Haimin Zhang

Yunxiao Shi, Xing Zi, Zijing Shi et al.

Retrieval-augmented generation (RAG) techniques leverage the in-context learning capabilities of large language models (LLMs) to produce more accurate and relevant responses. Originating from the simple 'retrieve-then-read' approach, the RAG framework has evolved into a highly flexible and modular paradigm. A critical component, the Query Rewriter module, enhances knowledge retrieval by generating a search-friendly query. This method aligns input questions more closely with the knowledge base. Our research identifies opportunities to enhance the Query Rewriter module to Query Rewriter+ by generating multiple queries to overcome the Information Plateaus associated with a single query and by rewriting questions to eliminate Ambiguity, thereby clarifying the underlying intent. We also find that current RAG systems exhibit issues with Irrelevant Knowledge; to overcome this, we propose the Knowledge Filter. These two modules are both based on the instruction-tuned Gemma-2B model, which together enhance response quality. The final identified issue is Redundant Retrieval; we introduce the Memory Knowledge Reservoir and the Retriever Trigger to solve this. The former supports the dynamic expansion of the RAG system's knowledge base in a parameter-free manner, while the latter optimizes the cost for accessing external knowledge, thereby improving resource utilization and response efficiency. These four RAG modules synergistically improve the response quality and efficiency of the RAG system. The effectiveness of these modules has been validated through experiments and ablation studies across six common QA datasets. The source code can be accessed at https://github.com/Ancientshi/ERM4.

Yunxiao Shi, Wujiang Xu, Haimin Zhang et al.

Modeling high-order feature interactions is crucial for click-through rate (CTR) prediction, yet traditional approaches typically predefine a maximum interaction order and exhaustively enumerate feature combinations up to that order. This paradigm depends heavily on prior domain knowledge to delimit the interaction space and incurs substantial computational overhead. As a result, conventional CTR models face a persistent tension between enriching representations with complex high-order interactions and keeping computation tractable. To address this dual challenge, this study introduces the Kolmogorov-Arnold Represented Sparse Efficient Interaction Network (KarSein). Drawing inspiration from the learnable activation mechanism in the Kolmogorov-Arnold Network (KAN), KarSein leverages this mechanism to adaptively transform low-order basic features into high-order feature interactions, offering a novel approach to feature interaction modeling. KarSein extends the capabilities of KAN by introducing a more efficient architecture that significantly reduces computational costs while accommodating two-dimensional embedding vectors as feature inputs. Furthermore, it overcomes the limitation of KAN's its inability to spontaneously capture multiplicative relationships among features. Extensive experiments highlight the superiority of KarSein, demonstrating its ability to surpass not only the vanilla implementation of KAN in CTR prediction tasks but also other baseline methods. Remarkably, KarSein achieves exceptional predictive accuracy while maintaining a highly compact parameter size and minimal computational overhead. Moreover, KarSein exhibits strong interpretability and structural sparsity. As the first systematic adaptation of KAN to CTR prediction, KarSein offers a practical, parameter-efficient, and interpretable alternative for modeling complex feature interactions in CTR prediction.

Jiahao Xia, Wenjian Huang, Min Xu et al.

Object parts serve as crucial intermediate representations in various downstream tasks, but part-level representation learning still has not received as much attention as other vision tasks. Previous research has established that Vision Transformer can learn instance-level attention without labels, extracting high-quality instance-level representations for boosting downstream tasks. In this paper, we achieve unsupervised part-specific attention learning using a novel paradigm and further employ the part representations to improve part discovery performance. Specifically, paired images are generated from the same image with different geometric transformations, and multiple part representations are extracted from these paired images using a novel module, named PartFormer. These part representations from the paired images are then exchanged to improve geometric transformation invariance. Subsequently, the part representations are aligned with the feature map extracted by a feature map encoder, achieving high similarity with the pixel representations of the corresponding part regions and low similarity in irrelevant regions. Finally, the geometric and semantic constraints are applied to the part representations through the intermediate results in alignment for part-specific attention learning, encouraging the PartFormer to focus locally and the part representations to explicitly include the information of the corresponding parts. Moreover, the aligned part representations can further serve as a series of reliable detectors in the testing phase, predicting pixel masks for part discovery. Extensive experiments are carried out on four widely used datasets, and our results demonstrate that the proposed method achieves competitive performance and robustness due to its part-specific attention.

