Minghui Jia, Qichao Zhang, Ali Luo, Linjing Li, Shuo Ye, Hailing Lu, Wen Hou, Dongbin Zhao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The offline datasets for imitation learning (IL) in multi-agent games typically contain player trajectories exhibiting diverse strategies, which necessitate measures to prevent learning algorithms from acquiring undesirable behaviors. Learning representations for these trajectories is an effective approach to depicting the strategies employed by each demonstrator. However, existing learning strategies often require player identification or rely on strong assumptions, which are not appropriate for multi-agent games. Therefore, in this paper, we introduce the Strategy Representation for Imitation Learning (STRIL) framework, which (1) effectively learns strategy representations in multi-agent games, (2) estimates proposed indicators based on these representations, and (3) filters out sub-optimal data using the indicators. STRIL is a plug-in method that can be integrated into existing IL algorithms. We demonstrate the effectiveness of STRIL across competitive multi-agent scenarios, including Two-player Pong, Limit Texas Hold'em, and Connect Four. Our approach successfully acquires strategy representations and indicators, thereby identifying dominant trajectories and significantly enhancing existing IL performance across these environments.

Large language models (LLMs) have achieved significant progress in mathematical reasoning, especially in elementary math. However, they remain indisposed on tackling complex questions at high-school or college levels, which put forward a more advanced requirement of mastering relevant mathematical theorems. For we humans, whether selecting the appropriate theorems according to the provided question is a crucial factor affecting the quality of the ultimate solutions, yet which has been neglected by previous research in the field of LLM reasoning. In this paper, we propose a novel approach to enhance the LLM's capability of utilizing the mathematical theorems to specific problems, which we refer to as Theorem Rationale (TR). To this end, a new dataset encompassing problem-theorem-solution triples is deliberately established for transferring principles of TR. Furthermore, we develop an evolving strategy to boost hierarchical instructions oriented on the theorems to alleviate difficulty in acquiring the curated data and facilitate the digestion of theorem application from various perspectives. Evaluations on a wide range of public datasets exhibit that the model fine-tuned with our dataset achieves consistent improvements at varying mathematical levels compared to the backbone. And further ablation studies illustrate the effectiveness of our proposed evolutionary strategies on enhancing the model's capability of math problem-solving. Overall, extensive experiments reveal the potential of our proposed method which highlights the significance of aligning the problems with the concrete theorems for LLMs to alleviate hallucination and improve the models' mathematical reasoning capabilities.

Transformer-based language models (LMs) have demonstrated remarkable performance on many natural language tasks, yet to what extent LMs possess the capability of generalizing to unseen logical rules remains not explored sufficiently. In classical logic category, abductive, deductive and inductive (ADI) reasoning are defined as the fundamental reasoning types, sharing the identical reasoning primitives and properties, and some research have proposed that there exists mutual generalization across them. However, in the field of natural language processing, previous research generally study LMs’ ADI reasoning capabilities separately, overlooking the generalization across them. To bridge this gap, we propose UniADILR, a novel logical reasoning dataset crafted for assessing the generalization capabilities of LMs across different logical rules. Based on UniADILR, we conduct extensive investigations from various perspectives of LMs’ performance on ADI reasoning. The experimental results reveal the weakness of current LMs in terms of extrapolating to unseen rules and inspire a new insight for future research in logical reasoning.

Positional bias (PB), manifesting as nonuniform sensitivity across different contextual locations, significantly impairs long-context comprehension and processing capabilities.Previous studies have addressed PB either by modifying the underlying architectures or by employing extensive contextual awareness training.However, the former approach fails to effectively eliminate the substantial performance disparities, while the latter imposes significant data and computational overhead.To address PB effectively, we introduce Pos2Distill, a position to position knowledge distillation framework.Pos2Distill transfers the superior capabilities from advantageous positions to less favorable ones, thereby reducing the huge performance gaps.The conceptual principle is to leverage the inherent, position-induced disparity to counteract the PB itself.We identify distinct manifestations of PB under Retrieval and Reasoning paradigms, thereby designing two specialized instantiations: Pos2Distill-R 1 and Pos2Distill-R 2 respectively, both grounded in this core principle.By employing our approach, we achieve enhanced uniformity and significant performance gains across all contextual positions in long-context retrieval and reasoning tasks.Crucially, both specialized systems exhibit strong cross-task generalization mutually, while achieving superior performance on their respective tasks.

Offline meta-reinforcement learning (OMRL) enables reinforcement learning (RL) agents to adapt to unseen tasks without interacting with the environment. However, OMRL faces challenges such as Q-function overestimation and difficulties in inferring tasks correctly and robustly due to distribution discrepancy. In this paper, we introduce ConseRvative q-learning and task similarity mEAsuremenT for Offline meta-Reinforcement learning (CREATOR), a method to address these challenges using only offline datasets, without requiring additional interactions. To mitigate Q-function overestimation, we incorporate conservative Q-learning during training. We also propose a novel task similarity-based distance metric to improve the robustness of task inference. Experimental results demonstrate that the proposed CREATOR effectively reduces Q-function estimation errors, enhances task inference accuracy, and improves generalization performance across a range of challenging domains compared to existing methods.

