Sequential recommendation has become indispensable in modern digital services. Prevalent recommendation techniques formulate the recommendation task with a language instruction fed into large language models (LLMs) to generate recommendations. However, the implicit interaction scenario of recommendation task cannot provide explicit reasoning supervision to activate LLM's multi-step reasoning capability. Besides, the manner of reasoning for enhancing recommendation is still underexplored. Therefore, we investigate activating multi-step reasoning with users' interactions and propose a multi-step reasoning-enhanced LLM (MSR-Rec), which tightly integrates reasoning with recommendation from designing reasoning chain to reasoning-based recommendation. A task-decomposed reasoning chain is elaborately designed to imitate users' thinking process, seamlessly involving reasoning into recommendation. Following the reasoning chain, MSR-Rec synthesizes reasoning supervision and fine-tunes LLM to adapt for task-specific reasoning. In inference, bidirectional reasoning is implemented from user and item sides, performing a closed-loop reasoning for recommendation. Comprehensive experiments demonstrate that MSR-Rec achieves the state-of-the-art performance in both recommendation quality and reasoning interpretability, advancing the integration of reasoning and recommendation in LLM-based systems.

In the digital age, the overwhelming amount of information necessitates advanced recommendation systems to deliver personalized content. However, these systems face significant challenges, such as sparse user-item interactions and long-tail bias. Recent studies construct structural learning or self-supervised learning on the interaction graph achieving a positive impact on alleviating the problems, but the interaction data itself may be far too little to solve the problems. While knowledge graphs (KGs) offer a promising solution by providing semantic depth to recommendations, their integration often introduces noise from redundant knowledge. Addressing these critical gaps, this study proposes a knowledge-guided interest contrast (KGIC) to enhance recommendations, which innovatively harmonizes collaborative filtering with semantic insights from KG. The KGIC model introduces three key innovations: (1) a knowledge filtering mechanism that selectively leverages interest-relevant signals from the knowledge graph to encode interest and avoid redundant knowledge interference; (2) an adaptive graph augmentation strategy that enhances the interaction graph based on semantic-aware interest propagation and interaction intensity estimation; and (3) a self-supervised contrastive learning task that mitigates long-tail bias and sparsity issues by homogenizing the embedding distribution between augmented views. The extensive evaluation reveals the superiority of KGIC with knowledge filtering and graph augmentation for recommendation.

Natural language-guided drone geo-localization (DGL) provides an intuitive and scalable mode of human-drone interaction for tasks such as search, rescue, and surveillance. Recent Vision-Language Models (VLMs) can learn semantic correspondences between text and images during fine-tuning. However, their performance in DGL tasks remains constrained, as complex instructions and cluttered scenes often cause semantic dilution and granularity mismatch, leading to weak cross-modal alignment. Consequently, the models struggle with ambiguous targets and suffer from reduced localization accuracy. To address these challenges, we propose SAA-DGL, a framework for interpretable language-guided Drone Geo-Localization that enriches Semantic Attribute Alignment (SAA) with large language models (LLMs). It introduces two parameter-free cross-modal fusion modules: (1) the LLM-driven Cross-modal Semantic Attribute Enrichment (LCSAE) module, which extracts fine-grained attributes (e.g., color, shape, position) from text and embeds them into visual features as explicit semantic anchors, producing semantically enriched cross-modal representations; and (2) the Bidirectional Feature Alignment (BFA) module, which builds fusion relationships between visual and textual features via similarity-driven mechanisms, enabling effective integration of enriched visual and textual information. This design improves cross-modal consistency and interpretability while preserving pretrained alignment priors and enhancing training stability. Experiments on the GeoText-1652 benchmark show that SAA-DGL achieves state-of-the-art performance and strong robustness under complex visual and linguistic disturbances, validating its effectiveness for challenging geo-localization scenarios. We will release the code.

Xinyu Zhang, Linmei Hu, Luhao Zhang, Wentao Cheng, Yashen Wang, Ge Shi, Chong Feng, Liqiang Nie. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Existing video LLMs typically excel at capturing the overall description of a video but lack the ability to demonstrate an understanding of temporal dynamics and a fine-grained grasp of localized content within the video. In this paper, we propose a Time-Perception Enhanced Video Grounding via Boundary Perception and Temporal Reasoning aimed at mitigating LLMs’ difficulties in understanding the discrepancies between video and text temporality. Specifically, to address the inherent biases in current datasets, we design a series of boundary-perception tasks to enable LLMs to capture accurate video temporality. To tackle LLMs’ insufficient understanding of temporal information, we develop specialized tasks for boundary perception and temporal relationship reasoning to deepen LLMs’ perception of video temporality. Our experimental results show significant improvements across three datasets: ActivityNet, Charades, and DiDeMo (achieving up to 11.2% improvement on [email protected]), demonstrating the effectiveness of our proposed temporal awareness-enhanced data construction method.

