CoaKG: A Contextualized Knowledge Graph Approach for Exploratory Search and Decision Making

Information needs can range from basic information used in short-term actions, such as the weekly weather forecast, to information that explains a broader phenomenon, such as the relation between the dollar exchange rate variation and gasoline prices. For instance, decisions involving health and well-being, such as getting vaccines or vacation trips, are also influenced by information-seeking behaviors. Such behavior is part of people’s daily lives [8].

Different information needs require different search strategies that specialized computational tools can support [26]. More recent developments on conversational Artificial Intelligence (AI) have made Large Language Models (LLMs) part of information-seeking tools [20]. Although LLMs are trained to incorporate language patterns, grammar, and semantics effectively, this does not make them reliable sources of information. The appearance of conversational fluency may improve the perception of trustworthiness, but this does not guarantee their reliability [34]. Indeed, users must be aware that the response provided by LLM and machine learning (ML) methods is in fact the most likely (probable) answer (based on its training corpus), regardless of its actual truthfulness.

The central problem addressed in this article is that traditional KGs used for decision-making and exploratory searches often lack the proper qualification of relations and attibutes that affect the accuracy of the answers for the task at hand. Another issue is that users frequently submit ambiguous queries or omit critical contextual elements, such as temporal or spatial information, leading to responses that may not fully address their information needs. This gap creates a significant challenge in ensuring the trustworthiness and usefulness of the answers provided, particularly when users make critical decisions based on the retrieved information. Finally, when the KG contains contextual information, it is still necessary to differentiate this from other types of meta-information.

The primary contribution of this paper is to propose the concept of a Concept Augmented Knowledge Graph (CoaKG) that explicitly represents contexts, coupled with a novel framework that enhances decision-making by supporting context-aware exploratory search. This framework presupposes the existence of a Trust Layer, that filters trusted information according to the context, the purposes, and the policies of the agent making the decision. It does so by addressing ambiguity and incompleteness in queries submitted to the underlying CoaKG to retrieve relevant information for the task at hand. We briefly discuss the decision support architecture, but actual details are out of the scope of this paper.

We argue for the feasability of the approach by presenting a realistic PoC based on a real-world KG, Wikidata. We demonstrate how the CoaKG Query Engine can infer implicit context and expand graph queries to retrieve all relevant contextualized answers. This approach significantly improves the relevance of information, making it more useful for complex decision-making scenarios.

The paper is organized as follows. This section has introduced the problem, motivation, and research goals, highlighting the challenges of traditional KGs in decision-making contexts. Section 2 provides background information, discussing key concepts such as uncertainty in information-seeking, trust in knowledge-based systems, and the limitations of current KG structures. Section 3 details the proposed framework for CoaKG, including the architecture of the Knowledge Layer, Trust Layer, and Decision Layer. Section 4 presents the PoC, demonstrating the effectiveness of the proposed approach through experiments using a subset of Wikidata. Next, Section  5 discusses related work compared with our approach. Finally, Section  6 concludes the paper, summarizing the contributions and outlining future research directions.

Uncertainty is an important concept both in information-seeking and decision-making. Users are engaged in reducing the uncertainty associated with information to make a decision. From philosophical decision theory, decision-making under uncertainty involves concepts related to expected utility and maximizing rational choice regarding the likelihood of different outcomes arising from a particular choice [36]. However, simply consuming more information is not guaranteed to reduce uncertainty [8]. Contextualized information can help reduce uncertainty.

An analysis of the current status of Wikidata [31] concluded that given the characteristics of the Wikidata data structures (such as the lack of references for qualifiers, incorrect specification of property scope in addition to omissions and ambiguities of definitions in constraints) and in the data (violations of constraints that generate incompleteness and possible interpretation errors), it is necessary to have an additional and separate layer to adequately support the trust process.

Given the premises that:

• $\mathrm{■}$

  Crowd-sourced KGs, such as Wikidata, or KGs generated through the automatic integration of multiple information sources with varying degrees of quality and trustworthiness, must be used via a Trust Layer [32, 33] in decision-making scenarios and

• $\mathrm{■}$

  Trust is typically assessed using policies based on hard evidence or reputation-based estimation and these mechanisms should also account for the contextual information, as well as the circumstances and associations related to the decision-making goal, to allow a well-founded trust assessment [3].

This research proposes a theoretical and abstract layered architecture of DSS supported by KG, named CoaKG4DSS, which incorporates a Trust Layer placed between Knowledge and Decision Layers, as shown in Figure 4.