Social media platforms has long served as open arenas where individuals discuss and change their opinions on various events, subsequently influencing the progression of these events. Public opinion, recognized as an important social signal, is instrumental in understanding the developmental patterns of social events and in guiding more informed responses. In light of this, we propose a sociologically-informed opinion prediction model, which integrates rich social interaction data with time series forecasting techniques using a graph neural network framework. This model, enriched by a sociological theoretical model, reflects the real-world dynamics of opinion evolution. Experimental results derived from three synthetic datasets and two real-world datasets indicate that incorporating user interaction data, along with more effective utilization of historical information, has led to a large improvement in the accuracy of opinion predictions. The source code and sample data for our study are available at https://github.com/RiikkaYang/SIGNN.

Board game offers a unique platform for exploring the capabilities of artificial intelligence in decision-making. It demands long-term strategic planning and opponent behaviors to refine decision-making. Since the success of AlphaGo family, learning agents have become pivotal methods for board game. However, current learning agents rarely incorporate planners or build interactive loops with opponents’ behaviors in decision-making. This paper proposes a novel planning-based model (BG-Planner) for strategic decision-making and long-term planning in board game. We propose a Graphplan-style network with alternating action and proposition layers to predict actions and assess wining rate. Further, an opponent modeling strategy is incorporated to predict opponent behaviors, assist decision-making and reduce competitive uncertainty. We also introduce a knowledge-based search tactic to enhance BG-Planner’s learning. Experimental results demonstrate that BG-Planner enhances the quality and efficiency of decision-making in the Gomoku game. It shows potential to improve deep planning strategies in decision-making intelligence.

Continual Pre-Training (CPT) has become a popular and effective method to apply strong foundation models to specific downstream tasks. In this work, we explore the learning dynamics throughout the CPT process for large language models (LLMs). We specifically focus on how general and downstream domain performance evolves at each training step, with domain performance measured via validation losses. We have observed that the CPT loss curve fundamentally characterizes the transition from one curve to another hidden curve, and could be described by decoupling the effects of distribution shift and learning rate (LR) annealing. We derive a CPT scaling law that combines the two factors, enabling the prediction of loss at any (continual) training steps and across learning rate schedules (LRS) in CPT. Our formulation presents a comprehensive understanding of several critical factors in CPT, including the learning rate, the training steps, and the distribution distance between PT and CPT datasets. Moreover, our approach can be adapted to customize training hyper-parameters to different CPT goals such as balancing general and domain-specific performance. Extensive experiments demonstrate that our scaling law holds across various CPT datasets and training hyper-parameters.

Agent-based modeling is a widely adopted approach for forecasting human opinion evolution trends. The rapid development of large language models (LLMs) introduces new perspectives to agent-based modeling, enabling sophisticated descriptions of agents’ personalities and facilitating researchers’ control over the interaction background. In this study, we propose a novel approach that leverages LLM-empowered agents to simulate the opinion distribution trends of given social issues. Our approach involves assigning agents with different personality traits based on the Five Factor Model, varying levels of background knowledge, and diverse initial opinions. We use two datasets to evaluate the performance of our framework. The results suggest that by employing our framework with appropriate prompt engineering, LLM agents can effectively forecast opinion evolution trends. The results of our framework outperform that of traditional numerical opinion dynamics methods. Our framework can effectively simulate key attributes of different social groups with agents, make agent interactions more reflective of real-world scenarios, and enable the automatic emergence of specific social psychological phenomena, such as the "backfire effect", through autonomous interactions among multiple agents.

Large reasoning models (LRMs) are proficient at generating explicit, step-by-step reasoning sequences before producing final answers. However, such detailed reasoning can introduce substantial computational overhead and latency, particularly for simple problems. To address this over-thinking problem, we explore how to equip LRMs with adaptive thinking capabilities—enabling them to dynamically decide whether or not to engage in explicit reasoning based on problem complexity. Building on R1-style distilled models, we observe that inserting a simple ellipsis ("...") into the prompt can stochastically trigger either a thinking or no-thinking mode, revealing a latent controllability in the reasoning behavior. Leveraging this property, we propose AutoThink, a multi-stage reinforcement learning (RL) framework that progressively optimizes reasoning policies via stage-wise reward shaping. AutoThink learns to invoke explicit reasoning only when necessary, while defaulting to succinct responses for simpler tasks. Experiments on five mainstream mathematical benchmarks demonstrate that AutoThink achieves favorable accuracy–efficiency trade-offs compared to recent prompting and RL-based pruning methods. It can be seamlessly integrated into any R1-style model, including both distilled and further fine-tuned variants. Notably, AutoThink improves relative accuracy by 6.4\% while reducing token usage by 52\% on DeepSeek-R1-Distill-Qwen-1.5B, establishing a scalable and adaptive reasoning paradigm for LRMs. Project Page: https://github.com/ScienceOne-AI/AutoThink.

Knowledge reasoning has been widely researched in knowledge graphs (KGs), but there has been relatively less research on hyper-relational KGs, which also plays an important role in downstream tasks. Existing reasoning methods on hyper-relational KGs are based on representation learning. Though this approach is effective, it lacks interpretability and ignores the graph structure information. In this paper, we make the first attempt at symbolic reasoning on hyper-relational KGs. We introduce rule extraction methods based on both individual facts and paths, and propose a rule-based symbolic reasoning approach, HyperPath. This approach is simple and interpretable, it can serve as a baseline model for symbolic reasoning in hyper-relational KGs. We provide experimental results on almost all datasets, including five large-scale datasets and seven sub-datasets of them. Experiments show that the expressive power of the proposed model is similar to simple neural networks like convolutional networks, but not as advanced as more complex networks such as Transformer and graph convolutional networks, which is consistent with the performance of symbolic methods on KGs. Furthermore, we also analyze the impact of rule length and hyperparameters on the model's performance, which can provide insights for future research in hypergraph symbolic reasoning.