While automated CT report generation (CTRG) systems promise to enhance clinical workflow efficiency, current solutions, even those based on advanced multi-modal large language models (MLLMs), face fundamental challenges in ensuring report quality and reliability. Through systematic analysis of representation dynamics in MLLM-based CTRG models, we discover that existing systems lack robust quality discrimination capabilities, manifesting in two critical limitations: representation entanglement between reports of varying quality levels in the feature space, and quality-insensitive generation due to conventional training paradigms focusing solely on ground-truth reports. To address these limitations, we propose CT-ClearView, a quality-aware cross-modal alignment framework consisting of two key innovations: (1) a systematic methodology for constructing clinically-relevant hard negative examples using GPT-4, which introduces subtle but significant clinical errors while maintaining report structure and plausibility, and (2) a contrastive learning framework that leverages these examples to effectively disentangle representations of varying quality reports and enhance the model's sensitivity to clinical details. Extensive experiments on CTRG-Chest-548K and CTRG-Brain-263K datasets demonstrate significant performance improvements in natural language generation (NLG) metrics compared to existing approaches (e.g., increases of 11% in BLEU-1 and 16% in both BLEU-4 and ROUGE-L, on the CTRG-Chest-548K datasets).

The prevalent recommendation techniques explore the graph structure of interactions to alleviate the interaction sparsity issue for inferring users’ interests. These graph models focus on extracting local structural signals to model users’ interests, introducing grid-like distortion and ignoring the hierarchical tree-like structure when learning from the interaction graph. The learned interests lack significant hierarchical signals, resulting in suboptimal recommendation performance. In this article, we investigate geometric-augmented graph learning with hyperbolic and Euclidean geometries to delve into local structural and hierarchical knowledge from the interaction graph. A self-teaching network called geometric-augmented self-distillation (GASD) is proposed to transfer hierarchical knowledge from hyperbolic to Euclidean space. The transfer learning enables shrinking of the network into a primary student to implement effective and efficient inference in Euclidean space, preventing computational burden in hyperbolic space. Experiments on publicly available datasets demonstrate that the proposed GASD outperforms the state-of-the-art models, verifying the effectiveness and efficiency of knowledge transfer by self-distillation to aggregate knowledge adaptively for personalized recommendation.

Move structures have been studied in English for Specific Purposes (ESP) and English for Academic Purposes (EAP) for decades. However, there are few move annotation corpora for Research Article (RA) abstracts. In this paper, we introduce RAAMove, a comprehensive multi-domain corpus dedicated to the annotation of move structures in RA abstracts. The primary objective of RAAMove is to facilitate move analysis and automatic move identification. This paper provides a thorough discussion of the corpus construction process, including the scheme, data collection, annotation guidelines, and annotation procedures. The corpus is constructed through two stages: initially, expert annotators manually annotate high-quality data; subsequently, based on the human-annotated data, a BERT-based model is employed for automatic annotation with the help of experts’ modification. The result is a large-scale and high-quality corpus comprising 33,988 annotated instances. We also conduct preliminary move identification experiments using the BERT-based model to verify the effectiveness of the proposed corpus and model. The annotated corpus is available for academic research purposes and can serve as essential resources for move analysis, English language teaching and writing, as well as move/discourse-related tasks in Natural Language Processing (NLP).

Image sentiment analysis is a domain fraught with the dual challenges of interpreting complex visual content and discerning the subtle emotional undertones it may convey. Despite the notable successes of existing visual language pretraining (VLP) models in a variety of visual tasks, they fall short in the nuanced realm of sentiment analysis. This shortfall is primarily due to their inadequate processing of sentiment-specific cues—most notably, the oversight of localized sentioment cues within images and the intricate interplay of these signals. Furthermore, these models inadequately harness the rich sentiment cues often embedded in accompanying text. In response to these shortcomings, we introduce the Specific Sentiment Mask Auto-encoder (S2MA) model, which is expressly designed to integrate sentiment information during the pretraining process. S2MA is meticulously engineered to focuse on both intermodal and intramodal sentiment cue, thereby augmenting the model’s proficiency in anlysising the sentiment knowledge within visual content. Rigorous comparative evaluations of S2MA against the CLIP model, across a spectrum of downstream datasets in zero-shot and supervised learning scenarios, have validated the superiority of our approach. The empirical outcomes affirm S2MA’s capacity to significantly enhance the analytical landscape of image sentiment analysis.

