The interpretative efficacy of large language models (LLMs) fundamentally hinges on the intricate alignment between user inputs and model-specific linguistic priors. Existing methodologies predominantly employ static input optimization strategies, failing to account for the empirically observed divergence in linguistic preference spaces across distinct LLM architectures, including variations in syntactic parsing heuristics, semantic grounding mechanisms, and knowledge retrieval pathways. We propose QueryAligner, an adaptive rewriting system implementing dynamic model-aware input transformation through architecture-specific preference modeling. Our framework introduces two pivotal innovations: 1) A dual-phase optimization engine integrating supervised learning on reverse-engineered cross-architectural training data with reinforcement learning driven by multi-objective reward signals, ensuring simultaneous preservation of semantic integrity and maximization of target model compatibility; 2) An architecture-informed rewriting protocol that automatically discovers latent alignment patterns encoded within distinct LLMs' parametric configurations. Experimental results demonstrate that our method achieves superior performance compared to conventional input optimization techniques.

Empathy relies on the cognitive capacity to relate to similar past experiences. Consequently, retrieval-based approaches utilize analogous exemplars to guide empathetic dialogue generation. However, existing methods prioritize semantic similarity over emotion characteristics, often leading to unempathetic responses. To address this, we propose REG, a framework that integrates four Emotion Attributes into the retrieval process to ensure explicit emotional alignment. Furthermore, to mitigate the noise and limited diversity caused by coarse-grained sentence-level attributes, we incorporate Token-level Retrieval for finer granularity and a Retrieval Candidate Augmentation strategy to enhance diversity. Empirical results on the EmpatheticDialogues dataset demonstrate that REG significantly outperforms baselines, offering a robust solution for empathetic generation.

Role-playing systems powered by large language models (LLMs) have become increasingly influential in emotional communication applications. However, these systems are susceptible to character hallucinations, where the model deviates from predefined character roles and generates responses that are inconsistent with the intended persona. This paper presents the first systematic analysis of character hallucination from an attack perspective, introducing the RoleBreak framework. Our framework identifies two core mechanisms—query sparsity and role-query conflict—as key factors driving character hallucination. Leveraging these insights, we construct a novel dataset, RoleBreakEval, to evaluate existing hallucination mitigation techniques. Our experiments reveal that even enhanced models trained to minimize hallucination remain vulnerable to attacks. To address these vulnerabilities, we propose a novel defence strategy, the Narrator Mode, which generates supplemental context through narration to mitigate role-query conflicts and improve query generalization. Experimental results demonstrate that Narrator Mode significantly outperforms traditional refusal-based strategies by reducing hallucinations, enhancing fidelity to character roles and queries, and improving overall narrative coherence.

Empathy improves human-machine dialogue systems by enhancing the user’s experience. While traditional models have aimed to detect and express users’ emotions from dialogue history, they neglect the crucial and complex interactions among emotion, emotion causes, and commonsense. To address this, we introduce the ECC (Emotion, Cause, and Commonsense) framework, which leverages specialized encoders to capture the key features of emotion, cause, and commonsense and collaboratively models these through a Conditional Variational Auto-Encoder. ECC further employs novel loss functions to refine the interplay of three factors and generates empathetic responses using an energy-based model supported by ODE sampling. Empirical results on the EmpatheticDialogues dataset demonstrate that ECC outperforms existing baselines, offering a robust solution for empathetic dialogue generation.

Current research shows that providing instructions to guide Large Language Models (LLMs) improves reasoning tasks, but existing methods struggle to balance accuracy and generalization. Manually crafted instructions tailored to specific LLMs and tasks improve performance but reduce generalizability, while more general instructions lack detail and lower performance. To address this, we propose a dynamic instruction-generation method using an Instruction-Generation Prompt (IGP). IGP categorizes problems into domains and integrates the model's capabilities to generate detailed task-specific instructions, resulting in a comprehensive plan. This approach achieves high precision with general prompts without requiring in-depth knowledge of LLMs or tasks. We validated our method across five LLMs and ten datasets in three task categories. Our dynamically generated instructions outperformed traditionally handcrafted, LLM-specific instructions across various LLMs and tasks.

Emotion recognition in conversation (ERC) is a field that aims to classify the emotion of each utterance within conversational contexts. This presents significant challenges, particularly in handling emotional ambiguity across various speakers and contextual factors. Existing ERC approaches have primarily focused on modeling conversational contexts while incorporating only superficial speaker attributes such as names, memories, and interactions. Recent works introduce personality as an essential deep speaker factor for emotion recognition, but relies on static personality, overlooking dynamic variability during conversations. Advances in personality psychology conceptualize personality as dynamic, proposing that personality states can change across situations. In this paper, we introduce ERC-DP, a novel model considering the dynamic personality of speakers during conversations. ERC-DP accounts for past utterances from the same speaker as situation impacting dynamic personality. It combines personality modeling with prompt design and fine-grained classification modules. Through a series of comprehensive experiments, ERC-DP demonstrates superior performance on three benchmark conversational datasets.

The few-shot tasks require the model to have the ability to generalize from a few samples. However, due to the lack of cognitive ability, the current works cannot fully utilize limited samples to expand the sample space and still suffer from overfitting issues. To address the problems, we propose a LLM-Augmented Unsupervised Contrastive Learning Framework (LA-UCL), which introduces a cognition-enabled Large Language Model (LLM) for efficient data augmentation, and presents corresponding contrastive learning strategies. Specifically, in the self-augmented contrastive learning module, we construct a retrieval-based in-context prompt scheme by retrieving similar but different category data from the original samples, guiding the LLM to generate more discriminative augmented data. Then, by designing group-level contrastive loss to enhance the model’s discriminative ability. In the external-augmented contrastive learning module, we utilize web knowledge retrieval to expand the sample space and leverage LLM to generate more diverse data, and introduce sample-level contrastive loss for unlabeled data to improve the model’s generalization. Experimental results on six datasets show that our model exceeds the baseline models.

