A5100436613- Topic Modeling
- Natural Language Processing Techniques
- Multimodal Machine Learning Applications
- Advanced Text Analysis Techniques
- Text and Document Classification Technologies
- Domain Adaptation and Few-Shot Learning
- Sentiment Analysis and Opinion Mining
- Advanced Graph Neural Networks
- Advanced Image and Video Retrieval Techniques
- Machine Learning and Data Classification
- Recommender Systems and Techniques
- Advanced Neural Network Applications
- Image and Video Quality Assessment
- Computer Graphics and Visualization Techniques
- Speech and dialogue systems
- Data Stream Mining Techniques
- Music and Audio Processing
- Anomaly Detection Techniques and Applications
- Human Pose and Action Recognition
- Generative Adversarial Networks and Image Synthesis
- Machine Learning and ELM
- Game Theory and Voting Systems
- Educational Technology and Pedagogy
- Green IT and Sustainability
- Reinforcement Learning in Robotics
Nanjing University of Science and Technology
2023-2024
Hunan Software Vocational Institute
2024
Institute of Information Engineering
2024
Nanyang Technological University
2010-2023
University of Chinese Academy of Sciences
2023
Liaoning University
2023
Minzu University of China
2022
Space Engineering University
2020
Zhejiang Financial College
2020
Shanghai Center for Brain Science and Brain-Inspired Technology
2020
Data quality, or sometimes referred to as data credibility, is a critical issue in mobile crowd sensing (MCS) and more generally Internet of Things (IoT). While candidate solutions, such incentive mechanisms mining have been well explored the literature, power crowds has largely overlooked under-exploited. In this paper, we propose cross validation approach which seeks validating ratify contributing terms sensor contributed by latter, uses result reshape into credible posterior belief ground...
We present a novel deep reinforcement learning method to learn construction heuristics for vehicle routing problems. In specific, we propose Multi-Decoder Attention Model (MDAM) train multiple diverse policies, which effectively increases the chance of finding good solutions compared with existing methods that only one policy. A customized beam search strategy is designed fully exploit diversity MDAM. addition, an Embedding Glimpse layer in MDAM based on recursive nature construction, can...
In this paper, we explore how to adapt a general Hidden Markov Model-based named entity recognizer effectively biomedical domain. We integrate various features, including simple deterministic morphological POS features and semantic trigger capture evidences especially for evaluate their contributions. also present algorithm solve the abbreviation problem rule-based method deal with cascaded phenomena in Our experiments on GENIA V3.0 V1.1 achieve 66.1 62.5 F-measure respectively, which...
In this work, we construct the largest dataset for multimodal pretraining in Chinese, which consists of over 1.9TB images and 292GB texts that cover a wide range domains. We propose cross-modal method called M6, referring to Multi-Modality Multitask Mega-transformer, unified on data single modality multiple modalities. scale model size up 10 billion 100 parameters, build pretrained Chinese. apply series downstream applications, demonstrate its outstanding performance comparison with strong...
Imbalanced data is a perennial problem that impedes the learning abilities of current machine learning-based classification models. One approach to address it leverage augmentation expand training set. For image data, there are number suitable techniques have proven effective in previous work. textual however, due discrete units inherent natural language, randomly perturb signal may be ineffective. Additionally, substantial discrepancy between different datasets (e.g., domains), an...
Large language models (LLMs) provide a promising way for accurate session-based recommendation (SBR), but they demand substantial computational time and memory. Knowledge distillation (KD)-based methods can alleviate these issues by transferring the knowledge to small student, which trains student based on predictions of cumbersome teacher. However, encounter difficulties LLM-based KD in SBR. 1) It is expensive make LLMs predict all instances KD. 2) may ineffective some KD, e.g., incorrect...
Customer services are critical to all companies, as they may directly connect the brand reputation. Due a great number of customers, e-commerce companies often employ multiple communication channels answer customers' questions, for example, Chatbot and Hotline. On one hand, each channel has limited capacity respond requests; on other customers have different preferences over these channels. The current production systems mainly built based business rules that merely consider tradeoffs...
