Dingling Xu, Ruobing Wang, Qingfei Zhao, Yukun Yan, Zhichun Wang, Daren Zha, Shi Yu, Zhenghao Liu, Shuo Wang, Xu Han, Maosong Sun. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The rapid spread of multimodal fake news poses a serious societal threat, as its evolving nature and reliance on timely factual details challenge existing detection methods. Dynamic Retrieval-Augmented Generation provides a promising solution by triggering keyword-based retrieval and incorporating external knowledge, thus enabling both efficient and accurate evidence selection. However, it still faces challenges in addressing issues such as redundant retrieval, coarse similarity, and irrelevant evidence when applied to deceptive content. In this paper, we propose ExDR—an Explanation-driven Dynamic Retrieval-Augmented Generation framework for Multimodal Fake News Detection. Our framework systematically leverages model-generated explanations in both the retrieval triggering and evidence retrieval modules. It assesses triggering confidence from three complementary dimensions, constructs entity-aware indices by fusing deceptive entities, and retrieves contrastive evidence based on deception-specific features to challenge the initial claim and enhance the final prediction. Experiments on two benchmark datasets, AMG and MR2, demonstrate that ExDR consistently outperforms previous methods in retrieval triggering accuracy, retrieval quality, and overall detection performance, highlighting its effectiveness and generalization capability.

Event prediction is a critical task in natural language processing, aimed at reasoning and forecasting future events based on known event texts. This paper introduces EventPuzzle, a benchmark designed to evaluate the event prediction capabilities of large language models based on event arguments. By introducing argument points, we design tasks and evaluation methods to assess models' ability to predict events from different argument perspectives. EventPuzzle consists of both closed-ended and open-ended tasks. In the closed-ended task, models select the correct argument point from causal chains, while in the open-ended task, models generate event descriptions using two strategies: Argument-based Generation and Direct Generation. We construct an argument point dataset and evaluate multiple LLMs, demonstrating the models' performance across various tasks. Our experimental analysis reveals the strengths and limitations of current models and suggests future directions for improving event prediction.

Event Argument Extraction is a critical task of Event Extraction, focused on identifying event arguments within text. This paper presents a novel Fusion Selection-Generation-Based Approach, by combining the precision of selective methods with the semantic generation capability of generative methods to enhance argument extraction accuracy. This synergistic integration, achieved through fusion prompt, element-based extraction, and fusion learning, addresses the challenges of input, process, and output fusion, effectively blending the unique characteristics of both methods into a cohesive model. Comprehensive evaluations on the RAMS and WikiEvents demonstrate the model’s state-of-the-art performance and efficiency.

Graph representation learning (GRL) in hyperbolic space has gradually emerged as a promising approach. Meanwhile, masking and reconstruction-based (MR-based) methods lead to state-of-the-art self-supervised graph representation. However, existing MR-based methods do not fully consider deep node and structural information. Inspired by the recent active and emerging field of self-supervised learning, we propose a novel node and edge dual-masked self-supervised graph representation learning framework in hyperbolic space, named HDM-GAE. We have designed a graph dual-masked module and a hyperbolic structural self-attention encoder module to mask nodes or edges and perform node aggregation within hyperbolic space, respectively. Comprehensive experiments and ablation studies on real-world multi-category datasets, demonstrate the superiority of our method in downstream tasks such as node classification and link prediction.

