Wang Cai, Yilin Wen, Jinchang Hou, Du Su, Guoqiu Wang, Zhonghou Lv, Chenfu Bao, Yunfang Wu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Reinforcement learning with verifiable rewards (RLVR) has significantly boosted the reasoning capability of language models (LMs).However, existing RLVR approaches train LMs based on their own on-policy responses and are constrained by the initial capability of LMs, thus prone to exploration stagnation, in which LMs fail to solve more training problems and cannot further learn from the training data.Some approaches try to address this by leveraging off-policy solutions to training problems, but rely on external expert guidance that is limited in availability and scalability.In this work, we propose LTE (Learning to reason from Trial and Error), an approach that hints LMs with their previously self-made mistakes, not requiring any external expert guidance.Experiments validate the effectiveness of LTE, which outperforms the normal group relative policy optimization (GRPO) by 5.02 in Pass@1 and 9.96 in Pass@k on average across six mathematical reasoning benchmarks for Qwen3-8B-Base and even performs better than methods that require external guidance.Further analysis confirms that LTE successfully mitigates exploration stagnation and enhances both exploitation and exploration during training.Our code is available at https://github.com/ JamyDon/LTE.17 × 13 = ?17 × 13 = 170 + 41 = 211.17 × 13 = 260 -39 = 231.17 × 13 = 100 + 21 = 121.17 × 13 = ?Hint: possible incorrect answers include 211, 231, 121 … 17 × 13 = 170 + 41 = 211.Wait, 211 is wrong.Must be 51! 17 × 13 = 170 + 51 = 221.17 × 13 = 260 -39 = 231.But the hint says 231 is incorrect … Aha, 260 -39 is 221! 17 × 13 = 100 + 21 = 121.But the hint refutes 121 … Oh, 17 × 13 = 100 + 70 + 30 + 21 = 221.Normal rollouts "die" here.LTE "revives" with hints!

Knowledge editing aims to modify outdated knowledge in language models efficiently while retaining their original capabilities.Mainstream datasets for knowledge editing are predominantly static and fail to keep in pace with the evolving real-world knowledge.In this work, we introduce CRAFT, an everevolving real-world dataset for knowledge editing.It evaluates models on temporal locality, common-sense locality, composite portability and alias portability, providing a comprehensive and challenging evaluation for knowledge editing, on which previous methods hardly achieve balanced performance.Towards flexible real-time knowledge editing, we propose KEDAS, a novel paradigm of knowledge editing alignment featuring diverse edit augmentation and self-adaptive post-alignment inference, exhibiting significant performance gain on both CRAFT and traditional datasets compared to previous methods.We hope this work may serve as a catalyst for shifting the focus of knowledge editing from static update to dynamic evolution.1

In-Context derived Vector (ICV) methods extract task-relevant representations from large language models (LLMs) and reinject them during inference, achieving comparable performance to few-shot In-Context Learning (ICL) without repeated demonstration processing.However, existing ICV methods remain sensitive to ICL-specific factors, often use coarse or semantically fragmented representations as the source of the vector, and rely on heuristicbased injection positions, limiting their applicability.To address these issues, we propose Dynamic Vector (DyVec), which incorporates an Exhaustive Query Rotation (EQR) strategy to extract robust semantically aggregated latent representations by mitigating variance introduced by ICL.It then applies Dynamic Latent Segmentation and Injection to adaptively partition representations based on task complexity and leverages REINFORCE-based optimization to learn optimal injection positions for each segment.Experiments results show that DyVec outperforms few-shot ICL, LoRA, and prior ICV baselines.Further analysis highlights the effectiveness of dynamically segmenting and injecting semantically aggregated latent representations.DyVec provides a lightweight and data-efficient solution for inference-time task adaptation.

