Knowledge graphs are a structured form of knowledge representation that connects entities to other entities and to literal data values via relations. Combined with an ontology that captures relationships between concepts, knowledge graphs can provide a flexible structure that supports logical inference, aligns data, and adapts to a variety of knowledge tasks. A number of these tasks, such as question answering and entity summarization, often begin with understanding the context of one entity of interest (either an instance or a class). Querying the multi-hop neighborhood of entity to three or more hops can quickly become intractable in large knowledge graphs, but many existing ontology design patterns create meaningful entity relationships that span more than two relation hops in the graph. This research aims to use both ontology reasoning and graph analysis to build a contextually relevant subgraph around an entity of interest. This targeted subgraph extraction will have useful applications for a variety of upstream tasks related to human interaction with the knowledge graph via exploration and question answering.
This research project investigates methods for extracting a subset of the neighborhood around one
instance entity in a knowledge graph that contains both data (instance axioms i.e. ABox) and an ontology
(TBox). The primary motivation and application for this task is to delineate the relevant context for a
spatial entity in a geospatial knowledge graph. A geospatial knowledge graph, especially one that utilizes
spatial entities like those in a discrete global grid, can contain a multitude of information that describes
the environmental, social, and economic attributes of a region. For example, the SAWGraph project
connects information about chemical contamination by PFAS in the environment of the continental U.S.,
with hydrology, soil properties, crop cover, census data, etc. . The graph uses topological enrichment
[
Subgraph extraction is a useful task for a variety of additional challenges related to knowledge graphs,
including question answering on knowledge graphs, ontology exploration, entity summarization and
semantic similarity between entities. These challenges often require a starting graph on which to
run statistics, perform reasoning, or render a visualization to a user that is a less dense graph. An
initial subgraph extraction can speed-up and simplify further processing. Filtering to only the most
important statements around a focal point in the graph, and hiding other parts of the graph that are less
strongly relevant to a task indeed has uses both for human interaction with knowledge graphs (ontology
exploration and other graph visualization) as well as computation (eficient question answering from a
starting entity, entity summarization, etc.). Complex question answering is a more challenging variation
on the question answering task that can involve entities that are many hops (relations) away from the
starting entity, and many question answering pipelines use the neighborhood of the starting entity
as candidates from which to select the answer [
Many common ontology design patterns create multiple sub-entities to represent what in a traditional
tabular database would constitute one record, or one data view in a relational database. For example
PROV-O [
What is needed is a principled method to extract a subgraph around an entity of interest, reaching multiple hops away from the starting entity, but terminating at diferent hop distances in each direction once the connected entities can no longer be considered an extension of the description of the entity of interest. The goal of this research is to extract a meaningful subgraph that captures the distinguishing relationships of a starting entity, without traversing too far into neighboring entities, by leveraging both the ontology and the graph structure. One strategy would be to capture the pattern of the ontology in reference to a starting class and map it to a query on the instances to build a subgraph. Alternately, querying the instances directly in the full graph can also reveal a diferent relationship between types of entities that might not be explicit in the ontology or may span multiple ontologies. Another strategy is to handle diferent semantic relationships between entities in diferent ways, for relationship types like mereology, subsumption, topology, and functional relations. Additionally, comparing graph statistics calculated on classes in the ontology to statistics calculated on instances in the ABox can inform the role that some related entities may play.
Entity summarization focuses on providing a short summary of an entity in a Knowledge Graph,
mostly for the purpose of presenting a user with a digestible amount of information [
RDF graph, Knowledge Graph and ontology summarization tasks tackle summarization of an entire
graph to aid in user exploration and understanding [
The entity subgraph extraction task is very similar to ontology modularization, especially those
modularization methods that extract a section of an ontology to perform a specific task, referred to
as a knowledge hiding motivation [
Another closely related concept is sub-graph extraction for question answering. In order to constrain
the search space for question answering on knowledge graphs, methods such as personal page rank
information propagation are applied to the graph to prune to the most relevant parts for the question
[
The proposed approach to entity subgraph extraction is a hybrid logical and graphical methodology. The method performs both reasoning on the ontologies and instances, and also computing graph statistics on both the TBox and ABox graphs. These methods are combined to arrive at a subgraph around an entity of interest in a eficient and repeatable manner.
