Large language models (LLMs) may generate harmful outputs on malicious inputs. Existing safety methods, including prompt engineering and model editing, rely on hand-crafted templates or target-driven parameter modifications, limiting their generalizability in unseen harmful scenarios. Post-training aims to ensure LLM safety in general domains via supervised fine-tuning (SFT) or reinforcement learning (RL) on diverse malicious inputs. SFT needs annotated refusal samples while RL learns to refuse risk by exploring diverse harmful inputs. However, these methods tend to harshly refuse over any possible risks, sacrificing potentially useful information and degrading model utility. We argue that realistic malicious inputs often mix both harmful and helpful semantics (i.e., entities and relations), and LLMs should identify and remove only harmful relations while preserving useful ones. Thus, the original malicious user inputs can shift into safe queries, to which LLMs can respond safely and helpfully. In this paper, we propose WALKSAFE, a graph-based risk-aware training framework that enables LLMs to identify potential risks of key semantics (entities and relations) in user inputs via graph structure. By filtering harmful relations, LLMs can respond to safe input queries and then generate their corresponding safe and helpful responses. First, we model all entities and relations in the inputs with a graph structure. Second, we adopt a risk-aware random walk on the graph to quantify potential risk under multiple entities and relations. Then, we reconstruct safe queries by filtering harmful relations to promote the LLM to answer safely and helpfully rather than with direct refusals. Finally, we propose Bi-GRPO to post-train LLMs. As vanilla GRPO conducts only the intra-group comparison, Bi-GRPO performs both intra-group and inter-group comparisons between different response groups. The extra inter-group rewards encourage the model to distinguish harmful and safe semantics, and thus prefer safe and helpful responses. Experiments on three LLMs show that our models obtain SOTA results.

In multi-hop question answering (MHQA) tasks, Chain of Thought (CoT) improves the quality of generation by guiding large language models (LLMs) through multi-step reasoning, and Knowledge Graphs (KGs) reduce hallucinations via semantic matching. Outcome Reward Models (ORMs) provide feedback after generating the final answers but fail to evaluate the process for multi-step reasoning. Traditional Process Reward Models (PRMs) evaluate the reasoning process but require costly human annotations or rollout generation. While implicit PRM is trained only with outcome signals and derives step rewards through reward parameterization without explicit annotations, it is more suitable for multi-step reasoning in MHQA tasks. However, existing implicit PRM has only been explored for plain text scenarios. When adapting to MHQA tasks, it cannot handle the graph structure constraints in KGs and capture the potential inconsistency between CoT and KG paths. To address these limitations, we propose the DPRM (Dual Implicit Process Reward Model). It trains two implicit PRMs for CoT and KG reasoning in MHQA tasks. Both PRMs, namely KG-PRM and CoT-PRM, derive step-level rewards from outcome signals via reward parameterization without additional explicit annotations. Among them, KG-PRM uses preference pairs to learn structural constraints from KGs. DPRM further introduces a consistency constraint between CoT and KG reasoning steps, making the two PRMs mutually verify and collaboratively optimize the reasoning paths. We also provide a theoretical demonstration of the derivation of process rewards. Experimental results show that our method outperforms 13 baselines on multiple datasets with up to 16.6% improvement on Hit@1.

Xinyue Fang, Zhiliang Tian, Zhen Huang, Ziyi Pan, Zhihua Wen, Xi Wang, Quntian Fang, Dongsheng Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Tianle Liu, Zhiliang Tian, Zhen Huang, Tianlun Liu, Jingyuan Huang, Zhaoning Zhang, Chengcheng Shao, Dongsheng Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Tianlun Liu, Zhiliang Tian, Zhen Huang, Xingzhi Zhou, Wanlong Yu, Tianle Liu, Feng Liu, Dongsheng Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

When fact-checking methods based on large language models (LLMs) use external evidence to validate claims, knowledge conflicts often arise. These conflicts typically stem from inconsistencies between the external evidence and LLMs' internal pre-existing knowledge. Such an inconsistency could lead LLMs to draw incorrect answers when validating claims, especially when they are overly confident in their internal incorrect knowledge. Previous works on LLM-based fact-checking have overlooked this issue. This paper, for the first time, proposes a framework (namely KnowFC) to navigate this issue. Our key insight is dividing and adaptively utilizing the knowledge that LLMs know and do not know, thereby avoiding conflicts while enhancing the correctness and efficiency of fact-checking. Specifically, in KnowFC, we propose an adaptive retrieval method, where we train an LLM using a reinforcement learning algorithm coupled with the Dunning-Kruger effect-inspired reward mechanism to identify its knowledge boundaries through confidence calibration, thereby realizing adaptive evidence retrieval. Besides, we propose a reliable and debiased fact verification method, where we organize and construct reasoning graphs using retrieved evidence to verify claims, followed by a causal intervention method using causal mediation analysis to mitigate internal knowledge interference. Experimental results on both FEVEROUS and AVeriTeC datasets show that our method outperforms baseline methods in terms of accuracy and F1 score, while also improving fact-checking efficiency.