Recent advances in cross-prompt automated essay scoring typically train models jointly on all available source domains, often requiring simultaneous access to unlabeled target domain samples.However, using all sources can lead to suboptimal transfer and high computational cost.Moreover, repeatedly accessing the source essays for continual adaptation raises privacy concerns.We propose a source-free adaptation approach that selectively merges the parameters of individually trained source models without further access to the source datasets.In particular, we mix the task vectors-the parameter updates from finetuning-via a weighted sum to efficiently simulate selective joint-training.We use Bayesian optimization to determine the mixing weights using our proposed Prior-encoded Information Maximization (PIM), an unsupervised objective which promotes score discriminability by leveraging useful priors pre-computed from the sources.Experimental results with LLMs on in-dataset and cross-dataset adaptation show that our method (1) consistently outperforms joint-training on all sources, (2) maintains superior robustness compared to other merging methods, (3) excels under severe distribution shifts where recent leading cross-prompt methods struggle, all while retaining computational efficiency. 1

Sanwoo Lee, Jiahao Liu, Qifan Wang, Jingang Wang, Xunliang Cai, Yunfang Wu. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Deep multimodal semantic understanding that goes beyond the mere superficial content relation mining has received increasing attention in the realm of artificial intelligence. The challenges of collecting and annotating high-quality multi-modal data have underscored the significance of few-shot learning. In this paper, we focus on two critical tasks under this context: few-shot multi-modal sarcasm detection (MSD) and multi-modal sentiment analysis (MSA). To address them, we propose Mixture-of-Prompt-Experts with Block-Aware Prompt Fusion (MoPE-BAF), a novel multi-modal soft prompt framework based on the unified vision-language model (VLM). Specifically, we design three experts of soft prompts: a text prompt and an image prompt that extract modality-specific features to enrich the single-modal representation, and a unified prompt to assist multi-modal interaction. Additionally, we reorganize Transformer layers into several blocks and introduce cross-modal prompt attention between adjacent blocks, which smoothens the transition from single-modal representation to multi-modal fusion. On both MSD and MSA datasets in few-shot setting, our proposed model not only surpasses the 8.2B model InstructBLIP with merely 2% parameters (150M), but also significantly outperforms other widely-used prompt methods on VLMs or task-specific methods.

Multi-modal semantic understanding requires integrating information from different modalities to extract users’ real intention behind words. Most previous work applies a dual-encoder structure to separately encode image and text, but fails to learn cross-modal feature alignment, making it hard to achieve cross-modal deep information interaction. This paper proposes a novel CLIP-guided contrastive-learning-based architecture to perform multi-modal feature alignment, which projects the features derived from different modalities into a unified deep space. On multi-modal sarcasm detection (MMSD) and multi-modal sentiment analysis (MMSA) tasks, the experimental results show that our proposed model significantly outperforms several baselines, and our feature alignment strategy brings obvious performance gain over models with different aggregating methods and models even enriched with knowledge. More importantly, our model is simple to implement without using task-specific external knowledge, and thus can easily migrate to other multi-modal tasks. Our source codes are available at https://github.com/ChangKe123/CLFA.

In-context learning (ICL) greatly improves the performance of large language models (LLMs) on various down-stream tasks, where the improvement highly depends on the quality of demonstrations.In this work, we introduce syntactic knowledge to select better in-context examples for machine translation (MT).We propose a new strategy, namely Syntax-augmented COverage-based In-context example selection (SCOI), leveraging the deep syntactic structure beyond conventional word matching.Specifically, we measure the set-level syntactic coverage by computing the coverage of polynomial terms with the help of a simplified treeto-polynomial algorithm, and lexical coverage using word overlap.Furthermore, we devise an alternate selection approach to combine both coverage measures, taking advantage of syntactic and lexical information.We conduct experiments with two multi-lingual LLMs on six translation directions.Empirical results show that our proposed SCOI obtains the highest average COMET score among all learning-free methods, indicating that combining syntactic and lexical coverage successfully helps to select better in-context examples for MT.Our code is available at https://github.com/ JamyDon/SCOI.