While extensive work has examined the explicit and implicit biases in large language models (LLMs), little research explores the relation between these two types of biases. This paper presents a comparative study of the explicit and implicit biases in LLMs grounded in social psychology. Social psychology distinguishes between explicit and implicit biases by whether the bias can be self-recognized by individuals. Aligning with this conceptualization, we propose a self-evaluation-based two-stage measurement of explicit and implicit biases within LLMs. First, the LLM is prompted to automatically fill templates with social targets to measure implicit bias toward these targets, where the bias is less likely to be self-recognized by the LLM. Then, the LLM is prompted to self-evaluate the templates filled by itself to measure explicit bias toward the same targets, where the bias is more likely to be self-recognized by the LLM. Experiments conducted on state-of-the-art LLMs reveal human-like inconsistency between explicit and implicit occupational gender biases. This work bridges a critical gap where prior studies concentrate solely on either explicit or implicit bias. We advocate that future work highlight the relation between explicit and implicit biases in LLMs.

Personalized Dialogue Generation (PDG) aims to create coherent responses according to roles or personas.Traditional PDG relies on external role data, which can be scarce and raise privacy concerns.Approaches address these issues by extracting role information from dialogue history, which often fail to generically model roles in continuous space.To overcome these limitations, we introduce a novel framework MOdels Roles from Personalized Dialogue History by Exploring and Utilizing Latent Space (MOR-PHEUS) through a three-stage training process.Specifically, we create a persona codebook to represent roles in latent space compactly, and this codebook is used to construct a posterior distribution of role information.This method enables the model to generalize across roles, allowing the generation of personalized dialogues even for unseen roles.Experiments on both Chinese and English datasets demonstrate that MORPHEUS enhances the extraction of role information, and improves response generation without external role data.Additionally, MORPHEUS can be considered an efficient fine-tuning for large language models.* Morpheus excels at imitating others' forms, effortlessly embodying their physicality, posture, facial expressions, and vocal tone.He flawlessly embodies every intricate detail, including attire and unique mannerisms.

The capability to disentangle underlying factors hidden in the observable data, thereby obtaining their abstract representations, is considered one important ingredient for the subsequent success of deep networks in various application scenarios. Recently, numerous practical measures and learning strategies have been established for disentanglement, showcasing their potential in improving the model’s explainability, controlability, and robustness. However, when the downstream tasks come to the classification issues, there is still no consensus in the community on the definition or measurement for disentanglement, and its connection to the generalization capacity remains not very clear. Aiming at this, we explore the highly non-linear effect of a specified hidden layer on the generalization capacity from an information perspective and obtain a tight bound. Upon decompsing the bound, we find that besides the unsupervised disentanglement measure term in the conventional sense, a new supervised disentanglement term also emerges with a nonnegligible effect on the generality. Consequently, a novel label-based disentanglement measure (LDM) is naturally introduced as the discrepancy between these two terms under the supervised learning settings to substitute the commonly used unsupervised disentanglement measure. The theoretical analysis reveals an inverse relationship between the defined LDM and the generalization capacity. Finally, using LDM as regularizer, the experiments show that the deep neural networks (DNNs) can effectively reduce generalization error while improving classification accuracy when noise is added to the data features or labels, which strongly supports our points.

Sentiment and emotion, which correspond to long-term and short-lived human feelings, are closely linked to each other, leading to the fact that sentiment analysis and emotion recognition are also two interdependent tasks in natural language processing (NLP). One task often leverages the shared knowledge from another task and performs better when solved in a joint learning paradigm. Conversational context dependency, multi-modal interaction, and multi-task correlation are three key factors that contribute to this joint paradigm. However, none of the recent approaches have considered them in a unified framework. To fill this gap, we propose a multi-modal, multi-task interactive graph attention network, termed M3GAT, to simultaneously solve the three problems. At the heart of the model is a proposed interactive conversation graph layer containing three core sub-modules, which are: (1) local-global context connection for modeling both local and global conversational context, (2) cross-modal connection for learning multi-modal complementary and (3) cross-task connection for capturing the correlation across two tasks. Comprehensive experiments on three benchmarking datasets, MELD, MEISD, and MSED, show the effectiveness of M3GAT over state-of-the-art baselines with the margin of 1.88%, 5.37%, and 0.19% for sentiment analysis, and 1.99%, 3.65%, and 0.13% for emotion recognition, respectively. In addition, we also show the superiority of multi-task learning over the single-task framework.

Yihong Tang, Bo Wang, Miao Fang, Dongming Zhao, Kun Huang, Ruifang He, Yuexian Hou. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

For human-like dialogue systems, it is significant to inject the empathetic ability or elicit the opposite’s positive emotions, while existing studies mostly only focus on either of the above two research lines. In this work, we propose a novel and grafted task named Empathetic Emotion Elicitation Dialog to make a dialog system able to possess both aspects of ability simultaneously. We do not train an empathetic dialog system and an emotion elicitation dialog system separately and then simply concatenate the responses generated by these two systems, which will cause illogical and repetitive responses. Instead, we propose a unified solution: (1) To generate empathetic responses and emotion elicitation responses within the same semantic space, we design a unified framework. (2) The unified framework has three stages which first retrieve the empathetic and emotion elicitation exemplars as external knowledge, then fine-tune the emotion/action prediction on a pre-trained language model to enhance the empathetic ability, and finally model the user feedback by reinforcement learning to enhance the emotion elicitation ability. Experiments show that our method outperforms the baselines in the response generation quality and simultaneously empathizes with the user and elicits their positive emotions.

Empathetic dialogue is a human-like behavior that requires the perception of both affective factors (e.g., emotion status) and cognitive factors (e.g., cause of the emotion). Besides concerning emotion status in early work, the latest approaches study emotion causes in empathetic dialogue. These approaches focus on understanding and duplicating emotion causes in the context to show empathy for the speaker. However, instead of only repeating the contextual causes, the real empathic response often demonstrate a logical and emotion-centered transition from the causes in the context to those in the responses. In this work, we propose an emotion cause transition graph to explicitly model the natural transition of emotion causes between two adjacent turns in empathetic dialogue. With this graph, the concept words of the emotion causes in the next turn can be predicted and used by a specifically designed concept-aware decoder to generate the empathic response. Automatic and human experimental results on the benchmark dataset demonstrate that our method produces more empathetic, coherent, informative, and specific responses than existing models.