Retrieval-augmented generation (RAG) has become a cornerstone for enhancing large language models (LLMs) with real-time information from the Web, but its performance often heavily depends on the quality of the retrieved documents. Given that RAG systems frequently draw from vast and often noisy Web corpora, ensuring the reliability of retrieved content is paramount. While rerankers improve the factual accuracy of the RAG system by elevating the proportion of ground-truth documents (GD) in high-ranked results, the shifts of document type distributions during reranking remain unclear, hindering the understanding of the reranker's behavior. To bridge this gap, we conduct an empirical study to categorize documents and compare their distribution before and after reranking. We reveal a counterintuitive finding: though rerankers improve the proportion of GD, they also significantly increase the proportion of harmful documents (HD) in top-ranked retrieved documents. It not only narrows the potential context window for ranking the GD higher but also increases the risk of HD misleading the LLMs, potentially leading to the generation and propagation of misinformation across Web platforms. Motivated by this finding, we propose a risk-aware reranking method for RAG with LLMs, which balances the risk and benefit during reranking. Given a query, the RAG framework first retrieves relevant documents. Then, our approach quantifies the potential beneficial and harmful impacts of various documents on the LLMs' generation. To estimate the impacts, we conduct a dual-aspect document impact assessment via information gain, which employs a risk clipping to avoid the numerical fluctuations in the estimation. Finally, we conduct the reranking according to the potential impact of each document, enabling the reranker to significantly reduce the HD proportion. Experiments and analysis across multiple models and datasets, including Wikipedia, web news, and research papers, show the effectiveness of our method. Our code is available at https://github.com/lzz335/hidden_risk_of_reranking.

Rumor detection is essential for building a responsible web and internet ecosystem, which has attracted significant attention from the research community. However,emerging topic rumor detection, i.e., identify rumors at the early stages of a topic's emergence where only limited discussions can be observed, still remains a challenge. Technically, this scenario is accompanied by the issues ofdata scarcityon emerging topics and thedata distribution discrepancybetween old topics and emerging new topic. In this paper, we propose a new framework termedLLM-drivenADversarialExampleSynthesis (LADES) for emerging topic rumor detection. LADES utilizes Large Language Models (LLMs) for generating readable and contextually coherent adversarial examples. The generated adversarial examples not only expand the training set to tackle the data scarcity issue, but also act as a bridge to connect the data distribution of old and new topics. To overcome training instability in adversarial example generation, LADES introduces a gradient-free Markov Chain Monte Carlo (MCMC) sampling method. This method ensures adversarial examples are readable and contextually coherent by harnessing LLMs, while promoting effective attacks through entropy-based sampling that targets model uncertainty. To mitigate the impact of potential mislabeling in synthetic data, LADES implements a meta-mixed-learning mechanism. This mechanism dynamically adjusts the weights of synthetic adversarial examples, guided by limited labeled data from emerging topics, thereby alleviating the data noise.

Test-time adaptation (TTA) deals with domain shifts during inference by training models based on only unlabeled test samples. Test samples may include noisy samples, which degrade domain adaptation. Existing methods rely on the model's output prediction to detect and filter noisy samples, and further search for flat regions during optimization, which makes the optimization more robust on noisy samples. However, there are two issues: (1) the output prediction tends to be inaccurate due to domain shifts, weakening noisy-sample detection; (2) current approaches for searching flat regions focus on optimization to enhance the worst case, which ignores achieving flatness by avoiding the quick changing of losses. To address these challenges, we propose a model pruning-based test-time adaptation model for noisy data streams, named MoTTA, which leverages a new proposed filtering, output difference under pruning (ODP)-based filtering, and a flatness-aware entropy minimization (FlatEM). Specifically, to reduce the impact of inaccurate output predictions, ODP-based filtering measures the output difference of a sample before and after model pruning, which works even under inaccurate output. To improve the search for flat loss surfaces, FlatEM integrates zeroth-order flatness and first-order flatness (minimize the maximal gradient normalization with a weight perturbation constrained in a small Euclidean ball) on entropy minimization. To solve these hard maximum problems, we leverage Taylor expansion to obtain approximated results for optimization. FlatEM also adopts a parameter regularization to mitigate incorrect updates from noisy samples. The experiments show our advantages in dealing with noisy data streams at TTA comparable to existing baselines.

LLMs obtain remarkable performance but suffer from hallucinations. Most research on detecting hallucination focuses on questions with short and concrete correct answers that are easy to check faithfulness. Hallucination detections for text generation with open-ended answers are more hard. Some researchers use external knowledge to detect hallucinations in generated texts, but external resources for specific scenarios are hard to access. Recent studies on detecting hallucinations in long texts without external resources conduct consistency comparison among multiple sampled outputs. To handle long texts, researchers split long texts into multiple facts and individually compare the consistency of each pair of facts. However, these methods (1) hardly achieve alignment among multiple facts; (2) overlook dependencies between multiple contextual facts. In this paper, we propose a graph-based context-aware (GCA) hallucination detection method for text generations, which aligns facts and considers the dependencies between contextual facts in consistency comparison. Particularly, to align multiple facts, we conduct a triple-oriented response segmentation to extract multiple knowledge triples. To model dependencies among contextual triples (facts), we construct contextual triples into a graph and enhance triples’ interactions via message passing and aggregating via RGCN. To avoid the omission of knowledge triples in long texts, we conduct an LLM-based reverse verification by reconstructing the knowledge triples. Experiments show that our model enhances hallucination detection and excels all baselines.

