Multimodal recommendation systems increasingly adopt GNNs to perform message passing on the user–item interaction graph and content-based item graphs to learn user and item representations. However, existing methods that use static-graph GNNs in multimodal recommendation face two key limitations: (i) representation homogenization, where repeated aggregation on a fixed topology makes node embeddings overly similar and weakens personalized signals; and (ii) incomplete context modeling, where similarity-based item graphs mainly capture local pairwise resemblance but miss group- or scenario-level relations that are far apart in the similarity space. Together, these limitations reduce the quality of learned node representations on static graphs and lead to suboptimal recommendations. To alleviate these limitations, we propose ESG-Rec, a unified multi-view framework that constructs three complementary structural views: degree-aware structural diversification to preserve individualized collaborative signals, content-based semantic modeling to provide stable local relations, and group-level semantic modeling to capture higher-order contextual information. ESG-Rec further performs multi-view contrastive alignment to learn consistent representations across views. Extensive experiments on benchmark datasets demonstrate that ESG-Rec consistently improves recommendation performance over strong multimodal baselines and effectively mitigates the above two limitations.

Sequential recommendation systems play a key role in modern information retrieval. However, existing intent-related work fails to adequately capture long-term dependencies in user behavior, i.e., the influence of early user behavior on current behavior, and also fails to effectively utilize item relevance. To this end, we propose a novel sequential recommendation framework to overcome the above limitations, called ICMA. Specifically, we combine temporal variability with position encoding that has extrapolation properties to encode sequences, thereby expanding the model’s view of user behavior and capturing long-term user dependencies more effectively. Additionally, we design a multi-view data augmentation method, i.e., based on random data augmentation methods (e.g., crop, mask, and reorder), and further introduce insertion and substitution operations to augment the sequence data from different views by utilizing item relevance. Within this framework, clustering is performed to learn intent distributions, and these learned intents are integrated into the sequential recommendation model via contrastive SSL, which maximizes consistency between sequence views and their corresponding intents. The training process alternates between the Expectation (E) step and the Maximization (M) step. Experiments on three real datasets show that our approach improves by 0.8% to 14.7% compared to most baselines.

Multimodal sarcasm detection aims to identify whether utterances express sarcastic intentions contrary to their literal meaning based on multimodal information. However, existing methods fail to explore the model’s "ability to understand" the semantics expressed by sentences in the image context from semantic diversity perspectives. In this paper, we propose a multi-view semantic awareness method, which concretizes semantics from multiple perspectives to improve the model’s ability to capture different semantic features. Specifically, two learnable prefixes are attached to the text representation respectively to construct semantic representations from both the literal meaning and sarcastic intention perspectives. Then, image-text information is further fused through cross-attention to guide the semantic representation of different perspectives in the image context. Finally, the semantics expressed by prefixes are strengthened through KL divergence, thereby encouraging the model to capture two distinctive semantic features. Experiments on benchmark datasets demonstrate the effectiveness of our method.

Personalized chatbots concentrate on learning human personalities, making them act similar to real users. When it is authorized to respond to other people’s messages, it has the same way of speaking as the user. Many personalized methods have been proposed to use several persona descriptions or key-value-based persona information to assign a personality for dialogue chatbots. Most of them employ explicit user profiles. However, obtaining generous explicit user profiles are extremely time-consuming and requires tremendous manual labor. In addition, explicit user profiles cannot be updated as the user’s interests change. In this article, we propose a generation-based personalized chatbot model, IMDPchat, that learns latent user representation from the abundant users’ dialogue history. Specially, we train a personalized language model to build a global user profile using dialogue responses. To take full advantage of users’ information used in the historical dialogue, we establish a key-value memory network and construct a post-sensitive personalized selection module. The above two parts is context-aware: we endow higher weights to historical post-response pairs that are connected to the current post. To predict more personalized responses, we design a personalized response decoder that can well integrate two decoding modes, including generating tokens and copying personalized words. Experimental results indicate that the IMDPchat model outperforms previous baselines remarkably.