Target-oriented dialog aims to reach a global target through multi-turn conversation. The key to the task is the global planning towards the target, which flexibly guides the dialog concerning the context. However, existing target-oriented dialog works take a local and greedy strategy for response generation, where global planning is absent. In this work, we propose global planning for target-oriented dialog on a commonsense knowledge graph (KG). We design a global reinforcement learning with the planned paths to flexibly adjust the local response generation model towards the global target. We also propose a KG-based method to collect target-oriented samples automatically from the chit-chat corpus for model training. Experiments show that our method can reach the target with a higher success rate, fewer turns, and more coherent responses.

Conversational recommendation systems (CRS) aim to determine a goal item by sequentially tracking users’ interests through multi-turn conversation. In CRS, implicit patterns of user interest sequence guide the smooth transition of dialog utterances to the goal item. However, with the convenient explicit knowledge of knowledge graph (KG), existing KG-based CRS methods over-rely on the explicit separate KG links to model the user interests but ignore the rich goal-aware implicit interest sequence patterns in a dialog. In addition, interest sequence is also not fully used to generate smooth transited utterances. We propose CR-GIS with a parallel star framework. First, an interest-level star graph is designed to model the goal-aware implicit user interest sequence. Second, a hierarchical Star Transformer is designed to guide the multi-turn utterances generation with the interest-level star graph. Extensive experiments verify the effectiveness of CR-GIS in achieving more accurate recommended items with more fluent and coherent dialog utterances.

Human conversations of recommendation naturally involve the shift of interests which can align the recommendation actions and conversation process to make accurate recommendations with rich explanations.However, existing conversational recommendation systems (CRS) ignore the advantage of user interest shift in connecting recommendation and conversation, which leads to an ineffective loose coupling structure of CRS.To address this issue, by modeling the recommendation actions as recommendation paths in a knowledge graph (KG), we propose DICR (Dual Imitation for Conversational Recommendation), which designs a dual imitation to explicitly align the recommendation paths and user interest shift paths in a recommendation module and a conversation module, respectively.By exchanging alignment signals, DICR achieves bidirectional promotion between recommendation and conversation modules and generates high-quality responses with accurate recommendations and coherent explanations.Experiments demonstrate that DICR outperforms the state-of-theart models on recommendation and conversation performance with automatic, human, and novel explainability metrics.

The performance of Feedforward Neural Networks (FNNs) in specific application areas is often higher than the upper limit allowed by (classical) theory, and therefore is called “Alchemy”, that is, the lack of interpretability of FNNs. In this paper, based on the Quantum Weak Measurement theory (QWM), we propose an interpretable network with complex-valued representation to deal with the insufficient interpretability of FNNs and the Semantic Interpretability of Complex Numbers (SICN) from a novel perspective. For the specific role of complex-valued representation in capturing semantic information, we give an in-depth analysis that the real and imaginary parts of complex numbers can simultaneously represent two different types of complementary information. Finally, we propose a QWM inspired interpretable Network for text representation learning (i.e., QWMN) and implement two frameworks for Question Answering (QWMN-QA) and Text Classification (QWMN-TC) respectively. Experimental results prove the feasibility, effectiveness of QWMN.

Recently, a Quantum Probability Drive Network (QPDN) is proposed to model different levels of semantic units by extending word embedding to complex-valued representation (CR). The extended complex-valued embeddings are still insensitive to polarity causing that they generalize badly in sentiment analysis (SA). To solve it, we propose a method of encoding sentiment information into sentiment words for SA. Attention mechanism and an auxiliary task are introduced to help learn the CR of sentiment words with the help of the sentiment lexicon. We use the amplitude part to represent the distributional information and the phase part to represent the sentiment information of the language. Experiments on three popular SA datasets show that our method is effective.

In deep learning models, most of network architectures are designed artificially and empirically. Although adding new structures such as convolution kernels in CNN is widely used, there are few methods to design new structures and mathematical tools to evaluate feature representation capabilities of new structures. Inspired by ensemble learning, we propose an interpretable regularization method named Minimize Mutual Information Method(MMIM), which minimize the generalization error by minimizing the mutual information of hidden neurons. The experimental results also verify the effectiveness of our proposed MMIM.

Although paths of user interests shift in knowledge graphs (KGs) can benefit conversational recommender systems (CRS), explicit reasoning on KGs has not been well considered in CRS, due to the complex of high-order and incomplete paths.We propose CRFR, which effectively does explicit multi-hop reasoning on KGs with a conversational context-based reinforcement learning model.Considering the incompleteness of KGs, instead of learning single complete reasoning path, CRFR flexibly learns multiple reasoning fragments which are likely contained in the complete paths of interests shift.A fragments-aware unified model is then designed to fuse the fragments information from item-oriented and concept-oriented KGs to enhance the CRS response with entities and words from the fragments.Extensive experiments demonstrate CRFR's SOTA performance on recommendation, conversation and conversation interpretability.

Social influence is essential to social recommendation. Current influence-based social recommendation focuses on the explicit influence on observed social links. However, in real cases, implicit social influence can also impact users' preference in an unobserved way. In this work, we concern two kinds of implicit influence: Local Implicit Influence of persons on unobserved interpersonal relations, and Global Implicit Influence of items broadcasted to users. We improve the state-of-the-art GNN-based social recommendation methods by modeling two kinds of implicit influences separately. Local implicit influence is involved by predicting unobserved social relationships. Global implicit influence is involved by defining global popularity of each item and personalize the impact of the popularity on each user. In a GCN network, explicit and implicit influence are integrated to learn the social embedding of users and items in social recommendation. Experimental results on Yelp initially demonstrate the effectiveness of proposed model.