Multi-agent collaborative tasks exhibit exceptional capabilities in natural language applications and generation. By prompting agents to assign clear roles, it is possible to facilitate cooperation and achieve complementary capabilities among LLMs. A common strategy involves adopting a relatively general role assignment mechanism, such as introducing a “judge” or a “summarizer”. However, these approaches lack task-specific role customization based on task characteristics. Another strategy involves decomposing the task based on domain knowledge and task characteristics, followed by assigning appropriate roles according to LLMs’ respective strengths, such as programmers and testers. However, in some given tasks, obtaining domain knowledge related to task characteristics and getting the strengths of different LLMs is hard. To solve these problems, we propose a Multi-LLM Cooperation (MLC) framework with automatic role assignment capabilities. The core idea of the MLC is to initialize role assignments randomly and then allow the role embeddings to be learned jointly with the downstream task. To capture the state transitions of multiple LLMs during turn-based speaking, the role embedding is sequence-aware. At the same time, to avoid role convergence, the role differentiation module in MLC encourages behavioral differentiation between LLMs while ensuring the LLM team consistency, guiding different LLMs to develop complementary strengths from the optimization level. Our experiments on seven datasets demonstrate that MLC significantly enhances collaboration and expertise, which collaboratively addresses multi-agent tasks.

Shilong Pan, Zhiliang Tian, Zhen Huang, Wanlong Yu, Zhihua Wen, Xinwang Liu, Kai Lu, Minlie Huang, Dongsheng Li. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Zhiliang Tian, Jingyuan Huang, Zejiang He, Zhen Huang, Menglong Lu, Linbo Qiao, Songzhu Mei, Yijie Wang, Dongsheng Li. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Haiyang Wang, Zhiliang Tian, Yuchen Pan, Xin Song, Xin Niu, Minlie Huang, Bin Zhou. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Qiunan Du, Zhiliang Tian, Zhen Huang, Kailun Bian, Tianlun Liu, Zhaoning Zhang, Xinwang Liu, Feng Liu, Dongsheng Li. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

With the emergence of new topics on social media as sources of rumor propagation, addressing the domain shift between the source and target domain and the target domain samples scarcity remains a crucial task in cross-domain rumor detection.Traditional deep learningbased methods and LLM-based methods are mostly focused on the in-domain condition, thus having poor performance in cross-domain setting.Existing domain adaptation rumor detection approaches ignore the data generalization differences and rely on a large amount of unlabeled target domain samples to achieve domain adaptation, resulting in less effective on emerging topic rumor detection.In this paper, we propose a Gradient Coherence guided Meta-Learning approach (GCML) for emerging topics rumor detection.Firstly, we calculate the task generalization score of each source task (sampled from source domain) from a gradient coherence perspective, and selectively learn more "generalizable" tasks that are more beneficial in adapting to the target domain.Secondly, we leverage meta-learning to alleviate the target domain samples scarcity, which utilizes task generalization scores to re-weight metatest gradients and adaptively updates learning rate.Extensive experimental results on realworld datasets show that our method substantially outperforms SOTA baselines.

With the continuous advancement in the performance of large language models (LLMs), their demand for computational resources and memory has significantly increased, which poses major challenges for efficient inference on consumer-grade devices and legacy servers.These devices typically feature relatively weaker GPUs and stronger CPUs.Although techniques such as parameter offloading and partial offloading can alleviate GPU memory pressure to some extent, their effectiveness is limited due to communication latency and suboptimal hardware resource utilization.To address this issue, we propose Dovetail 1 , a lossless inference acceleration method that leverages the complementary characteristics of heterogeneous devices and the advantages of speculative decoding.Dovetail deploys a draft model on the GPU to perform preliminary predictions, while a target model running on the CPU validates these outputs.By reducing the granularity of data transfer, Dovetail significantly minimizes communication overhead.To further improve efficiency, we optimize the draft model specifically for heterogeneous hardware environments by reducing the number of draft tokens to lower parallel verification latency, increasing model depth to enhance predictive capabilities, and introducing a Dynamic Gating Fusion (DGF) mechanism to improve the integration of feature and embedding information.We conduct comprehensive evaluations of Dovetail across various consumergrade GPUs, covering multiple tasks and mainstream models.Experimental results on 13B models demonstrate that Dovetail achieves inference speedups ranging from 1.79x to 10.1x across different devices, while maintaining consistency and stability in the distribution of generated texts.* indicates equal contribution.† indicates corresponding authors. 1

Test-time domain adaptation adjusts a source domain model to accommodate previously unseen domain shifts in a target domain during inference. In real-world scenarios, domain shifts continually evolve, and test data are often non-independent and identically distributed (non-i.i.d.). Existing methods update batch normalization (BN) statistics (mean and variance) based on test batch statistics to mitigate domain shifts and use a memory bank to provide approximate i.i.d. sampling by selectively storing samples. However, excessive updates to BN statistics lead to overfitting to specific domain shifts. To address this issue, we propose a resilient practical test-time adaptation method (ResiTTA), employing soft constraints on the BN statistics and a low-entropy sampling strategy, which reduces overfitting on domain shifts and enables rapid adaptation. Specifically, we develop a resilient batch normalization (BN) with estimated statistics and soft constraints between the source and the estimated statistics. The soft constraints regularize the estimated statistics to mitigate overfitting caused by the excessive updates. To avoid overfitting, we design a low-entropy memory bank that accounts for sample uncertainty and class balance. We adapt the source domain model via a teacher-student self-training adaptation on the samples from the memory, incorporating the soft constraints’ updates to BN. Our ResiTTA obtains state-of-the-art results on various benchmarks. We release our code1.