In this paper, we investigate whether Large Language Models (LLMs) actively recall or retrieve their internal repositories of factual knowledge when faced with reasoning tasks.Through an analysis of LLMs' internal factual recall at each reasoning step via Knowledge Neurons, we reveal that LLMs fail to harness the critical factual associations under certain circumstances.Instead, they tend to opt for alternative, shortcut-like pathways to answer reasoning questions.By manually manipulating the recall process of parametric knowledge in LLMs, we demonstrate that enhancing this recall process directly improves reasoning performance whereas suppressing it leads to notable degradation.Furthermore, we assess the effect of Chain-of-Thought (CoT) prompting, a powerful technique for addressing complex reasoning tasks.Our findings indicate that CoT can intensify the recall of factual knowledge by encouraging LLMs to engage in orderly and reliable reasoning.Furthermore, we explored how contextual conflicts affect the retrieval of facts during the reasoning process to gain a comprehensive understanding of the factual recall behaviors of LLMs.

Performing logical reasoning based on prior knowledge is a crucial human cognitive ability and has been a long-standing objective in the field of artificial intelligence. Large language models based on transformer architecture have been a common approach for logical reasoning over text. However, the current language models often struggle to learn semantic information from logical expressions, resulting in underwhelming performance on logical reasoning tasks. In this paper, we propose a novel method to convert first-order logic (FOL) expressions to the form of a graph and integrate it with embeddings from language models to enhance their reasoning ability. The proposed method is designed to learn directly from FOL formulas and is able to generalize to any scenarios involving logical expressions. Experimental results demonstrate that the proposed method enhances the model’s ability of learning logical semantic representations, and thus it brings a significant improvement on the performance of complex reasoning tasks. The code is available at https://github.com/FOL-GNN.

In recent years, few-shot knowledge graph completion (FKGC) emerged as a prominent research problem, focused on utilizing a limited number of reference entity pairs to complete triples with unseen relations. Recent studies have attempted addressing this problem by modeling interactions between head and tail entities. However, existing FKGC methods represent semantics predominantly based on the neighborhood information of entities in the knowledge graph, thus can only infer the hidden and unobserved relations within the knowledge graph, limiting their reasoning capabilities. To overcome these limitations, we introduce text descriptions to FKGC and propose BERT-FKGC, a model capable of learning the integrated distribution of both the entity text descriptions and neighborhood information. By using a gating network that allows the model to dynamically select weights, our method can flexibly combine neighborhood information and textual descriptions. Besides addressing the prediction of unseen relations, our method is also capable of representing unseen entities. To validate the effectiveness of our model, we introduce a new dataset, FB15K-237-One, which includes textual descriptions for entities. We conduct extensive experiments on the FB15K-237-One dataset to validate the superiority of BERTFKGC.

Few-shot, especially one-shot learning is a prominent research area in the field of knowledge graphs (KGs), aiming to utilize a limited number of triples with unseen relations as reference information for inferring missing knowledge. Recent research focuses on improving the semantic representation of entity pairs using interactions between their head and tail entities. However, this method only considers the reference information as the measurement criterion without taking into account the potential impact of it on the reasoning process of the model. In this paper, we propose a novel method that utilizes factual information interactions. Firstly, we learn static representations of entities based on their neighborhood information. Subsequently, we learn relation adaptive representations by incorporating the reference information. This interactive modeling strengthens the association between entity representations and task relations while suppressing irrelevant relations. Extensive experiments demonstrate that our model outperforms state-of-the-art methods on two public datasets. Remarkably, on the NELL-One dataset for one-shot link prediction, our model achieves an improvement of 11.8% in MRR compared to the best baseline model.

Large language models(LLMs) have demonstrated remarkable planning and reasoning abilities, particularly as few-shot learners, when utilizing in-context learning. However, since LLMs are not grounded during training, they still encounter difficulties when acting as agents in tasks that require interaction with the environment, especially in scenarios that involve long-term and multistep interactions. Even when provided with a complete game trajectory as context, LLMs struggle to comprehend the meaning of each interaction step, and may easily hallucinate and fail. To address these challenges, we introduce the RRdE (Reasoning and Replanning during Exploration), a framework inspired by planning theories, designed for reasoning about actions and planning subgoals in complex interactive environments. The RRdE method can integrate the long-term planning ability and tooluse ability of LLMs, and transform the long-term sequential decision problem into a relatively simple reasoning problem, thereby reducing the error behavior caused by excessive context. We devise a reflection-based goal decomposition and replanning scheme, which enables the agent to overcome the strict sub-goal dependency problem caused by long-term goal planning. Consequently, RRdE achieves state-of-the-art performance in the few-shot learning setting in both AlfWorld and ScienceWorld environments, accomplishing 132 out of 134 test tasks in AlfWorld, and obtaining an average score of 82.16 in the 30 more complex and challenging scientific tasks in ScienceWorld, successfully completing 7 tasks with a full score of 100.