Chenming Tang, Fanyi Qu, Yunfang Wu. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Ziyang Wang, Sanwoo Lee, Hsiu-Yuan Huang, Yunfang Wu. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

The conventional wisdom behind learning deep classification models is to focus on bad-classified examples and ignore well-classified examples that are far from the decision boundary. For instance, when training with cross-entropy loss, examples with higher likelihoods (i.e., well-classified examples) contribute smaller gradients in back-propagation. However, we theoretically show that this common practice hinders representation learning, energy optimization, and margin growth. To counteract this deficiency, we propose to reward well-classified examples with additive bonuses to revive their contribution to the learning process. This counterexample theoretically addresses these three issues. We empirically support this claim by directly verifying the theoretical results or significant performance improvement with our counterexample on diverse tasks, including image classification, graph classification, and machine translation. Furthermore, this paper shows that we can deal with complex scenarios, such as imbalanced classification, OOD detection, and applications under adversarial attacks because our idea can solve these three issues. Code is available at https://github.com/lancopku/well-classified-examples-are-underestimated.

Today the pre-trained language models achieve great success for question generation (QG) task and significantly outperform traditional sequence-to-sequence approaches. However, the pre-trained models treat the input passage as a flat sequence and are thus not aware of the text structure of input passage. For QG task, we model text structure as answer position and syntactic dependency, and propose answer localness modeling and syntactic mask attention to address these limitations. Specially, we present localness modeling with a Gaussian bias to enable the model to focus on answer-surrounded context, and propose a mask attention mechanism to make the syntactic structure of input passage accessible in question generation process. Experiments on SQuAD dataset show that our proposed two modules improve performance over the strong pre-trained model ProphetNet, and combing them together achieves very competitive results with the state-of-the-art pre-trained model.

We introduce a novel position offset label prediction subtask to the encoder-decoder architecture for grammatical error correction (GEC) task. To keep the meaning of the input sentence unchanged, only a few words should be inserted or deleted during correction, and most of tokens in the erroneous sentence appear in the paired correct sentence with limited position movement. Inspired by this observation, we design an auxiliary task to predict position offset label (POL) of tokens, which is naturally capable of integrating different correction editing operations into a unified framework. Based on the predicted POL, we further propose a new copy mechanism (P-copy) to replace the vanilla copy module. Experimental results on Chinese, English and Japanese datasets demonstrate that our proposed POL-Pc framework obviously improves the performance of baseline models. Moreover, our model yields consistent performance gain over various data augmentation methods. Especially, after incorporating synthetic data, our model achieves a 38.95 F-0.5 score on Chinese GEC dataset, which outperforms the previous state-of-the-art by a wide margin of 1.98 points.

Automatic medical question summarization can significantly help the system to understand consumer health questions and retrieve correct answers. The Seq2Seq model based on maximum likelihood estimation (MLE) has been applied in this task, which faces two general problems: the model can not capture well question focus and and the traditional MLE strategy lacks the ability to understand sentence-level semantics. To alleviate these problems, we propose a novel question focus-driven contrastive learning framework (QFCL). Specially, we propose an easy and effective approach to generate hard negative samples based on the question focus, and exploit contrastive learning at both encoder and decoder to obtain better sentence level representations. On three medical benchmark datasets, our proposed model achieves new state-of-the-art results, and obtains a performance gain of 5.33, 12.85 and 3.81 points over the baseline BART model on three datasets respectively. Further human judgement and detailed analysis prove that our QFCL model learns better sentence representations with the ability to distinguish different sentence meanings, and generates high-quality summaries by capturing question focus.

For readability assessment, traditional methods mainly employ machine learning classifiers with hundreds of linguistic features.Although the deep learning model has become the prominent approach for almost all NLP tasks, it is less explored for readability assessment.In this paper, we propose a BERT-based model with feature projection and length-balanced loss (BERT-FP-LBL) for readability assessment.Specially, we present a new difficulty knowledge guided semi-supervised method to extract topic features to complement the traditional linguistic features.From the linguistic features, we employ projection filtering to extract orthogonal features to supplement BERT representations.Furthermore, we design a new length-balanced loss to handle the greatly varying length distribution of data.Our model achieves state-of-theart performances on two English benchmark datasets and one dataset of Chinese textbooks, and also achieves the near-perfect accuracy of 99% on one English dataset.Moreover, our proposed model obtains comparable results with human experts in consistency test.