Controllable text-to-image generation synthesizes visual text and objects in images with certain conditions. However, existing visual text rendering and layout-to-image generation tasks focus on single modality generation or rendering, leaving yet-to-be-bridged gaps between the approaches correspondingly designed for each of the tasks. In this paper, we introduce a unified task called layout-controllable text-object synthesis (LTOS), which merges text rendering and layout-based object generation. We construct a layout-aware text-object synthesis dataset for the LTOS task, containing well-aligned labels visual text, and object information. We also propose a layout-controllable text-object adaptive fusion (TOF) model, which generates images with legible visual text and plausible objects. Specifically, we construct a visual-text rendering to synthesize text and employ an object-layout control module to generate objects while integrating the two modules to harmoniously generate and integrate text and objects in images. To enhance the image-text integration, we propose a self-adaptive cross-attention fusion that forces the image generation to attend more to important text spans. Within such a fusion, we use a self-adaptive learnable factor to learn to control the influence of cross-attention outputs. Experiments show that our model excels the strong baselines and obtain SOTA in LTOS. We release our dataset and codes1.

Large language models (LLMs) are now pivotal in real-world applications. Model editing has emerged as a promising paradigm for efficiently modifying LLMs without full retraining. However, current editing approaches face significant limitations due to parameter drift, which stems from inconsistencies between newly edited knowledge and the model's existing knowledge. In sequential editing scenarios, cumulative drifts progressively lead to model collapse characterized by general capability degradation and balance between acquiring new knowledge and catastrophic forgetting of existing knowledge. Drawing inspiration from the hippocampal trisynaptic circuit for continual memorizing and forgetting, we propose a Hippocampal-like Sequential Editing (HSE) framework that designs the unlearning of obsolete knowledge, domain-specific knowledge update separation and replay for edited knowledge. Specifically, the HSE framework designs three core mechanisms: (1) Machine unlearning selectively erases outdated knowledge to facilitate integration of new information, (2) Fisher Information Matrix-guided parameter updates prevents cross-domain knowledge interference, and (3) Parameter replay consolidates long-term editing memory through lightweight and global replay of editing data in a parametric form. Theoretical analysis demonstrates that HSE achieves smaller generalization error bounds, more stable convergence and higher computational efficiency. Experimental results validate its effective balance between acquiring new knowledge and mitigating catastrophic forgetting, maintaining or even slightly enhancing general capabilities. In practical applications, experiments confirm its effectiveness in multi-domain hallucination mitigation, healthcare knowledge injecting, and societal bias reduction.

Large language models (LLMs) demonstrate significant potential in various applications; however, they are susceptible to generating hallucinations, which can lead to the spread of online misinformation. Existing studies address hallucination detection by (1) employing reference-based methods that consult external resources for verification or (2) utilizing reference-free methods that mainly estimate answer uncertainty based on LLM's internal states. However, reference-based methods incur significant costs and can be infeasible for obtaining reliable external references. Besides, existing uncertainty estimation (UE) methods often overlook the impact of scenario backgrounds inherited from the query's lexical resources, leading to noise in UE. In almost all real-world applications, users care about the uncertainty concerning semantics or facts instead of the query's scenario information. Therefore, we argue that mitigating scenario-related noise and focusing on semantic information can yield a more desirable UE. In this paper, we introduce a plug-and-play scenario-independent framework to enhance unsupervised UE in LLMs by removing scenario-related noise and focusing on semantic information. This framework is compatible with most existing UE methods, as it leverages only the existing UE methods' outputs. Specifically, we design a scenario-specific sampling to paraphrase queries, maintaining their common semantics while diversifying the scenario distribution. Subsequently, to estimate the contribution of the common semantics, we design a factor analysis (FA) model to disentangle the UE score obtained from the given UE method into a combination of multiple latent factors, which represent the contribution of the common semantics and scenario-related noise. By solving the FA model, we decompose the impact of the most significant factor to approximate the uncertainty caused by the common semantics, thus achieving scenario-independent UE. Extensive experiments and analysis across multiple models and datasets demonstrate the effectiveness of our approach.

Event extraction (EE) is a complex natural language processing (NLP) task that aims at identifying and classifying triggers and arguments in raw text. The polysemy of triggers and arguments stands out as one of the key challenges affecting the precise extraction of events. Existing approaches commonly consider the semantic distribution of triggers and arguments to be balanced. However, the sample quantities of different semantics in the same trigger or argument vary in real-world scenarios, leading to a biased semantic distribution. The bias introduces two challenges: (1) low-frequency semantics is difficult to identify; (2) high-frequency semantics is often mistakenly identified. To tackle these challenges, we propose an adaptive learning method with the reward–penalty mechanism for balancing the semantic distribution in polysemous triggers and arguments. The reward–penalty mechanism balances the semantic distribution by enlarging the gap between the target and nontarget semantics by rewarding correct classifications and penalizing incorrect classifications. Additionally, we propose a sentence-level event situation awareness (SA) mechanism to guide the encoder to accurately learn the knowledge of events mentioned in the sentence, thereby enhancing target event semantics in the distribution of polysemous triggers and arguments. Finally, for various semantics in different tasks, we propose task-specific semantic decoders to precisely identify the boundaries of the predicted triggers and arguments for the semantics. Our experimental results on ACE2005 and its variants, along with the rich Entities, Relations, and Events (ERE), demonstrate the superiority of our approach over single-task and multi-task EE baselines.

