With the rise of linguistic steganography, the robustness to withstand minor perturbations like textual edits or tokenization shifts remains a critical yet crucial challenge. To address this, we propose a novel robustness-enhancing method, Position-Agnostic Probabilistic Generation, which combines semantic clustering with probabilistic generation control. During decoding, the model is softly guided to prefer or avoid specific token sets through time-aware probability boosting, enabling robust bit embedding without relying on fixed token positions. These token sets are constructed from semantically coherent clusters derived from the language model's vocabulary and expanded to ensure fluency. A dynamic, time-aware boosting strategy is then applied to gradually amplify the likelihood of valid tokens throughout the generation. Experimental results demonstrate that our method consistently outperforms baselines in preserving hidden information under 10 types of perturbations.

Difficulty Controllable Question Generation (DCQG) for reading comprehension learns to generate questions for measuring the reading abilities of examinees, playing a crucial role in educational scenarios. This work studies answer-aware DCQG, a challenging task that requires the generated questions to remain faithful to the assigned answer and match the desired difficulty level at the same time. To this end, we first propose an effective two-stage framework, Asking Questions with Thoughts (AQT), to guide a backbone large language model (LLM) to generate questions that are both faithful and difficulty-aware through conducting in-depth self-thinking. Then, we introduce a novel Posterior Knowledge Distillation (PKD) to efficiently fine-tune AQT by distilling knowledge from posterior inference. Finally, to address the scarcity of DCQG datasets, we use an efficient LLM Pretest-based Difficulty Estimation (LP-DE) to automatically construct DCQG datasets from common QG/QA datasets. Extensive experiments prove that our methods have promising results in terms of both faithfulness and difficulty awareness.

Traditional conversational systems can only access the given query during the response generation, leading to meaningless responses. To this end, researchers proposed to enhance dialogue generation by integrating external knowledge. Although such methods have achieved remarkable gains, the use of only single-source knowledge often makes existing knowledge-enhanced methods degenerate into traditional models in real scenarios because of the insufficient knowledge coverage of single-source knowledge. To improve the applicability of knowledge-enhanced methods, we propose two novel frameworks to use heterogeneous knowledge from multiple sources. We first propose an MHKD-Seq2Seq framework, which can use different heterogeneous knowledge by identifying abstract-level knowledge behaviors; meanwhile, a Diffuse-Aggregate scheme is used to process multiple knowledge simultaneously and produce a unified result. The next framework MHKD-ARPLM can leverage the advantages of pretrained language models with Knowledge Linearization techniques. In experiments, we collected dialogues from previously open-released datasets and built a multi-source knowledge-aligned dataset TriKE-Weibo, which involves three knowledge sources: commonsense, texts, and infobox tables. Extensive evaluations demonstrate the performance leadership of our approaches against competitive baseline models.

Recent studies have shown that graph learning models are highly vulnerable to adversarial attacks, and network alignment methods are no exception. How to enhance the robustness of network alignment against adversarial attacks remains an open research problem. In this paper, we propose a robust network alignment solution, RNA, for offering preemptive protection of existing network alignment algorithms, enhanced with the guidance of effective adversarial attacks. First, we analyze how popular iterative gradient-based adversarial attack techniques suffer from gradient vanishing issues and show a fake sense of attack effectiveness. Based on dynamical isometry theory, an attack signal scaling (ASS) method with established upper bound of feasible signal scaling is introduced to alleviate the gradient vanishing issues for effective adversarial attacks while maintaining the decision boundary of network alignment. Second, we develop an adversarial perturbation elimination (APE) model to neutralize adversarial nodes in vulnerable space to adversarial-free nodes in safe area, by integrating Dirac delta approximation (DDA) techniques and the LSTM models. Our proposed APE method is able to provide proactive protection to existing network alignment algorithms against adversarial attacks. The theoretical analysis demonstrates the existence of an optimal distribution for the APE model to reach a lower bound. Last but not least, extensive evaluation on real datasets presents that RNA is able to offer the preemptive protection to trained network alignment methods against three popular adversarial attack models.

