Generating with citations is crucial for trustworthy Large Language Models (LLMs), yet even advanced LLMs often produce mismatched or irrelevant citations. Existing methods over-optimize citation fidelity while overlooking relevance to the user query, which degrades answer quality and robustness in real-world settings with noisy or irrelevant retrieved content. Moreover, the prevailing single-pass paradigm struggles to deliver optimal answers in long-form generation that requiring multiple citations. To address these limitations, we propose FineRef, a framework based on Fine-grained error Reflection, which explicitly teaches the model to self-identify and correct two key citation errors—mismatch and irrelevance—on a per-citation basis. FineRef follows a two-stage training strategy. The first stage instills an “attempt–reflect–correct” behavioral pattern via supervised fine-tuning, using fine-grained and controllable reflection data constructed by specialized lightweight models. An online self-reflective bootstrapping strategy is designed to improve generalization by iteratively enriching training data with verified, self-improving examples. To further enhance the self-reflection and correction capability, the second stage applies process-level reinforcement learning with a multi-dimensional reward scheme that promotes reflection accuracy, answer quality, and correction gain. Experiments on the ALCE benchmark demonstrate that FineRef significantly improves both citation performance and answer accuracy. Our 7B model outperforms GPT-4 by up to 18% in Citation F1 and 4% in EM Recall, while also surpassing the state-of-the-art model across key evaluation metrics. FineRef also exhibits strong generalization and robustness in domain transfer settings and noisy retrieval scenarios.

Training Large Language Models (LLMs) for chain-of-thought reasoning presents a significant challenge: supervised fine-tuning on a single "golden" rationale hurts generalization as it penalizes equally valid alternatives, whereas reinforcement learning with verifiable rewards struggles with credit assignment and prohibitive computational cost. To tackle these limitations, we introduce InTRO (In-Token Rationality Optimization), a new framework that enables both token-level exploration and self-feedback for accurate and concise reasoning. Instead of directly optimizing an intractable objective over all valid reasoning paths, InTRO leverages correction factors—token-wise importance weights estimated by the information discrepancy between the generative policy and its answer-conditioned counterpart, for informative next-token selection. This approach allows the model to perform token-level exploration and receive self-generated feedback within a single forward pass, ultimately encouraging accurate and concise rationales. Across six math-reasoning benchmarks, InTRO consistently outperforms other baselines, raising solution accuracy by up to 20% relative to the base model. Its chains of thought are also notably more concise, exhibiting reduced verbosity. Beyond this, InTRO enables cross-domain transfer, successfully adapting to out-of-domain reasoning tasks that extend beyond the realm of mathematics, demonstrating robust generalization.

The widespread deployment of Large Language Models (LLMs) has spurred significant progress in the detection of LLM-generated text.However, existing detection methods often rely on statistical features that are insufficient for reliable detection; for example, even though LLM-generated and humanwritten texts exhibit different probability distributions in surrogate models, they can produce nearly identical entropy values, thereby conflating the two types of text.In this paper, we propose that modulating the decoding temperature and monitoring how the probability distributions respond can better probe the intrinsic discrepancies between two types of text.Building upon this insight, we introduce a new feature termed Temperature Sensitivity (TS) and demonstrate that LLM-generated text tends to exhibit higher TS than humanwritten text.Finally, we propose NTS, a novel and simple zero-shot detector built upon normalized temperature sensitivity.Extensive experiments across three datasets, multiple domains, and various source models demonstrate the superior effectiveness and robustness of our proposed approach.Code avaliable at

Detecting LLM-generated code is crucial for ensuring software provenance, security, reliability, and licensing compliance. Existing training-free detectors, mostly adapted from text-based methods, rely on global statistics of the Token Perplexity Sequence (TPS) and struggle with code. We reveal a key insight: despite the convergence of global statistics, LLM-generated and human-written code differ fundamentally in their local TPS dynamics: the former shows narrow transient spikes while the latter exhibits broad sustained fluctuations. To capture this distinction, we introduce CodeRipple, a novel training-free detection framework that employs wavelet analysis to characterize TPS morphology across scales. It jointly leverages the Stationary Wavelet Transform to model fluctuation shape and the Discrete Wavelet Transform to quantify cross-scale energy distribution. Evaluated on three challenging benchmarks spanning diverse programming languages, multiple generating LLMs, and various evasion strategies, CodeRipple consistently outperforms existing training-free methods, demonstrating its superior effectiveness and generalizability without any model training. Code available at: https://github.com/yaoxingyu77/CodeRipple.

Existing generative image transformers follow a two-stage generation paradigm, where the first stage learns a codebook to encode images into discrete codes via vector quantization, and the second stage completes the image generation based on the learned codebook. However, existing methods ignore the naturally varying information densities across different image regions and indiscriminately encode fixed-size regions into fixed-length codes, resulting in insufficient encoding in important regions and redundant encoding in unimportant ones, which degrades both the image generation quality and speed. To address this challenge, we propose a novel information-density-based variable-length image coding and generation framework. In the first stage, our Dynamic Quantization VAE++ (DQVAE++) performs information-adaptive encoding by assigning variable-length codes to image regions according to their information densities, yielding more accurate and robust code representations. In the second stage, the Dynamic Generative Image Transformer (DGiT) enables information-adaptive image generation in both autoregressive and non-autoregressive manners. Specifically, for autoregressive (AR) generation, DGiT-AR generates images autoregressively from coarse-grained regions (smooth areas with fewer codes) to fine-grained regions (detailed areas with more codes). This is accomplished through a novel stacked-transformer architecture that alternately models the position and content of image codes, and a novel heterogeneous embedding scheme to distinguish codes of different granularities. Similarly, for non-autoregressive (NAR) generation, DGiT-NAR introduces a novel information-prioritized mask scheduling mechanism, prioritizing the generation of key structural regions with higher information density. This enables more coherent modeling of global structures initially, followed by a more effective synthesis of local details subsequently. Comprehensive experiments on unconditional and conditional image generation validate the superiority of our proposed variable-length coding in both effectiveness and efficiency.