Lu Huo, Jiahao Xia, Leijie Zhang et al.

Existing multiple modality fusion methods, such as concatenation, summation, and encoder-decoder-based fusion, have recently been employed to combine modality characteristics of Hyperspectral Image (HSI) and Light Detection And Ranging (LiDAR). However, these methods consider the relationship of HSI-LiDAR signals from limited perspectives. More specifically, they overlook the contextual information across modalities of HSI and LiDAR and the intra-modality characteristics of LiDAR. In this paper, we provide a new insight into feature fusion to explore the relationships across HSI and LiDAR modalities comprehensively. An Interconnected Fusion (IF) framework is proposed. Firstly, the center patch of the HSI input is extracted and replicated to the size of the HSI input. Then, nine different perspectives in the fusion matrix are generated by calculating self-attention and cross-attention among the replicated center patch, HSI input, and corresponding LiDAR input. In this way, the intra- and inter-modality characteristics can be fully exploited, and contextual information is considered in both intra-modality and inter-modality manner. These nine interrelated elements in the fusion matrix can complement each other and eliminate biases, which can generate a multi-modality representation for classification accurately. Extensive experiments have been conducted on three widely used datasets: Trento, MUUFL, and Houston. The IF framework achieves state-of-the-art results on these datasets compared to existing approaches.

Haimin Zhang, Jiahao Xia, Min Xu

Combining the message-passing paradigm with the global attention mechanism has emerged as an effective framework for learning over graphs. The message-passing paradigm and the global attention mechanism fundamentally generate node embeddings based on information aggregated from a node's local neighborhood or from the whole graph. The most basic and commonly used aggregation approach is to take the sum of information from a node's local neighbourhood or from the whole graph. However, it is unknown if the dominant information is from a node itself or from the node's neighbours (or the rest of the graph nodes). Therefore, there exists information lost at each layer of embedding generation, and this information lost could be accumulated and become more serious when more layers are used in the model. In this paper, we present a differential encoding method to address the issue of information lost. The idea of our method is to encode the differential representation between the information from a node's neighbours (or the rest of the graph nodes) and that from the node itself. The obtained differential encoding is then combined with the original aggregated local or global representation to generate the updated node embedding. By integrating differential encodings, the representational ability of generated node embeddings is improved. The differential encoding method is empirically evaluated on different graph tasks on seven benchmark datasets. The results show that it is a general method that improves the message-passing update and the global attention update, advancing the state-of-the-art performance for graph representation learning on these datasets.

Jiahao Xia, Min Xu, Wenjian Huang et al.

Despite the similar structures of human faces, existing face alignment methods cannot learn unified knowledge from multiple datasets with different landmark annotations. The limited training samples in a single dataset commonly result in fragile robustness in this field. To mitigate knowledge discrepancies among different datasets and train a task-agnostic unified face alignment (TUFA) framework, this paper presents a strategy to unify knowledge from multiple datasets. Specifically, we calculate a mean face shape for each dataset. To explicitly align these mean shapes on an interpretable plane based on their semantics, each shape is then incorporated with a group of semantic alignment embeddings. The 2D coordinates of these aligned shapes can be viewed as the anchors of the plane. By encoding them into structure prompts and further regressing the corresponding facial landmarks using image features, a mapping from the plane to the target faces is finally established, which unifies the learning target of different datasets. Consequently, multiple datasets can be utilized to boost the generalization ability of the model. The successful mitigation of discrepancies also enhances the efficiency of knowledge transferring to a novel dataset, significantly boosts the performance of few-shot face alignment. Additionally, the interpretable plane endows TUFA with a task-agnostic characteristic, enabling it to locate landmarks unseen during training in a zero-shot manner. Extensive experiments are carried on seven benchmarks and the results demonstrate an impressive improvement in face alignment brought by knowledge discrepancies mitigation. The code is available at https://github.com/Jiahao-UTS/TUFA.

Sarthak Anand, Debanjan Mahata, Kartik Aggarwal et al.