Despite their widespread applications, deep neural networks often tend to overfit the training data. Here, we propose a measure called VECA (Variance of Eigenvalues of Covariance matrix of Activation matrix) and demonstrate that VECA is a good predictor of networks' generalization performance during the training process. Experiments performed on fully-connected networks and convolutional neural networks trained on benchmark image datasets show a strong correlation between test loss and VECA, which suggest that we can calculate the VECA to estimate generalization performance without sacrificing training data to be used as a validation set.

Learning a classifier from positive and unlabeled data may occur in various applications. It differs from the standard classification problems by the absence of labeled negative examples in the training set. So far, two main strategies have typically been used for this issue: the likely negative examplesbased strategy and the class prior-based strategy, in which the likely negative examples or the class prior is required to be obtained in a preprocessing step. In this paper, a new strategy based on the Bhattacharyya coefficient is put forward, which formalizes this learning problem as an optimization problem and does not need a preprocessing step. We first show that with the given positive class conditional probability density function (PDF) and the mixture PDF of both the positive class and the negative class, the class prior can be estimated by minimizing the Bhattacharyya coefficient of the positive class with respect to the negative class. We then show how to use this result in an implicit mixture model of restricted Boltzmann machines to estimate the positive class conditional PDF and the negative class conditional PDF directly to obtain a classifier without the explicit estimation of the class prior. Many experiments on real and synthetic datasets illustrated the superiority of the proposed approach.

In supervised learning, the generative approach is an important one, which obtains the generative model by learning the joint probability between features and categories. In quantum mechanics, Quantum Entanglement (QE) can provide a statistical correlation between subsystems (or attributes) that is stronger than what classical systems are able to produce. It inspires us to use entangled systems (states) to characterize this strong statistical correlation between features and categories, that is, to use the joint probability derived from QE to model the correlation. Based on the separability of the density matrix of entangled systems, this paper formally clarifies the manifestation of the strong statistical correlation revealed by QE, and implements a classification algorithm (called ECA) to verify the feasibility and superiority of the correlation in specific tasks. Since QE arises from the measurement process of entangled systems, the core of ECA is quantum measurement operations. In this paper, we use the GHZ [25] and W [22] states to prepare the entangled system and use a fully connected network layer to learn the measurement operator. It can also be understood as replacing the output layer of the Multi-Layer Perceptron (MLP) with a quantum measurement operation. The experimental results show that ECA is superior to most representative classification algorithms in multiple evaluation metrics.

Recently, recurrent neural networks (RNNs) based methods have achieved profitable performance on mining temporal characteristics in user behavior. However, user preferences are changing over time and have not been fully exploited in e-commerce scenarios. To fill in the gap, we propose an approach, called quantum inspired preference interactive networks (QPIN), which leverages the mathematical formalism of quantum theory (QT) and the long short term memory (LSTM) network, to interactively learn user preferences. Specifically, the tensor product operation is used to model the interaction among a single user's own preferences, i.e. individual preferences. A quantum many-body wave function (QMWF) is employed to model interaction among all users' preferences, i.e. group preferences. Further, we bridge them by deriving a rigorous projection, and thus take the interplay between them into account. Experiments on an Amazon dataset as well as a real-world e-commerce dataset demonstrate the effectiveness of QPIN, which achieves superior performances compared with the state-of-the-art methods in terms of AUC and F1-score.

Deep Neural Networks (DNNs) with a large number of parameters are very powerful machine learning systems. However, overfitting is a serious problem in such networks. Till now, many regularizers such as dropout, data augmentation have been proposed to prevent overfitting. Motivated by ensemble learning, we treat each hidden layer in neural networks as an ensemble of some base learners by dividing hidden units into some non-overlapping groups and each group is considered as a base learner. Based on the theoretical analysis of generalization error of ensemble estimators (bias-variance-covariance decomposition), we find the variance of each base learner plays an important role in preventing overfitting and propose a novel regularizer---\emphEnsemble-based Low-Level Sample-Variances Method (ELSM) to encourage each base learner of hidden layers to have a low-level sample-variance. Experiments across a number of datasets and network architectures show that ELSM can effectively reduce overfitting and improve the generalization ability of DNNs.

Conversational sentiment analysis (CSA) is emergent research field in natural language processing (NLP). This brings a lot of new issues worth studying and directions worth exploring. At the same time, there are many difficulties that need to be overcome. There are lot of challenges, such as lacking of effective deep learning models and being short of appropriate datasets. Inspired by the concept of density matrix in quantum mechanics, we propose a novel attention mechanism called DMATT and apply it to conversational sentiment analysis tasks. In the experiment, we find that deep learning model combined with DMATT has a great improvement in test results compared to the model with traditional attention mechanism. Recurrent neural networks (RNN) and their variants LSTM and GRU are very effective choices in solving time series problem such as conversational sentiment analysis tasks. In this paper, we propose a new model combining GRU and DMATT called DMATT-BiGRU. We experiment in multiple datasets, one of which is called ScenarioSA collected by ourselves.

Latest development of neural models has connected the encoder and decoder through a self-attention mechanism. In particular, Transformer, which is solely based on self-attention, has led to breakthroughs in Natural Language Processing (NLP) tasks. However, the multi-head attention mechanism, as a key component of Transformer, limits the effective deployment of the model to a resource-limited setting. In this paper, based on the ideas of tensor decomposition and parameters sharing, we propose a novel self-attention model (namely Multi-linear attention) with Block-Term Tensor Decomposition (BTD). We test and verify the proposed attention method on three language modeling tasks (i.e., PTB, WikiText-103 and One-billion) and a neural machine translation task (i.e., WMT-2016 English-German). Multi-linear attention can not only largely compress the model parameters but also obtain performance improvements, compared with a number of language modeling approaches, such as Transformer, Transformer-XL, and Transformer with tensor train decomposition.

Quantum Theory (QT) has been applied in a number of fields outside physics, e.g. Information Retrieval (IR). A series of pioneering works have verified the necessity to employ QT in IR user models. In this paper, we explore the process of relevance judgement from a novel perspective of the two state vector quantum weak measurement (WM) by considering context information in time domain. Experiments are carried out to verify our arguments.