Shilong Pan, Zhiliang Tian, Liang Ding, Haoqi Zheng, Zhen Huang, Zhihua Wen, Dongsheng Li. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Counterspeech is an effective way to combat online hate speech. Considering the multifaceted nature of online hate speech, counterspeech with varying intents (e.g., denouncing or empathy) has significant potential to mitigate hate speech effectively. Recently, controlled approaches based on large language models (LLMs) have been explored to generate intent-specific counterspeech. Due to the lack of attention to intent-specific information by LLMs during the decoding process, those methods cater more to the semantic information rather than matching with the desired intents. Further, there are still limitations in quantitatively evaluating the effectiveness of counterspeech with different intents in mitigating hate speech. In this paper, to address the above issues, we propose DART, an LLMs-based DuAl-discRiminaTor guided framework for counterspeech generation. We employ an intent-aware discriminator and hate-mitigating discriminator to jointly guide the decoding preferences of LLMs, which facilitates the model towards generating counterspeech catering to specific intent and hate mitigation. We apply a maximum-margin relative objective for training discriminators. This objective leverages the distance between counterspeech aligned with the desired target (such as specific intent or effectiveness in hate mitigation) and undesired as an effective learning signal. Extensive experiments show that DART achieves excellent performances in matching the desired intent and mitigating hate.

Unsupervised text style transfer aims to modify the style of a sentence while preserving its content without parallel corpora. Existing approaches attempt to separate content from style, but some words contain both content and style information. It makes them difficult to disentangle, where unsatisfactory disentanglement results in the loss of the content information or the target style. To address this issue, researchers adopted a “cycle reconstruction” mechanism to maintain content information, but it is still hard to achieve satisfactory content preservation due to incomplete disentanglement. In this paper, we propose a new disentanglement-based method, StyleFlow, which effectively avoids the loss of contents through a better cycle reconstruction via a reversible encoder. The reversible encoder is a normalizing flow that can not only produce output given input but also infer the exact input given the output reversely. We design a stack of attention-aware coupling layers, where each layer is reversible and adopts the attention mechanism to improve the content-style disentanglement. Moreover, we propose a data augmentation method based on normalizing flow to enhance the training data. Our experiments on sentiment transfer and formality transfer tasks show that StyleFlow outperforms strong baselines on both content preservation and style transfer.

Watermarking enables people to determine whether the text is generated by a specific model.It injects a unique signature based on the "green-red" list that can be tracked during detection, where the words in green lists are encouraged to be generated.Recent researchers propose to fix the green/red lists or increase the proportion of green tokens to defend against paraphrasing attacks.However, these methods cause degradation of text quality due to semantic disparities between the watermarked text and the unwatermarked text.In this paper, we propose a semantic-aware watermark method that considers contexts to generate a semantic-aware key to split a semantically balanced green/red list for watermark injection.The semantic balanced list reduces the performance drop due to adding bias on green lists.To defend against paraphrasing attacks, we generate the watermark key considering the semantics of contexts via locally sensitive hashing.To improve the text quality, we propose to split green/red lists considering semantics to enable the green list to cover almost all semantics.We also dynamically adapt the bias to balance text quality and robustness.The experiments show our advantages in both robustness and text quality comparable to existing baselines.

Large Language Models (LLMs) are widely used for knowledge-seeking purposes yet suffer from hallucinations. The knowledge boundary of an LLM limits its factual understanding, beyond which it may begin to hallucinate. Investigating the perception of LLMs' knowledge boundary is crucial for detecting hallucinations and LLMs' reliable generation. Current studies perceive LLMs' knowledge boundary on questions with concrete answers (close-ended questions) while paying limited attention to semi-open-ended questions that correspond to many potential answers. Some researchers achieve it by judging whether the question is answerable or not. However, this paradigm is not so suitable for semi-open-ended questions, which are usually ``partially answerable questions'' containing both answerable answers and ambiguous (unanswerable) answers. Ambiguous answers are essential for knowledge-seeking, but it may go beyond the knowledge boundary of LLMs. In this paper, we perceive the LLMs' knowledge boundary with semi-open-ended questions by discovering more ambiguous answers. First, we apply an LLM-based approach to construct semi-open-ended questions and obtain answers from a target LLM. Unfortunately, the output probabilities of mainstream black-box LLMs are inaccessible to sample more low-probability ambiguous answers. Therefore, we apply an open-sourced auxiliary model to explore ambiguous answers for the target LLM. We calculate the nearest semantic representation for existing answers to estimate their probabilities, with which we reduce the generation probability of high-probability existing answers to achieve a more effective generation. Finally, we compare the results from the RAG-based evaluation and LLM self-evaluation to categorize four types of ambiguous answers that are beyond the knowledge boundary of the target LLM. Following our method, we construct a dataset to perceive the knowledge boundary for GPT-4. We find that GPT-4 performs poorly on semi-open-ended questions and is often unaware of its knowledge boundary. Besides, our auxiliary model, LLaMA-2-13B, is effective in discovering many ambiguous answers, including correct answers neglected by GPT-4 and delusive wrong answers GPT-4 struggles to identify.

Document-level event extraction endeavors to automatically extract structural events from a given document. Many existing approaches focus on modeling entity interactions and decoding these interactions into events, assigning each entity as an event argument. However, these approaches encounter two primary limitations: they exclusively capture semantic dependencies to model entity interactions, overlooking the indication of the spatial distribution features of entities; they decode interactions imprecisely with a hard binary-classification boundary, potentially failing to calibrate micro differences in interactions. To overcome these limitations, we introduce a novel approach termed the S patiality-augmented I nteraction Model with A daptive T hresholding (SIAT). Our method addresses the first limitation by calculating the relative position encoding of entities to represent spatial interaction features. These features are then integrated with multi-granularity semantic interactions, enhancing the modeling of entity interactions for each entity pair. Furthermore, we introduce an adaptive event decoding mechanism, which establishes a more flexible decision boundary for different entity interactions. Additionally, an adaptive loss function for threshold learning is designed to further refine the model. Experimental results demonstrate that our proposed method achieves competitive performance compared to state-of-the-art methods on two public event extraction datasets while maintaining considerable training efficiency.