Event schema generation is crucial for understanding the structure and temporal relationships of complex events. In this paper, we introduce a novel Directed Acyclic Graph Diffusion Model (DAGDM) that integrates DAG characteristics within a diffusion framework to enhance the effectiveness of schema generation. Our method leverages DAG positional embeddings to capture the hierarchical structure of nodes within graphs, while employing a reachability-based attention to better extract structural relationships between events. To this end, we design a cross-generation strategy that separately generates event sequence and adjacency matrix. Experiments show that our model effectively captures long-range event sequences, significantly enhancing schema generation for complex events.1

Optimizing the efficiency and adaptability of large language models (LLMs) for diverse downstream tasks remains a critical challenge. We propose FlexFFN (Flexible Feedforward Network), a novel framework that introduces hierarchical dynamic selection to enhance computational efficiency, flexibility, and performance in LLMs. At the macro level, FlexFFN leverages a Mixture of Experts (MoE) architecture to dynamically activate distinct FFN modules based on input characteristics. At the micro level, within each FFN module, a dynamic switching mechanism selects between KAN and traditional MLP, combining the rapid convergence capabilities of MLPs with the compositional learning and interpretability strengths of KAN. Additionally, FlexFFN integrates QLoRA (Quantized Low-Rank Adaptation) to significantly reduce memory requirements and computational costs during fine-tuning. By introducing these innovations, FlexFFN achieves a fine-grained balance between computational cost and model expressiveness, making it well-suited for large-scale training and deployment. Experimental results demonstrate that FlexFFN outperforms traditional architectures by reducing computational overhead while improving task-specific adaptability and model efficiency.

Graph Neural Networks (GNNs) are popular for learning node representations in complex graph structures. Traditional methods use Euclidean space but struggle to capture hierarchical structures in real-world graphs. Besides, it’s important to note that in practical applications, many graphs are dynamic and undergo continuous evolution over time. To investigate the characteristics of complex temporal networks, we have introduced a dynamic graph embedding model in the Lorentz space, building upon the foundation of the previously proposed DynHAT model. More specially, our model divides the dynamic graph into multiple discrete static graphs, maps each static graph to the Lorentz space, and then learns informative node representations over time using a self-attention mechanism. We have conducted link prediction experiments on two types of graphs: communication networks and rating networks. Through comprehensive experiments conducted on five real-world datasets, we have demonstrated the superiority of our model in embedding dynamic graphs within Lorentz space.

Long-Context Question Answering (LCQA), a challenging task, aims to reason over longcontext documents to yield accurate answers to questions.Existing long-context Large Language Models (LLMs) for LCQA often struggle with the "lost in the middle" issue.Retrieval-Augmented Generation (RAG) mitigates this issue by providing external factual evidence.However, its chunking strategy disrupts the global long-context information, and its low-quality retrieval in long contexts hinders LLMs from identifying effective factual details due to substantial noise.To this end, we propose LongRAG, a general, dual-perspective, and robust LLM-based RAG system paradigm for LCQA to enhance RAG's understanding of complex long-context knowledge (i.e., global information and factual details).We design LongRAG as a plug-and-play paradigm, facilitating adaptation to various domains and LLMs.Extensive experiments on three multihop datasets demonstrate that LongRAG significantly outperforms long-context LLMs (up by 6.94%), advanced RAG (up by 6.16%), and Vanilla RAG (up by 17.25%).Furthermore, we conduct quantitative ablation studies and multidimensional analyses, highlighting the effectiveness of the system's components and finetuning strategies.Data and code are available at https://github.com/QingFei1/LongRAG.

Next point-of-interest (POI) recommendation has been a trending task to provide next POI suggestions. Most existing sequential-based and graph-based methods have endeavored to model user visiting behaviors and achieved considerable performances. However, they have either modeled user interests at a coarse-grained interaction level or ignored complex high-order feature interactions through general heuristic message passing scheme, making it challenging to capture complementary effects. To tackle these challenges, we propose a novel framework Adaptive Spatial-Temporal Hypergraph Fusion Learning (ASTHL) for next POI recommendation. Specifically, we design disentangled POI-centric learning to decouple spatial-temporal factors and utilize cross-view contrastive learning to enhance the quality of POI representations. Furthermore, we propose multi-semantic enhanced hypergraph learning to adaptively fuse spatial-temporal factors through well-designed aggregation and propagation scheme. Extensive experiments on three real-world datasets validate the superiority of our proposal over various state-of-the-arts. To facilitate future research, our code is available at https://github.com/icmpnorequest/ICASSP2024_ASTHL.