Real-world graphs are usually very sparse in terms of inadequate edges and labels as well as have poor quality due to a large amount of noisy data. In this paper, we propose a classification-based local vertex ranking architecture through dual generative adversarial networks in the semi-supervised setting, DQGAN, for analyzing sparse noisy graphs with rarely labeled data. First, we develop a quadruple generative adversarial ClassNet model to address the noisy data and data sparsity issues as well as to classify each vertex into K classes by automatically creating imaginary/real-looking supplementary labeled vertices with the quite different/similar distributions as real vertices, without the high cost of multi-step graph propagation, heterogeneous graph mining, and iterative weight learning. In addition, the vertex label vicinity is incorporated into the classification model to capture the pairwise vertex closeness based on the labeling and align the vertex label vicinity with the well-known vertex homophily for preserving the original structural semantics in the classification space. Second, we present a quintuple generative adversarial RankNet framework to locally rank each vertex on each of K classes by designing the game of multiple competitors utilizing the mix of real and noisy data to fight against each other, for improving the robustness of local vertex ranking to noisy data with few help from human efforts. The cycle ranking consistency strategy is designed to make the ranking quality verifiable through the bidirectional information-lossless translations between the original features and the ranking features. We propose to utilize the relaxed local PageRank property to produce high-quality local vertex ranking results in the context of information networks. Third but last, extensive evaluation on real graph datasets demonstrates that DQGAN outperforms existing representative methods in terms of both classification and ranking in the semi-supervised setting.

Traditional network representation learning techniques aim to learn latent low-dimensional representation of vertices in graphs. This paper presents a novel edge representation learning framework, GANDLERL, that combines generative adversarial network based multi-label classification with density-adaptive local edge representation learning for producing high-quality low-dimensional edge representations. First, we design a generative adversarial network based multi-label edge classification model to classify rarely labeled edges in graphs with a large amount of noise data into K classes. A four-player zero-sum game model, with the mixed training of true and real-looking fake edges as well as a contrastive loss containing a similar-loss and a dissimilar-loss, is proposed to improve the classification quality of unlabeled edges. Second, a local autoencoder edge representation learning method is developed to design K local representation learning models, each with individual parameters and structure to perform local representation learning on each of K classification-based subgraphs with unique local characteristics and jointly optimize the loss functions within and across classes. Third but last, we propose a density-adaptive edge representation learning method with the optimization at both edge and subgraph levels to address the representation learning of graph data with highly imbalanced vertex degree and edge distribution.

Sentence-level sentiment classification is important to understand users' fine-grained opinions. Existing methods for sentence-level sentiment classification are mainly based on supervised learning. However, it is difficult to obtain sentiment labels of sentences since manual annotation is expensive and time-consuming. In this paper, we propose an approach for sentence-level sentiment classification without the need of sentence labels. More specifically, we propose a unified framework to incorporate two types of weak supervision, i.e., document-level and word-level sentiment labels, to learn the sentence-level sentiment classifier. In addition, the contextual information of sentences and words extracted from unlabeled sentences is incorporated into our approach to enhance the learning of sentiment classifier. Experiments on benchmark datasets show that our approach can effectively improve the performance of sentence-level sentiment classification.

Sentiment domain adaptation is widely studied to tackle the domain-dependence problem in sentiment analysis field. Existing domain adaptation methods usually train a sentiment classifier in a source domain and adapt it to the target domain using transfer learning techniques. However, when the sentiment feature distributions of the source and target domains are significantly different, the adaptation performance will heavily decline. In this paper, we propose a new sentiment domain adaptation approach by adapting the sentiment knowledge in general-purpose sentiment lexicons to a specific domain. Since the general sentiment words of general-purpose sentiment lexicons usually convey consistent sentiments in different domains, they have better generalization performance than the sentiment classifier trained in a source domain. In addition, we propose to extract various kinds of contextual sentiment knowledge from massive unlabeled samples in target domain and formulate them as sentiment relations among sentiment expressions. It can propagate the sentiment information in general sentiment words to massive domain-specific sentiment expressions. Besides, we propose a unified framework to incorporate these different kinds of sentiment knowledge and learn an accurate domain-specific sentiment classifier for target domain. Moreover, we propose an efficient optimization algorithm to solve the model of our approach. Extensive experiments on benchmark datasets validate the effectiveness and efficiency of our approach.