Textual data is a key signal for modeling relationships between news articles in recommendation systems. However, exaggerated and biased texts often hinder the accurate understanding of textual content, negatively affecting recommendation quality. Multimodal approaches have been proposed to address this issue by incorporating associated news images, but they often overlook the complementary role of visual information. To bridge this gap, we propose an Image-Guided Knowledge (IGKnow) distillation framework that transfers image-guided debiasing capability from a multimodal LLM into a text-only LLM. This design allows the distilled model to extract factual, debiased triples from text alone while preserving the benefits of multimodal learning. To further enhance stability, we introduce Triple-level Supervised Fine-Tuning (TriSFT), a permutation-invariant training for triple structures. Moreover, we refine the model through preference alignment to ensure objective, coherent knowledge extraction across news histories. We evaluate the extracted knowledge in sequential news recommendation and show that debiased triples improve recommendation performance.

Entity alignment (EA) is crucial for integrating knowledge graphs (KGs) constructed from diverse sources. Conventional unsupervised EA approaches attempt to eliminate human intervention but often suffer from accuracy limitations. With the rise of large language models (LLMs), leveraging their capabilities for EA presents a promising direction. However, it introduces new challenges: formulating the LLM-based EA problem and extracting the background knowledge in LLMs to realize EA without human intervention. This paper proposes HLMEA, a novel hybrid language model-based unsupervised EA method. HLMEA formulates the EA task into a filtering and single-choice problem and synergistically integrates small language models (SLMs) and LLMs. Specifically, SLMs filter candidate entities based on textual representations generated from KG triples. Then, LLMs refine this selection to identify the most semantically aligned entities. An iterative self-training mechanism allows SLMs to distill knowledge from LLM outputs, enhancing the EA ability of hybrid language models in subsequent rounds cooperatively. We also conducted extensive experiments on benchmark datasets to evaluate HLMEA's performance. The results demonstrate that HLMEA significantly outperforms unsupervised and even supervised EA baselines, proving its potential for scalable and effective EA across large KGs. The code and data are available at \url{https://github.com/xnjin-ai/HLMEA}.

Joint information extraction aims to convert unstructured text into structured knowledge by identifying entities and their relations. However, existing methods often rely on static span formation and relation-agnostic validation, limiting their ability to capture dynamic, context-sensitive semantics. We present RePooL, a hierarchical validation framework that performs fine-grained token-level filtering followed by coarse-grained span-level validation, enabling robust multi-granular semantic modeling. RePooL constructs a dual-view knowledge graph that models tokens and relations as distinct node types. It leverages auxiliary structural relations to encode token-relation semantic compatibility via subject and object roles and to compose multi-token spans dynamically, thereby enabling relation-aware validation across multiple granularities. To further strengthen semantic grounding, RePooL incorporates LLM-guided alignment, which evaluates candidate triples against the input text to specifically reinforce coherent extractions. Extensive experiments on standard IE benchmarks show that RePooL achieves superior performance, demonstrating its effectiveness in modeling fine-grained entity-relation interactions.

Numerous large-scale knowledge graphs (KGs) fundamentally represent two-view KGs: an ontology-view KG with abstract classes in ontology and an instance-view KG with specific collections of entities instantiated from ontology classes. Two-view KG embedding aims to jointly learn continuous vector representations of entities and relations in the aforementioned two-view KGs. In essence, an ontology schema exhibits a tree-like structure guided by class hierarchies, which leads classes to form inheritance hierarchies. However, existing two-view KG embedding models neglect those hierarchies, which provides the necessity to reflect class inheritance. On the other hand, KG is constructed based on a pre-defined ontology schema that includes heterogeneous relations between classes. Furthermore, these relations are defined within the scope of those among classes since instances inherit all the properties of their corresponding classes, which reveals structural similarity between two multi-relational networks. Despite the consideration to bridge the gap among two-view KG representations, existing methods ignore the existence of structural similarity between two-view KGs. To address these issues, we propose a novel two-view KG embedding model, CISS, considering Class Inheritance and Structural Similarity between two-view KGs. To deal with class inheritance, we utilize class sets, each of which is composed of sibling classes, to learn fine-grained class representations. In addition, we configure virtual instance-view KG from clustered instances and compare subgraph representations of two-view KGs to enhance structural similarity between them. Experimental results show our superior performance compared to existing models.