Next point-of-interest (POI) recommendation has been a prominent and trending task to provide next suitable POI suggestions for users. Most existing sequential-based and graph neural network-based methods have explored various approaches to modeling user visiting behaviors and have achieved considerable performances. However, two key issues have received less attention: i) Most previous studies have ignored the fact that user preferences are diverse and constantly changing in terms of various aspects, leading to entangled and suboptimal user representations. ii) Many existing methods have inadequately modeled the crucial cooperative associations between different aspects, hindering the ability to capture complementary recommendation effects during the learning process. To tackle these challenges, we propose a novel framework Disentangled Contrastive Hypergraph Learning (DCHL) for next POI recommendation. Specifically, we design a multi-view disentangled hypergraph learning component to disentangle intrinsic aspects among collaborative, transitional and geographical views with adjusted hypergraph convolutional networks. Additionally, we propose an adaptive fusion method to integrate multi-view information automatically. Finally, cross-view contrastive learning is employed to capture cooperative associations among views and reinforce the quality of user and POI representations based on self-discrimination. Extensive experiments on three real-world datasets validate the superiority of our proposal over various state-of-the-arts. To facilitate future research, our code is available at https://github.com/icmpnorequest/SIGIR2024_DCHL.

When the unknown degradation is mixed with unknown blurry kernels, how to perform super-resolution operation is an open issue. The mean idea of the existing zero-shot and non-zero-shot methods is to estimate blurry kernel. The effects of these methods depend on the accuracy of the deduced blurry kernel. In this paper, we propose Randomly initialized Zero-Shot Super-Resolution (RZSR) training strategy. RZSR is a zero-shot training method and it allows the network to extract low-resolution image features and generate its counterpart high-resolution images under the interference of degradation algorithms. We further propose two model-agnostic modules which are Adaptive Information Extraction Module (AIEM) and knowledge dictionary. They respectively assist the network to extract features and well fit the data distribution of clear images. RZSR can be applied to any single image super-resolution and video super-resolution models. We prove the generalization ability and superiority of RZSR through a series of experiments.

Multimodal Conversational Emotion recognition (MMCER) aims to detect the muti-emotion label for each utterance from heterogeneous visual, text and audio modalities. In this paper, we focus on applying multi-modal graph data structures to conversational emotion recognition and use a novel and efficient graph—MMDCGs to better integrate multi-modal contextual information into conversations. MMDCG provides a new way of encoding intrinsic structural connectivity. Besides, inspired by time series analysis, we set a rolling time window as the receptive field, which can reduce the interference of remote information on the current utterances detection and achieve the purpose of data enhancement. We innovatively ensemble such graph structures with transformers, named rolling-windows MMDCGs (RW-MMDCG). Comprehensive experiments are performed on two representative multi-modal datasets, IEMOCAP and MELD, and we compare them with existing baselines, demonstrating the great advantages and effectiveness of RW-MMDCG.

Graph neural networks (GNNs) have recently become increasingly popular due to their ability to learn node representations in complex graphs. Existing graph representation learning methods mainly target static graphs in Euclidean space, whereas many graphs in practical applications are dynamic and evolve continuously over time. Recent work has demonstrated that real-world graphs exhibit hierarchical properties. Unfortunately, many methods typically do not account for these latent hierarchical structures. In this work, we propose a dynamic network in hyperbolic space via self-attention, referred to as DynHAT, which leverages both the hyperbolic geometry and attention mechanism to learn node representations. More specifically, DynHAT captures hierarchical information by mapping the structural graph onto hyperbolic space, and time-varying dynamic evolution by flexibly weighting historical representations. Through extensive experiments on three real-world datasets, we show the superiority of our model in embedding dynamic graphs in hyperbolic space and competing methods in a link prediction task. In addition, our results show that embedding dynamic graphs in hyperbolic space has competitive performance when necessitating low dimensions.