Large language models (LLMs) require model editing to efficiently update specific knowledge within them and avoid factual errors. Most model editing methods are solely designed for single-time use and result in a significant forgetting effect in lifelong editing scenarios, where sequential edits are conducted over time. Previous approaches manage sequential edits by freezing original parameters and discretely allocating new parameters for each knowledge update. However, these methods lack robustness to minor input variations due to the discrete mapping between data and parameters. To overcome this challenge, we propose ELDER, a novel approach to create a continuous association between data and adapters. ELDER integrates multiple LoRAs through a router network and is trained to establish a smooth data-adapter association, thereby enhancing the edit robustness and generalization of semantically equivalent inputs. To ensure inputs containing the same knowledge will be processed by the same LoRAs, we design a novel loss to guide the model link LoRA allocations with edit knowledge. Furthermore, we propose a deferral mechanism to retain the original LLM capabilities post-edit. Extensive experiments on GPT-2 XL and LLaMA2-7B demonstrate that ELDER effectively edits models in the lifelong setting, outperforming eight baselines while exhibiting strong scalability and preserving LLMs' general abilities on downstream tasks.

The rapid advancement of large language models (LLMs) has brought about increased concerns regarding their safety, especially as adversaries develop jailbreak techniques to bypass LLMs' safety mechanism.Although recent work on safety training with modules such as low-rank adaptation (LoRA) to resist jailbreaks shows promise, these approaches can inadvertently degrade a model's general utility.In this paper, we propose a novel plugand-play method that mitigates the impact of safety training on model utility by explicitly locating and leveraging safety-critical singular vectors, which only contribute to safety, within the model's parameter space.We quantify the safety-criticality of each singular vector as the difference of their importance for safety and utility measured by a corresponding low-rank projection.The top scored singular vectors are located as safety-critical and are used to initialize the LoRA modules within existing safety training methods in a plug-and-play manner, thereby constraining the training updates within safety-critical parameters.Additionally, we propose a dynamic rank number determination strategy to further reduce parameter overhead.Experiments on HarmBench with multiple jailbreak methods validate the effectiveness of our approach in safety training, while evaluations on several utility benchmarks demonstrate that our method successfully mitigates the adverse impact of safety training on model utility, enhancing the utility performance of the evaluated safety training baselines.

The rapid development of large language models (LLMs) raises concerns about their potential misuse. Accurately identifying and tracing the origin of LLM-generated content is crucial for accountability and transparency. Previous methods typically frame origin tracing as multi-class classification with a fixed label set, thus struggle to adapt to new LLMs without frequent retraining. This paper introduces a new task, continual origin tracing of LLM-generated text, which frames origin tracing in a continual learning or, more precisely, class-incremental learning manner, where new LLMs continuously emerge, and a model incrementally learns to identify new LLMs without forgetting old ones. A novel training-free method is further devised for the task, which continually extracts prototypes for emerging LLMs using a frozen pre-trained model, and conducts global and local prototype decorrelation to improve prototype matching, thus favoring more accurate tracing. To facilitate evaluation on the new task, we construct a benchmark comprising text generated by 19 recently released LLMs from 12 vendors that simulates a real-world scenario where these LLMs emerge over time and need to be recognized incrementally across 8 diverse domains. Rigorous evaluations on this benchmark highlight the effectiveness and potential of the proposed method in the new task, offering a promising direction for future research.

While large language models (LLMs) have exhibited impressive instruction-following capabilities, it is still unclear whether and to what extent they can respond to explicit constraints that might be entailed in various instructions. As a significant aspect of LLM alignment, it is thus important to formulate such a specialized set of instructions as well as investigate the resulting behavior of LLMs. To address this vacancy, we propose a new benchmark CoDI-Eval to systematically and comprehensively evaluate LLMs' responses to instructions with various constraints. We construct a large collection of constraints-attributed instructions as a test suite focused on both generalization and coverage. Specifically, we advocate an instruction diversification process to synthesize diverse forms of constraint expression and also deliberate the candidate task taxonomy with even finer-grained sub-categories. Finally, we automate the entire evaluation process to facilitate further developments. Different from existing studies on controllable text generation, CoDI-Eval extends the scope to the prevalent instruction-following paradigm for the first time. We provide extensive evaluations of representative LLMs (e.g., ChatGPT, Vicuna) on CoDI-Eval, revealing their limitations in following instructions with specific constraints and there is still a significant gap between open-source and commercial closed-source LLMs. We believe this benchmark will facilitate research into improving the controllability of LLMs' responses to instructions. Our data and code are available at https://github.com/Xt-cyh/CoDI-Eval.

Text style transfer (TST) is an important task in natural language generation, which aims to transfer the text style (e.g., sentiment) while keeping its semantic information.Due to the absence of parallel datasets for supervision, most existing studies have been conducted in an unsupervised manner, where the generated sentences often suffer from high semantic divergence and thus low semantic preservation.In this paper, we propose a novel disentanglementbased framework for TST named DisenTrans, where disentanglement means that we separate the attribute and content components in the natural language corpus and consider this task from these two perspectives.Concretely, we first create a disentangled Chain-of-Thought prompting procedure to synthesize parallel data and corresponding attribute components for supervision.Then we develop a disentanglement learning method with synthetic data, where two losses are designed to enhance the focus on attribute properties and constrain the semantic space, thereby benefiting style control and semantic preservation respectively.Instructed by the disentanglement concept, our framework creates valuable supervised information and utilizes it effectively in TST tasks.Extensive experiments on mainstream datasets present that our framework achieves significant performance with great sample efficiency.