Recently, contrastive loss is adopted as a main objective of recommender systems. InfoNCE-like losses penalize hard negative items more and control the strength of penalties with a temperature, called hardness-aware sensitivity. However, since they leverageuser->item patterns in a non-symmetric way, negative items are pushed away from anchor users and attract semantically-similar items to each other, focusing on the distribution of item embeddings. We point out that user embeddings also have inherent semantic structures that can be captured fromitem->user patterns. This paper presents Dual-oriented Contrast(DuCo), a novel symmetric learning objective for recommendation to learn more comprehensive representations fromusereftrightarrowitem patterns. DuCo controls user-/item-centric hardness-aware sensitivities and simultaneously optimizes the score distributions over sampled items (user-oriented contrast) and users (item-oriented contrast). This aims to explore ideal user and item distributions that are locally clustered and globally uniform. However, since user-/item-side temperatures are interdependent, naive control over temperatures may break the underlying semantic structures of the other side. To this end, we employ a stop-gradient operation to preserve the individual characteristics of user/item embedding distributions. Furthermore, we balance user-/item-oriented contrasts during learning to maintain consistent high-rank performance (e.g., recall@1). Empirical results show that DuCo contributes to the top-k user and item prediction simultaneously, and outperforms state-of-the-art learning objectives across different backbones from ID-based to neighbor-based encoders.

Both entity typing and relation extraction from text corpora are widely used to identify the semantic types of an entity and a relation in a knowledge graph (KG). Most existing approaches rely on a pre-defined set of entity types and relation types in a KG. They thus cannot map entity mentions (relation mentions) to unseen entity types (relation types). To fundamentally overcome the limitations, we should add new semantic types of entities and relations to a KG schema. However, schema expansion traditionally requires manual conceptualization through a user’s observation on the text corpus while assuming the existence of suitable target KG schemas. In this work, we propose anActive learning framework forKnowledge graphSchemaExpansion (AKSE), which can generate a new semantic type for KG schemas, without depending on a set of target schemas and human users’ observation. Specifically, a granularity based active learning algorithm determines whether a KG schema requires new semantic types or not. We also introduce a KG schema attention-based neural method which assigns semantic types to the entities and relationships extracted. To the best of our knowledge, our work is the first study to expand a KG schema with active learning.

The accurate prediction of stress in a person’s life has a significant effect on improving personal health and the national economy. Since individuals have different historical circumstances and personality traits, stress symptoms and levels may vary from person to person. Thus, most studies on stress prediction pay attention to personalized models, which determine the personal stress level using user-specific information and heterogeneous stress-related data. However, these models cannot elaborately handle the uncertainty caused by the sparsity, data imbalance, irregularity, and high-dimensionality of user-specific information. In particular, out-of-sample users increase uncertainty. To cope with the problem, we propose a personalized stress-level prediction model with inductive Gaussian representation (PSP-IGR), which exploits heterogeneous inputs with a unified end-to-end approach. PSP-IGR extracts feature vectors from the heterogeneous inputs via Gaussian sampling, domain rules, and deep learning, depending on the characteristics of each input. Especially, PSP-IGR inductively generates a Gaussian feature vector called IGR by Gaussian sampling from the shared contents of user-specific information. Thus, PSP-IGR not only generalizes to both in-sample and out-of-sample users effectively but also deals with the uncertainty problem caused by limitations of healthcare datasets. Also, since we fuse the extracted feature vectors considering their characteristics (Gaussian and point vectors), we can preserve the expressiveness of each feature vector. Experiments on a real-world dataset, including survey results, wearable sensor signals, and contexts, demonstrate that PSP-IGR shows higher accuracy in predicting individual stress-level than previous models.