Link prediction is to predict missing relations between entities for Knowledge Graph Completion (KGC). Convolution neural network has been used in much previous work on link prediction to capture fundamental data pattern of knowledge graph. However, because these models use low-dimensional convolution operation, which limits their performance, they learn fewer expressive features. Further more, they do not have the the capability of keeping the translation property of knowledge triplet, which is an important property for knowledge reasoning. Focusing on these problems, we propose Conv3D (3D Convolution Embedding), a neural network model for link prediction that uses 3D convolution. To capture deeper feature interactions in the knowledge graph, we employ 3D convolution instead of shallow 1D or 2D convolution for generating triplet scores. We conduct link prediction experiments on four general datasets (WN18, WN18RR, FB15k, FB15k-237) and get state-of-the-art (SOTA) results on WN18 and WN18RR. We also explore the influence of convolution parameters (reshaping dimension, number of filters, kernel size) on FB15k-237 and obtain quantitative analytical findings.

In recent years, PowerShell has been widely used in cyber attacks and malicious PowerShell scripts can easily evade the detection of anti-virus software through obfuscation. Existing deobfuscation tools often fail to recover obfuscated scripts correctly due to imprecise obfuscation identification, improper recovery and wrong replacement. In this paper, we propose an AST-based and semantics-preserving deobfuscation approach, Invoke-Deobfuscation. It utilizes recoverable nodes of Abstract Syntax Tree to identify obfuscated pieces precisely, simulates the recovery process through Invoke function and variable tracing, and replaces obfuscated pieces in place to keep the original semantics. We build a large evaluation dataset containing 39,713 wild PowerShell scripts. Compared with the state-of-the-art tools, the experimental results show Invoke-Deobfuscation performs most efficiently. It recovers much more key information than others and significantly reduces samples’ obfuscation score, on average, by 46%. Moreover, 100% of Invoke-Deobfuscation’s results have the same network behavior as the original scripts.

The real-world image degradation in the super-resolution task is recently considered as a combination of Gaussian blur, down-sampling, and additional white Gaussian noise. To han-dle this degradation, previous methods estimate the Gaussian blur kernel or model the degradation based on a randomly selected image patch. However, these methods cannot han-dle degradations with high-level noise well as they ignore the spatial variability or even the existence of noise. Moreover, using image denoising networks to preprocess low-resolution images also fails due to the loss of important high-frequency information. In this paper, we propose a framework called EASE to flexibly handle real-world degradations. Specifi-cally, we develop a lightweight module to erase noise and blur simultaneously by learning from an image denoising and an image restoration network, which adapts to existing net-works that focus on handling bicubic down-sampling. Exten-sive experiments prove the superiority of our method, espe-cially when handling degradations with high-level noise.

Blind super-resolution task aims to restore low-resolution im“ages with unknown degradations to high-resolution counter-parts. Existing methods rely on degradations estimation to re-construct high-resolution images. However, they need human involvement to obtain the best results as they treat unknown types of degradations as known conditions and manually select corresponding trained models. Moreover, they cannot fully use estimated degradations and generate blurry artifacts as they ignore that the impact of degradations on images is re-lated to images contents. In this paper, we propose HIS-NEST which contains an automatic search strategy HIS and a net-work structure NEST. Specifically, to bypass manual partici-pation, HIS automatically selects the clearest image by esti-mating the qualities of generated images. Furthermore, NEST protects the connection between degradations and images by using no loss functions to limit the degradations estimation and analyzing degradations from the perspective of channel and space. Extensive experiments show that our method out-performs state-of-the-art methods.