In-context learning (ICL), which promotes inference with several demonstrations, has become a widespread paradigm to stimulate LLM capabilities for downstream tasks.Due to context length constraints, it cannot be further improved in spite of more training data, and general features directly from LLMs in ICL are not adaptive to the specific downstream task.In this paper, we propose a feature-adaptive and datascalable in-context learning framework (FADS-ICL), which can leverage task-adaptive features to promote inference on the downstream task, with the supervision of beyond-context samples.Specifically, it first extracts general features of beyond-context samples via the LLM with ICL input form one by one, and introduces a task-specific modulator to perform feature refinement and prediction after fitting a specific downstream task.We conduct extensive experiments on FADS-ICL under varying data settings (4∼128 shots) and LLM scale (0.8∼70B) settings.Experimental results show that FADS-ICL consistently outperforms previous state-of-the-art methods by a significant margin under all settings, verifying the effectiveness and superiority of FADS-ICL.For example, under the 1.5B and 32 shots setting, FADS-ICL can achieve +14.3 average accuracy from feature adaptation over vanilla ICL on 10 datasets, with +6.2 average accuracy over the previous state-of-the-art method, and the performance can further improve with increasing training data.

Tianqi Zhong, Zhaoyi Li, Quan Wang, Linqi Song, Ying Wei, Defu Lian, Zhendong Mao. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

The effective detection of AI-generated text is a vital principle to ensure responsible use of large language models (LLMs). Previous studies mainly focused on discovering and utilizing internal evidences contained in the text itself to perform the detection, while ignoring external evidences implicated in an established knowledge graph (KG) which may also be key discriminative factors between AI-generated and human-written text. To address this deficiency, we propose IDEATE, a novel hierarchical graph network that utilizes both internal and external factual structures to detect AI-generated text. IDEATE consists of a mention-level subgraph at the bottom to describe internal factual structures of mentioned entities reflected in the input text, and an entity-level subgraph at the top to describe external factual structures of mentioned entities reflected in an external KG. Hierarchical graph convolution is then applied successively on the two subgraphs, through which the two types of factual structures will be embedded into the output and used for the final detection. Extensive experiments on four benchmarking datasets show that IDEATE consistently outperforms current state-of-the-art methods in detecting text generated by various LLMs, ranging from GPT-2 to the more powerful ChatGPT, verifying the necessity and superiority of introducing external evidences for AI-generated text detection.

Recent breakthroughs in preference alignment have significantly improved Large Language Models' ability to generate texts that align with human preferences and values.However, current alignment metrics typically emphasize the post-hoc overall improvement, while overlooking a critical aspect: regression, which refers to the backsliding on previously correctly-handled data after updates.This potential pitfall may arise from excessive fine-tuning on already well-aligned data, which subsequently leads to over-alignment and degeneration.To address this challenge, we propose FlipGuard, a constrained optimization approach to detect and mitigate update regression with focal attention.Specifically, FlipGuard identifies performance degradation using a customized reward characterization and strategically enforces a constraint to encourage conditional congruence with the pre-aligned model during training.Comprehensive experiments demonstrate that FlipGuard effectively alleviates update regression while demonstrating excellent overall performance, with the added benefit of knowledge preservation while aligning preferences.

Although the pre-trained language models have achieved great success on machine reading comprehension task, they often rely on large-scale annotated data, while only a little amount of data is available in the most real-world scenarios. To enhance the PTLMs' capabilities in low-resource scenario, we propose a curriculum learning driven domain adaptation method for low-resource machine reading comprehension, the basic paradigm of which is to train a source model with sufficient data and then adaptive it to our target domain. In the adapting procedure, we introduce the curriculum learning strategy, the core idea of which is arranging training examples from easy to difficult, to bridge the gap between source and target domains and enable the source model adapting to the target domain progressively. Specifically, before fine-tuning the well-trained source model using target data, we firstly calculate the loss of each target example using the source model to evaluating the example difficulty accurately. After that, we sample suitable batches based on an increasing sampling function at each fine-tuning step, allowing the source model to start learning from easy examples in the target domain and gradually transition to difficult ones. Experiments conducted on two public datasets have demonstrated the effectiveness of our method.

Document-level relation extraction (RE) aims to simultaneously predict relations (including no-relation cases denoted as NA) between all entity pairs in a document. It is typically formulated as a relation classification task with entities pre-detected in advance and solved by a hard-label training regime, which, however, neglects the divergence of the NA class and the correlations among other classes. This article introduces progressive self-distillation (PSD), a new training regime that employs online, self-knowledge distillation (KD) to produce and incorporate soft labels for document-level RE.The key idea of PSD is to gradually soften hard labels using past predictions from an RE model itself, which are adjusted adaptively as training proceeds. As such, PSD has to learn only one RE model within a single training pass, requiring no extra computation or annotation to pretrain another high-capacity teacher. PSD is conceptually simple, easy to implement, and generally applicable to various RE models to further improve their performance, without introducing additional parameters or significantly increasing training overheads into the models. It is also a general framework that can be flexibly extended to distilling various types of knowledge, rather than being restricted to soft labels themselves. Extensive experiments on four benchmarking datasets verify the effectiveness and generality of the proposed approach. The code is available at https://github.com/GaoJieCN/psd

