Large language models (LLMs) have demonstrated significant potential in clinical decision support. Yet LLMs still suffer from hallucinations and lack fine-grained contextual medical knowledge, limiting their high-stake healthcare applications such as clinical diagnosis. Traditional retrieval-augmented generation (RAG) methods attempt to address these limitations but frequently retrieve sparse or irrelevant information, undermining prediction accuracy. We introduce KARE, a novel framework that integrates knowledge graph (KG) community-level retrieval with LLM reasoning to enhance healthcare predictions. KARE constructs a comprehensive multi-source KG by integrating biomedical databases, clinical literature, and LLM-generated insights, and organizes it using hierarchical graph community detection and summarization for precise and contextually relevant information retrieval. Our key innovations include: (1) a dense medical knowledge structuring approach enabling accurate retrieval of relevant information; (2) a dynamic knowledge retrieval mechanism that enriches patient contexts with focused, multi-faceted medical insights; and (3) a reasoning-enhanced prediction framework that leverages these enriched contexts to produce both accurate and interpretable clinical predictions. Extensive experiments demonstrate that KARE outperforms leading models by up to 10.8-15.0\% on MIMIC-III and 12.6-12.7\% on MIMIC-IV for mortality and readmission predictions. In addition to its impressive prediction accuracy, our framework leverages the reasoning capabilities of LLMs, enhancing the trustworthiness of clinical predictions.

Reinforcement learning (RL) has become a powerful approach for improving the reasoning capabilities of large language models (LLMs), as evidenced by recent successes such as OpenAI's o1 and Deepseek-R1. However, applying RL at scale remains intimidatingly resource-intensive, requiring multiple model copies and extensive GPU workloads. On the other hand, while being powerful, recent studies suggest that RL does not fundamentally endow models with new knowledge; rather, it primarily reshapes the model's output distribution to activate reasoning capabilities latent in the base model. Building on this insight, we hypothesize that the changes in output probabilities induced by RL are largely model-size invariant, opening the door to a more efficient paradigm: training a small model with RL and transferring its induced probability shifts to larger base models. To verify our hypothesis, we conduct a token-level analysis of decoding trajectories and find high alignment in RL-induced output distributions across model scales, validating our hypothesis. Motivated by this, we propose RAST, a simple yet effective method that transfers reasoning behaviors by injecting RL-induced probability adjustments from a small RL-trained model into larger models. Experiments across multiple mathematical reasoning benchmarks show that RAST substantially and consistently enhances the reasoning capabilities of base models while requiring significantly lower GPU memory than direct RL training, sometimes even yielding better performance than the RL-trained counterparts. Our findings offer new insights into the nature of RL-driven reasoning and practical strategies for scaling its benefits without incurring its full computational cost. The project page of RAST is available at https://ozyyshr.github.io/RAST/.

Fine-grained entity typing (FET), which assigns entities in text with context-sensitive, fine-grained semantic types, is a basic but important task for knowledge extraction from unstructured text. FET has been studied extensively in natural language processing and typically relies on human-annotated corpora for training, which is costly and difficult to scale. Recent studies explore the utilization of pre-trained language models (PLMs) as a knowledge base to generate rich and context-aware weak supervision for FET. However, a PLM still requires direction and guidance to serve as a knowledge base as they often generate a mixture of rough and fine-grained types, or tokens unsuitable for typing. In this study, we vision that an ontology provides a semantics-rich, hierarchical structure, which will help select the best results generated by multiple PLM models and head words. Specifically, we propose a novel annotation-free, ontology-guided FET method, OntoType, which follows a type ontological structure, from coarse to fine, ensembles multiple PLM prompting results to generate a set of type candidates, and refines its type resolution, under the local context with a natural language inference model. Our experiments on the Ontonotes, FIGER, and NYT datasets using their associated ontological structures demonstrate that our method outperforms the state-of-the-art zero-shot fine-grained entity typing methods as well as a typical LLM method, ChatGPT. Our error analysis shows that refinement of the existing ontology structures will further improve fine-grained entity typing.