Heterogeneous information network (HIN) embedding, aiming to project HIN into a low-dimensional space, has attracted considerable research attention.Existing heterogeneous graph representation learning methods also take temporal evolution into consideration in Euclidean space which, however, underestimates the inherent complex and hierarchical properties in many real-world temporal networks, leading to sub-optimal embeddings.To explore these properties of a dynamic heterogeneous network, we propose a dynamic hyperbolic heterogeneous embedding(DyHHE) model that fully takes advantage of the hyperbolic geometry and structural heterogeneity.More specially, to capture the structure and semantic relations between nodes, we employ the meta-path guided random walk to sample the sequences for each node.Then DyHHE maps the temporal graph into hyperbolic space, and capture the structural heterogeneity and evolving behaviors by facilitating the proximity measurement.Experimental results on two real-world datasets demonstrate the superiority of DyHHE, as it consistently outperforms competing methods in link prediction task.

Abstract DPA Contest is a world-famous side-channel competition aiming at analyzing and evaluating the implementing security of some latest countermeasures. Improved Rotating S-box Masking Scheme (RSM2.0) is one of the most popular countermeasures designed during DPA Contest V4.2, which arms with both Low Entropy Masking Schemes and shuffling strategy to ensure the software security of AES-128, particularly the non-profiled security. Up to now, conducting high efficient non-profiled attacking scheme with low resource costs is still a challenge. In this paper, we first propose general and non-profiled leakage fingerprint attacks (named NP-LFA) for secret cracking and make use of it to crack RSM2.0 random masks with almost 100% accuracy. Further, we analyze the hidden vulnerabilities embedded in RSM2.0 implementation, and utilize them to bypass the shuffling defense and perform the master key recovery. Official evaluation results show that NP-LFA is capable of compromising RSM2.0 within 14 traces, each of which only costs 60 ms processing time. Such result validates the high efficiency and light-weighted characteristics of our attacking scheme, which has ranked the first in the official website till now. In addition, we discuss and put forward some possible strategies to mitigate our NP-LFA threats.

Traditional X.509 public key infrastructures (PKIs) depend on trusted certification authorities (CAs) to sign certificates, used in SSL/TLS to authenticate web servers and establish secure channels. However, recent security incidents indicate that CAs may (be compromised to) sign fraudulent certificates. In this article, we propose blockchain-based certificate transparency (CT) and revocation transparency (RT) to balance the absolute authority of CAs. Our scheme is compatible with X.509 PKIs but significantly reinforces the security guarantees of a certificate. The CA-signed certificates and their revocation status information of an SSL/TLS web server are published by the subject (i.e., the web server) as a transaction in the global certificate blockchain. The certificate blockchain acts as append-only public logs to monitor CAs’ certificate signing and revocation operations, and an SSL/TLS web server is granted with the cooperative control on its certificates. A browser compares the certificate received in SSL/TLS negotiations with the ones in the public certificate blockchain, and accepts it only if it is published and not revoked. We implement the prototype system with Firefox and Nginx, and the experimental results show that it introduces reasonable overheads.

In recent years, more and more researchers pay attention to the recommendation system based on heterogeneous information network(HIN), because HIN is rich in various kinds of information, which can significantly improve the performance of the recommendation system. But the HIN based recommendation system faces the following problems: how to leverage high-order information to get semantic-level interaction characteristics between users and items; How to deeply fuse first-order and high-order information to enhance the representation ability of the system. To address these issues, we propose a novel model: First-order and High-order Information Fusion over Heterogeneous Information Network for Top-N Recommender System(FHRec). For first-order information, we use graph neural networks to generate the latent vectors of users and items. And for high-order information, we use a meta-path based semantic-level aggregation to get the interaction between users and items. Then we deeply integrate the first-order and high-order information and use neural collaborative filtering to improve the recommendation performance. Finally, we conduct comparative experiments of our model with other baseline algorithms on three real world datasets, and the experimental results prove the superiority of our model.