Graph representation learning is an emerging area for graph analysis and inference. However, existing approaches for large-scale graphs either sample nodes in sequential walks or manipulate the adjacency matrices of graphs. The former approach can cause sampling bias against less-connected nodes, whereas the latter may suffer from sparsity that exists in many real-world graphs. To learn from structural information in a graph more efficiently and comprehensively, this paper proposes a new graph representation learning approach inspired by the cognitive model of spreading-activation mechanisms in human memory. This approach learns node embeddings by adopting a graph activation model that allows nodes to “activate” their neighbors and spread their own structural information to other nodes through the paths simultaneously. Comprehensive experiments demonstrate that the proposed model performs better than existing methods on several empirical datasets for multiple graph inference tasks. Meanwhile, the spreading-activation-based model is computationally more efficient than existing approaches–the training process converges after only a small number of iterations, and the training time is linear in the number of edges in a graph. The proposed method works for both homogeneous and heterogeneous graphs.

Script event prediction aims to infer subsequent events given an incomplete script. It requires a deep understanding of events, and can provide support for a variety of tasks. Existing models rarely consider the relational knowledge between events, they regard scripts as sequences or graphs, which cannot capture the relational information between events and the semantic information of script sequences jointly. To address this issue, we propose a new script form, relational event chain, that combines event chains and relational graphs. We also introduce a new model, relational-transformer, to learn embeddings based on this new script form. In particular, we first extract the relationship between events from an event knowledge graph to formalize scripts as relational event chains, then use the relational-transformer to calculate the likelihood of different candidate events, where the model learns event embeddings that encode both semantic and relational knowledge by combining transformers and graph neural networks (GNNs). Experimental results on both one-step inference and multistep inference tasks show that our model can outperform existing baselines, indicating the validity of encoding relational knowledge into event embeddings. The influence of using different model structures and different types of relational knowledge is analyzed as well.

Spiking Neural Networks (SNNs), also referred to as the third generation of artificial neural networks, are highly prized for their biological realism, robustness, and low power requirements. SNNs are crucial in fields such as object detection, image recognition, etc. The classification of short text plays an significant role in the development of chatbots and intent detection. It is also an important task that is widely used in many downstream tasks. However, studies applying SNNs to short text classification are limited. This paper provides a new model that uses SNNs to classify short texts. SNNs are difficult to train directly when using deep models and cannot employ large-scale language models to learn good embeddings. To resolve the challenge, we apply the character-level encoding method and convert analog neural networks into SNNs. To begin with, we represent character-level text using a temporal-and-rate joint horizontal encoding method. Then we develop a tailored deep Convolutional Neural Network (CNN) model for classifying texts. At the inference stage, we convert the tailored CNN model into an SNN model. To test the effectiveness of the proposed method, we conduct text encoding experiments on the NAMES dataset and short text classification experiments on both the 20-newsgroups dataset and the emoji-mult dataset. Experiments demonstrate that the proposed method can obtain classification accuracies that are better than or comparable to other methods.

Transformer models have been widely used in the filed of natural language processing due to their powerful learning ability. Nevertheless, recent studies have shown that transformer models are vulnerable to the maliciously crafted adversarial examples. In the challenging black box setting, main stream textual adversarial attacks typically consist of two steps: Word Importance Ranking (WIR) and word transformation. The attack performance is highly dependent on the ranking of words. Existing WIR methods are designed with heuristic rules, which lack theoretical guarantee and require a large amount of queries. To address this issue, we design a textual coalitional game and propose PWSHAP, which is a plug-and-in WIR method employing Shapley value to determine the significance of each word based on its impact on the classification. Through extensive experiments on three benchmark datasets and model architectures, we illustrate that the proposed PWSHAP achieve the-state-of-the-art attack success rate with significant fewer queries to the classification model. Meanwhile, the generated adversarial examples are more natural and coherent compared to the strong baselines.

Multimodal Named Entity Recognition (MNER) in social media posts plays an important role in both security and natural language processing domains. Existing approaches mainly include extracting useful visual features from images, and integrating them into text representation for NER via multimodal fusion. Nevertheless, there is potential correlation among samples in the dataset, but is ignored by most of the existing studies. In this paper, we propose a potential correlation-enhanced network (PCEN) for MNER. Specifically, we (1) consider the potential correlation as an important visual feature for MNER, and (2) utilize it to guide the final recognition of entities. To tackle the first issue, we employ unsupervised clustering to divide the images of training samples into clusters, and take the trainable embedding of each cluster label as a visual feature because samples with the same cluster label have higher potential correlation. To tackle the second issue, we argue that the samples in the same cluster are more likely to have similar distributions of entity types in their text. We design an inconsistency loss to encourage the consistency between the entity recognition result of each sample and the pre-trained entity type distribution of the corresponding cluster this sample belongs to. Experiments on two MNER benchmarks demonstrate the effectiveness of our proposed method.