In the ad hoc teamwork setting, a team of agents needs to perform task without prior coordination. The most advanced approach learns policies based on previous experiences and reuses one interact with new teammates. However, selected policy in many cases is sub-optimal. Switching between adapt teammates' behaviour takes time, which threatens successful performance task. this paper, we propose AATEAM – method that uses attention-based neural networks cope real-time. We train attention network...
Generating questions based on answers and relevant contexts is a challenging task. Recent work mainly pays attention to the quality of single generated question. However, question generation actually one-to-many problem, as it possible raise with different focuses various means expression. In this paper, we explore diversity come up methods from these two aspects. Specifically, relate contextual content selectors, which are modeled by continuous latent variable technique conditional...
Recent expeditious developments in deep learning algorithms, distributed training, and even hardware design for large models have enabled training extreme-scale models, say GPT-3 Switch Transformer possessing hundreds of billions or trillions parameters. However, under limited resources, model that requires enormous amounts computes memory footprint suffers from frustratingly low efficiency convergence. In this paper, we propose a simple strategy called "Pseudo-to-Real"...
In recent years, efforts have been made to use text information for better user profiling and item characterization in recommendations. However, can sometimes be of low quality, hindering its effectiveness real-world applications. With knowledge reasoning capabilities capsuled Large Language Models (LLMs), utilizing LLMs emerges as a promising way description improvement. existing ways prompting with raw texts ignore structured user-item interactions, which may lead hallucination problems...
In light of the millions households that have adopted intelligent assistant powered devices, multi-turn dialogue has become an important field inquiry. Most current methods identify underlying intent in using opaque classification techniques fail to provide any interpretable basis for classification. To address this, we propose a scheme interpret based on specific characteristics text. We rely policy-guided reinforcement learning paths graph confirm concrete inference serve as explanations....
Stream classification models for non-stationary environments often assume the immediate availability of data labels. However, in a practical scenario, it is quite natural that labels are available only after some temporal lag. This paper explores how stream classifier model can be made adaptive to such label latency scenario. We propose SkipE-RNN, self-evolutionary recurrent neural network with dynamically evolving skipped-recurrent-connection best utilization previously observed information...
Utterance classification is a key component in many conversational systems. However, classifying real-world user utterances challenging, as people may express their ideas and thoughts manifold ways, the amount of training data for some categories be fairly limited, resulting imbalanced distributions. To alleviate these issues, we conduct comprehensive survey regarding augmentation approaches text classification, including simple random resampling, word-level transformations, neural...
Multi-task learning (MTL) has been widely utilized in various industrial scenarios, such as recommender systems and search engines. MTL can improve efficiency prediction accuracy by exploiting commonalities differences across tasks. However, is sensitive to relationships among tasks may have performance degradation real-world applications, because existing neural-based models often share the same network structures original input features. To address this issue, we propose a novel multi-task...
Synthetic training data has gained prominence in numerous learning tasks and scenarios, offering advantages such as dataset augmentation, generalization evaluation, privacy preservation. Despite these benefits, the efficiency of synthetic generated by current methodologies remains inferior when advanced deep models exclusively, limiting its practical utility. To address this challenge, we analyze principles underlying synthesis for supervised elucidate a principled theoretical framework from...
We develop a state-of-the-art fraud prediction model using machine learning approach. demonstrate the value of combining domain knowledge and method in building. select our input based on existing accounting theories, but we differ from prior research by raw numbers rather than financial ratios. employ one most powerful methods, ensemble learning, commonly used logistic regression. To assess performance models, introduce new evaluation metric ranking problems that is more appropriate for...
The key challenge of generative Visual Dialogue (VD) systems is to respond human queries with informative answers in natural and contiguous conversation flow. Traditional Maximum Likelihood Estimation-based methods only learn from positive responses but ignore the negative responses, consequently tend yield safe or generic responses. To address this issue, we propose a novel training scheme conjunction weighted likelihood estimation method. Furthermore, an adaptive multi-modal reasoning...
Document-level event extraction (DEE) aims at extracting records from given documents. Existing DEE methods handle troublesome challenges by using multiple encoders and casting the task into a multi-step paradigm. However, most of previous approaches ignore missing feature mean pooling or max operations in different encoding stages have not explicitly modeled interdependency features between input tokens, thus long-distance problem cannot be solved effectively. In this study, we propose...