Aspect-level sentiment analysis aims to distinguish the sentiment polarity of each specific aspect term in a given sentence. Both industry and academia have realized the importance of the relationship between aspect term and sentence, and made attempts to model the relationship by designing a series of attention models. However, most existing methods usually neglect the fact that the position information is also crucial for identifying the sentiment polarity of the aspect term. When an aspect term occurs in a sentence, its neighboring words should be given more attention than other words with long distance. Therefore, we propose a position-aware bidirectional attention network (PBAN) based on bidirectional GRU. PBAN not only concentrates on the position information of aspect terms, but also mutually models the relation between aspect term and sentence by employing bidirectional attention mechanism. The experimental results on SemEval 2014 Datasets demonstrate the effectiveness of our proposed PBAN model.

Although Deep Neural Networks (DNNs) have achieved excellent performance in many tasks, improving the generalization capacity of DNNs still remains a challenge. In this work, we propose a novel regularizer named Ensemble-based Decorrelation Method (EDM), which is motivated by the idea of the ensemble learning to improve generalization capacity of DNNs. EDM can be applied to hidden layers in fully connected neural networks or convolutional neural networks. We treat each hidden layer as an ensemble of several base learners through dividing all the hidden units into several non-overlap groups, and each group will be viewed as a base learner. EDM encourages DNNs to learn more diverse representations by minimizing the covariance between all base learners during the training step. Experimental results on MNIST and CIFAR datasets demonstrate that EDM can effectively reduce the overfitting and improve the generalization capacity of DNNs

Typical dimensionality reduction (DR) methods are data-oriented, focusing on directly reducing the number of random variables (or features) while retaining the maximal variations in the high-dimensional data. Targeting unsupervised situations, this paper aims to address the problem from a novel perspective and considers model-oriented DR in parameter spaces of binary multivariate distributions. Specifically, we propose a general parameter reduction criterion, called confident-information-first (CIF) principle, to maximally preserve confident parameters and rule out less confident ones. Formally, the confidence of each parameter can be assessed by its contribution to the expected Fisher information distance within a geometric manifold over the neighborhood of the underlying real distribution. Then, we demonstrate two implementations of CIF in different scenarios. First, when there are no observed samples, we revisit the Boltzmann machines (BMs) from a model selection perspective and theoretically show that both the fully visible BM and the BM with hidden units can be derived from the general binary multivariate distribution using the CIF principle. This finding would help us uncover and formalize the essential parts of the target density that BM aims to capture and the nonessential parts that BM should discard. Second, when there exist observed samples, we apply CIF to the model selection for BM, which is in turn made adaptive to the observed samples. The sample-specific CIF is a heuristic method to decide the priority order of parameters, which can improve the search efficiency without degrading the quality of model selection results as shown in a series of density estimation experiments.

Tang poetry, as one of the most typical ways for ancient Chinese to express their emotions, has been continuously inherited for millenniums in China. Nowadays, though our way of life has changed dramatically, the traditional culture about Tang poetry is still affecting us deeply. However, the psychological effect of Tang poetry remains unclear currently. Motivated by this, we aim to investigate the impact of Tang poetry on human emotional state through designing effective experimental paradigm. We gathered 16 Tang poetry video clips with different genres, and played these to 18 subjects, meanwhile recorded their brain neural responses via Electroencephalogram (EEG). A questionnaire was designed to record subjects' subjective ratings of the emotional experience when watching the Tang poetry videos. Through analyzing the questionnaires, we found that Tang poetry can indeed induce the subjects' specific emotions. Finally, we performed a pilot analysis of the recorded EEG signals and the subjective ratings, for exploring the correlation of the brain activity with the emotional states. The results indicate that the Tang poetry can be used as a kind of emotive stimuli in affective computing research, and the EEG induced by Tang poetry can be utilized to probe the human's internal emotions.

Moving object detection is an important part in intelligent video surveillance under the banner of Internet of things. The detection of moving target’s shadow is also an important step in moving object detection. On the accuracy of shadow detection will affect the detection results of the object directly. Based on the variety of shadow detection method, we find that only using one feature can’t make the result of detection accurately. Then we present a new method for shadow detection which contains colour information, the invariance of optical and texture feature. Through the comprehensive analysis of the detecting results of three kinds of information, the shadow was effectively determined. It gets ideal effect in the experiment when combining advantages of various methods.

In many research and application areas, such as information retrieval and machine learning, we often encounter dealing with a probability distribution that is mixed by one distribution that is relevant to our task in hand and the other that is irrelevant and that we want to get rid of. Thus, it is an essential problem to separate the irrelevant distribution from the mixture distribution. This article is focused on the application in Information Retrieval, where relevance feedback is a widely used technique to build a refined query model based on a set of feedback documents. However, in practice, the relevance feedback set, even provided by users explicitly or implicitly, is often a mixture of relevant and irrelevant documents. Consequently, the resultant query model (typically a term distribution) is often a mixture rather than a true relevance term distribution, leading to a negative impact on the retrieval performance. To tackle this problem, we recently proposed a Distribution Separation Method (DSM), which aims to approximate the true relevance distribution by separating a seed irrelevance distribution from the mixture one. While it achieved a promising performance in an empirical evaluation with simulated explicit irrelevance feedback data, it has not been deployed in the scenario where one should automatically obtain the irrelevance feedback data. In this article, we propose a substantial extension of the basic DSM from two perspectives: developing a further regularization framework and deploying DSM in the automatic irrelevance feedback scenario. Specifically, in order to avoid the output distribution of DSM drifting away from the true relevance distribution when the quality of seed irrelevant distribution (as the input to DSM) is not guaranteed, we propose a DSM regularization framework to constrain the estimation for the relevance distribution. This regularization framework includes three algorithms, each corresponding to a regularization strategy incorporated in the objective function of DSM. In addition, we exploit DSM in automatic (i.e., pseudo) irrelevance feedback, by automatically detecting the seed irrelevant documents via three different document reranking methods. We have carried out extensive experiments based on various TREC datasets, in order to systematically evaluate the proposed methods. The experimental results demonstrate the effectiveness of our proposed approaches in comparison with various strong baselines.