By first reasoning on the core ontology in the graph to which the instance belongs, we can obtain
graph closure under logical entailment, and begin to identify walks through the ontology from the
starting class via domain and range restrictions, universal and existential quantification, and property
chain axioms. The starting class is defined by the entity of interest, and there may be more than one
depending on the ontology complexity. Transitive reduction may also assist in simplifying the possible
pathways. Next we turn to a sampling of the instances in the graph to do further reasoning. Preliminary
testing shows this needs to be limited to a subset of instance data for performance reasons. Reasoning
on the instances may identify pathways that connect diferent ontologies, such as instances from a
class in one ontology connected by relations from another ontology. Additionally instance data can
potentially predict some domain and range values for relations that either cross ontologies or are not
explicitly defined in the TBox of one ontology. The method for querying the full graph is similar to that
employed in [
The last proposed step in the approach is using graph analysis techniques such as calculating the degree of nodes in both the ontology only graph (TBox as a graph) and the instances graph (ABox sample as a graph). By comparing graph statistics between the ontology class, and a summary of instances of the class we may be able to further identify triples which should be included in or excluded from the subgraph. One such example is controlled vocabularies that are implemented as class instances. These are akin to rdf:type / ’is a’ statements, and traversing the graph past them generally leads to ‘sister’ instances that may not be particularly useful for a summative understanding of the entity of interest. For example in SAWGraph, a PFAS sample has an object property relation to an instance of the sample material type class, which if followed further would lead to all other samples of the same type. This implementation of a controlled vocabulary is useful for practical querying of the graph rather than directly translating every type attribute to a subclass structure, however it becomes one of the key challenges for delineating an appropriate subgraph.
This research is part of SAWGraph project, to build an open knowledge network for environmental contamination data about PFAS contamination. It represents an essential task in support of a couple of diferent future work goals related to the project.
The first goal relates to a user interface for the graph, which allows users to build interactive queries and provides map-based visualization of the results. We would like to present filters to the user based on selected starting classes of entities, and then allow them to further refine the query. The interface is similar to faceted search in terms of the filtering, but allows connecting two diferent entities via a spatial relation. However we don’t want to present the user with all possible data values as filter options, and therefore allow them to build a complex query which has no result in the knowledge graph. We need a mechanism to dynamically update the filters relating to the possible range of values of connected attributes to the set of selected features, and achieve this result dynamically at near real-time to create a responsive interface. Creating an entity subgraph for the possible relevant connections to instances of entities will facilitate building an eficient query programmatically for any starting entity class, rather than manually for each possible starting entity type in the interface. It would also allow the interface to adapt as the graph grows.
The second use of entity subgraphs is for future work related to multi-hop entity summarization. For a knowledge graph on environmental data, one goal of the graph is to elucidate the environmental context in which the observed contamination is occurring, and potentially to be able to compare and pinpoint the diferences in environmental context. In the geospatial context especially, single-hop entity summaries don’t capture suficient contextual information, where there is often a nesting of spatial entities. Measurement are associated with spatial entities, but represent their own distinct entities. Design patterns for representing observations and measurements, as described above, also have complex multi-hop contexts. An ideal contextual summary would include not just a filtering of directly related facts, but also an aggregation of associated entities and their range of values. For example the range of PFAS concentration for all samples in an area, together with the range of hydrology features and their properties in the same area, and properties of facilities in the area that are emitting or suspected of using PFAS chemicals together builds a picture of the environmental context. This can be compared to the environmental context at another location, to help elucidate what properties might be impacting the observations. This can lead to new hypothesis about what environmental factors in these complex systems could be contributing to higher or lower accumulation of PFAS concentrations. Acknowledgements This work and the development of SAWGraph have been supported by the National Science Foundation (NSF) under Grant No. 2333782 as part of the Proto-OKN initiative (https://www.proto-okn.net/) and by a Non-Assistance Cooperative Agreement with the USDA-ARS New England Center for Sustained Soil and Water Health. I would like to thank my advisor, Torsten Hahmann, and the entire SKAI Lab and SAWGraph teams.
Declaration on Generative AI The author has not employed any Generative AI tools