Geospatial topological relation extraction (GeoTopoRE) aims to extract topological relations between named geospatial entities (i.e., geo-entities) in text. It is a domain-specific relation extraction (RE) task essential in geospatial knowledge graph construction and spatial reasoning. Unlike general-purpose RE, which primarily depends on semantic and syntactic cues, GeoTopoRE requires integrating geometric knowledge about geo-entities. This is essential for accurately capturing or inferring the complex geospatial relationships among entities. GeoTopoRE is not studied systematically and lacks dedicated datasets for evaluation, posing significant challenges to developing and assessing effective models. This study presents two major contributions: (i) the introduction of a high-quality, human-labeled dataset WikiTopo for the GeoTopoRE task, and (ii) a novel framework GeoWISE designed to adapt existing RE models to the GeoTopoRE task, With Integrated Semantic and External geospatial domain knowledge. We leverage coarse-to-fine-grained natural language inference (NLI) to align externally sourced knowledge with the semantic text context, enhanced by geospatial expertise. This integrated knowledge is then conveyed to language models as geospatial cues, enabling a nuanced understanding of topological relations. Empirical results demonstrate the efficacy of our framework in few-shot settings, showing significant and consistent improvements in the GeoTopoRE task for diverse state-of-the-art RE models.

Weakly-supervised text classification trains a classifier using the label name of each target class as the only supervision, which largely reduces human annotation efforts.Most existing methods first use the label names as static keyword-based features to generate pseudo labels, which are then used for final classifier training.While reasonable, such a commonly adopted framework suffers from two limitations: (1) keywords can have different meanings in different contexts and some texts may not explicitly contain any keyword, so keyword matching can induce noisy and inadequate pseudo labels; (2) the errors made in the pseudo label generation stage will directly propagate to the classifier training stage without a chance of being corrected.In this paper, we propose a new method, PIEClass, consisting of two modules: (1) a pseudo label acquisition module that uses zero-shot prompting of pre-trained language models (PLM) to get pseudo labels based on contextualized text understanding beyond static keyword matching, and (2) a noise-robust iterative ensemble training module that iteratively trains classifiers and updates pseudo labels by utilizing two PLM fine-tuning methods that regularize each other.Extensive experiments show that PIEClass achieves overall better performance than existing strong baselines on seven benchmark datasets and even achieves similar performance to fully-supervised classifiers on sentiment classification tasks. 1

Today a tremendous amount of geospatial knowledge is hidden in massive volumes of text data. To facilitate flexible and powerful geospatial analysis and applications, we introduce a new architecture: geospatial knowledge hypercube, a multi-scale, multidimensional knowledge structure that integrates information from geospatial dimensions, thematic themes and diverse application semantics, extracted and computed from spatial-related text data. To construct such a knowledge hypercube, weakly supervised language models are leveraged for automatic, dynamic and incremental extraction of heterogeneous geospatial data, thematic themes, latent connections and relationships, and application semantics, through combining a variety of information from unstructured text, structured tables, and maps. The hypercube lays a foundation for many knowledge discovery and in-depth spatial analysis, and other advanced applications. We have deployed a prototype web application of proposed geospatial knowledge hypercube for public access at: https://hcwebapp.cigi.illinois.edu/.

We propose REACTCLASS that automatically maps the low-level concrete chemical entities into the high-level reactant groups without human effort for training data annotation. REACTCLASS is designed to take two special characteristics of the chemical molecules into consideration. The first characteristic is that each chemical molecule can be represented in two modalities: a chemical name in the text and a molecular structure in the graph. We propose to use cross-modal supervision to automatically create the training data for chemical name classification in the text via molecular structure matching in the graph. The second characteristic is that there is a knowledge-aware subword correlation between the surface names of the chemical entities to be classified and that of the reactant groups as class labels. We propose to train a classification model based on the subword cross-attention map between each chemical name and the corresponding reaction group. Experiments demonstrate that REACTCLASS is highly effective, achieving state-of-the-art performance in classifying the chemical names into human-defined reactant groups without requiring human effort for training data annotation.