Events such as crises, public opinion issues, and social hotspots usually follow certain patterns. From a large amount of historical data, we can extract these patterns to predict future events. This valuable task can be viewed as the Temporal Knowledge Graph (TKG) inference problem, as TKGs are widely employed to sketch ongoing events. However, most of the traditional TKG inference methods mainly focus only on entity prediction and do not take into consideration the variable length of information summarized from events at different periods. To address these challenges, we propose a new collaborative entity- relation prediction method called Predicting the Future Without Forgetting (PFWF). PFWF introduces historical representation to deal with the issue posed by the variability of information length. We also treat the TKG prediction task as a continual learning problem that prevents training new models from scratch when new data are added, as real-world knowledge graphs are constantly evolving. We validated the effectiveness of PFWF on four public TKG datasets related to crisis events in offline and online continual learning settings.

Representation learning is a fundamental task in knowledge graph-related research and applications. Most existing approaches learn representations for entities and relations only based on static facts, where temporal information has been ignored completely. This paper aims to learn time-aware representations for entities and relations in knowledge graphs. Based on how temporal information affects the learned embeddings, we propose three assumptions and build three different models, BTS, ETS, and RTS, respectively. In these models, we build two separate embedding spaces for entities and relations, the standard translation condition is checked after projecting embedding vectors between these spaces by model-specific transformations. As to the performance, the proposed RTS model achieves state-of-the-art results in three experiments conducted on two datasets: YAGO11k and Wikidata12k, which validates the effectiveness of our model. Comparing the results of all three models, we find that relation embeddings are time-sensitive and form natural ordering, while the effects of time on entity embeddings can be safely ignored for translation-based methods. Experiments also show that our findings can be used to simplify other existing models like HyTE.

Natural Language Inference (NLI) is a vital task in natural language processing.It aims to identify the logical relationship between two sentences.Most of the existing approaches make such inference based on semantic knowledge obtained through training corpus.The adoption of background knowledge is rarely seen or limited to a few specific types.In this paper, we propose a novel Knowledge Graph-enhanced NLI (KGNLI) model to leverage the usage of background knowledge stored in knowledge graphs in the field of NLI.KGNLI model consists of three components: a semantic-relation representation module, a knowledge-relation representation module, and a label prediction module.Different from previous methods, various kinds of background knowledge can be flexibly combined in the proposed KGNLI model.Experiments on four benchmarks, SNLI, MultiNLI, SciTail, and BNLI, validate the effectiveness of our model.

Knowledge graph completion, one of the most important research questions in knowledge graphs, aims at predicting missing links in a given graph. Current mainstream approaches adopt high-quality embeddings of entities and relations of the graph to improve their performances. However, it is not easy to devise a universal embedding learner that can fit various scenarios. In this paper, we propose a general-purpose framework which can be employed to improve the performance of knowledge graph completion. Specifically, given an arbitrary knowledge graph completion model, we first run the original model to get a ranked entity list. Then, we combine the query and the top ranked entities with attention mechanism, re-rank all these entities by feeding the combined vector into a neural network. The proposed re-ranking phase can be conveniently added to a variety of models to improve their performance without substantial modification. We conduct experiments on four datasets: WN18, FB15k, WN18RR, and FB15k-237. We choose TransE, TransH, TransD, DistMult, and ANALOGY as base models. Experiments on these datasets and models validate the effectiveness of the proposed re-ranking framework. We further explore the influence of the number of top ranked entities used in the re-ranking phase. We also test other attention mechanism to determine the most effective one, and found that vanilla attention mechanism can balance accuracy and complexity.

Social networks have significantly altered the behavior patterns of netizens all around the world. Therefore, accurate and expressive model of social users is increasingly demanded as it pose great value in a variety of scenarios, such as e-commerce, cyber security, and entertainment to name a few. In this paper, we propose the Pop Culture Attention Writing Model (PAWM) to explore the writing patterns of social users by explicitly capturing the influence of Internet pop culture ingredients with an attention mechanism. The writing pattern representations are learned by a memory network through storing and updating historical latent patterns. We then develop the Deep Social User Model via jointly modeling basic properties of social users, temporal contents, and the learned writing patterns based on PAWM. This paper is the first trial, to the best of our knowledge, which captures Internet pop culture information and applies deep neural network to model user writing pattern. A series of experiments conducted on social bot detection and social user identification demonstrate and validate the effectiveness of the proposed models.

Knowledge graph, or knowledge base, plays an important role in a variety of applications in the field of artificial intelligence. In both research and application of knowledge graph, knowledge representation learning is one of the fundamental tasks. Existing representation learning approaches are mainly based on structural knowledge between entities and relations, while knowledge among entities per se is largely ignored. Though a few approaches integrated entity knowledge while learning representations, these methods lack the flexibility to apply to multimodalities. To tackle this problem, in this paper, we propose a new representation learning method, TransAE, by combining multimodal autoencoder with TransE model, where TransE is a simple and effective representation learning method for knowledge graphs. In TransAE, the hidden layer of autoencoder is used as the representation of entities in the TransE model, thus it encodes not only the structural knowledge, but also the multimodal knowledge, such as visual and textural knowledge, into the final representation. Compared with traditional methods based on only structural knowledge, TransAE can significantly improve the performance in the sense of link prediction and triplet classification. Also, TransAE has the ability to learn representations for entities out of knowledge base in zero-shot. Experiments on various tasks demonstrate the effectiveness of our proposed TransAE method.