Knowledge facts are typically represented by relational triples, while we observe that some commonsense facts are represented by triples whose forms are inconsistent with the corresponding language expressions. For commonsense mining tasks, this inconsistency raises a challenge for the prevailing methods using pre-trained language models that learn the expression of language. However, there are few studies which focus on this inconsistency issue. To fill this empty, in this paper, we term the commonsense knowledge whose triple form is heavily inconsistent with the language expression asdeep commonsense knowledgeand first conduct extensive exploratory experiments to study deep commonsense knowledge. We show that deep commonsense knowledge occupies a significant part of commonsense knowledge, while the conventional methods based on pre-trained language models fail to capture it effectively. We further propose a novel method to mine the deep commonsense knowledge from raw text that is exactly language expression, alleviating the reliance of conventional methods on the triple representation form. Experiments demonstrate that our proposed method substantially improves the performance in mining deep commonsense knowledge.

Question Generation (QG) is an essential component of the automatic intelligent tutoring systems, which aims to generate high-quality questions for facilitating the reading practice and assessments. However, existing QG technologies encounter several key issues concerning the biased and unnatural language sources of datasets which are mainly obtained from the Web (e.g. SQuAD). In this paper, we propose an innovative Examination-type Question Generation approach (EQG-RACE) to generate exam-like questions based on a dataset extracted from RACE. Two main strategies are employed in EQG-RACE for dealing with discrete answer information and reasoning among long contexts. A Rough Answer and Key Sentence Tagging scheme is utilized to enhance the representations of input. An Answer-guided Graph Convolutional Network (AG-GCN) is designed to capture structure information in revealing the inter-sentences and intra-sentence relations. Experimental results show a state-of-the-art performance of EQG-RACE, which is apparently superior to the baselines. In addition, our work has established a new QG prototype with a reshaped dataset and QG method, which provides an important benchmark for related research in future work. We will make our data and code publicly available for further research.

Online advertising is an important revenue source for many IT companies. In the search advertising scenario, advertisement text that meets the need of the search query would be more attractive to the user. However, the manual creation of query-variant advertisement texts for massive items is expensive. Traditional text generation methods tend to focus on the general searching needs with high frequency while ignoring the diverse personalized searching needs with low frequency. In this paper, we propose the query-variant advertisement text generation task that aims to generate candidate advertisement texts for different web search queries with various needs based on queries and item keywords. To solve the problem of ignoring low-frequency needs, we propose a dynamic association mechanism to expand the receptive field based on external knowledge, which can obtain associated words to be added to the input. These associated words can serve as bridges to transfer the ability of the model from the familiar high-frequency words to the unfamiliar low-frequency words. With association, the model can make use of various personalized needs in queries and generate query-variant advertisement texts. Both automatic and human evaluations show that our model can generate more attractive advertisement text than baselines.

Generating high quality question-answer pairs is a hard but meaningful task.Although previous works have achieved great results on answer-aware question generation, it is difficult to apply them into practical application in the education field.This paper for the first time addresses the question-answer pair generation task on the real-world examination data, and proposes a new unified framework on RACE.To capture the important information of the input passage we first automatically generate (rather than extracting) keyphrases, thus this task is reduced to keyphrase-question-answer triplet joint generation.Accordingly, we propose a multi-agent communication model to generate and optimize the question and keyphrases iteratively, and then apply the generated question and keyphrases to guide the generation of answers.To establish a solid benchmark, we build our model on the strong generative pre-training model.Experimental results show that our model makes great breakthroughs in the question-answer pair generation task.Moreover, we make a comprehensive analysis on our model, suggesting new directions for this challenging task.