Text style transfer (TST) is an important task in natural language generation, which aims to alter the stylistic attributes (e.g., sentiment) of a sentence and keep its semantic meaning unchanged.Most existing studies mainly focus on the transformation between styles, yet ignore that this transformation can be actually carried out via different hidden transfer patterns.To address this problem, we propose a novel approach, contrastive transfer pattern mining (CTPM), which automatically mines and utilizes inherent latent transfer patterns to improve the performance of TST.Specifically, we design an adaptive clustering module to automatically discover hidden transfer patterns from the data, and introduce contrastive learning based on the discovered patterns to obtain more accurate sentence representations, and thereby benefit the TST task.To the best of our knowledge, this is the first work that proposes the concept of transfer patterns in TST, and our approach can be applied in a plug-andplay manner to enhance other TST methods to further improve their performance.Extensive experiments on benchmark datasets verify the effectiveness and generality of our approach.1

Current relation extraction methods suffer from the inadequacy of large-scale annotated data.While distant supervision alleviates the problem of data quantities, there still exists domain disparity in data qualities due to its reliance on domain-restrained knowledge bases. In this work, we propose S2ynRE, a framework of two-stage Self-training with Synthetic data for Relation Extraction.We first leverage the capability of large language models to adapt to the target domain and automatically synthesize large quantities of coherent, realistic training data.We then propose an accompanied two-stage self-training algorithm that iteratively and alternately learns from synthetic and golden data together.We conduct comprehensive experiments and detailed ablations on popular relation extraction datasets to demonstrate the effectiveness of the proposed framework.

Existing autoregressive models follow the two-stage generation paradigm that first learns a codebook in the latent space for image reconstruction and then completes the image generation autoregressively based on the learned codebook. However, existing codebook learning simply models all local region information of images without distinguishing their different perceptual importance, which brings redundancy in the learned codebook that not only limits the next stage's autoregressive model's ability to model important structure but also results in high training cost and slow generation speed. In this study, we borrow the idea of importance perception from classical image coding theory and propose a novel two-stage framework, which consists of Masked Quantization VAE (MQVAE) and Stackformer, to relieve the model from modeling redundancy. Specifically, MQ-VAE incorporates an adaptive mask module for masking redundant region features before quantization and an adaptive de-mask module for recovering the original grid image feature map to faithfully reconstruct the original images after quantization. Then, Stackformer learns to predict the combination of the next code and its position in the feature map. Comprehensive experiments on various image generation validate our effectiveness and efficiency. Code will be released at https://github.com/CrossmodalGroup/MaskedVectorQuantization.

Achieving empathy is a crucial step toward humanized dialogue systems.Current approaches for empathetic dialogue generation mainly perceive an emotional label to generate an empathetic response conditioned on it, which simply treat emotions independently, but ignore the intrinsic emotion correlation in dialogues, resulting in inaccurate emotion perception and unsuitable response generation.In this paper, we propose a novel emotion correlation enhanced empathetic dialogue generation framework, which comprehensively realizes emotion correlation learning, utilization, and supervising.Specifically, a multi-resolution emotion graph is devised to capture context-based emotion interactions from different resolutions, further modeling emotion correlation.Then we propose an emotion correlation enhanced decoder, with a novel correlation-aware aggregation and soft/hard strategy, respectively improving the emotion perception and response generation.Experimental results on the benchmark dataset demonstrate the superiority of our model in both empathetic perception and expression.

Representation Learning on Knowledge Graphs (KGs) is essential for downstream tasks.The dominant approach, KG Embedding (KGE), represents entities with independent vectors and faces the scalability challenge.Recent studies propose an alternative way for parameter efficiency, which represents entities by composing entity-corresponding codewords matched from predefined small-scale codebooks.We refer to the process of obtaining corresponding codewords of each entity as entity quantization, for which previous works have designed complicated strategies.Surprisingly, this paper shows that simple random entity quantization can achieve similar results to current strategies.We analyze this phenomenon and reveal that entity codes, the quantization outcomes for expressing entities, have higher entropy at the code level and Jaccard distance at the codeword level under random entity quantization.Therefore, different entities become more easily distinguished, facilitating effective KG representation.The above results show that current quantization strategies are not critical for KG representation, and there is still room for improvement in entity distinguishability beyond current strategies.The code to reproduce our results is available here.

Controllable text generation (CTG) aims to generate text with desired attributes, and decodingtime-based methods have shown promising performance on this task.However, in this paper, we identify the phenomenon of Attribute Collapse for the first time.It causes the fluency of generated text to rapidly decrease when the control strength exceeds a critical value, rendering the text completely unusable.This limitation hinders the effectiveness of decoding methods in achieving high levels of controllability.To address this problem, we propose a novel lightweight decoding framework named Air-Decoding.Its main idea is reconstructing the attribute distributions to balance the weights between attribute words and nonattribute words to generate more fluent text.Specifically, we train prefixes by prefix-tuning to obtain attribute distributions.Then we design a novel attribute distribution reconstruction method to balance the obtained distributions and use the reconstructed distributions to guide language models for generation, effectively avoiding the issue of Attribute Collapse.Experiments on multiple CTG tasks prove that our method achieves a new state-of-the-art control performance 1 .

Relation prediction on knowledge graphs (KGs) is a key research topic. Dominant embedding-based methods mainly focus on the transductive setting and lack the inductive ability to generalize to new entities for inference. Existing methods for inductive reasoning mostly mine the connections between entities, i.e., relational paths, without considering the nature of head and tail entities contained in the relational context. This paper proposes a novel method that captures both connections between entities and the intrinsic nature of entities, by simultaneously aggregating RElational Paths and cOntext with a unified hieRarchical Transformer framework, namely REPORT. REPORT relies solely on relation semantics and can naturally generalize to the fully-inductive setting, where KGs for training and inference have no common entities. In the experiments, REPORT performs consistently better than all baselines on almost all the eight version subsets of two fully-inductive datasets. Moreover. REPORT is interpretable by providing each element’s contribution to the prediction results.