In recent years, the knowledge graph based recommendation system is a research hotspot and scholars propose a propagation-based method, which combines graph neural networks with knowledge graph. But the previous work faces two problems: The first is that exist propagation methods generate neighours of target entity through random sampling strategy which will bring noise to the system. The second problem is that exist models directly aggregate the neighbors information of the target entity at each step, while ignoring the fact that the propagation of high-order information also needs to be selective and memorable. To solve these problems, we propose a novel model: Gated Knowledge Graph Neural Networks for Top-N Recommendation System(GKGNN). This model uses pretrain technique to generate neighbor set with high priority of the target entity in the graph. At the same time, this model introduces the gated mechanism into the propagation process and the valuable information is remembered and unimportant information is forgotten during the high-order information propagation. Finally, we conduct comparative experiments of our model with other baseline algorithms on three real world datasets, and the experimental results prove the superiority of our model.

Recently, top-k recommendation system is getting more and more attention from researchers and unlike the rating prediction task, the purpose of top-k recommendation is to present the user with a list of items they are most interested in. Many rating prediction models do not produce good ranking results, so the top-k recommendation faces two challenges: the first is how to accurately obtain user and item latent vector from the user-item interaction rating matrix, and the second challenge is how to combine the ranking strategy with the recommendation model deeply. The booming deep learning technology, such as autoencoder and graph neural networks, brings us new solutions. So in this paper, we propose a novel model: Graph Attention Autoencoder for Collaborative Pair-wise Ranking. The autoencoder consists of graph attention encoder and collaborative neural decoder, which is used to generate user and item latent vector accurately. And then we use the pairwise ranking learning process to ensure the rating value accuracy and rating value pairwise ranking consistency. Finally, experimental results on three real-world datasets demonstrate the superiority of our model.

Video super resolution (SR) is an important task, which recovers high-resolution (HR) frames from consecutive low-resolution (LR) couterparts. The most advanced works achieved good performance to this day. However, most of them has largely focussed on making a breakthrough in accuracy and speed, which has neglect that how to recover the finer texture details. Therefore, in this paper, we first present an Video SR model combined generative adversarial network and recurrent neural network (GAN-RNN) structure. It is forced by the self-attention mechanism to pay great attention to the high-frequency information of the LR frames. The perceptual loss is introduced to retain the high-frequency detail which is different from other video SR network. A great deal of evaluations and comparisons with previous methods have confirmed the merits of the proposed framework which can significantly outperform the current state of the art.

Next Point-Of-Interest (POI) recommendation has attracted extensive attention recently, benefiting from the large volumes of check-in records on location-based social networks (LBSNs). Most of the existing works either consider user's long- and short-term preferences or only one of them, ignoring users' preferences have multiple different resolutions. Moreover, these methods suffer from the high sparsity issue. Although hierarchical category information contributes to alleviating this problem, it has not been fully exploited and integrated with spatiotemporal contexts in next POI recommendation. To this end, we propose a novel method named Hierarchical and Multi-Resolution Preference Modeling (HMRPM), which simultaneously models hierarchical personalized preferences at three different resolutions, i.e., we learn long-term, short-term and transient preferences. HMRPM consists of three corresponding preference modeling modules: (1) The long-term module captures general preferences of users by modeling user-POI interactions and hierarchical category-level interactions. (2) The short-term module utilizes a hierarchical category-aware self-attention mechanism to capture users' recent preferences. Spatiotemporal position embeddings and user-aware attention are also proposed. (3) The transient module captures hierarchical current preferences by modeling hierarchical transient transition patterns. Extensive experimental results on two public datasets demonstrate the proposed method significantly outperforms other state-of-the-art methods.