Trading volume movement prediction is the key in a variety of financial applications. Despite its importance, there is few research on this topic because of its requirement for comprehensive understanding of information from different sources. For instance, the relation between multiple stocks, recent transaction data and suddenly released events are all essential for understanding trading market. However, most of the previous methods only take the fluctuation information of the past few weeks into consideration, thus yielding poor performance. To handle this issue, we propose a graph-based approach that can incorporate multi-view information, i.e., long-term stock trend, short-term fluctuation and sudden events information jointly into a temporal heterogeneous graph. Besides, our method is equipped with deep canonical analysis to highlight the correlations between different perspectives of fluctuation for better prediction. Experiment results show that our method outperforms strong baselines by a large margin.

Procedural text describes dynamic state changes during a step-by-step natural process (e.g., photosynthesis). In this work, we focus on the task of procedural text understanding, which aims to comprehend such documents and track entities’ states and locations during a process. Although recent approaches have achieved substantial progress, their results are far behind human performance. Two challenges, the difficulty of commonsense reasoning and data insufficiency, still remain unsolved, which require the incorporation of external knowledge bases. Previous works on external knowledge injection usually rely on noisy web mining tools and heuristic rules with limited applicable scenarios. In this paper, we propose a novel KnOwledge-Aware proceduraL text understAnding (KoaLa) model, which effectively leverages multiple forms of external knowledge in this task. Specifically, we retrieve informative knowledge triples from ConceptNet and perform knowledge-aware reasoning while tracking the entities. Besides, we employ a multi-stage training schema which fine-tunes the BERT model over unlabeled data collected from Wikipedia before further fine-tuning it on the final model. Experimental results on two procedural text datasets, ProPara and Recipes, verify the effectiveness of the proposed methods, in which our model achieves state-of-the-art performance in comparison to various baselines.1

In reading comprehension, generating sentence-level distractors is a significant task, which requires a deep understanding of the article and question. The traditional entity-centered methods can only generate word-level or phrase-level distractors. Although recently proposed neural-based methods like sequence-to-sequence (Seq2Seq) model show great potential in generating creative text, the previous neural methods for distractor generation ignore two important aspects. First, they didn't model the interactions between the article and question, making the generated distractors tend to be too general or not relevant to question context. Second, they didn't emphasize the relationship between the distractor and article, making the generated distractors not semantically relevant to the article and thus fail to form a set of meaningful options. To solve the first problem, we propose a co-attention enhanced hierarchical architecture to better capture the interactions between the article and question, thus guide the decoder to generate more coherent distractors. To alleviate the second problem, we add an additional semantic similarity loss to push the generated distractors more relevant to the article. Experimental results show that our model outperforms several strong baselines on automatic metrics, achieving state-of-the-art performance. Further human evaluation indicates that our generated distractors are more coherent and more educative compared with those distractors generated by baselines.

Given a sentence and its relevant answer, how to ask good questions is a challenging task, which has many real applications.Inspired by human's paraphrasing capability to ask questions of the same meaning but with diverse expressions, we propose to incorporate paraphrase knowledge into question generation(QG) to generate human-like questions.Specifically, we present a two-hand hybrid model leveraging a self-built paraphrase resource, which is automatically conducted by a simple back-translation method.On the one hand, we conduct multi-task learning with sentence-level paraphrase generation (PG) as an auxiliary task to supplement paraphrase knowledge to the task-share encoder.On the other hand, we adopt a new loss function for diversity training to introduce more question patterns to QG. Extensive experimental results show that our proposed model obtains obvious performance gain over several strong baselines, and further human evaluation validates that our model can ask questions of high quality by leveraging paraphrase knowledge.

