Verified
Journal Article
Open access
JZJiabin Zheng
Jiabin Zheng
FXFangyi Xu
Fangyi Xu
WCWei Chen
Wei Chen
ZFZihao Fang
Zihao Fang
YJYao Jiahui
Yao Jiahui
Cross-domain Consultation Systems have become essential in numerous critical applications, for instance, an online citizen complaint system. However, addressing complaints with distinct orality characteristics often necessitates retrieving and integrating knowledge from diverse professional domains. This scenario represents a typical cross-domain problem. Nevertheless, the prevailing approach of utilizing generative large language models to tackle this problem presents challenges including model scale and drawbacks like hallucination and limited interpretability. To address these challenges, we proposed a novel approach called the Core-View Contrastive Learning (CVCL) network. Leveraging contrastive learning techniques with an integrated core-adaptive augmentation module, the CVCL network achieves accuracy in cross-domain information matching. Our objective is to construct a lightweight, precise, and interpretable cross-domain consultation system, overcoming the limitations encountered with large language models in addressing such challenges. Empirical validation of our proposed method using real-world datasets demonstrates its effectiveness. Our experiments show that the proposed method achieves comparable performance to large language models in terms of accuracy in text-matching tasks and surpasses the best baseline model by over 24 percentage points in F1-score for classification tasks. Additionally, our lightweight model achieved a performance level of 96% compared to the full model, while utilizing only 6% of the parameters.