In this paper we present our approach and the system description for Sub Task A of SemEval 2019 Task 9: Suggestion Mining from Online Reviews and Forums. Given a sentence, the task asks to predict whether the sentence consists of a suggestion or not. Our model is based on Universal Language Model Fine-tuning for Text Classification. We apply various pre-processing techniques before training the language and the classification model. We further provide detailed analysis of the results obtained using the trained model. Our team ranked 10th out of 34 participants, achieving an F1 score of 0.7011. We publicly share our implementation at https://github.com/isarth/SemEval9_MIDAS

Haimin Zhang, Min Xu

Studies continually find that message-passing graph convolutional networks suffer from the over-smoothing issue. Basically, the issue of over-smoothing refers to the phenomenon that the learned embeddings for all nodes can become very similar to one another and therefore are uninformative after repeatedly applying message passing iterations. Intuitively, we can expect the generated embeddings become smooth asymptotically layerwisely, that is each layer of graph convolution generates a smoothed version of embeddings as compared to that generated by the previous layer. Based on this intuition, we propose RandAlign, a stochastic regularization method for graph convolutional networks. The idea of RandAlign is to randomly align the learned embedding for each node with that of the previous layer using randomly interpolation in each graph convolution layer. Through alignment, the smoothness of the generated embeddings is explicitly reduced. To better maintain the benefit yielded by the graph convolution, in the alignment step we introduce to first scale the embedding of the previous layer to the same norm as the generated embedding and then perform random interpolation for aligning the generated embedding. RandAlign is a parameter-free method and can be directly applied without introducing additional trainable weights or hyper-parameters. We experimentally evaluate RandAlign on different graph domain tasks on seven benchmark datasets. The experimental results show that RandAlign is a general method that improves the generalization performance of various graph convolutional network models and also improves the numerical stability of optimization, advancing the state of the art performance for graph representation learning.

Yongze Wang, Haimin Zhang, Qiang Wu et al.

Graph neural networks (GNNs) have shown great success in learning from graph-based data. The key mechanism of current GNNs is message passing, where a node's feature is updated based on the information passing from its local neighbourhood. A limitation of this mechanism is that node features become increasingly dominated by the information aggregated from the neighbourhood as we use more rounds of message passing. Consequently, as the GNN layers become deeper, adjacent node features tends to be similar, making it more difficult for GNNs to distinguish adjacent nodes, thereby, limiting the performance of GNNs. In this paper, we propose conditional local feature encoding (CLFE) to help prevent the problem of node features being dominated by the information from local neighbourhood. The idea of our method is to extract the node hidden state embedding from message passing process and concatenate it with the nodes feature from previous stage, then we utilise linear transformation to form a CLFE based on the concatenated vector. The CLFE will form the layer output to better preserve node-specific information, thus help to improve the performance of GNN models. To verify the feasibility of our method, we conducted extensive experiments on seven benchmark datasets for four graph domain tasks: super-pixel graph classification, node classification, link prediction, and graph regression. The experimental results consistently demonstrate that our method improves model performance across a variety of baseline GNN models for all four tasks.

Yunxiao Shi, Shuo Yang, Haimin Zhang et al.

Neural Collaborative Filtering models are widely used in recommender systems but are typically trained under static settings, assuming fixed data distributions. This limits their applicability in dynamic environments where user preferences evolve. Incremental learning offers a promising solution, yet conventional methods from computer vision or NLP face challenges in recommendation tasks due to data sparsity and distinct task paradigms. Existing approaches for neural recommenders remain limited and often lack generalizability. To address this, we propose MEGG, Replay Samples with Maximally Extreme GGscore, an experience replay based incremental learning framework. MEGG introduces GGscore, a novel metric that quantifies sample influence, enabling the selective replay of highly influential samples to mitigate catastrophic forgetting. Being model-agnostic, MEGG integrates seamlessly across architectures and frameworks. Experiments on three neural models and four benchmark datasets show superior performance over state-of-the-art baselines, with strong scalability, efficiency, and robustness. Implementation will be released publicly upon acceptance.

Yunxiao Shi, Xing Zi, Zijing Shi et al.