Multi-dataset question answering (QA) aims to combine multiple QA datasets to build models that not only perform well on training distributions, but also transfer and generalize well to new distributions. Some prior work considered building a collection of dataset-specific experts upon a shared Transformer, so as to simultaneously encode both regularities across datasets and specificities of each dataset. This approach, however, has its limitations when generalized to an unseen new distribution, and the number of extra parameters will increase with the number of training datasets. In this paper, we devise Self-ADaptive Expert (SADE), the key idea of which is to train a single expert that can be automatically adapted to each individual instance according to its gradients. This gradient-based, instance-level modulation scheme makes our approach easily adaptable to any instance from unseen new distributions, and keeps the number of extra parameters as a constant. We further design a contrastive learning mechanism to enhance the discriminability of modulation signals across different datasets. Experimental results on twelve QA datasets demonstrate that SADE consistently outperforms previous state-of-the-art in all the three settings including in-domain learning, few-shot transfer learning, and zero-shot generalization.

To counter the online abusive language in social media, it is desirable to develop automated detection methods. Previous research has primarily formulated this problem as a sentence-level classification task, ignoring the crucial role of abusive lexicons that can strengthen the model explainability and enable more faithful predictions. Although a few methods have introduced the abusive lexicons for detection, the lexicons they use are either externally provided or labeled by human annotators, suffering from two limitations: (1) lack adaptability to diverse and evolving offensive scenarios; (2) require large human efforts to annotate the words.This paper overcomes the limitations of prior work with a multi-task abusive language detection framework. It combines sentence-level and word-level classification tasks, based on dependency tree based graph attention networks (GAT). With the two tasks, it is encouraged to capture both global and local data properties to produce better sentence representations. It is also advantageous in automatic lexicon construction during the learning process, without human annotations. Extensive experiments on two public datasets exhibit that our proposal can outperform the state-of-the-art baselines. Case studies show that the model explainability can be strengthened with the abusive parts identified by our framework. Our code is released to public.1

Image-text matching, as a fundamental task, bridges the gap between vision and language. The key of this task is to accurately measure similarity between these two modalities. Prior work measuring this similarity mainly based on matched fragments (i.e., word/region with high relevance), while underestimating or even ignoring the effect of mismatched fragments (i.e., word/region with low relevance), e.g., via a typical LeaklyReLU or ReLU operation that forces negative scores close or exact to zero in attention. This work argues that mismatched textual fragments, which contain rich mismatching clues, are also crucial for image-text matching. We thereby propose a novel Negative-Aware Attention Framework (NAAF), which explicitly exploits both the positive effect of matched fragments and the negative effect of mismatched fragments to jointly infer image-text similarity. NAAF (1) delicately designs an iterative optimization method to maximally mine the mismatched fragments, facilitating more discriminative and robust negative effects, and (2) devises the two-branch matching mechanism to precisely calculate similarity/dissimilarity degrees for matched/mismatched fragments with different masks. Extensive experiments on two benchmark datasets, i.e., Flickr30K and MSCOCO, demonstrate the superior effectiveness of our NAAF, achieving state-of-the-art performance. Code will be released at: https://github.com/CrossmodalGroup/NAAF.

Chinese spelling check (CSC) is a fundamental NLP task that detects and corrects spelling errors in Chinese texts.As most of these spelling errors are caused by phonetic similarity, effectively modeling the pronunciation of Chinese characters is a key factor for CSC.In this paper, we consider introducing an auxiliary task of Chinese pronunciation prediction (CPP) to improve CSC, and, for the first time, systematically discuss the adaptivity and granularity of this auxiliary task.We propose SCOPE which builds on top of a shared encoder two parallel decoders, one for the primary CSC task and the other for a fine-grained auxiliary CPP task, with a novel adaptive weighting scheme to balance the two tasks.In addition, we design a delicate iterative correction strategy for further improvements during inference.Empirical evaluation shows that SCOPE achieves new state-of-theart on three CSC benchmarks, demonstrating the effectiveness and superiority of the auxiliary CPP task.Comprehensive ablation studies further verify the positive effects of adaptivity and granularity of the task.Code and data used in this paper are publicly available at https: //github.com/jiahaozhenbang/SCOPE.

Rumor spreaders are increasingly taking advantage of multimedia content to attract and mislead news consumers on social media. Although recent multimedia rumor detection models have exploited both textual and visual features for classification, they do not integrate the social structure features simultaneously, which have shown promising performance for rumor identification. It is challenging to combine the heterogeneous multi-modal data in consideration of their complex relationships. In this work, we propose a novel Multi-modal Feature-enhanced Attention Networks (MFAN) for rumor detection, which makes the first attempt to integrate textual, visual, and social graph features in one unified framework. Specifically, it considers both the complement and alignment relationships between different modalities to achieve better fusion. Moreover, it takes into account the incomplete links in the social network data due to data collection constraints and proposes to infer hidden links to learn better social graph features. The experimental results show that MFAN can detect rumors effectively and outperform state-of-the-art methods.

