Detoxification in large language models (LLMs) remains a significant research challenge.Existing decoding detoxification methods are all based on external constraints, which require additional resource overhead and lose generation fluency.This work innovatively proposes Detoxification with Self-Constrained Decoding (DSCD), a novel method for LLMs detoxification without parameter fine-tuning.DSCD strengthens the inner next-token distribution of the safety layer while weakening that of hallucination and toxic layer during output generation.This effectively diminishes toxicity and enhances output safety.DSCD offers lightweight, high compatibility, and plug-andplay capabilities, readily integrating with existing detoxification methods for further performance improvement.Extensive experiments on representative open-source LLMs and public datasets validate DSCD's effectiveness, demonstrating state-of-the-art (SOTA) performance in both detoxification and generation fluency, with superior efficiency compared to existing methods.These results highlight DSCD's potential as a practical and scalable solution for safer LLM deployments.For more details, please refer to the project repository: https://github.com/ZHANGJINKUI/DSCD.

Link prediction in multilayer networks aims to predict missing links at the target layer by incorporating structural information from both auxiliary layers and the target layer. Existing methods tend to learn layer-specific knowledge to maximize the link prediction performance on a specific network layer. However, they have difficulty incorporating multilayer structural information to improve the link prediction performance. Therefore, we propose a Multiple Attention Layer-shareable Method (MALM) for link prediction in multilayer networks, which consists of a feature encoder, a knowledge learner, and a fusion predictor. The feature encoder introduces multiple attention mechanisms to encode the feature representations of links by differentiating the importance of structural information for each link. In cooperation with the feature encoder, the knowledge learner splits the link prediction tasks into different layers and employs meta-learning to learn layer-shareable knowledge from these link prediction tasks. Finally, the fusion predictor combines the learned layer-shareable knowledge with the layer-specific knowledge at the target layer for link prediction. Experiments on real-world datasets demonstrate that the proposed MALM outperforms existing state-of-the-art baselines in link prediction in multilayer networks.

Link prediction has received increased attention in social network analysis. One of the unique challenges in heterogeneous social networks is link prediction in new link types without verified link information, such as recommending products to new overseas groups. Existing link prediction models tend to learn type-specific knowledge on specific link types and predict missing or future links on the same link types. However, because of the uncertainty of new link types in the evolving process of social networks, it is difficult to collect sufficient verified link information in new link types. Therefore, we propose the Transferable Domain Adversarial Network ( TDAN ) based on transfer learning to handle the challenge. TDAN exploits transferable type-shared knowledge in historical link types to help predict the unobserved links in new link types. TDAN mainly comprises a structural encoder, a domain discriminator, and an optimization decoder. The structural encoder learns the link representations in a heterogeneous social network. Subsequently, to learn transferable type-shared knowledge, the domain discriminator distinguishes link representations into different link types while minimizing the differences between type-specific knowledge in adversarial training. Inspired by the denoising auto-encoder, the optimization decoder reconstructs the learned type-shared knowledge to eliminate the noise generated during the adversarial training. Extensive experiments on Facebook and YouTube show that TDAN can outperform the state-of-the-art models.

Nearest neighbor machine translation (kNN-MT), which interpolates target token probabilities with estimates derived from additional examples, has achieved significant improvements and attracted extensive interest in recent years.However, existing research does not explicitly consider the source context when retrieving similar examples, potentially leading to suboptimal performance.To address this, we comprehensively revisit the role of source context and propose a simple and effective method for improving neural machine translation via source context enhancement, demonstrating its crucial role in both retrieving superior examples and determining more suitable interpolation coefficients.Furthermore, we reveal that the probability estimation can be further optimized by incorporating a source-aware distance calibration module.Comprehensive experiments show that our proposed approach can be seamlessly integrated with representative kNN-MT baselines, resulting in substantial improvements over these strong baselines across a number of settings and domains.Remarkably, these improvements can reach up to 1.6 BLEU points. 1

Event detection, which aims to identify instances of specific event types from pieces of text, is a fundamental task in information extraction. Most existing approaches leverage syntactic knowledge with a set of syntactic relations to enhance event detection. However, a side effect of these syntactic-based approaches is that they may confuse different syntactic relations and tend to introduce redundant or noisy information, which may lead to performance degradation. To this end, we propose a simple yet effective model named DualGAT (Dual Relational Graph Attention Networks), which exploits the complementary nature of syntactic and semantic relations to alleviate the problem. Specifically, we first construct a dual relational graph that both aggregates syntactic and semantic relations to the key nodes in the graph, so that event-relevant information can be comprehensively captured from multiple perspectives (i.e., syntactic and semantic views). We then adopt augmented relational graph attention networks to encode the graph and optimize its attention weights by introducing contextual information, which further improves the performance of event detection. Extensive experiments conducted on the standard ACE2005 benchmark dataset indicate that our method significantly outperforms the state-of-the-art methods and verifies the superiority of DualGAT over existing syntactic-based methods.