Retrieval-augmented generation (RAG) for language models significantly improves language understanding systems. The basic retrieval-then-read pipeline of response generation has evolved into a more extended process due to the integration of various components, sometimes even forming loop structures. Despite its advancements in improving response accuracy, challenges like poor retrieval quality for complex questions that require the search of multifaceted semantic information, inefficiencies in knowledge re-retrieval during long-term serving, and lack of personalized responses persist. Motivated by transcending these limitations, we introduce ERAGent, a cutting-edge framework that embodies an advancement in the RAG area. Our contribution is the introduction of the synergistically operated module: Enhanced Question Rewriter and Knowledge Filter, for better retrieval quality. Retrieval Trigger is incorporated to curtail extraneous external knowledge retrieval without sacrificing response quality. ERAGent also personalizes responses by incorporating a learned user profile. The efficiency and personalization characteristics of ERAGent are supported by the Experiential Learner module which makes the AI assistant being capable of expanding its knowledge and modeling user profile incrementally. Rigorous evaluations across six datasets and three question-answering tasks prove ERAGent's superior accuracy, efficiency, and personalization, emphasizing its potential to advance the RAG field and its applicability in practical systems.

Haimin Zhang, Min Xu

Message passing has become the dominant framework in graph representation learning. The essential idea of the message-passing framework is to update node embeddings based on the information aggregated from local neighbours. However, most existing aggregation methods have not encoded neighbour-level message interactions into the aggregated message, resulting in an information lost in embedding generation. And this information lost could be accumulated and become more serious as more layers are added to the graph network model. To address this issue, we propose a neighbour-level message interaction information encoding method for improving graph representation learning. For messages that are aggregated at a node, we explicitly generate an encoding between each message and the rest messages using an encoding function. Then we aggregate these learned encodings and take the sum of the aggregated encoding and the aggregated message to update the embedding for the node. By this way, neighbour-level message interaction information is integrated into the generated node embeddings. The proposed encoding method is a generic method which can be integrated into message-passing graph convolutional networks. Extensive experiments are conducted on six popular benchmark datasets across four highly-demanded tasks. The results show that integrating neighbour-level message interactions achieves improved performance of the base models, advancing the state of the art results for representation learning over graphs.

Rakesh Gosangi, Ravneet Arora, Mohsen Gheisarieha et al.

In this paper, we study the importance of context in predicting the citation worthiness of sentences in scholarly articles. We formulate this problem as a sequence labeling task solved using a hierarchical BiLSTM model. We contribute a new benchmark dataset containing over two million sentences and their corresponding labels. We preserve the sentence order in this dataset and perform document-level train/test splits, which importantly allows incorporating contextual information in the modeling process. We evaluate the proposed approach on three benchmark datasets. Our results quantify the benefits of using context and contextual embeddings for citation worthiness. Lastly, through error analysis, we provide insights into cases where context plays an essential role in predicting citation worthiness.

Jiahao Xia, Haimin Zhang, Shiping Wen et al.

While Convolutional Neural Networks (CNNs) have significantly boosted the performance of face related algorithms, maintaining accuracy and efficiency simultaneously in practical use remains challenging. The state-of-the-art methods employ deeper networks for better performance, which makes it less practical for mobile applications because of more parameters and higher computational complexity. Therefore, we propose an efficient multitask neural network, Alignment & Tracking & Pose Network (ATPN) for face alignment, face tracking and head pose estimation. Specifically, to achieve better performance with fewer layers for face alignment, we introduce a shortcut connection between shallow-layer and deep-layer features. We find the shallow-layer features are highly correspond to facial boundaries that can provide the structural information of face and it is crucial for face alignment. Moreover, we generate a cheap heatmap based on the face alignment result and fuse it with features to improve the performance of the other two tasks. Based on the heatmap, the network can utilize both geometric information of landmarks and appearance information for head pose estimation. The heatmap also provides attention clues for face tracking. The face tracking task also saves us the face detection procedure for each frame, which also significantly boost the real-time capability for video-based tasks. We experimentally validate ATPN on four benchmark datasets, WFLW, 300VW, WIDER Face and 300W-LP. The experimental results demonstrate that it achieves better performance with much less parameters and lower computational complexity compared to other light models.