While doing summarization, human needs to understand the whole document, rather than separately understanding each sentence in the document. However, inter-sentence features within one document are not adequately modeled by previous neural network-based models that almost use only one layer recurrent neural network as document encoder. To learn high quality context-aware representation, we propose a shortcut-stacked document encoder for extractive summarization. We use multiple stacked bidirectional long short-term memory (LSTM) layers and add shortcut connections between LSTM layers to increase representation capacity. The shortcut-stacked document encoder is built on a temporal convolutional neural network-based sentence encoder to capture the hierarchical structure of the document. Then sentence representations encoded by document encoder are fed to a sentence selection classifier for summary extraction. Experiments on the well-known CNN/Daily Mail dataset show that the proposed model outperforms several recently proposed strong baselines, including both extractive and abstractive neural network-based models. Furthermore, the ablation analysis and position analysis also demonstrate the effectiveness of the proposed shortcut-stacked document encoder.

Cognitive dissonance is a ubiquitous phenomenon which can be applied in various fields potentially. In this paper, we study cognitive dissonance through empirical analysis on social media platforms. Our study focuses on a recent “reversal event” - a topic or event experiencing a reversed development trend because of the new facts. Through statistical analysis and semantic analysis based methods, we found that (1) after the event is revised, the performance of the original followers were abnormal, which is consistent with the existence of cognitive dissonance; (2) the followers' attitude afterwards usually tended to maintain their previous behaviors. This research provides a primary building block towards the mental inference based behavior prediction for social media users, which is of great value for security related research issues.

Nowadays, it is common that a netizen creates multiple accounts across social platforms. Mapping accounts across platforms could facilitate various applications in security. Existing methods usually focus on profile and network based features. In this paper, we concentrate on capturing dynamic information of social users and present a deep dynamic user mapping model to identify the accounts across platforms. The proposed model captures dynamic latent features from three aspects including posting pattern, writing pattern, and emotional fluctuation. We also develop a matching network that fuses dynamic and traditional features to identify accounts. To the best knowledge of ourselves, this is the first trial that applies deep neural network in mapping users with dynamic information. Experiments on real world dataset demonstrated the effectiveness of the proposed method.

With the great success of neural networks, it is important to improve the information security of application systems based on them. This paper investigates a scenario where an attacker eavesdrops the intermediate representation computed by the encoder layers and tries to recover the private information of the input text. We propose a new metric to evaluate the encoder's ability to protect privacy and evaluate the Transformer-based encoder, which is the first privacy research conducted on Transformer-based neural networks. We also propose an adversarial training method to enhance the privacy of Transformer-based neural networks.

Cryptocurrency is a cutting-edge Fintech innovation and currently a worldwide hotspot. However, the high-speed evolution of it has already caused a series of public security related events all around the world. Cryptocurrency was built initially as a possible implementation of digital currency, then various derivatives were created in a variety of fields such as financial transactions, capital management, and even nonmonetary applications. This paper aims to offer analytical insights to help understand cryptocurrency by treating it as a financial asset. We position cryptocurrency by comparing its dynamic characteristics with two traditional and massively adopted financial assets: foreign exchange and stock. Based on the daily close prices about four years, we first construct the correlation matrices and asset trees of all three markets, then conduct comparisons on five properties: volatility, centrality, clustering structure, robustness, and risk. Our investigation suggests that the dynamics of cryptocurrency are more similar to stock. As to the robustness and clustering structure, our analysis shows cryptocurrency market is more fragile than stock market, thus it is currently a high-risk financial market. Our work is the first to study cryptocurrency with the help of well-understood financial assets and may shed some light on investment decisions, regulation, and legislation.

The anonymity and decentralization of Bitcoin make it widely accepted in illegal transactions, such as money laundering, drug and weapon trafficking, gambling, to name a few, which has already caused significant security risk all around the world. The obvious de-anonymity approach that matches transaction addresses and users is not possible in practice due to limited annotated data set. In this paper, we divide addresses into four types, exchange, gambling, service, and general, and propose targeted addresses identification algorithms with high fault tolerance which may be employed in a wide range of applications. We use network representation learning to extract features and train imbalanced multi-classifiers. Experimental results validated the effectiveness of the proposed method.

Nowadays, more and more sexual harassment personal stories have been shared on social media. To better monitor and analyze the extent of sexual harassment based on these social media data, we need to automatically categorize different forms of sexual harassment personal stories. Existing methods apply convolutional neural network (CNN) with different convolution window sizes to this text classification task. However, the previous CNN models do not provide an effective way to synthesize window size-related local representations, but simply concatenate all local representations together. To address this problem, we propose a new density matrix encoder, inspired by quantum mechanics, to encode local representations as particles in quantum state and generate a global representation as quantum mixed system for each story. Experiment on SafeCity dataset shows that our model outperforms CNN baseline and achieves better performance than the state-of-the-art model when considering both accuracy and speed, demonstrating the effectiveness of the proposed density matrix encoder.