Image emotion classification (IEC) aims to extract the abstract emotions evoked in images. Recently, language-supervised methods such as contrastive language-image pretraining (CLIP) have demonstrated superior performance in image understanding. However, the underexplored task of IEC presents three major challenges: a tremendous training objective gap between pretraining and IEC, shared suboptimal prompts, and invariant prompts for all instances. In this study, we propose a general framework that effectively exploits the language-supervised CLIP method for the IEC task. First, a prompt-tuning method that mimics the pretraining objective of CLIP is introduced, to exploit the rich image and text semantics associated with CLIP. Subsequently, instance-specific prompts are automatically composed, conditioning them on the categories and image content of instances, diversifying the prompts, and thus avoiding suboptimal problems. Evaluations on six widely used affective datasets show that the proposed method significantly outperforms state-of-the-art methods (up to 9.29% accuracy gain on the EmotionROI dataset) on IEC tasks with only a few trained parameters. The code is publicly available at https://github.com/dsn0w/PT-DPC/for research purposes .

Domain adaptation proves to be an effective solution for addressing inadequate translation performance within specific domains. However, the straightforward approach of mixing data from multiple domains to obtain the multi-domain neural machine translation (NMT) model can give rise to the parameter interference between domains problem, resulting in a degradation of overall performance. To address this, we introduce a multi-domain adaptive NMT method aimed at learning domain specific sub-layer latent variable and employ the Gumbel-Softmax reparameterization technique to concurrently train both model parameters and domain specific sub-layer latent variable. This approach facilitates learning private domain-specific knowledge while sharing common domain-invariant knowledge, effectively mitigating the parameter interference problem. The experimental results show that our proposed method significantly improved by up to 7.68 and 3.71 BLEU compared with the baseline model in English-German and Chinese-English public multi-domain datasets, respectively.

Image Emotion Classification (IEC) is an essential research area, offering valuable insights into user emotional states for a wide range of applications, including opinion mining, recommendation systems, and mental health treatment. The challenges associated with IEC are mainly attributed to the complexity and ambiguity of human emotions, the lack of a universally accepted emotion model, and excessive dependence on prior knowledge. To address these challenges, we propose a novel Unified Generative framework for Image Emotion Classification (UGRIE), which is capable of simultaneously modeling various emotion models and capturing intricate semantic relationships between emotion labels. Our approach employs a flexible natural language template, converting the IEC task into a template-filling process that can be easily adapted to accommodate a diverse range of IEC tasks. To further enhance the performance, we devise a mapping mechanism to seamlessly integrate the multimodal pre-training model CLIP with the text generation pre-training model BART, thus leveraging the strengths of both models. A comprehensive set of experiments conducted on multiple public datasets demonstrates that our proposed method consistently outperforms existing approaches to a large margin in supervised settings, exhibits remarkable performance in low-resource scenarios, and unifies distinct emotion models within a single, versatile framework.

Recently, recommender systems have witnessed the fast evolution of Internet services. However, it suffers hugely from inherent bias and sparsity issues in interactions. The conventional uniform embedding learning policies fail to utilize the imbalanced interaction clue and produce suboptimal representations to users and items for recommendation. Towards the issue, this work is dedicated to bias-aware embedding learning in a decomposed manner and proposes a counterfactual graph convolutional learning (CGCL) model for personalized recommendation. Instead of debiasing with uniform interaction sampling, we follow the natural interaction bias to model users’ interests with a counterfactual hypothesis. CGCL introduces bias-aware counterfactual masking on interactions to distinguish the effects between majority and minority causes on the counterfactual gap. It forms multiple counterfactual worlds to extract users’ interests in minority causes compared to the factual world. Concretely, users and items are represented with a causal decomposed embedding of majority and minority interests for recommendation. Experiments show that the proposed CGCL is superior to the state-of-the-art baselines. The performance illustrates the rationality of the counterfactual hypothesis in bias-aware embedding learning for personalized recommendation.

Weakly supervised group activity recognition deals with the dependence on individual-level annotations during understanding scenes involving multiple individuals, which is a challenging task. Existing methods either take the trained detectors to extract individual features or utilize the attention mechanisms for partial context encoding, followed by integration to form the final group-level representations. However, the detectors require individual-level annotations during the training phase and have a mis-detection issue, and the partial contexts extracted immediately from the whole complex scene are too ambiguous without the guidance of concrete semantics. In this paper, we investigate the hierarchical structure inherent in group-level labels to extract the fine-grained semantics without using detectors for weakly supervised group activity recognition. A multi-hot encoding strategy combined with a semantic encoder is first adopted to get the label semantics embeddings. The semantic and visual scene information are then fused through a semantic decoder to obtain activity-specific features. Lastly, we employ the multi-label classification and integrate the scores of hierarchical activity labels. Experimental results show that our proposed method achieves the state-of-the-art performance on three benchmarks, and the accuracy on the Volleyball dataset exceeds the second-best method by 2%.