Text-to-image generation aims at generating realistic images which are semantically consistent with the given text. Previous works mainly adopt the multi-stage architecture by stacking generator-discriminator pairs to engage multiple adversarial training, where the text semantics used to provide generation guidance remain static across all stages. This work argues that text features at each stage should be adaptively re-composed conditioned on the status of the historical stage (\emphi.e., historical stage's text and image features) to provide diversified and accurate semantic guidance during the coarse-to-fine generation process. We thereby propose a novel Dynamical Semantic Evolution GAN (DSE-GAN) to re-compose each stage's text features under a novel single adversarial multi-stage architecture. Specifically, we design (1) Dynamic Semantic Evolution (DSE) module, which first aggregates historical image features to summarize the generative feedback, and then dynamically selects words required to be re-composed at each stage as well as re-composed them by dynamically enhancing or suppressing different granularity subspace's semantics. (2) Single Adversarial Multi-stage Architecture (SAMA), which extends the previous structure by eliminating complicated multiple adversarial training requirements and therefore allows more stages of text-image interactions, and finally facilitates the DSE module. We conduct comprehensive experiments and show that DSE-GAN achieves 7.48% and 37.8% relative FID improvement on two widely used benchmarks, i.e., CUB-200 and MSCOCO, respectively.

Benfeng Xu, Quan Wang, Yajuan Lyu, Yabing Shi, Yong Zhu, Jie Gao, Zhendong Mao. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Deep metric learning aims to learn a deep embedding space, where similar objects are pushed towards together and different objects are repelled against. Existing approaches typically use inter-class characteristics, e.g. class-level information or instance-level similarity, to obtain semantic relevance of data points and get a large margin between different classes in the embedding space. However, the intra-class characteristics, e.g. local manifold structure or relative relationship within the same class, are usually overlooked in the learning process. Hence the data structure cannot be fully exploited and the output embeddings have limitation in retrieval. More importantly, retrieval results lack in a good ranking. This paper presents a novel self-supervised ranking auxiliary framework, which captures intra-class characteristics as well as inter-class characteristics for better metric learning. Our method defines specific transform functions to simulates the local structure change of intra-class in the initial image domain, and formulates a self-supervised learning procedure to fully exploit this property and preserve it in the embedding space. Extensive experiments on three standard benchmarks show that our method significantly improves and outperforms the state-of-the-art methods on the performances of both retrieval and ranking by 2%-4%.

Entities, as the essential elements in relation extraction tasks, exhibit certain structure. In this work, we formulate such entity structure as distinctive dependencies between mention pairs. We then propose SSAN, which incorporates these structural dependencies within the standard self-attention mechanism and throughout the overall encoding stage. Specifically, we design two alternative transformation modules inside each self-attention building block to produce attentive biases so as to adaptively regularize its attention flow. Our experiments demonstrate the usefulness of the proposed entity structure and the effectiveness of SSAN. It significantly outperforms competitive baselines, achieving new state-of-the-art results on three popular document-level relation extraction datasets. We further provide ablation and visualization to show how the entity structure guides the model for better relation extraction. Our code is publicly available.

With the notable success of pretrained language models, the pretraining-fine-tuning paradigm has become a dominant solution for natural language understanding (NLU) tasks. Typically, the training instances of a target NLU task are introduced in a completely random order and treated equally at the fine-tuning stage. However, these instances can vary greatly in difficulty, and similar to human learning procedures, language models can benefit from an easy-to-difficult curriculum. Based on this concept, we propose a curriculum learning (CL) framework. Our framework consists of two stages, Review and Arrange, targeting the two main challenges in curriculum learning, i.e., how to define the difficulty of instances and how to arrange a curriculum based on the difficulty, respectively. In the first stage, we devise a cross-review (CR) method to train several teacher models first and then review the training set in a crossed manner to distinguish easy instances from difficult instances. In the second stage, two sampling algorithms, a coarse-grained arrangement (CGA) and a fine-grained arrangement (FGA), are proposed to arrange a curriculum for language models in which the learning materials start from the easiest instances, and more difficult instances are gradually added into the training procedure. Compared to previous heuristic CL methods, our framework can avoid the errors caused by a gap in difficulty between humans and machines and has strong generalization ability. We conduct comprehensive experiments, and the results show that our curriculum learning framework, without any manual model architecture design or use of external data, obtains significant and universal performance improvements on a wide range of NLU tasks in different languages.

With the great success of pre-trained language models, the pretrain-finetune paradigm now becomes the undoubtedly dominant solution for natural language understanding (NLU) tasks.At the fine-tune stage, target task data is usually introduced in a completely random order and treated equally.However, examples in NLU tasks can vary greatly in difficulty, and similar to human learning procedure, language models can benefit from an easy-to-difficult curriculum.Based on this idea, we propose our Curriculum Learning approach.By reviewing the trainset in a crossed way, we are able to distinguish easy examples from difficult ones, and arrange a curriculum for language models.Without any manual model architecture design or use of external data, our Curriculum Learning approach obtains significant and universal performance improvements on a wide range of NLU tasks.

Machine reading comprehension (MRC) is a crucial and challenging task in NLP.Recently, pre-trained language models (LMs), especially BERT, have achieved remarkable success, presenting new state-of-the-art results in MRC.In this work, we investigate the potential of leveraging external knowledge bases (KBs) to further improve BERT for MRC.We introduce KT-NET, which employs an attention mechanism to adaptively select desired knowledge from KBs, and then fuses selected knowledge with BERT to enable context-and knowledgeaware predictions.We believe this would combine the merits of both deep LMs and curated KBs towards better MRC.Experimental results indicate that KT-NET offers significant and consistent improvements over BERT, outperforming competitive baselines on ReCoRD and SQuAD1.1 benchmarks.Notably, it ranks the 1st place on the ReCoRD leaderboard, and is also the best single model on the SQuAD1.1 leaderboard at the time of submission (March 4th, 2019). 1