• $\mathrm{■}$

  Decision Layer: Given the goal of performing some action, decision-making occurs above the Trust Layer, which can operate under uncertain environments employing a variety of approaches: explicit programming (anticipate all scenarios), supervised learning (based on training examples), optimization (specify possible decision strategies and a performance measure to be maximized); planning (known models and deterministic problems); and reinforcement learning (strategy is learned according to a performance measure) [28]. The Decision Layer is already present in KBDSS architectures found in the literature [37].

• $\mathrm{■}$

  Trust Layer: The Trust Layer, with trust rules and policies established according to users and task requirements, is essential in ensuring the accuracy and reliability of the information provided by KGs. Trust policies take contextualized answers A as input and generate trusted information A’ that satisfies the task and agent criteria. The Trust Layer transforms information into trusted information by applying task- and user-dependent trust policies using both information and contextual information.

• $\mathrm{■}$

  Knowledge Layer: Enables explicit context representation, retrieval, and interpretation enhancements to support decision-making since context is essential for identifying the scope of a claim that is considered valid and useful. Furthermore, to provide all the information the Trust Layer requires, the knowledge retrieved from the KG should be explicitly contextualized, at least in terms of Provenance, Temporal, and Location dimensions. KGs can incorporate claims representing different perspectives that are apparently contradictory. Therefore, decision-makers must rely on their trust policies to decide what is true and useful in the context of their information needs and tasks.

Due to these characteristics and differences from KBDSS, the Knowledge Layer operates under DOWA. Contrary to OWA, under DOWA, the mere presence of a statement in a KG does not imply its truth. Therefore, DOWA posits that truthfulness is user-dependent trust and is determined by a trust evaluation that depends on the contexts of claims and tasks being carried out or intended.

A CoaKG definition as shown in Definition 1 and the mapping schema as shown in Figure 5 characterize the elements necessary to model explicit contextual information and extract implicit contextual information from a standard KG. The mapping schema demonstrates how various components interact to enhance query responses with both explicit and inferred contexts. The knowledge context, represented by attributes such as provenance, location, and temporal information, is mapped to entity types and concepts through context labels and claim predicates.

A common way to contextualize claims (statements) is by adding property-value pairs as qualifiers to statements. We must distinguish between additive qualifiers, which represent n-ary relationships and do not affect the assessment of the claim’s truthfulness, and contextual qualifiers, which can restrict the contexts in which the underlying claim is considered true and may modify the claim itself [29].

For example, consider the Wikidata statement “Born on the Fourth of July (Q471159) has cast member (P161) Tom Cruise (Q37079) with qualification character role (P453) Ron Kovic (Q530602)”. The character role (P453) qualifier complements the cast member (P161) relationship with additional information but does not affect its truthfulness evaluation. In this case, reference qualifiers such as the reference URL (P854) can help verify the veracity of the information. An example is the link to the catalog of this film on the IMDB platform, which is a popular website about films. On this site, it is possible to consult the entire cast, including their roles. Another reference qualifier associated with reference URL (P854) is retrieved (P813) which specifies when the access to this site was made. Reference qualifiers add provenance context to the statements in Wikidata.

Since explicit contextual information is part of the KG, context mappings specify how it is represented as qualifiers and predicates among claims and metadata. According to the proposed Conceptual Schema, context mappings added to a standard KG transform it into a CoaKG. Similarly to Contextualized Ontologies [6], in CoaKGs, mappings establish the role of each element, i.e., whether it functions as an entity in the standard KG or as context information. The context specification C works as a meta schema above the KG schema, adding semantics to its meta-information.

Three context types are usually found in the literature (Temporal, Location, and Provenance) and are common in several application domains.

• $\mathrm{■}$

  Temporal contexts encompass dates, periods, events, and so on, allowing time-related interpretations. It is possible to differentiate a claim, an entity, or a concept from the others according to the period that it is valid. WHEN-type questions that may arise during the exploratory search can also be answered.

• $\mathrm{■}$

  Location contexts encompass places, coordinates, regions, spatial shapes, etc, enabling spatially-related interpretations. Using location context values, it is possible to differentiate a claim, an entity, or a concept from others about the place where it occurred or belongs. The place can be an entity that represents an object in space, e.g., PUC-Rio Campus Gávea or a class of places, e.g., University. With this dimension, it is possible to answer the WHERE-type questions that may arise during the exploratory search.

• $\mathrm{■}$

  Provenance contexts encompass information sources, agents, processes, methods, and so on, allowing lineage-related interpretations. Using provenance context values, it is possible to differentiate a claim, an entity, or a concept from others regarding its source, origin, or how it was produced, obtained, and calculated. In addition to identifying the source, the user may wish to obtain more information about this source, exploring aspects that may contribute to assessing the veracity and usefulness of the answers. With this contextual information, it is possible to answer questions such as WHO stated, WHEN stated, WHERE stated, HOW the claim was generated, etc.