The restricted Boltzmann machine (RBM) has been used as building blocks for many successful deep learning models, e.g., deep belief networks (DBN) and deep Boltzmann machine (DBM) etc. The training of RBM can be extremely slow in pathological regions. The second order optimization methods, such as quasi-Newton methods, were proposed to deal with this problem. However, the non-convexity results in many obstructions for training RBM, including the infeasibility of applying second order optimization methods. In order to overcome this obstruction, we introduce an em-like iterative project quasi-Newton (IPQN) algorithm. Specifically, we iteratively perform the sampling procedure where it is not necessary to update parameters, and the sub-training procedure that is convex. In sub-training procedures, we apply quasi-Newton methods to deal with the pathological problem. We further show that Newton's method turns out to be a good approximation of the natural gradient (NG) method in RBM training. We evaluate IPQN in a series of density estimation experiments on the artificial dataset and the MNIST digit dataset. Experimental results indicate that IPQN achieves an improved convergent performance over the traditional CD method.

Automatic emotion recognition based on multi-channel neurophysiological signals, as a challenging pattern recognition task, is becoming an important computer-aided method for emotional disorder diagnoses in neurology and psychiatry. Traditional approaches require designing and extracting a range of features from single or multiple channel signals based on extensive domain knowledge. This may be an obstacle for non-domain experts. Moreover, traditional feature fusion method can not fully utilize correlation information between different channels. In this paper, we propose a preprocessing method that encapsulates the multi-channel neurophysiological signals into grid-like frames through wavelet and scalogram transform. We further design a hybrid deep learning model that combines the `Convolutional Neural Network (CNN)' and `Recurrent Neural Network (RNN)', for extracting task-related features, mining inter-channel correlation and incorporating contextual information from those frames. Experiments are carried out, in a trial-level emotion recognition task, on the DEAP benchmarking dataset. Our results demonstrate the effectiveness of the proposed methods, with respect to the emotional dimensions of Valence and Arousal.

For sentiment classification, it is often recognized that embedding based on distributional hypothesis is weak in capturing sentiment contrast-contrasting words may have similar local context.Based on broader context, we propose to incorporate Theta Pure Dependence (TPD) into the Paragraph Vector method to reinforce topical and sentimental information.TPD has a theoretical guarantee that the word dependency is pure, i.e., the dependence pattern has the integral meaning whose underlying distribution can not be conditionally factorized.Our method outperforms the state-of-the-art performance on text classification tasks.

Session search aims to improve ranking effectiveness by incorporating user interaction information, including short-term interactions within one session and global interactions from other sessions (or other users). While various session search models have been developed and a large number of interaction features have been used, there is a lack of a systematic investigation on how different features would influence the session search. In this paper, we propose to classify typical interaction features into four categories (current query, current session, query change, and collective intelligence). Their impact on the session search performance is investigated through a systematic empirical study, under the widely used Learning-to-Rank framework. One of our key findings, different from what have been reported in the literature, is: features based on current query and collective intelligence have a more positive influence than features based on query change and current session. This would provide insights for development of future session search techniques.

Query language modeling based on relevance feedback has been widely applied to improve the effectiveness of information retrieval. However, intra‐query term dependencies (i.e., the dependencies between different query terms and term combinations) have not yet been sufficiently addressed in the existing approaches. This article aims to investigate this issue within a comprehensive framework, namely the Aspect Query Language Model (AM). We propose to extend the AM with a hidden Markov model (HMM) structure to incorporate the intra‐query term dependencies and learn the structure of a novel aspect HMM (AHMM) for query language modeling. In the proposed AHMM, the combinations of query terms are viewed as latent variables representing query aspects. They further form an ergodic HMM, where the dependencies between latent variables (nodes) are modeled as the transitional probabilities. The segmented chunks from the feedback documents are considered as observables of the HMM. Then the AHMM structure is optimized by the HMM, which can estimate the prior of the latent variables and the probability distribution of the observed chunks. Our extensive experiments on three large‐scale text retrieval conference (TREC) collections have shown that our method not only significantly outperforms a number of strong baselines in terms of both effectiveness and robustness but also achieves better results than the AM and another state‐of‐the‐art approach, namely the latent concept expansion model. © 2014 Wiley Periodicals, Inc.

Recent research has shown that the improvement of mean retrieval effectiveness (e.g., MAP) may sacrifice the retrieval stability across queries, implying a tradeoff between effectiveness and stability. The evaluation of both effectiveness and stability are often based on a baseline model, which could be weak or biased. In addition, the effectiveness-stability tradeoff has not been systematically or quantitatively evaluated over TREC participated systems. The above two problems, to some extent, limit our awareness of such tradeoff and its impact on developing future IR models. In this paper, motivated by a recently proposed bias-variance based evaluation, we adopt a strong and unbiased "baseline", which is a virtual target model constructed by the best performance (for each query) among all the participated systems in a retrieval task. We also propose generalized bias-variance metrics, based on which a systematic and quantitative evaluation of the effectiveness-stability tradeoff is carried out over the participated systems in the TREC Ad-hoc Track (1993-1999) and Web Track (2010-2012). We observe a clear effectiveness-stability tradeoff, with a trend of becoming more obvious in more recent years. This implies that when we pursue more effective IR systems over years, the stability has become problematic and could have been largely overlooked.

Twitter, as one of the most popular social media platforms, provides a convenient way for people to communicate and interact with each other. It has been well recognized that influence exists during users' interactions. Some pioneer studies on finding influential users have been reported in the literature, but they do not distinguish different influence roles, which are of great value for various marketing purposes. In this paper, we move a step forward trying to further distinguish influence roles of Twitter users in a certain topic. By defining three views of features relating to topic, sentiment and popularity respectively, we propose a Multi-view Influence Role Clustering (MIRC) algorithm to group Twitter users into five categories. Experimental results show the effectiveness of the proposed approach in inferring influence roles.

