Yuhan Chen, Zizhuo Shen, Miaomiao Cheng, Xu Han, Jiefu Gong, Shijin Wang, Wei Song. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Prompt tuning vision-language models like CLIP has shown great potential in learning transferable representations for various downstream tasks. The main issue is how to mitigate the over-fitting problem on downstream tasks with limited training samples. While knowledge-guided context optimization has been proposed by constructing consistency constraints to handle catastrophic forgetting in the pre-trained backbone, it also introduces a bias toward pre-training. This paper proposes a novel and simple Divergence-enhanced Knowledge-guided Prompt Tuning (DeKg) method to address this issue. The key insight is that the bias toward pre-training can be alleviated by encouraging the independence between the learnable and the crafted prompt. Specifically, DeKg employs the Hilbert-Schmidt Independence Criterion (HSIC) to regularize the learnable prompts, thereby reducing their dependence on prior general knowledge, and enabling divergence induced by target knowledge. Comprehensive evaluations demonstrate that DeKg serves as a plug-and-play module can seamlessly integrate with existing knowledge-guided context optimization methods and achieves superior performance in three challenging benchmarks. We make our code available at https://github.com/cnunlp/DeKg.

Large language models (LLMs) have demonstrated remarkable performance in various natural language processing tasks. However, they still face significant challenges in automated reasoning, particularly in scenarios involving multi-step reasoning. In this paper, we focus on the logical reasoning problem. The main task is to answer a question based on a set of available facts and rules. A lot of work has focused on guiding LLMs to think logically by generating reasoning paths, ignoring the structure among available facts. In this paper, we propose a simple approach MindMap by introducing evidence chains for supporting reasoning. An evidence chain refers to a set of facts that involve the same subject. In this way, we can organize related facts together to avoid missing important information. MindMap can be integrated with existing reasoning framework, such as Chain-of-Thought (CoT) and Selection-Inference (SI), by letting the model select relevant evidence chains instead of independent facts. The experimental results on the bAbI and ProofWriterOWA datasets demonstrate the effectiveness of MindMap.It can significantly improve CoT and SI, especially in multi-step reasoning tasks.

Cross-modal retrieval is an important yet challenging task due to the semantic discrepancy between visual content and language. To measure the correlation between images and text, most existing research mainly focuses on learning global or local correspondence, failing to explore fine-grained local-global alignment. To infer more accurate similarity scores, we introduce a novel High Order Semantic Alignment (HOSA) model that can provide complementary and comprehensive semantic clues. Specifically, to jointly learn global and local alignment and emphasize local-global interaction, we employ tensor-product (t-product) operation to reconstruct one modal’s representation based on another modal’s information in a common semantic space. Such a cross-modal reconstruction strategy would significantly enhance inter-modal correlation learning in a fine-grained manner. Extensive experiments on two benchmark datasets validate that our model significantly outperforms several state-of-the-art baselines, especially in retrieving the most relevant results.

This paper proposes to adapt self-attention to discourse level for modeling discourse elements in argumentative student essays.Specifically, we focus on two issues.First, we propose structural sentence positional encodings to explicitly represent sentence positions.Second, we propose to use inter-sentence attentions to capture sentence interactions and enhance sentence representation.We conduct experiments on two datasets: a Chinese dataset and an English dataset.We find that (i) sentence positional encodings can lead to a large improvement for identifying discourse elements; (ii) a structural relative positional encoding of sentences shows to be most effective; (iii) inter-sentence attention vectors are useful as a kind of sentence representation for identifying discourse elements.

This paper proposes a pre-training based automated Chinese essay scoring method.The method involves three components: weakly supervised pre-training, supervised crossprompt fine-tuning and supervised targetprompt fine-tuning.An essay scorer is first pretrained on a large essay dataset covering diverse topics and with coarse ratings, i.e., good and poor, which are used as a kind of weak supervision.The pre-trained essay scorer would be further fine-tuned on previously rated essays from existing prompts, which have the same score range with the target prompt and provide extra supervision.At last, the scorer is fine-tuned on the target-prompt training data.The evaluation on four prompts shows that this method can improve a state-of-the-art neural essay scorer in terms of effectiveness and domain adaptation ability, while in-depth analysis also reveals its limitations.