With the popularity of location-based social networks (LBSNs), Point-of-Interest (POI) recommendation has become an essential location-based service to help people explore novel locations. Although the massive check-in data bring a good opportunity, there are still many challenges in building personalized POI recommender systems based on geographical information. First, current coarse-grained geographical models provide considerably limited improvements on POI recommendations and fail to capture the overall impact of fine-grained geographical characteristics in LBSNs. Second, previous methods such as matrix factorization always give equal weight to each positive example and may not distinguish between their different contributions in learning the objective function. To cope with these challenges, we develop a fine-grained POI recommendation framework that makes full use of the geographical characteristics from both users’ and locations’ perspectives. For capturing the fine-grained geographical influence, we present a unified probability distribution model based on four key geographical characteristics. For mining more contribution information from positive examples, we assign a higher weight to highlight the contribution of a higher check-in frequency by employing a logistic matrix factorization. Finally, experimental results on two real-world datasets demonstrate the effectiveness and superiority of the proposed method.

Point-of-interest (POI) recommendation has become an important service to help users discover attractive locations. A variety of available check-in data make it possible to build a personalized POI recommender system, but the extreme sparsity of check-in data poses a severe challenge for POI recommendation. Recent studies mainly utilize social information, categorical information and/or geographical information to supplement the highly sparse check-in data. However, these studies often apply shallow methods for the extra information and provide considerably limited improvements on POI recommendation. In this paper, we propose a fine-grained POI recommendation framework, called FGRec to capture the intrinsic influences of social, categorical and geographical information on the check-in behaviors of users. First, we study the social influence in depth by exploiting the multi-hop social friends and top-n nearest neighbor friends, not only the direct friends (i.e., 1-hop friends). Second, we investigate the categorical influence by factorizing both user-POI and user-category matrices simultaneously over the same user embedding space, rather than simply using the popularity of POI categories. Third, we explore the geographical influence by integrating two types of distance (i.e., the distance between user homes and POIs and the distance among POIs) into a unified probability distribution over check-in POIs, instead of modeling them separately. Finally, experimental results on two large-scale real-world datasets demonstrate the effectiveness and superiority of the proposed method.

Accurate intention understanding of the user inputs is the key to human-computer interaction (HCI). At present, more and more studies just focus on the improvement of algorithm efficiency and ignore the nature exploration of intention understanding. In humans, working memory is regarded as a cognitive system for handling a range of neuro-cognitive tasks. Because the intention understanding is a kind of human cognitive ability, in this paper we will explore the human cognitive execution mechanism and try to apply it to improve the machines' intention understanding level. First, we demonstrated a cognitive learning model called Cortico-Basal ganglia-Cerebella (CBC) loops plays an important role in the process of working memory. Then, based on the full understanding of the loops operation mechanism, we put forward a new model of intension understanding. Finally, we applied this model on speech data and compared it with other two methods. The results showed that the new model could help to get task-specific vectors and offer further gains in performance on intention understanding.

In recent years, social recommendation is a research hotspot because it contains social network information which can effectively solve the problem of data sparsity and cold start. But the social recommendation task faces two problems: one is that how to accurately learn user latent vector and item latent vector from user-item interaction graph and social graph, the other is that how to depict the intrinsic and complex interaction between users and items. With the development of graph neural networks, node embedding is becoming more and more accurate on the graph. Besides neural collaborative filtering explores the interaction of users and items deeply. So in this paper, we propose a novel model: graph attention networks for neural social recommendation (GAT-NSR). This model adopts multi-head attention mechanism for message passing on the two graphs, which get user & item latent vector from different perspectives. And also we design a neural collaborative recommendation module to capture the inherent characteristics of user-item interaction behavior for recommendation. Finally, detailed experimental results on two real-world datasets clearly prove the effectiveness of our proposed model.