Self-training emerges as an important research line on domain adaptation.By taking the model's prediction as the pseudo labels of the unlabeled data, self-training bootstraps the model with pseudo instances in the target domain.However, the prediction errors of pseudo labels (label noise) challenge the performance of self-training.To address this problem, previous approaches only use reliable pseudo instances, i.e., pseudo instances with high prediction confidence, to retrain the model.Although these strategies effectively reduce the label noise, they are prone to miss the hard examples.In this paper, we propose a new self-training framework for domain adaptation, namely Domain adversarial learning enhanced Self-Training Framework (DaMSTF).Firstly, DaMSTF involves meta-learning to estimate the importance of each pseudo instance, so as to simultaneously reduce the label noise and preserve hard examples.Secondly, we design a meta constructor for constructing the meta validation set, which guarantees the effectiveness of the meta-learning module by improving the quality of the meta validation set.Thirdly, we find that the meta-learning module suffers from the training guidance vanishment and tends to converge to an inferior optimal.To this end, we employ domain adversarial learning as a heuristic neural network initialization method, which can help the meta-learning module converge to a better optimal.Theoretically and experimentally, we demonstrate the effectiveness of the proposed DaMSTF.On the cross-domain sentiment classification task, DaMSTF improves the performance of BERT with an average of nearly 4%.

The recent progress in graph representation learning boosts the development of many graph classification tasks, such as protein classification and social network classification. One of the mainstream approaches for graph representation learning is the hierarchical pooling method. It learns the graph representation by gradually reducing the scale of the graph, so it can be easily adapted to large-scale graphs. However, existing graph pooling methods discard the original graph structure during downsizing the graph, resulting in a lack of graph topological structure. In this paper, we propose a multi-scale graph neural network (MSGNN) model that not only retains the topological information of the graph but also maintains the key-subgraph for better interpretability. MSGNN gradually discards the unimportant nodes and retains the important subgraph structure during the iteration. The key subgraphs are first chosen by experience and then adaptively evolved to tailor specific graph structures for downstream tasks. The extensive experiments on seven datasets show that MSGNN improves the SOTA performance on graph classification and better retains key subgraphs.

Text classification tasks often encounter fewshot scenarios with limited labeled data, and addressing data scarcity is crucial.Data augmentation with mixup merges sample pairs to generate new pseudos, which can relieve the data deficiency issue in text classification.However, the quality of pseudo-samples generated by mixup exhibits significant variations.Most of the mixup methods fail to consider the varying degree of learning difficulty in different stages of training.And mixup generates new samples with one-hot labels, which encourages the model to produce a high prediction score for the correct class that is much larger than other classes, resulting in the model's over-confidence.In this paper, we propose a self-evolution learning (SE) based mixup approach for data augmentation in text classification, which can generate more adaptive and model-friendly pseudo samples for the model training.SE caters to the growth of the model learning ability and adapts to the ability when generating training samples.To alleviate the model over-confidence, we introduce an instance-specific label smoothing regularization approach, which linearly interpolates the model's output and one-hot labels of the original samples to generate new soft labels for label mixing up.Through experimental analysis, experiments show that our SE brings consistent and significant improvements upon different mixup methods.In-depth analyses demonstrate that SE enhances the model's generalization ability.

Neural architecture search (NAS) has made tremendous progress in the automatic design of effective neural network structures but suffers from a heavy computational burden. One-shot NAS significantly alleviates the burden through weight sharing and improves computational efficiency. Zero-shot NAS further reduces the cost by predicting the performance of the network from its initial state, which conducts no training. Both methods aim to distinguish between "good" and "bad" architectures, i.e., ranking consistency of predicted and true performance. In this paper, we propose Ranking Distillation one-shot NAS (RD-NAS) to enhance ranking consistency, which utilizes zero-cost proxies as the cheap teacher and adopts the margin ranking loss to distill the ranking knowledge. Specifically, we propose a margin subnet sampler to distill the ranking knowledge from zero-shot NAS to one-shot NAS by introducing Group distance as margin. Our evaluation of the NAS-Bench-201 and ResNet-based search space demonstrates that RD-NAS achieve 10.7% and 9.65% improvements in ranking ability, respectively. Our codes are available at https://github.com/pprp/CVPR2022-NAS-competition-Track1-3th-solution

Practical networks for edge devices adopt shallow depth and small convolutional kernels to save memory and computational cost, which leads to a restricted receptive field. Conventional efficient learning methods focus on lightweight convolution designs, ignoring the role of the receptive field in neural network design. In this paper, we propose the Meta-Pooling framework to make the receptive field learnable for a lightweight network, which consists of parameterized pooling-based operations. Specifically, we introduce a parameterized spatial enhancer, which is composed of pooling operations to provide versatile receptive fields for each layer of a lightweight model. Then, we present a Progressive Meta-Pooling Learning (PMPL) strategy for the parameterized spatial enhancer to acquire a suitable receptive field size. The results on the ImageNet dataset demonstrate that MobileNetV2 using Meta-Pooling achieves top1 accuracy of 74.6%, which outperforms MobileNetV2 by 2.3%.