Cross-modal retrieval, given the data of one specific modality as a query, aims to search the relevant data in other modalities. Recently, cross-modal hashing has attracted much attention due to its high efficiency and low storage cost. Its main idea is to approximate the cross-modality similarity via binary codes. This kind of method works well when the cross-modal data is completely observed. However, the real-world application usually avoids this situation, where part of the information is unobserved in some modality. Such partial multimodal data will result in the lack of pairwise information and then destroy the performance of cross-modal hashing. In this paper, we proposed a novel unsupervised cross-modal hashing approach, named as Unsupervised Deep Imputed Hashing (UDIH). It is a two-stage learning strategy. Firstly, the unobserved pairwise data is imputed by the proposed generators. Then a neural network with weighted triplet loss is applied on the correlation graph to learn the hashing code in the Hamming space for each modality, where the correlation graph is constructed with the aid of augmented data. UDIH has the ability to preserve the semantic consistency and difference among data objects. The extensive experimental results have shown that the proposed method outperforms the state-of-the-art methods on two benchmark datasets (MIRFlickr and NUS-WIDE). The source code could be available at https://github.com/AkChen/UDIH.

With the quick development of social websites, there are more opportunities to have different media types (such as text, image, video, etc.) describing the same topic from large-scale heterogeneous data sources. To efficiently identify the inter-media correlations for multimedia retrieval, unsupervised cross-modal hashing (UCMH) has gained increased interest due to the significant reduction in computation and storage. However, most UCMH methods assume that the data from different modalities are well paired. As a result, existing UCMH methods may not achieve satisfactory performance when partially paired data are given only. In this article, we propose a new-type of UCMH method called robust unsupervised cross-modal hashing ( RUCMH ). The major contribution lies in jointly learning modal-specific hash function, exploring the correlations among modalities with partial or even without any pairwise correspondence, and preserving the information of original features as much as possible. The learning process can be modeled via a joint minimization problem, and the corresponding optimization algorithm is presented. A series of experiments is conducted on four real-world datasets (Wiki, MIRFlickr, NUS-WIDE, and MS-COCO). The results demonstrate that RUCMH can significantly outperform the state-of-the-art unsupervised cross-modal hashing methods, especially for the partially paired case, which validates the effectiveness of RUCMH.

As a new occurring unsupervised method, multi-view clustering offers a good way to investigate the hidden structure from multi-view data and attracts extensive attention in the community of machine learning and data mining. One popular approach is to identify a common latent subspace for capturing the multi-view information. However, these methods are still limited due to the unsupervised learning process and suffer from considerable noisy information from different views. To address this issue, we present a novel multi-view subspace clustering method, named self-supervised deep multi-view subspace clustering (\textbf{S2DMVSC}). It seamlessly integrates spectral clustering and affinity learning into a deep learning framework. \textbf{S2DMVSC} has two main merits. One is that the clustering results can be sufficiently exploited to supervise the latent representation learning for each view (via a classification loss) and the common latent subspace learning (via a spectral clustering loss) for multiple views. The other is that the affinity matrix among data objects is automatically computed according to the high-level and cluster-driven representation. Experiments on two scenarios, including original features and multiple hand-crafted features, demonstrate the superiority of the proposed approach over the state-of-the-art baselines.

With the development of data collection techniques, multi-view clustering becomes an emerging research direction to improve the clustering performance. Existing work has shown that leveraging multi-view information is able to provide a rich and comprehensive description. One of the core problems is how to sufficiently represent multi-view data in the analysis. In this paper, we introduce a tensor-based Representation Learning method for Multi-view Clustering (tRLMvC) that can unify heterogenous and high-dimensional multi-view feature spaces to a low-dimensional shared latent feature space and improve multi-view clustering performance. To sufficiently capture plenty multi-view information, tRLMvC represents multi-view data as a third-order tensor, expresses each tensorial data point as a sparse t-linear combination of all data points with t-product, and constructs a self-expressive tensor through reconstruction coefficients. The low-dimensional multi-view data representation in the shared latent feature space can be obtained via Tucker decomposition on the self-expressive tensor. These two parts are iteratively performed so that the interaction between selfexpressive tensor learning and its factorization can be enhanced and the new representation can be effectively generated for clustering purpose. We conduct extensive experiments on eight multi-view data sets and compare the proposed model with the state-of-the-art methods. Experimental results have shown that tRLMvC outperforms the baselines in terms of various evaluation metrics.

The spatial and temporal characters of NO2over Yangtze River Delta (YZD) are analyzed using monthly averaged tropospheric & column densities from GOME (1996-2002) and SCIAMACHY (2003-2010) measurements. In addition, with the land cover map of YZD, the characters of NO2concentration over different ecosystem have been analyzed. The results indicated that the tropospheric NO2column densities in YZD increased distinctly from 1996 to 2010 and it showed distinct regional and seasonal variation. It has the highest concentration in winter while lowest in summer. Under the influence of the urbanization process, the highest NO2column densities appeared at cities and metropolis while the lowest column densities appeared in forest. The sharp increasing of tropospheric NO2column appeared in water bodies, followed urban land and woodland. Furthermore, we compared with the NO2column and the mass concentration of N dry deposition from the control experiments, they performed significant correlation and the high R value (R2=0.86, P<;0.001).