Haimin Zhang, Min Xu

Deep neural networks have been successfully applied in various machine learning tasks. However, studies show that neural networks are susceptible to adversarial attacks. This exposes a potential threat to neural network-based intelligent systems. We observe that the probability of the correct result outputted by the neural network increases by applying small first-order perturbations generated for non-predicted class labels to adversarial examples. Based on this observation, we propose a method for counteracting adversarial perturbations to improve adversarial robustness. In the proposed method, we randomly select a number of class labels and generate small first-order perturbations for these selected labels. The generated perturbations are added together and then clamped onto a specified space. The obtained perturbation is finally added to the adversarial example to counteract the adversarial perturbation contained in the example. The proposed method is applied at inference time and does not require retraining or finetuning the model. We experimentally validate the proposed method on CIFAR-10 and CIFAR-100. The results demonstrate that our method effectively improves the defense performance of several transformation-based defense methods, especially against strong adversarial examples generated using more iterations.

Haimin Zhang, Min Xu, Guoqiang Zhang et al.

Graph convolutional networks have been successfully applied in various graph-based tasks. In a typical graph convolutional layer, node features are updated by aggregating neighborhood information. Repeatedly applying graph convolutions can cause the oversmoothing issue, i.e., node features at deep layers converge to similar values. Previous studies have suggested that oversmoothing is one of the major issues that restrict the performance of graph convolutional networks. In this paper, we propose a stochastic regularization method to tackle the oversmoothing problem. In the proposed method, we stochastically scale features and gradients (SSFG) by a factor sampled from a probability distribution in the training procedure. By explicitly applying a scaling factor to break feature convergence, the oversmoothing issue is alleviated. We show that applying stochastic scaling at the gradient level is complementary to that applied at the feature level to improve the overall performance. Our method does not increase the number of trainable parameters. When used together with ReLU, our SSFG can be seen as a stochastic ReLU activation function. We experimentally validate our SSFG regularization method on three commonly used types of graph networks. Extensive experimental results on seven benchmark datasets for four graph-based tasks demonstrate that our SSFG regularization is effective in improving the overall performance of the baseline graph networks.

Sarthak Anand, Pradyumna Gupta, Hemant Yadav et al.

This paper presents our submission to the SemEval 2020 - Task 10 on emphasis selection in written text. We approach this emphasis selection problem as a sequence labeling task where we represent the underlying text with various contextual embedding models. We also employ label distribution learning to account for annotator disagreements. We experiment with the choice of model architectures, trainability of layers, and different contextual embeddings. Our best performing architecture is an ensemble of different models, which achieved an overall matching score of 0.783, placing us 15th out of 31 participating teams. Lastly, we analyze the results in terms of parts of speech tags, sentence lengths, and word ordering.

Dhruva Sahrawat, Debanjan Mahata, Mayank Kulkarni et al.

In this paper, we formulate keyphrase extraction from scholarly articles as a sequence labeling task solved using a BiLSTM-CRF, where the words in the input text are represented using deep contextualized embeddings. We evaluate the proposed architecture using both contextualized and fixed word embedding models on three different benchmark datasets (Inspec, SemEval 2010, SemEval 2017) and compare with existing popular unsupervised and supervised techniques. Our results quantify the benefits of (a) using contextualized embeddings (e.g. BERT) over fixed word embeddings (e.g. Glove); (b) using a BiLSTM-CRF architecture with contextualized word embeddings over fine-tuning the contextualized word embedding model directly, and (c) using genre-specific contextualized embeddings (SciBERT). Through error analysis, we also provide some insights into why particular models work better than others. Lastly, we present a case study where we analyze different self-attention layers of the two best models (BERT and SciBERT) to better understand the predictions made by each for the task of keyphrase extraction.

Avinash Swaminathan, Raj Kuwar Gupta, Haimin Zhang et al.

In this paper, we present a keyphrase generation approach using conditional Generative Adversarial Networks (GAN). In our GAN model, the generator outputs a sequence of keyphrases based on the title and abstract of a scientific article. The discriminator learns to distinguish between machine-generated and human-curated keyphrases. We evaluate this approach on standard benchmark datasets. Our model achieves state-of-the-art performance in generation of abstractive keyphrases and is also comparable to the best performing extractive techniques. We also demonstrate that our method generates more diverse keyphrases and make our implementation publicly available.

Ozan İrsoy, Rakesh Gosangi, Haimin Zhang et al.