In density estimation task, Maximum Entropy (Maxent) model can effectively use reliable prior information via nonparametric constraints, that is, linear constraints without empirical parameters. However, reliable prior information is often insufficient, and parametric constraints becomes necessary but poses considerable implementation complexity. Improper setting of parametric constraints can result in overfitting or underfitting. To alleviate this problem, a generalization of Maxent, under Tsallis entropy framework, is proposed. The proposed method introduces a convex quadratic constraint for the correction of (expected) quadratic Tsallis Entropy Bias (TEB). Specifically, we demonstrate that the expected quadratic Tsallis entropy of sampling distributions is smaller than that of the underlying real distribution with regard to frequentist, Bayesian prior, and Bayesian posterior framework, respectively. This expected entropy reduction is exactly the (expected) TEB, which can be expressed by the closed‐form formula and acts as a consistent and unbiased correction with an appropriate convergence rate. TEB indicates that the entropy of a specific sampling distribution should be increased accordingly. This entails a quantitative reinterpretation of the Maxent principle. By compensating TEB and meanwhile forcing the resulting distribution to be close to the sampling distribution, our generalized quadratic Tsallis Entropy Bias Compensation (TEBC) Maxent can be expected to alleviate the overfitting and underfitting. We also present a connection between TEB and Lidstone estimator. As a result, TEB–Lidstone estimator is developed by analytically identifying the rate of probability correction in Lidstone. Extensive empirical evaluation shows promising performance of both TEBC Maxent and TEB‐Lidstone in comparison with various state‐of‐the‐art density estimation methods.

With the tremendous amount of research publications, it has become increasingly important to provide a researcher with a rapid and accurate recommendation of a list of reference papers about a research field or topic. In this paper, we propose a unified graph model that can easily incorporate various types of useful information (e.g., content, authorship, citation and collaboration networks etc.) for efficient recommendation. The proposed model not only allows to thoroughly explore how these types of information can be better combined, but also makes personalized query-oriented reference paper recommendation possible, which as far as we know is a new issue that has not been explicitly addressed in the past. The experiments have demonstrated the clear advantages of personalized recommendation over non-personalized recommendation.

It has been recognized that, when an information retrieval (IR) system achieves improvement in mean retrieval effectiveness (e.g. mean average precision (MAP)) over all the queries, the performance (e.g., average precision (AP)) of some individual queries could be hurt, resulting in retrieval instability. Some stability/robustness metrics have been proposed. However, they are often defined separately from the mean effectiveness metric. Consequently, there is a lack of a unified formulation of effectiveness, stability and overall retrieval quality (considering both). In this paper, we present a unified formulation based on the bias-variance decomposition. Correspondingly, a novel evaluation methodology is developed to evaluate the effectiveness and stability in an integrated manner. A case study applying the proposed methodology to evaluation of query language modeling illustrates the usefulness and analytical power of our approach.

The classical bag-of-word models for information retrieval (IR) fail to capture contextual associations between words. In this article, we propose to investigate pure high-order dependence among a number of words forming an unseparable semantic entity, that is, the high-order dependence that cannot be reduced to the random coincidence of lower-order dependencies. We believe that identifying these pure high-order dependence patterns would lead to a better representation of documents and novel retrieval models. Specifically, two formal definitions of pure dependence—unconditional pure dependence (UPD) and conditional pure dependence (CPD)—are defined. The exact decision on UPD and CPD, however, is NP-hard in general. We hence derive and prove the sufficient criteria that entail UPD and CPD, within the well-principled information geometry (IG) framework, leading to a more feasible UPD/CPD identification procedure. We further develop novel methods for extracting word patterns with pure high-order dependence. Our methods are applied to and extensively evaluated on three typical IR tasks: text classification and text retrieval without and with query expansion.

Twitter, the most famous micro-blogging service and online social network, collects millions of tweets every day. Due to the length limitation, users usually need to explore other ways to enrich the content of their tweets. Some studies have provided findings to suggest that users can benefit from added hyperlinks in tweets. In this paper, we focus on the hyperlinks in Twitter and propose a new application, called hyperlink recommendation in Twitter. We expect that the recommended hyperlinks can be used to enrich the information of user tweets. A three-way tensor is used to model the user-tweet-hyperlink collaborative relations. Two tensor-based clustering approaches, tensor decomposition-based clustering (TDC) and tensor approximation-based clustering (TAC) are developed to group the users, tweets and hyperlinks with similar interests, or similar contexts. Recommendation is then made based on the reconstructed tensor using cluster information. The evaluation results in terms of Mean Absolute Error (MAE) shows the advantages of both the TDC and TAC approaches over a baseline recommendation approach, i.e., memory-based collaborative filtering. Comparatively, the TAC approach achieves better performance than the TDC approach.

We propose a new leaning method for Multiple Kernel Learning (MKL) based on the upper bounds of the leave-one-out error that is an almost unbiased estimate of the expected generalization error. Specifically, we first present two new formulations for MKL by minimizing the upper bounds of the leave-one-out error. Then, we compute the derivatives of these bounds and design an efficient iterative algorithm for solving these formulations. Experimental results show that the proposed method gives better accuracy results than that of both SVM with the uniform combination of basis kernels and other state-of-art kernel learning approaches.

Relation extraction is the task of finding semantic relations between two entities in text, and is often cast as a classification problem. In contrast to the significant achievements on English language, research progress in Chinese relation extraction is relatively limited. In this article, we present a novel Chinese relation extraction framework, which is mainly based on a 9-position structure. The design of this proposed structure is motivated by the fact that there are some obvious connections between relation types/subtypes and position structures of two entities. The 9-position structure can be captured with less effort than applying deep natural language processing, and is effective to relieve the class imbalance problem which often hurts the classification performance. In our framework, all involved features do not require Chinese word segmentation, which has long been limiting the performance of Chinese language processing. We also utilize some correction and inference mechanisms to further improve the classified results. Experiments on the ACE 2005 Chinese data set show that the 9-position structure feature can provide strong support for Chinese relation extraction. As well as this, other strategies are also effective to further improve the performance.