The task of RDF-to-text generation is to generate a corresponding descriptive text given a set of RDF triples. Most of the previous approaches either cast this task as a sequence-to-sequence problem or employ graph-based encoder for modeling RDF triples and decode a text sequence. However, none of these methods can explicitly model both local and global structure information between and within the triples. To address these issues, we propose to jointly learn local and global structure information via combining two new graph-augmented structural neural encoders (i.e., a bidirectional graph encoder and a bidirectional graph-based meta-paths encoder) for the input triples. Experimental results on two different WebNLG datasets show that our proposed model outperforms the state-of-the-art baselines. Furthermore, we perform a human evaluation that demonstrates the effectiveness of the proposed method by evaluating generated text quality using various subjective metrics.

Traditional information retrieval (IR) models, in which a document is normally represented as a bag of words and their frequencies, capture the term-level and document-level information. Topic models, on the other hand, discover semantic topic-based information among words. In this paper, we consider term-based information and semantic information as two features of query terms and propose a simple enhancement for ad-hoc IR via topic modeling. In particular, three topic-based hybrid models, LDA-BM25, LDA-MATF and LDA-LM, are proposed. A series of experiments on eight standard datasets show that our proposed models can always outperform significantly the corresponding strong baselines over all datasets in terms of MAP and most of datasets in terms of [email protected] and [email protected] A direct comparison on eight standard datasets also indicates our proposed models are at least comparable to the state-of-the-art approaches.

In recent years, online social networks are among the most popular websites with high PV (Page View) all over the world, as they have renewed the way for information discovery and distribution. Millions of users have registered on these websites and hence generate formidable amount of user-generated contents every day. The social networks become "giants", likely eligible to carry on any research tasks. However, we have pointed out that these giants still suffer from their "Achilles Heel", i.e., extreme sparsity. Compared with the extremely large data over the whole collection, individual posting documents such as microblogs seem to be too sparse to make a difference under various research scenarios, while actually these postings are different. In this paper we propose to tackle the Achilles Heel of social networks by smoothing the language model via influence propagation. To further our previously proposed work to tackle the sparsity issue, we extend the socialized language model smoothing with bi-directional influence learned from propagation. Intuitively, it is insufficient not to distinguish the influence propagated between information source and target without directions. Hence, we formulate a bi-directional socialized factor graph model, which utilizes both the textual correlations between document pairs and the socialized augmentation networks behind the documents, such as user relationships and social interactions. These factors are modeled as attributes and dependencies among documents and their corresponding users, and then are distinguished on the direction level. We propose an effective learning algorithm to learn the proposed factor graph model with directions. Finally we propagate term counts to smooth documents based on the estimated influence. We run experiments on two instinctive datasets of Twitter and Weibo. The results validate the effectiveness of the proposed model. By incorporating direction information into the socialized language model smoothing, our approach obtains improvement over several alternative methods on both intrinsic and extrinsic evaluations measured in terms of perplexity, nDCG and MAP measurements.

Community question answering (cQA) has become an important issue due to the popularity of cQA archives on the web. This paper is concerned with the problem of question retrieval. Question retrieval in cQA archives aims to find the existing questions that are semantically equivalent or relevant to the queried questions. However, the lexical gap problem brings about new challenge for question retrieval in cQA. In this paper, we propose to learn continuous word embeddings with metadata of category information within cQA pages for question retrieval. To deal with the variable size of word embedding vectors, we employ the framework of fisher kernel to aggregated them into the fixedlength vectors. Experimental results on large-scale real world cQA data set show that our approach can significantly outperform state-of-the-art translation models and topic-based models for question re-

Manifold ranking is one of the most competitive approaches for query-focused multi-document summarization. Despite its success for this task, it usually constructs a sentence affinity graph first based on inter-sentence content similarity, and then perform manifold ranking on the graph to score each sentence with the assumption that all the sentences live on a single manifold. Actually, for a document set to be summarized, the distribution of the sentences might form different, but related manifolds. This paper aims to generalize the basic manifold-ranking based approach to the more generic setting by introducing a novel affinity graph to estimate the similarity between sentences, which leverages both the local geometric structures and the contents of sentences jointly. Preliminary experimental results on the DUC datasets demonstrate the good effectiveness of the proposed approach.