Cryptocurrency is a rapid developing financial technology innovation which has attracted a large number of people around the world. The high-speed evolution, radical price fluctuations of cryptocurrency, and the inconsistent attitudes of monetary authorities in different countries have triggered panic and chain reactions towards the application and adoption of cryptocurrency and have caused public security related events. So far, a lot of researches and analyses have focused on just one or only a few number of cryptocurrencies, a comprehensive analysis of the whole cryptocurrency market and its systemic risk is still lacking. In this paper, we analyze the dynamics and systemic risk of the cryptocurrency market based on the public available price history. We first validated that the correlation matrix and asset tree are good tools to analyze the risk and stability of the cryptocurrency market. Furthermore, consistent with public perception, our quantitative analysis reveals that the cryptocurrency market is relatively fragile and unstable. Our work is the first to investigate the systemic risk of the whole cryptocurrency market and may shed some light on cryptocurrency related investment decision, regulation, and legislation.

Knowledge graph plays an important role in detection, prediction, early warning, and other security related applications. A fundamental task in applying knowledge graph is the so-called multi-hop reasoning, which focuses on inferring new relations between entities. In this paper, we introduce attention mechanism to the classic compositional method. After finding reasoning paths between entities, we aggregate these paths' embeddings into one according to their attentions, and infer the relation of entities based on the combined embedding. Two experiments on NELL-995 dataset, fact prediction and link prediction, validated that our method outperforms all baselines.

Bots are regarded as the most common kind of malwares in the era of Web 2.0. In recent years, Internet has been populated by hundreds of millions of bots, especially on social media. Thus, the demand on effective and efficient bot detection algorithms is more urgent than ever. Existing works have partly satisfied this requirement by way of laborious feature engineering. In this paper, we propose a deep bot detection model aiming to learn an effective representation of social user and then detect social bots by jointly modeling social behavior and content information. The proposed model learns the representation of social behavior by encoding both endogenous and exogenous factors which affect user behavior. As to the representation of content, we regard the user content as temporal text data instead of just plain text as be treated in other existing works to extract semantic information and latent temporal patterns. To the best of our knowledge, this is the first trial that applies deep learning in modeling social users and accomplishing social bot detection. Experiments on real world dataset collected from Twitter demonstrate the effectiveness of the proposed model.

Predicting meme burst is of great relevance to develop security-related detecting and early warning capabilities. In this paper, we propose a feature-based method for real-time meme burst predictions, namely “Semantic, Network, and Time” (SNAT). By considering the potential characteristics of bursty memes, such as the semantics and spatio-temporal characteristics during their propagation, SNAT is capable of capturing meme burst at the very beginning and in real time. Experimental results prove the effectiveness of SNAT in terms of both fixed-time and real-time meme burst prediction tasks.

Public sentiment permeated through social media is usually regarded as an important measure for public opinion monitoring, policy making, and so forth. However, the deluge of user-generated content in web, especially in social platform, causes great challenge to public sentiment analysis tasks. Therefore, Web-derived Emotional Word Detection (WEWD) is proposed as a fundamental tool aims to alleviate this problem. Most previous works on WEWD focus on rules, syntax, and sentence structures, a few utilize semantic information which has the potential to further increase the accuracy and efficiency of WEWD. In this paper, we propose a Global-Local Latent Semantic (GLLS) framework for WEWD to make a full use of latent semantic information with the help of multiple sense word embedding technology. We devise two computational WEWD models, called Ensemble GLLS (EGLLS) and Deep GLLS (DGLLS). EGLLS exploits an ensemble learning way to fuse the global and local latent semantics while DGLLS takes advantage of deep neural network. We also design an old-new corpus enrich technique to help increase the effectiveness of the overall training and detecting process. To the best of our knowledge, this is the first work which applies multiple sense word embedding and deep neural network in WEWD related tasks. Experiments on real datasets demonstrate the effectiveness of the proposed idea and methods.

Social bots are regarded as the most common kind of malwares in social platform. They can produce fake messages, spread rumours, and even manipulate public opinions. Recently, massive social bots are created and widely spread in social platform, they bring negative effects to public and netizen security. Bot detection aims to distinguish bots from human and it catches more and more attentions in recent years. In this paper, we propose a behavior enhanced deep model (BeDM) for bot detection. The proposed model regards user content as temporal text data instead of plain text to extract latent temporal patterns. Moreover, BeDM fuses content information and behavior information using deep learning method. To the best of our knowledge, this is the first trial that applies deep neural network in bot detection. Experiments on real world dataset collected from Twitter also demonstrate the effectiveness of our proposed model.

In this paper, we propose a novel text representation paradigm and a set of follow-up text representation models based on cognitive psychology theories. The intuition of our study is that the knowledge implied in a large collection of documents may improve the understanding of single documents. Based on cognitive psychology theories, we propose a general text enrichment framework, study the key factors to enable activation of implicit information, and develop new text representation methods to enrich text with the implicit information. Our study aims to mimic some aspects of human cognitive procedure in which given stimulant words serve to activate understanding implicit concepts. By incorporating human cognition into text representation, the proposed models advance existing studies by mining implicit information from given text and coordinating with most existing text representation approaches at the same time, which essentially bridges the gap between explicit and implicit information. Experiments on multiple tasks show that the implicit information activated by our proposed models matches human intuition and significantly improves the performance of the text mining tasks as well.

Topic modeling is a popular text mining technique for extracting latent semantics from text. It can be widely applied in intelligence analyzing, anti-terrorist, and various other security related tasks. Most existing topic models only focus exclusively on the text literally, and disregard rich contextual, cultural, and language background, hindering the understanding and discovering of the key clues implied in the text. Based on cognitive psychology theories, we justify the classical psychological activation theory named Adaptive Control of Thought from the perspective of information theory. Then, we propose a fast and loosely-coupled activation presentation of text for topic models. Our method mimics the aspect of human cognitive procedure when facing the activation of new concepts based on word correlations and word frequencies. Experimental results on multiple tasks show that our activation presentation models can significantly improve the performance of the topic models with linear time consumption.