Automatic article commenting is helpful in encouraging user engagement and interaction on online news platforms.However, the news documents are usually too long for traditional encoder-decoder based models, which often results in general and irrelevant comments.In this paper, we propose to generate comments with a graph-to-sequence model that models the input news as a topic interaction graph.By organizing the article into graph structure, our model can better understand the internal structure of the article and the connection between topics, which makes it better able to understand the story.We collect and release a large scale news-comment corpus from a popular Chinese online news platform Tencent Kuaibao. 1 Extensive experiment results show that our model can generate much more coherent and informative comments compared with several strong baseline models.2

Wenjie Zhou, Minghua Zhang, Yunfang Wu. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Wenjie Zhou, Minghua Zhang, Yunfang Wu. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Document-based Question Answering (DBQA) in Natural Language Processing (NLP) is important but difficult because of the long document and the complex question. Most of previous deep learning methods mainly focus on the similarity computation between two sentences. However, DBQA stems from the reading comprehension in some degree, which is originally used to train and test people's ability of reading and logical thinking. Inspired by the strategy of doing reading comprehension tests, we propose a unified model based on the human-like reading strategy. The unified model contains three major encoding layers that are consistent to different steps of the reading strategy, including the basic encoder, combined encoder and hierarchical encoder. We conduct extensive experiments on both the English WikiQA dataset and the Chinese dataset, and the experimental results show that our unified model is effective and yields state-of-the-art results on WikiQA dataset.

Question retrieval is a crucial subtask for community question answering. Previous research focus on supervised models which depend heavily on training data and manual feature engineering. In this paper, we propose a novel unsupervised framework, namely reduced attentive matching network (RAMN), to compute semantic matching between two questions. Our RAMN integrates together the deep semantic representations, the shallow lexical mismatching information and the initial rank produced by an external search engine. For the first time, we propose attention autoencoders to generate semantic representations of questions. In addition, we employ lexical mismatching to capture surface matching between two questions, which is derived from the importance of each word in a question. We conduct experiments on the open CQA datasets of SemEval-2016 and SemEval-2017. The experimental results show that our unsupervised model obtains comparable performance with the state-of-the-art supervised methods in SemEval-2016 Task 3, and outperforms the best system in SemEval-2017 Task 3 by a wide margin.

In order to learn universal sentence representations, previous methods focus on complex recurrent neural networks or supervised learning.In this paper, we propose a meanmax attention autoencoder (mean-max AAE) within the encoder-decoder framework.Our autoencoder rely entirely on the MultiHead self-attention mechanism to reconstruct the input sequence.In the encoding we propose a mean-max strategy that applies both mean and max pooling operations over the hidden vectors to capture diverse information of the input.To enable the information to steer the reconstruction process dynamically, the decoder performs attention over the mean-max representation.By training our model on a large collection of unlabelled data, we obtain highquality representations of sentences.Experimental results on a broad range of 10 transfer tasks demonstrate that our model outperforms the state-of-the-art unsupervised single methods, including the classical skip-thoughts (Kiros et al., 2015) and the advanced skip-thoughts+LN model (Ba et al., 2016).Furthermore, compared with the traditional recurrent neural network, our mean-max AAE greatly reduce the training time.

In this paper we focus on the problem of dialog act (DA) labelling. This problem has recently attracted a lot of attention as it is an important sub-part of an automatic question answering system, which is currently in great demand. Traditional methods tend to see this problem as a sequence labelling task and deals with it by applying classifiers with rich features. Most of the current neural network models still omit the sequential information in the conversation. Henceforth, we apply a novel multi-level gated recurrent neural network (GRNN) with non-textual information to predict the DA tag. Our model not only utilizes textual information, but also makes use of non-textual and contextual information. In comparison, our model has shown significant improvement over previous works on Switchboard Dialog Act (SWDA) task by over 6%.

We present a multi-sentence question segmentation strategy for community question answering services to alleviate the complexity of long sentences. We develop a complete scheme and make a solution to complex-question segmentation, including a question detector to extract question sentences, a question compression process to remove duplicate information, and a graph model to segment multi-sentence questions. In the graph model, we train a SVM classifier to compute the initial weight and we calculate the authority of a vertex to guide the propagating. The experimental results show that our method gets a good balance between completeness and redundancy of information, and significantly outperforms state-of-the-art methods.