Mixed-Precision Quantization (MQ) can achieve a competitive accuracy-complexity trade-off for models. Conventional training-based search methods require time-consuming candidate training to search optimized per-layer bit-width configurations in MQ. Recently, some training-free approaches have presented various MQ proxies and significantly improve search efficiency. However, the correlation between these proxies and quantization accuracy is poorly understood. To address the gap, we first build the MQ-Bench-101, which involves different bit configurations and quantization results. Then, we observe that the existing training-free proxies perform weak correlations on the MQ-Bench-101. To efficiently seek superior proxies, we develop an automatic search of proxies framework for MQ via evolving algorithms. In particular, we devise an elaborate search space involving the existing proxies and perform an evolution search to discover the best correlated MQ proxy. We proposed a diversity-prompting selection strategy and compatibility screening protocol to avoid premature convergence and improve search efficiency. In this way, our Evolving proxies for Mixed-precision Quantization (EMQ) framework allows the auto-generation of proxies without heavy tuning and expert knowledge. Extensive experiments on ImageNet with various ResNet and MobileNet families demonstrate that our EMQ obtains superior performance than state-of-the-art mixed-precision methods at a significantly reduced cost. The code will be released.

Text classification is a fundamental task for natural language processing, and adapting text classification models across domains has broad applications. Self-training generates pseudo-examples from the model's predictions and iteratively trains on the pseudo-examples, i.e., minimizes the loss on the source domain and the Gibbs entropy on the target domain. However, Gibbs entropy is sensitive to prediction errors, and thus, self-training tends to fail when the domain shift is large. In this paper, we propose Meta-Tsallis Entropy minimization (MTEM). MTEM uses an instance adaptive Tsallis entropy to replace the Gibbs entropy and a meta-learning algorithm to optimize the instance adaptive Tsallis entropy on the target domain. To reduce the computation cost of MTEM, we propose an approximation technique to approximate the second-order derivation involved in the meta-learning. To efficiently generate pseudo labels, we propose an annealing sampling mechanism for exploring the model's prediction probability. Theoretically, we prove the convergence of the meta-learning algorithm in MTEM and analyze the effectiveness of MTEM in achieving domain adaptation. Experimentally, MTEM improves the adaptation performance of BERT with an average of 4 percent on the benchmark dataset.

Social media classification tasks (e.g., tweet sentiment analysis, tweet stance detection) are challenging because social media posts are typically short, informal, and ambiguous. Thus, training on tweets is challenging and demands large-scale human-annotated labels, which are time-consuming and costly to obtain. In this paper, we find that providing hashtags to social media tweets can help alleviate this issue because hashtags can enrich short and ambiguous tweets in terms of various information, such as topic, sentiment, and stance. This motivates us to propose a novel Hashtag-guided Tweet Classification model (HashTation), which automatically generates meaningful hashtags for the input tweet to provide useful auxiliary signals for tweet classification. To generate high-quality and insightful hashtags, our hashtag generation model retrieves and encodes the post-level and entity-level information across the whole corpus. Experiments show that HashTation achieves significant improvements on seven low-resource tweet classification tasks, in which only a limited amount of training data is provided, showing that automatically enriching tweets with model-generated hashtags could significantly reduce the demand for large-scale human-labeled data. Further analysis demonstrates that HashTation is able to generate high-quality hashtags that are consistent with the tweets and their labels. The code is available at https://github.com/shizhediao/HashTation.

As massive conversation examples are easily accessible on the Internet, we are now able to organize large-scale conversation corpora to build chatbots in a data-driven manner. Multi-modal social chatbots produce conversational utterances according to both textual utterances and vision signals. Due to the difficulty of bridging different modalities, the dialogue generation model of chatbots falls into local minima that only capture the mapping between textual input and textual output, as a result, it almost ignores the non-textual signals. Further, similar to the dialogue model with plain text as input and output, the generated responses from multi-modal dialogue also lack diversity and informativeness. In this paper, to address the above issues, we propose a Multi-View Meta-Learning (MultiVML) algorithm that groups samples in multiple views and customizes generation models to different groups. We employ a multi-view clustering to group the training samples so as to attend more to the unique information in non-textual modality. Tailoring different sets of model parameters for each group boosts the genereation diversity via meta-learning. We evaluate MultiVML on two variants of the OpenViDial benchmark datasets. The experiments show that our model not only better explore the information from multiple modalities, but also excels baselines in both quality and diversity.

Yingxiu Zhao, Zhiliang Tian, Huaxiu Yao, Yinhe Zheng, Dongkyu Lee, Yiping Song, Jian Sun, Nevin Zhang. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

In knowledge distillation, a student model is trained with supervisions from both knowledge from a teacher and observations drawn from a training data distribution.Knowledge of a teacher is considered a subject that holds interclass relations which send a meaningful supervision to a student; hence, much effort has been put to find such knowledge to be distilled.In this paper, we explore a question that has been given little attention: "when to distill such knowledge."The question is answered in our work with the concept of model calibration; we view a teacher model not only as a source of knowledge but also as a gauge to detect miscalibration of a student.This simple and yet novel view leads to a hard gate knowledge distillation scheme that switches between learning from a teacher model and training data.We verify the gating mechanism in the context of natural language generation at both the token-level and the sentence-level.Empirical comparisons with strong baselines show that hard gate knowledge distillation not only improves model generalization, but also significantly lowers model calibration error.