Knowledge-based Visual Question Answering (KBVQA) is a challenging task, which aims to answer an image related question based on external knowledge. Most of the works describe the semantic distance using the actual Euclidean distance between two nodes, which leads to distortion in modeling knowledge graphs with hierarchical and scale-free structure in KBVQA, and limits the multi-hop reasoning capability of the model. In contrast, the hyperbolic space shows exciting prospects for low-distortion embedding of graphs with hierarchical and free-scale structure. In addition, we map the different stages of reasoning into multiple adjustable hyperbolic spaces, achieving low-distortion, fine-grained reasoning. Extensive experiments on the KVQA, PQ and PQL datasets demonstrate the effectiveness of HyperMR for strong-hierarchy knowledge graphs.

Multimodal Sentiment Analysis(MSA) mostly based on the assumption that all modalities are available. However, this assumption does not always hold in practice, and the performance of most multimodal sentiment analysis models can suffer significant degradation when some modalities are missing. To this end, we propose an Information Aggregation and Sentiment Enhance network(IASE) to handle missing modalities. Specifically, IASE models multimodal data as a bipartite graph structure that reduces the impact of missing modalities by aggregating complementary information to update out the representation of missing modalities. Aggregating samples of different sentiments leads to a decrease of sentiment intensity. The sentiment enhancement module is designed to enhance the sentiment intensity. Extensive experiments on several datasets show significant improvements of our method compared to several baselines.

Multi-hop Knowledge Graph Question Answering aims at finding an entity to answer natural language questions from knowledge graphs. When humans perform multi-hop reasoning, people tend to focus on specific relations across different hops and confirm the next entity. Therefore, most algorithms choose the wrong specific relation, which makes the system deviate from the correct reasoning path. The specific relation at each hop plays an important role in multi-hop question answering. Existing work mainly relies on the question representation as relation information, which cannot accurately calculate the specific relation distribution. In this article, we propose an interpretable assistance framework that fully utilizes the relation embeddings to assist in calculating relation distributions at each hop. Moreover, we employ the fusion attention mechanism to ensure the integrity of relation information and hence to enrich the relation embeddings. The experimental results on three English datasets and one Chinese dataset demonstrate that our method significantly outperforms all baselines. The source code of REAN will be available at https://github.com/2399240664/REAN

Social recommendation enhances the learning of user preferences by incorporating user social information. Recently, graph neural network models have gradually become the subject of the social recommendation. However, most graph neural network-based approaches fail to fully learn the high-order collaborative semantics of user interest and social domains, and ignore the unique self-supervised signals in user social domains. To alleviate these problems, we propose a novel lightweight GCN-based social recommendation method SGSR that jointly models the high-order collaborative relations of user/item nodes in both domains. Meanwhile, in the process of message transmission of the bipartite graph and social graph, we respectively introduce a self-attention mechanism to measure the contributions of different nodes. In particular, to take full advantage of the self-supervised signals between user node messages in the social domain, we innovatively incorporate contrastive learning into this system to enable user-side node features to self-learn and update. Extensive experiments conducted on two real datasets demonstrate the effectiveness and necessity of our proposed approach.

A significant goal in an open-domain dialogue system is to make chatbots generate more persona coherent responses given a context. To achieve this goal, some researchers attempt to introduce persona information into neural dialogue models. However, these neural dialogue models describe excessively persona traits during the conversation, which still suffer from the problem of generating boring and meaningful responses. In this paper, we divide the open-domain personalized dialogue generation task into two processes, persona recognition and persona fusion. In the persona recognition process, we use the pre-training model to encode the personas and conversation history independently, which is beneficial to the persona information fusion. Then, we design a dynamic persona fusion mechanism to effectively mine the relevance of dialogue context and persona information, and dynamically predict whether to incorporate persona features in the process of the dialogues. Our model outperforms with 1.23% in Acc., 0.36% in BLEU, 0.92% in F1, and 0.036% in Distinct than baseline models. The experimental results on the ConvAI2 dataset illustrate that the proposed model is superior to baseline approaches for generating more coherent and persona consistent responses.