Embedding knowledge graphs (KGs) into continuous vector spaces is a focus of current research. Combining such an embedding model with logic rules has recently attracted increasing attention. Most previous attempts made a one-time injection of logic rules, ignoring the interactive nature between embedding learning and logical inference. And they focused only on hard rules, which always hold with no exception and usually require extensive manual effort to create or validate. In this paper, we propose Rule-Guided Embedding (RUGE), a novel paradigm of KG embedding with iterative guidance from soft rules. RUGE enables an embedding model to learn simultaneously from 1) labeled triples that have been directly observed in a given KG, 2) unlabeled triples whose labels are going to be predicted iteratively, and 3) soft rules with various confidence levels extracted automatically from the KG. In the learning process, RUGE iteratively queries rules to obtain soft labels for unlabeled triples, and integrates such newly labeled triples to update the embedding model. Through this iterative procedure, knowledge embodied in logic rules may be better transferred into the learned embeddings. We evaluate RUGE in link prediction on Freebase and YAGO. Experimental results show that: 1) with rule knowledge injected iteratively, RUGE achieves significant and consistent improvements over state-of-the-art baselines; and 2) despite their uncertainties, automatically extracted soft rules are highly beneficial to KG embedding, even those with moderate confidence levels. The code and data used for this paper can be obtained from https://github.com/iieir-km/RUGE.

Embedding knowledge graphs (KGs) into continuous vector spaces is a focus of current research.Early works performed this task via simple models developed over KG triples.Recent attempts focused on either designing more complicated triple scoring models, or incorporating extra information beyond triples.This paper, by contrast, investigates the potential of using very simple constraints to improve KG embedding.We examine non-negativity constraints on entity representations and approximate entailment constraints on relation representations.The former help to learn compact and interpretable representations for entities.The latter further encode regularities of logical entailment between relations into their distributed representations.These constraints impose prior beliefs upon the structure of the embedding space, without negative impacts on efficiency or scalability.Evaluation on WordNet, Freebase, and DBpedia shows that our approach is simple yet surprisingly effective, significantly and consistently outperforming competitive baselines.The constraints imposed indeed improve model interpretability, leading to a substantially increased structuring of the embedding space.Code and data are available at https://github.com/i ieir-km/ComplEx-NNE_AER.

Learning to hash has proven to be an effective solution for indexing high-dimensional data by projecting them to similarity-preserving binary codes. However, most existing methods end up the learning scheme with a binarization stage, i.e. binary quantization, which inevitably destroys the neighborhood structure of original data. As a result, those methods still suffer from great similarity loss and result in unsatisfactory indexing performance. In this paper we propose a novel hashing model, namely Post Tuned Hashing (PTH), which includes a new post-tuning stage to refine the binary codes after binarization. The post-tuning seeks to rebuild the destroyed neighborhood structure, and hence significantly improves the indexing performance. We cast the post-tuning into a binary quadratic optimization framework and, despite its NP-hardness, give a practical algorithm to efficiently obtain a high-quality solution. Experimental results on five noted image benchmarks show that our PTH improves previous state-of-the-art methods by 13-58% in mean average precision.

Knowledge graphs (KGs) are often significantly incomplete, necessitating a demand for KG completion. Path-based relation inference is one of the most important approaches to this task. Traditional methods treat each path between entity pairs as an atomic feature, thus inducing sparsity. Recently, neural network models solve this problem by decomposing a path as the sequence of relations in the path, before modelling path representations with Recurrent Neural Network (RNN) architectures. In cases there are multiple paths between an entity pair, state-of-the-art neural models either select only one path, or make usage of simple score pooling methods like Top-K, Average, LogSumExp. Unfortunately, none of these methods can model the scenario where relations can only be inferred by considering multiple informative paths collectively. In this paper, we propose a novel path-based relation inference model that learns entity pair representations with attentive path combination. Given an entity pair and a set of paths connecting the pair, our model allows for integrating information from each informative path, and form a dynamic entity pair representation for each query relation. We empirically evaluate the proposed method on a real-world dataset. Experimental results show that the proposed model achieves better performance than state-of-the-art path-based relation inference methods.

In this paper, we formally define the problem of representing and leveraging abstract event causality to power downstream applications. We propose a novel solution to this problem, which build an abstract causality network and embed the causality network into a continuous vector space. The abstract causality network is generalized from a specific one, with abstract event nodes represented by frequently co-occurring word pairs. To perform the embedding task, we design a dual cause-effect transition model. Therefore, the proposed method can obtain general, frequent, and simple causality patterns, meanwhile, simplify event matching. Given the causality network and the learned embeddings, our model can be applied to a wide range of applications such as event prediction, event clustering and stock market movement prediction. Experimental results demonstrate that 1) the abstract causality network is effective for discovering high-level causality rules behind specific causal events; 2) the embedding models perform better than state-of-the-art link prediction techniques in predicting events; and 3) the event causality embedding is an easy-to-use and sophisticated feature for downstream applications such as stock market movement prediction.

This paper considers the problem of embedding Knowledge Graphs (KGs) consisting of entities and relations into low-dimensional vector spaces. Most of the existing methods perform this task based solely on observed facts. The only requirement is that the learned embeddings should be compatible within each individual fact. In this paper, aiming at further discovering the intrinsic geometric structure of the embedding space, we proposeSemantically Smooth Embedding(SSE). The key idea of SSE is to take full advantage of additional semantic information and enforce the embedding space to be semantically smooth, i.e., entities belonging to the same semantic category will lie close to each other in the embedding space. Two manifold learning algorithms Laplacian Eigenmaps and Locally Linear Embedding are used to model the smoothness assumption. Both are formulated as geometrically based regularization terms to constrain the embedding task. Two lines of embedding strategies are tested, i.e., strategies based on latent distance models and strategies based on tensor factorization techniques. We empirically evaluate SSE on two benchmark tasks of link prediction and triple classification, and achieve significant and consistent improvements over state-of-the-art methods. The results demonstrate the superiority and generality of SSE.