With the advancement of photo and video editing tools, it has become fairly easy to tamper with photos and videos. One common way is to insert visually plausible composites into target images and videos. In this paper, we propose an automatic fake region detection method based on the planar homography constraint, and an automatic extraction method using graph cut with online feature/parameter selection. Two steps are taken in our method: 1) the targeting step, and 2) the segmentation step. First, the fake region is located roughly by enforcing the planar homography constraint. Second, the fake object is segmented via graph cut with the initialization given by the targeting step. To achieve an automatic segmentation, the optimal features and parameters for graph cut are dynamically selected via the proposed online feature/parameter selection. Performance of this method is evaluated on both semisimulated and real images. Our method works efficiently on images as long as there are regions satisfying the planar homography constraint, including image pairs captured by the approximately cocentered cameras, image pairs photographing planar or distant scenes, and a single image with duplications.

A fundamental goal of unsupervised feature selection is denoising, which aims to identify and reduce noisy features that are not discriminative. Due to the lack of information about real classes, denoising is a challenging task. The noisy features can disturb the reasonable distance metric and result in unreasonable feature spaces, i.e., the feature spaces in which common clustering algorithms cannot effectively find real classes. To overcome the problem, we make a primary observation that the relevance of features is intrinsic and independent of any metric scaling on the feature space. This observation implies that feature selection should be invariant, at least to some extent, with respect to metric scaling. In this paper, we clarify the necessity of considering the metric invariance in unsupervised feature selection and propose a novel model incorporating metric invariance. Our proposed method is motivated by the following observations: if the statistic that guides the unsupervised feature selection process is invariant with respect to possible metric scaling, the solution of this model will also be invariant. Hence, if a metric-invariant model can distinguish discriminative features from noisy ones in a reasonable feature space, it will also work on the unreasonable counterpart transformed from the reasonable one by metric scaling. A theoretical justification of the metric invariance of our proposed model is given and the empirical evaluation demonstrates its promising performance.

Pseudo relevance feedback (PRF), which has been widely applied in IR, aims to derive a distribution from the top n pseudo relevant documents D. However, these documents are often a mixture of relevant and irrelevant documents. As a result, the derived distribution is actually a mixture model, which has long been limiting the performance of PRF. This is particularly the case when we deal with difficult queries where the truly relevant documents in D are very sparse. In this situation, it is often easier to identify a small number of seed irrelevant documents, which can form a seed irrelevant distribution. Then, a fundamental and challenging problem arises: solely based on the mixed distribution and a seed irrelevance distribution, how to automatically generate an optimal approximation of the true relevance distribution? In this paper, we propose a novel distribution separation model (DSM) to tackle this problem. Theoretical justifications of the proposed algorithm are given. Evaluation results from our extensive simulated experiments on several large scale TREC data sets demonstrate the effectiveness of our method, which outperforms a well respected PRF Model, the Relevance Model (RM), as well as the use of RM on D with the seed negative documents directly removed.

High-dimensional data is involved in many fields of information processing. However, sometimes, the intrinsic structures of these data can be described by a few degrees of freedom. To discover these degrees of freedom or the low-dimensional nonlinear manifold underlying a high-dimensional space, many manifold learning algorithms have been proposed. Here we describe a novel algorithm, locally linear inlaying (LLI), which combines simple geometric intuitions and rigorously established optimality to compute the global embedding of a nonlinear manifold. Using a divide-and-conquer strategy, LLI gains some advantages in itself. First, its time complexity is linear in the number of data points, and hence LLI can be implemented efficiently. Second, LLI overcomes problems caused by the nonuniform sample distribution. Third, unlike existing algorithms such as isometric feature mapping (Isomap), local tangent space alignment (LTSA), and locally linear coordination (LLC), LLI is robust to noise. In addition, to evaluate the embedding results quantitatively, two criteria based on information theory and Kolmogorov complexity theory, respectively, are proposed. Furthermore, we demonstrated the efficiency and effectiveness of our proposal by synthetic and real-world data sets.

A novel approach to construct accurate and interpretable fuzzy classification system based on the multi-objective cooperative coevolutionary algorithm (MOCCA) is proposed in this paper. First, feature selection is used to reduce the dimensionality of the data in order to both improve the performance and reduce computational effort. Then the fuzzy clustering algorithm is employed to identify the initial fuzzy system. Third, the MOCCA with three species is carried out to evolve the initial fuzzy system to optimize its structures and parameters. In MOCCA, the interpretability-driven simplification techniques are used to reduce the fuzzy system, thus the interpretability of the fuzzy system is improved; the number of rules, the antecedents of the fuzzy rules and the parameters of the antecedents are optimized simultaneously. Finally, the proposed approach is applied to six benchmark problems, and the results show its validity.

Recently several algorithms, e.g., Isomap, self-organizing isometric embedding (SIE), Locally linear embedding (LLE) and Laplacian eigenmap, were proposed to deal with the problem of learning low dimensional nonlinear manifold embedded in a high dimensional space. Motivated by these algorithms, there is a trend of exploiting the intrinsic manifold structure of the data to improve precision and/or efficiency of classification under the assumption that the high dimensional observable data resides on a low dimensional manifold of latten variables. But these methods suffer their flaws respectively. In this work, we unified the problems of supervised manifold learning in a kernel view and proposed a novel implicit kernel construction method, i. e. supervised locally principal direction preservation kernel (SLPDK) construction, to combine the advantages of current implicit kernel construction methods motivated by manifold learning and try to overcome their disadvantages. SLPDK uses class information and locally principal direction of manifold to implement an approximately symmetric embedding. Implicit kernels constructed by SLPDK have a natural geometrical explanation and can gain a considerable classification precision improvement when the condition of locally linear manifold separability (LLMS) holds.

Most current Web search engines generate search results by analyzing queries and relevance between queries and Web-pages. However, as the number of Web-pages grows, this approach appears to be less efficient in finding relevant information. In many situations, search engines cannot determine what kind of information users want. We propose a framework of feedback search engine (FSE), which not only analyzes the relevance between queries and Web-pages but also uses clickthrough data to evaluate page-to-page relevance and re-generate content relevant search results. The efficient algorithms facilitating the framework are described. Making use of dynamical re-generating search results, FSE can provide its users more accurate and personalized information