Taxonomies are fundamental to many real-world applications in various domains, serving as structural representations of knowledge. To deal with the increasing volume of new concepts needed to be organized as taxonomies, researchers turn to automatically completion of an existing taxonomy with new concepts. In this paper, we propose TaxoEnrich, a new taxonomy completion framework, which effectively leverages both semantic features and structural information in the existing taxonomy and offers a better representation of candidate position to boost the performance of taxonomy completion. Specifically, TaxoEnrichconsists of four components: (1) taxonomy-contextualized embedding which incorporates both semantic meanings of concept and taxonomic relations based on powerful pretrained language models; (2) a taxonomy-aware sequential encoder which learns candidate position representations by encoding the structural information of taxonomy; (3) a query-aware sibling encoder which adaptively aggregates candidate siblings to augment candidate position representations based on their importance to the query-position matching; (4) a query-position matching model which extends existing work with our new candidate position representations. Extensive experiments on four large real-world datasets from different domains show that TaxoEnrichachieves the best performance among all evaluation metrics and outperforms previous state-of-the-art methods by a large margin.

The most efficient known approach for shortest distance querying on road networks is via a tree decomposition based 2-hop labeling index. A major challenge here is how to reduce the query time by reducing the label size. To this end, we propose P2H with the novel ideas of projected vertex separators and optimized selection of vertex separators. We also introduce mechanisms for index maintenance for edge weight updating. Our experiments on multiple real road networks show that P2H can greatly reduce the effective label sizes and query time over existing algorithms. For larger datasets, P2H is around twice as efficient as the best known algorithm.

Similarity among entities in graphs plays a key role in data analysis and mining. SimRank is a widely used and popular measurement to evaluate the similarity among the vertices. In real-life applications, graphs do not only grow in size, requiring fast and precise SimRank computation for large graphs, but also change and evolve continuously over time, demanding an efficient maintenance process to handle dynamic updates. In this paper, we propose a random walk based indexing scheme to compute SimRank efficiently and accurately over large dynamic graphs. We show that our algorithm outperforms the state-of-the-art static and dynamic SimRank algorithms.

Top-k nearest keyword search has been of interest because of applications ranging from road network location search by keyword to search of information on an RDF repository. We consider the evaluation of a query with a given vertex and a keyword, and the problem is to find a set of $k$ nearest vertices that contain the keyword. The known algorithms for handling this problem only give approximate answers. In this paper, we propose algorithms for top-k nearest keyword search that provide exact solutions and which handle networks of very large sizes. We have also verified the performance of our solutions compared with the best-known approximation algorithms with experiments on real datasets.

We study the problem of point-to-point distance querying for massive scale-free graphs, which is important for numerous applications. Given a directed or undirected graph, we propose to build an index for answering such queries based on a novel hop-doubling labeling technique. We derive bounds on the index size, the computation costs and I/O costs based on the properties of unweighted scale-free graphs. We show that our method is much more efficient and effective compared to the state-of-the-art techniques, in terms of both querying time and indexing costs. Our empirical study shows that our method can handle graphs that are orders of magnitude larger than existing methods.

Bichromatic reverse nearest neighbor (BRNN) queries have been studied extensively in the literature of spatial databases. Given a set P of service-providers and a set O of customers, a BRNN query is to find which customers in O are "interested" in a given service-provider in P. Recently, it has been found that this kind of queries lacks the consideration of the capacities of service-providers and the demands of customers. In order to address this issue, some spatial matching problems have been proposed, which, however, cannot be used for some real-life applications like emergency facility allocation where the maximum matching cost (or distance) should be minimized. In this paper, we propose a new problem called Spatial Matching for Minimizing Maximum matching distance (SPM-MM). Then, we design two algorithms for SPM-MM, Threshold-Adapt and Swap-Chain. Threshold-Adapt is simple and easy to understand but not scalable to large datasets due to its relatively high time/space complexity. Swap-Chain, which follows a fundamentally different idea from Threshold-Adapt, runs faster than Threshold-Adapt by orders of magnitude and uses significantly less memory. We conducted extensive empirical studies which verified the efficiency and scalability of Swap-Chain.