News recommendation systems? purpose is to tackle the immense amount of news and offer personalized recommendations to users. A major issue in news recommendation is to capture the precise news representations for the efficacy of recommended items. Commonly, news contents are filled with well-known entities of different types. However, existing recommendation systems overlook exploiting external knowledge about entities and topical relatedness among the news. To cope with the above problem, in this paper, we propose Topic-Enriched Knowledge Graph Recommendation System(TEKGR). Three encoders in TEKGR handle news titles in two perspectives to obtain news representation embedding: (1) to extract meaning of news words without considering latent knowledge features in the news and (2) to extract semantic knowledge of news through topic information and contextual information from a knowledge graph. After obtaining news representation vectors, an attention network compares clicked news to the candidate news in order to get the user's final embedding. Our TEKGR model is superior to existing news recommendation methods by manipulating topical relations among entities and contextual features of entities. Experimental results on two public datasets show that our approach outperforms state-of-the-art deep recommendation approaches.

This paper presents an efficient method of extracting n-ary relations from multiple sentences which is called Entity-path and Discourse relation-centric Relation Extractor (EDCRE). Unlike previous approaches, the proposed method focuses on an entity link, which consists of dependency edges between entities, and discourse relations between sentences. Specifically, the proposed model consists of two main sub-models. The first one encodes sentences with a higher weight on the entity link while considering the other edges with an attention mechanism. To consider various latent discourse relations between sentences, the second sub-model encodes discourse relations between adjacent sentences considering the contents of each sentence. Experiment results on the cross-sentence relation extraction dataset, PubMed, and the document-level relation extraction dataset, DocRED, show that the proposed model outperforms state-of-the-art methods of extracting relations across sentences. Furthermore, ablation study proves that both the two main sub-models have noticeable effect on the relation extraction task.

A large number of spatial knowledge graphs (SKGs) are available from spatially enriched knowledge bases, e.g., DBpedia and YAGO2. This provides a great chance to understand valuable information about the regions surrounding us. However, it is hard to comprehend SKGs due to the explosively growing volume and the complication of the graph structures. Thus we study the problem of similar region search (SRS), which is an easy-to-use but effective way to explore spatial data. The effectiveness of SRS highly depends on how to measure the region similarity. However, existing approaches cannot make use of the rich information contained in SKGs thus may lead to incorrect results. In this paper, we propose a spatial knowledge representation learning method for region similarity, namely SKRL4RS. SKRL4RS firstly encodes the spatial entities of an SKG into a vector space to make it easier to extract useful features. Then regions are represented by 3-D tensors using the spatial entity embeddings together with geographical information. Finally, region tensors are fed into the conventional triplet network to learn the feature vectors of regions. The region similarity measure learned by SKRL4RS can capture the hierarchical types, semantic relatedness, and relative locations of spatial entities inside a region. Experimental results on two real-world datasets show that our SKRL4RS outperforms the state-of-the-art by a significant margin in terms of the accuracy of measuring region similarity.

Byungkook Oh, Seungmin Seo, Cheolheon Shin, Eunju Jo, Kyong-Ho Lee. 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.

The main focus of relational learning for knowledge graph completion (KGC) lies in exploiting rich contextual information for facts. Many state-of-the-art models incorporate fact sequences, entity types, and even textual information. Unfortunately, most of them do not fully take advantage of rich structural information in a KG, i.e., connectivity patterns around each entity. In this paper, we propose a context-aware convolutional learning (CACL) model which jointly learns from entities and their multi-hop neighborhoods. Since we directly utilize the connectivity patterns contained in each multi-hop neighborhood, the structural role similarity among entities can be better captured, resulting in more informative entity and relation embeddings. Specifically, CACL collects entities and relations from the multi-hop neighborhood as contextual information according to their relative importance and uniquely maps them to a linear vector space. Our convolutional architecture leverages a deep learning technique to represent each entity along with its linearly mapped contextual information. Thus, we can elaborately extract the features of key connectivity patterns from the context and incorporate them into a score function which evaluates the validity of facts. Experimental results on the newest datasets show that CACL outperforms existing approaches by successfully enriching embeddings with neighborhood information.