Lifelong learning (LL) is vital for advanced task-oriented dialogue (ToD) systems.To address the catastrophic forgetting issue of LL, generative replay methods are widely employed to consolidate past knowledge with generated pseudo samples.However, most existing generative replay methods use only a single taskspecific token to control their models.This scheme is usually not strong enough to constrain the generative model due to insufficient information involved.In this paper, we propose a novel method, prompt conditioned VAE for lifelong learning (PCLL), to enhance generative replay by incorporating tasks' statistics.PCLL captures task-specific distributions with a conditional variational autoencoder, conditioned on natural language prompts to guide the pseudo-sample generation.Moreover, it leverages a distillation process to further consolidate past knowledge by alleviating the noise in pseudo samples.Experiments on natural language understanding tasks of ToD systems demonstrate that PCLL significantly outperforms competitive baselines in building lifelong learning models.We release the code and data at GitHub.

Conditional sentence generation aims to generate proper target sentences with the given condition, and has shown great promise in many text generation applications such as dialogue systems and poetry generation. The ensemble of retrieval and generation-based models retrieve texts according to the input condition to assist the generation-based model. Those approaches obtain great performance tasks as they can absorb both merits to generate informative and coherent sentences. However, the input condition and its retrieved results are usually not highly consistent due to the quality of retrieval. It leads to a retrieval bias between the condition and its retrieved result, and then text generation augmented by such results becomes unreliable. To fix this issue, we propose RBAEM, a Retrieval Bias Aware Ensemble Model. RBAEM employs two CVAEs (Conditional variational Auto-encoder) to represent the retrieved target and the ground truth with latent vectors, and then diminishes the bias by decreasing the distance of two corresponding distributions. The extensive experiments on two tasks show that the proposed methods excel the existing state-of-the-art generation models.

Protecting the privacy of user data is crucial for text generation models, which can leak sensitive information during generation. Differentially private (DP) learning methods provide guarantees against identifying the existence of a training sample from model outputs. PATE is a recent DP learning algorithm that achieves high utility with strong privacy protection on training samples. However, text generation models output tokens sequentially in a large output space; the classic PATE algorithm is not customized for this setting. Furthermore, PATE works well to protect sample-level privacy, but is not designed to protect phrases in samples. In this paper, we propose SeqPATE, an extension of PATE to text generation that protects the privacy of individual training samples and sensitive phrases in training data. To adapt PATE to text generation, we generate pseudo-contexts and reduce the sequence generation problem to a next-word prediction problem. To handle the large output space, we propose a candidate filtering strategy to dynamically reduce the output space, and refine the teacher aggregation of PATE to avoid low agreement due to voting for a large number of candidates. To further reduce privacy losses, we use knowledge distillation to reduce the number of teacher queries. The experiments verify the effectiveness of SeqPATE in protecting both training samples and sensitive phrases.

Personalized conversation models (PCMs) generate responses according to speaker preferences. Existing personalized conversation tasks typically require models to extract speaker preferences from user descriptions or their conversation histories, which are scarce for newcomers and inactive users. In this paper, we propose a few-shot personalized conversation task with an auxiliary social network. The task requires models to generate personalized responses for a speaker given a few conversations from the speaker and a social network. Existing methods are mainly designed to incorporate descriptions or conversation histories. Those methods can hardly model speakers with so few conversations or connections between speakers. To better cater for newcomers with few resources, we propose a personalized conversation model (PCM) that learns to adapt to new speakers as well as enabling new speakers to learn from resource-rich speakers. Particularly, based on a meta-learning based PCM, we propose a task aggregator (TA) to collect other speakers' information from the social network. The TA provides prior knowledge of the new speaker in its meta-learning. Experimental results show our methods outperform all baselines in appropriateness, diversity, and consistency with speakers.

Dongkyu Lee, Zhiliang Tian, Lanqing Xue, Nevin L. Zhang. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Dialogue generation models face the challenge of producing generic and repetitive responses. Unlike previous augmentation methods that mostly focus on token manipulation and ignore the essential variety within a single sample using hard labels, we propose to promote the generation diversity of the neural dialogue models via soft embedding augmentation along with soft labels in this paper. Particularly, we select some key input tokens and fuse their embeddings together with embeddings from their semantic-neighbor tokens. The new embeddings serve as the input of the model to replace the original one. Besides, soft labels are used in loss calculation, resulting in multi-target supervision for a given input. Our experimental results on two datasets illustrate that our proposed method is capable of generating more diverse responses than raw models while remains a similar n-gram accuracy that ensures the quality of generated responses.

Neural conversation models are known to generate appropriate but non-informative responses in general.A scenario where informativeness can be significantly enhanced is Conversing by Reading (CbR), where conversations take place with respect to a given external document.In previous work, the external document is utilized by (1) creating a contextaware document memory that integrates information from the document and the conversational context, and then (2) generating responses referring to the memory.In this paper, we propose to create the document memory with some anticipated responses in mind.This is achieved using a teacher-student framework.The teacher is given the external document, the context, and the ground-truth response, and learns how to build a response-aware document memory from three sources of information.The student learns to construct a response-anticipated document memory from the first two sources, and the teacher's insight on memory creation.Empirical results show that our model outperforms the previous stateof-the-art for the CbR task.

Neural generative models for open-domain chit-chat conversations have become an active area of research in recent years.A critical issue with most existing generative models is that the generated responses lack informativeness and diversity.A few researchers attempt to leverage the results of retrieval models to strengthen the generative models, but these models are limited by the quality of the retrieval results.In this work, we propose a memory-augmented generative model, which learns to abstract from the training corpus and saves the useful information to the memory to assist the response generation.Our model clusters query-response samples, extracts characteristics of each cluster, and learns to utilize these characteristics for response generation.Experimental results show that our model outperforms other competitive baselines.