The increasing prevalence of location-based social networks (LBSNs) poses a wonderful opportunity to build per-sonalized point-of-interest (POI) recommendations, which aim at recommending a top-N ranked list of POIs to users according to their preferences. Although previous studies on collaborative filtering are widely applied for POI recommendation, there are two significant challenges have not been solved perfectly. (1) These approaches cannot effectively and efficiently exploit unobserved feedback and are also unable to learn useful information from it. (2) How to seamlessly integrate multiple types of context information into these models is still under exploration. To cope with the aforementioned challenges, we develop a new Personalized pairwise Ranking Framework based on Poisson Factor factorization (PRFPF) that follows the assumption that users’ preferences for visited POIs are preferred over potential POIs, unvisited POIs are less preferred than potential POIs. The framework PRFPF is composed of two modules: candidate module and ranking module. Specifically, the candidate module is used to generate a series of potential POIs from unvisited POIs by incorporating multiple types of context information (e.g., social and geographical information). The ranking module learns the ultimate order of users’ preference by leveraging the potential POIs. Experimental results evaluated on two large-scale real-world datasets show that our framework outperforms other state-of-the-art approaches in terms of various metrics.

In open-domain question answering system, the granularities of the answers vary with different types of questions. For example, for the questions asking about locations (Location type questions), their answers are usually short phrases. While for the questions asking about reasons (Description type questions), their answers are usually long clauses or sentences. This insight can be used to improve the performance of answer sentence selection, which is a crucial component of the open-domain QA system. In this paper, we propose a novel Multi-Granularity Neural Network (MGNN) model to better evaluate the semantic matching of questions and answers. First, MGNN has three classes of channels with each computing the similarity of question and answer pairs from one of the three granularity levels: clause level, phrase level and ngram level. Then, MGNN uses a parametrization weighting scheme which considers question types to combine these different granularity channels. We carry out experiments on a public available benchmark dataset for question answering. Empirical results show that our method outperforms state-of-the-art methods.

Network embedding, which aims to learn the low-dimensional representations of nodes, has attracted increasing attention in various fields such as social networks, paper citation networks and knowledge graphs. At present, most of the network embedding works are based on static networks, that is, the evolution of networks over time is not taken into account. It is more realistic to consider temporal information in network embedding and it could also make the embedding get more abundant information. In this paper, we propose a dynamic network embedding model DynSEM with semantic evolution, to train node embeddings in a sequence of networks over time. The advantage of our method is that it presents an effective inheritance of historical information. Our method uses nonrandom initialization and orthogonal procrustes method to align the node embeddings into common space which makes node embedding able to inheritance information. In particular, in the common space, we train a model to capture the dynamics information of the networks and smooth temporal node embeddings. We evaluate our method comparing it with other methods on three real-world datasets. The experimental results prove the effectiveness of dynamic network embeddings generated by DynSEM model.

Abstract Image steganography is the technique of hiding secret information within images. It is an important research direction in the security field. Benefitting from the rapid development of deep neural networks, many steganographic algorithms based on deep learning have been proposed. However, two problems remain to be solved in which the most existing methods are limited by small image size and information capacity. In this paper, to address these problems, we propose a high capacity image steganographic model named HidingGAN. The proposed model utilizes a new secret information preprocessing method and Inception‐ResNet block to promote better integration of secret information and image features. Meanwhile, we introduce generative adversarial networks and perceptual loss to maintain the same statistical characteristics of cover images and stego images in the high‐dimensional feature space, thereby improving the undetectability. Through these manners, our model reaches higher imperceptibility, security, and capacity. Experiment results show that our HidingGAN achieves the capacity of 4 bits‐per‐pixel (bpp) at 256 × 256 pixels, improving over the previous best result of 0.4 bpp at 32 × 32 pixels.

Since the day it was proposed, return-oriented programming has shown to be an effective and powerful attack technique against the write or execute only (W ⊕ X) protection. However, a general belief in the previous research is, systems deployed with address space randomization where the executables are also randomized at run-time are able to defend against return-oriented programming, as the addresses of all instructions are randomized. In this paper, we show that due to the weakness of current address space randomization technique, there are still ways of launching return-oriented programming attacks against those well-protected systems efficiently. We demonstrate and evaluate our attacks with existing typical web server applications and discuss possible methods of mitigating such threats.