Previous works build interest learning via mining deeply on interactions. However, the interactions come incomplete and insufficient to support interest modeling, even bringing severe bias into recommendations. To address the interaction sparsity and the consequent bias challenges, we propose a graph contrastive learning on complementary embedding (GCCE), which introduces negative interests to assist positive interests of interactions for interest modeling. To embed interest, we design a perturbed graph convolution by preventing embedding distribution from bias. Since negative samples are not available in the general scenario of implicit feedback, we elaborate a complementary embedding generation to depict users’ negative interests. Finally, we develop a new contrastive task to contrastively learn from the positive and negative interests to promote recommendation. We validate the effectiveness of GCCE on two real datasets, where it outperforms the state-of-the-art models for recommendation.

Image emotion classification is an important computer vision task to extract emotions from images. The methods for image emotion classification (IEC) are primarily based on label or distribution as a supervision signal, which neither has enough accessibility nor diversity, limiting the development of IEC research. Inspired by psychology research and the recent booming of large-scale pretrained language models. We figure out a language-supervised paradigm, which can cleverly combine the features of language and visual emotion to drive the visual model to gain stronger emotional discernment with language prompts. To practice the paradigm, we present a conceptually simple while empirically powerful framework for image emotion classification, SimEmotion. That we propose a prompt-based fine-tuning strategy to learn task-specific representations by composing a template with the emotion-level concept and entity-level information. Evaluations on four widely-used affective datasets, namely, Flickr and Instagram (FI), EmotionROI, Twitter I, and Twitter II, demonstrate that the proposed algorithm outperforms the state-of-the-art methods with a large margin (i.e.,$8.42\%$absolute accuracy gain on EmotionROI) on image emotion classification tasks. Our codes will be publicly available for research purposes.

We consider event extraction in a generative manner with template-based conditional generation. Although there is a rising trend of casting the task of event extraction as a sequence generation problem with prompts, these generation-based methods have several significant challenges, including using suboptimal prompts, static event type information, and the overwhelming number of irrelevant event types. In this article, we propose a generative template-based method with dynamic prefixes and a relevance retrieval framework for event extraction (GREE) by first integrating context information with type-specific prefixes to learn a context-specific prefix for each context, and then retrieving the relevant event types with an adaptive threshold. Experimental results show that our model achieves competitive results with the state-of-the-art classification-based modelOneIEon ACE 2005 and achieves the best performances on ERE. Additionally, our model is proven to be portable to new types of events effectively.

We consider event extraction in a generative manner with template-based conditional generation.Although there is a rising trend of casting the task of event extraction as a sequence generation problem with prompts, these generation-based methods have two significant challenges, including using suboptimal prompts and static event type information.In this paper, we propose a generative templatebased event extraction method with dynamic prefix (GTEE-DYNPREF) by integrating context information with type-specific prefixes to learn a context-specific prefix for each context.Experimental results show that our model achieves competitive results with the state-ofthe-art classification-based model ONEIE on ACE 2005 and achieves the best performances on ERE.Additionally, our model is proven to be portable to new types of events effectively.

Deep supervised hashing for Hamming space retrieval has recently attracted increasing attention because it enables large-scale image retrieval with constant-time cost. However, the existing Hamming space retrieval methods cannot effectively focus on different pairs simultaneously inside and outside the Hamming ball, making it difficult to push dissimilar pairs outside or pull similar pairs inside the Hamming ball. We propose a novel Boundary-Guided Probability Hashing (BGPH) method that introduces a boundary to guide probability distribution. It makes the probability of similar pairs within the Hamming ball greater than dissimilar pairs and vice versa, which fits the purpose of Hamming space retrieval well. Moreover, we propose a threshold weighting method to indicate when optimization should be stopped to avoid the problem that dissimilar data are pulled into the ball caused by over-optimization in multi-label retrieval scenarios. Comprehensive experiments on three benchmark datasets demonstrate that BGPH yields state-of-the-art retrieval performance.

Relation Extraction suffers from dramatical performance decrease when training a model on one genre and directly applying it to a new genre, due to the distinct feature distributions.Previous studies address this problem by discovering a shared space across genres using manually crafted features, which requires great human effort.To effectively automate this process, we design a genre-separation network, which applies two encoders, one genreindependent and one genre-shared, to explicitly extract genre-specific and genre-agnostic features.Then we train a relation classifier using the genre-agnostic features on the source genre and directly apply to the target genre.Experiment results on three distinct genres of the ACE dataset show that our approach achieves up to 6.1% absolute F1-score gain compared to previous methods.By incorporating a set of external linguistic features, our approach outperforms the state-of-the-art by 1.7% absolute F1 gain.We make all programs of our model publicly available for research purpose 1 .