In this paper, we propose a new role analyzing paradigm for social networks enlightened by topic modeling, which can be adopted as a primitive building block in various security related tasks, such as hidden community finding, important person recognizing and so on. We first present the social network under analyzing as a heterogeneous network constructed by both the users and the subjects discussed among them. We then view this network in a Bag-of-Users schema, which mimics its classical Bag-of-Words counterpart. In this schema, the subjects discussed are treated as “documents” while the users are treated as “words” which construct the “documents”. Based on this novel presentation, we finally apply topic modeling technology to perform the social role clustering. Experiments on a practical security-related social network dataset prove the effectiveness of our approach.

Public sentiment permeated through social media is usually regarded as an important measure for hot event detecting, policy making and so forth, hence many governments and intelligence agencies have been launching various initiatives to facilitate theories, technologies and systems toward monitoring its fluctuation. Recently, massive new words are created and widely spread in social media, and they pose a great influence on sentiment analysis. Facing this situation, most previous work still just add those new words into sentiment lexicon, none of the existed researches focuses on the role and influence of new words in emotional expression. In this paper, we pay more attention to the influence of new words and propose two novel new words based sentiment analysis methods, named NWLb and NWSA, the former only with the help of lexicon and the latter further incorporates machine learning, which utilize the distinctive role of new words to improve the effectiveness of sentiment analysis in social media. Experiments on real social media dataset demonstrate the effectiveness and performance of our methods.

Recently, there is a rapid growth of the online auctions in e-commerce platforms, in which small and medium-sized enterprises (SMEs) heavily depend on the advertising systems. We need to design flexible price mechanisms to reduce the competition of SMEs without affecting competitive large companies. In this paper, a probabilistic price mechanism design approach is investigated for online auctions. Utilizing this approach, we first introduce simple mechanisms as a tool for designing new mechanisms. Based on a simple and a classical mechanism probabilistic price mechanisms are designed for online auctions and their properties are analysed. Furthermore, two mechanism design algorithms are suggested for different online auction scenarios. Experiments are presented to demonstrate the flexibility and the effictiveness of the proposed probabilistic mechanism design approach.

Quality score (QS) plays a critical role in sponsored search advertising (SSA) auctions, and in practice is closely correlated to the historical click-through rate (CTR) of an advertisement. The CTR-QS correlation may impose great influence on advertisers' positioning strategies of selecting the targeting slots in the sponsored list. In the literature, however, QS is implicitly assumed to be an independent variable and exogenously assigned by Web search engines, so that little theoretical or managerial insights can be offered to help understand the positioning dynamics in SSA auctions with CTR-QS correlation. We strive to bridge this research gap in this paper. Based on a discrete time-dependent optimal control model, which explicitly captures the relationship between the historical CTR and QS, we determine the optimal strategy for revenue-maximizing advertisers' QS-based positioning decisions through a policy-iteration-based numerical approximation method. We also investigate two practically-used heuristic strategies, namely the greedy and farsighted positioning strategies, aiming to examine and help understand advertisers' real-world positioning dynamics. Our analysis indicates that both the optimal and greedy positioning strategies lead advertisers to monotonically increase or decrease their targeting slots over time, which may cause a polarization trend emerging in SSA markets. Meanwhile, the farsighted positioning strategy can accelerate the polarization. Our simulations show that both the greedy and farsighted strategies have good revenue performance. Our findings indicate that advertisers should monotonically adjust their targeting positions to maximize their revenue in CTR-QS correlated SSA auctions. Our findings also highlight the need for Web search engine companies to set a lowered weight for historical CTRs or use position-normalized CTRs in their QS measurements, so as to suppress the polarization trend.

This paper investigates the collaboration patterns and research topic trends in the publications of the IEEE Transactions on Intelligent Transportation Systems (T-ITS) over the past decade. We find that coauthorship is prevalent and that the coauthorship networks possess the scale-free property on high degree nodes. Collaborations usually occur within the same research institutions and countries. Interorganization/region collaboration structures are usually connected through a few productive/high-impact authors. Typical international collaborations are between the U.S. and other countries such as China, Germany, U.K., and Italy. Active topics studied in IEEE T-ITS publications in the past ten years include traffic management and machine vision, among others. Authors can be partitioned into common interest groups, of which machine vision and automatic vehicle control attract more researchers.

This paper presents a bibliographic analysis of the papers published in the IEEE Transactions on Intelligent Transportation Systems (T-ITS). We identify the most productive and high-impact authors, institutions, and countries/regions. We find that research on intelligent transportation systems is dominated by U.S. researchers and institutions and that China and Japan are the second most productive countries. According to this analysis, M. M. Trivedi, N. P. Papanikolopoulos, and P. A. Ioannou are the three most productive and influential authors in the IEEE T-ITS, whereas the Massachusetts Institute of Technology, Cambridge, the University of California, San Diego, and the University of Minnesota, Minneapolis, are three of the most productive and influential institutions in the IEEE T-ITS.