Dialogue act (DA) classification has been studied for the past two decades and has several key applications such as workflow automation and conversation analytics. Researchers have used, to address this problem, various traditional machine learning models, and more recently deep neural network models such as hierarchical convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. In this paper, we introduce a new model architecture, directed-acyclic-graph LSTM (DAG-LSTM) for DA classification. A DAG-LSTM exploits the turn-taking structure naturally present in a multi-party conversation, and encodes this relation in its model structure. Using the STAC corpus, we show that the proposed method performs roughly 0.8% better in accuracy and 1.2% better in macro-F1 score when compared to existing methods. The proposed method is generic and not limited to conversation applications.

Haimin Zhang, Debanjan Mahata, Simra Shahid et al.

In this paper we present our approach and the system description for Sub-task A and Sub Task B of SemEval 2019 Task 6: Identifying and Categorizing Offensive Language in Social Media. Sub-task A involves identifying if a given tweet is offensive or not, and Sub Task B involves detecting if an offensive tweet is targeted towards someone (group or an individual). Our models for Sub-task A is based on an ensemble of Convolutional Neural Network, Bidirectional LSTM with attention, and Bidirectional LSTM + Bidirectional GRU, whereas for Sub-task B, we rely on a set of heuristics derived from the training data and manual observation. We provide detailed analysis of the results obtained using the trained models. Our team ranked 5th out of 103 participants in Sub-task A, achieving a macro F1 score of 0.807, and ranked 8th out of 75 participants in Sub Task B achieving a macro F1 of 0.695.

Haimin Zhang, Min Xu

Motivated by the observation that humans can learn patterns from two given images at one time, we propose a dual pattern learning network architecture in this paper. Unlike conventional networks, the proposed architecture has two input branches and two loss functions. Instead of minimizing the empirical risk of a given dataset, dual pattern learning networks is trained by minimizing the empirical dual prediction loss. We show that this can improve the performance for single image classification. This architecture forces the network to learn discriminative class-specific features by analyzing and comparing two input images. In addition, the dual input structure allows the network to have a considerably large number of image pairs, which can help address the overfitting issue due to limited training data. Moreover, we propose to associate each input branch with a random interest value for learning corresponding image during training. This method can be seen as a stochastic regularization technique, and can further lead to generalization performance improvement. State-of-the-art deep networks can be adapted to dual pattern learning networks without increasing the same number of parameters. Extensive experiments on CIFAR-10, CIFAR- 100, FI-8, Google commands dataset, and MNIST demonstrate that our DPLNets exhibit better performance than original networks. The experimental results on subsets of CIFAR- 10, CIFAR-100, and MNIST demonstrate that dual pattern learning networks have good generalization performance on small datasets.

Haimin Zhang, Min Xu

With the widespread of user-generated Internet videos, emotion recognition in those videos attracts increasing research efforts. However, most existing works are based on framelevel visual features and/or audio features, which might fail to model the temporal information, e.g. characteristics accumulated along time. In order to capture video temporal information, in this paper, we propose to analyse features in frequency domain transformed by discrete Fourier transform (DFT features). Frame-level features are firstly extract by a pre-trained deep convolutional neural network (CNN). Then, time domain features are transferred and interpolated into DFT features. CNN and DFT features are further encoded and fused for emotion classification. By this way, static image features extracted from a pre-trained deep CNN and temporal information represented by DFT features are jointly considered for video emotion recognition. Experimental results demonstrate that combining DFT features can effectively capture temporal information and therefore improve emotion recognition performance. Our approach has achieved a state-of-the-art performance on the largest video emotion dataset (VideoEmotion-8 dataset), improving accuracy from 51.1% to 62.6%.

Haimin Zhang

Recently, video classification attracts intensive research efforts. However, most existing works are based on framelevel visual features, which might fail to model the temporal information, e.g. characteristics accumulated along time. In order to capture video temporal information, we propose to analyse features in frequency domain transformed by discrete Fourier transform (DFT features). Frame-level features are firstly extract by a pre-trained deep convolutional neural network (CNN). Then, time domain features are transformed and interpolated into DFT features. CNN and DFT features are further encoded by using different pooling methods and fused for video classification. In this way, static image features extracted from a pre-trained deep CNN and temporal information represented by DFT features are jointly considered for video classification. We test our method for video emotion classification and action recognition. Experimental results demonstrate that combining DFT features can effectively capture temporal information and therefore improve the performance of both video emotion classification and action recognition. Our approach has achieved a state-of-the-art performance on the largest video emotion dataset (VideoEmotion-8 dataset) and competitive results on UCF-101.