Knowledge graph (KG) embedding is to embed components of a KG including entities and relations into continuous vector spaces, so as to simplify the manipulation while preserving the inherent structure of the KG. It can benefit a variety of downstream tasks such as KG completion and relation extraction, and hence has quickly gained massive attention. In this article, we provide a systematic review of existing techniques, including not only the state-of-the-arts but also those with latest trends. Particularly, we make the review based on the type of information used in the embedding task. Techniques that conduct embedding using only facts observed in the KG are first introduced. We describe the overall framework, specific model design, typical training procedures, as well as pros and cons of such techniques. After that, we discuss techniques that further incorporate additional information besides facts. We focus specifically on the use of entity types, relation paths, textual descriptions, and logical rules. Finally, we briefly introduce how KG embedding can be applied to and benefit a wide variety of downstream tasks such as KG completion, relation extraction, question answering, and so forth.

Knowledge bases (KBs) are often greatly incomplete, necessitating a demand for KB completion. The path ranking algorithm (PRA) is one of the most promising approaches to this task. Previous work on PRA usually follows a single-task learning paradigm, building a prediction model for each relation independently with its own training data. It ignores meaningful associations among certain relations, and might not get enough training data for less frequent relations. This paper proposes a novel multi-task learning framework for PRA, referred to as coupled PRA (CPRA). It first devises an agglomerative clustering strategy to automatically discover relations that are highly correlated to each other, and then employs a multi-task learning strategy to effectively couple the prediction of such relations. As such, CPRA takes into account relation association and enables implicit data sharing among them. We empirically evaluate CPRA on benchmark data created from Freebase. Experimental results show that CPRA can effectively identify coherent clusters in which relations are highly correlated. By further coupling such relations, CPRA significantly outperforms PRA, in terms of both predictive accuracy and model interpretability.

Distant supervision is an efficient approach that automatically generates labeled data for relation extraction (RE). Traditional distantly supervised RE systems rely heavily on handcrafted features, and hence suffer from error propagation. Recently, a neural network architecture has been proposed to automatically extract features for relation classification. However, this approach follows the traditional expressed-at-least-once assumption, and fails to make full use of information across different sentences. Moreover, it ignores the fact that there can be multiple relations holding between the same entity pair. In this paper, we propose a multi-instance multi-label convolutional neural network for distantly supervised RE. It first relaxes the expressed-at-least-once assumption, and employs cross-sentence max-pooling so as to enable information sharing across different sentences. Then it handles overlapping relations by multi-label learning with a neural network classifier. Experimental results show that our approach performs significantly and consistently better than state-of-the-art methods.

Embedding knowledge graphs into continuous vector spaces has recently attracted increasing interest.Most existing methods perform the embedding task using only fact triples.Logical rules, although containing rich background information, have not been well studied in this task.This paper proposes a novel method of jointly embedding knowledge graphs and logical rules.The key idea is to represent and model triples and rules in a unified framework.Specifically, triples are represented as atomic formulae and modeled by the translation assumption, while rules represented as complex formulae and modeled by t-norm fuzzy logics.Embedding then amounts to minimizing a global loss over both atomic and complex formulae.In this manner, we learn embeddings compatible not only with triples but also with rules, which will certainly be more predictive for knowledge acquisition and inference.We evaluate our method with link prediction and triple classification tasks.Experimental results show that joint embedding brings significant and consistent improvements over stateof-the-art methods.Particularly, it enhances the prediction of new facts which cannot even be directly inferred by pure logical inference, demonstrating the capability of our method to learn more predictive embeddings.

Shu Guo, Quan Wang, Bin Wang, Lihong Wang, Li Guo. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2015.

We consider the problem of embedding knowledge graphs (KGs) into continuous vector spaces.Existing methods can only deal with explicit relationships within each triple, i.e., local connectivity patterns, but cannot handle implicit relationships across different triples, i.e., contextual connectivity patterns.This paper proposes context-dependent KG embedding, a twostage scheme that takes into account both types of connectivity patterns and obtains more accurate embeddings.We evaluate our approach on the tasks of link prediction and triple classification, and achieve significant and consistent improvements over state-of-the-art methods.

Estimating questions ’ difficulty levels is an important task in community question answering (CQA) services. Previous stud-ies propose to solve this problem based on the question-user comparisons extract-ed from the question answering threads. However, they suffer from data sparseness problem as each question only gets a lim-ited number of comparisons. Moreover, they cannot handle newly posted question-s which get no comparisons. In this pa-per, we propose a novel question difficul-ty estimation approach called Regularized Competition Model (RCM), which natu-rally combines question-user comparisons and questions ’ textual descriptions into a unified framework. By incorporating tex-tual information, RCM can effectively deal with data sparseness problem. We further employ a K-Nearest Neighbor approach to estimate difficulty levels of newly post-ed questions, again by leveraging textu-al similarities. Experiments on two pub-licly available data sets show that for both well-resolved and newly-posted question-s, RCM performs the estimation task sig-nificantly better than existing methods, demonstrating the advantage of incorpo-rating textual information. More interest-ingly, we observe that RCMmight provide an automatic way to quantitatively mea-sure the knowledge levels of words. 1

In this paper, we address the problem of estimating question difficulty in community question answering services.We propose a competition-based model for estimating question difficulty by leveraging pairwise comparisons between questions and users.Our experimental results show that our model significantly outperforms a PageRank-based approach.Most importantly, our analysis shows that the text of question descriptions reflects the question difficulty.This implies the possibility of predicting question difficulty from the text of question descriptions.