Note that other domain-specific context types can be added to the schema. Contextual information can also help reveal temporal, spatial, and lineage relations among claims, entities, or concepts not explicitly represented in the KG.

The schema includes agents, such as KG engineers or users who carry out explorations, who define semantic rules that infer an implicit context. Rules can be represented as an implication between an antecedent (or IF clause or BODY rule) and a consequent (or THEN or HEAD rule). The semantic rules consist of a set of claims, its explicit context, and operators ($\wedge$, v, $\mathrm{\nexists }$) as antecedents that infer a relative context as consequent. The inferred and default context is used to refine the retrieved information, ensuring that responses are accurate and contextually relevant.

The contents of the Knowledge Layer can be queried through the Context Augmented Query Engine, which uses a Graph Query Language (GQL) to query the underlying KG to retrieve both trust policies and contextualized knowledge (i.e., claims and metadata). Such an engine also includes an Inference Engine to execute Semantic Rules using Contextual mappings and Co-domain Algebra functions associated with each type of Context. These mechanisms are used to answer potentially incomplete graph queries concerning context information.

The Trust Layer receives the user query and converts it into a graph query K to be used as input to the CoaKG Query Engine as shown in figure 6. CoaKG Query Engine transforms the original graph query K, which is potentially contextually incomplete, into a contextually complete query. Any query K is considered contextually complete if all mappings created by the KG Engineer that reference elements mentioned in the original query are used to retrieve all context information represented in the KG. Furthermore, the query can also be modified to add hidden relationships using specific Co-domain Algebra operations generating an expanded graph query K.

A Co-domain Algebra defines a finite set of operations over co-domain values in various contexts, enabling the inference of relationships not directly materialized in the KG. These relationships between two elements of a KG (nodes or edges) can then be explicitly represented in the query response by adding new edges. An example of Co-domain Algebra for date values in the temporal context is Allen’s Interval Algebra  [1], which defines 13 basic relationships between time intervals, such as before, after, during, starts, ends, among others. Another example, pertaining to the location context, is the Region Connection Calculus [12], which specifies spatial relationships such as intersection, inclusion, disconnection, touch, and more. Specific algebraic operations enable the explicit representation of implicit relationships across contextual information. Examples of such implicit relations are claims co-occurrence in time or spatial overlap between entities.

The CoaKG query engine generates two versions of the expanded graph query K: exact and approximated, to be executed in steps B5 and B6. The exact graph query K’ assumes that KG is complete and all mandatory context information are represented and can be retrieved by this query. The approximated graph query K” deals with KG incompleteness and optional (not mandatory) context information. Three types of modification can be applied: (i) OPTIONAL clause for context qualifiers, properties, and relations, (ii) replace constant with variables for context values, and (iii) check context values for NULL to substitute by DEFAULT VALUES in WHERE clauses. After executing the queries, it is possible to apply semantic rules defined by the KG engineer or by the user performing the exploration. The rules are applied to infer the implicit context among the set of responses to produce all possible contextualized responses. We illustrate these steps in Section 4.

We present more details about the CoaKG schema and the CoaKG Query Engine in [13]. In that article we illustrated the transformation of a contextual incomplete query into a complete one through an example of exploratory search about capital cities of Brazil during the colonial period³³3The previous work focused on presenting the CoaKG schema and the internal mechanisms of the CoaKG Query Engine. In this paper, we substantially broaden the scope by introducing the complete architectural framework - including the Trust Layer - and by providing a PoC to assess its applicability in knowledge-based decision support scenarios..

This section details our PoC for the CoaKG at the Knowledge Layer, showcasing its application in enhancing exploratory search and decision-making. We outline the methodology for building the CoaKG from a subset of Wikidata, including the profiling procedure to identify contextual predicates (properties or relationships). Competency questions were used to evaluate the Knowledge Layer operation and demonstrate how the CoaKG Query Engine retrieves and contextualizes answers, highlighting its effectiveness in providing context-aware information for informed decision-making.

Since anyone can build a CoaKG by extending an existing standard KG, we used a subset of Wikidata to demonstrate this pathway. We chose Wikidata because it serves as a rich knowledge repository, consolidating information from primary sources along with their references and connections to other databases. For example, Wikidata includes data about historical events, linking details from official records and academic publications to related entities and concepts, facilitating a comprehensive understanding of the topic.

Although Wikidata contains contextual information, its ontology does not explicitly identify which qualifier represents each type of context. Also, Wikidata is designed to represent provenance as a reference, but there is also Provenance information among qualifiers. For our PoC, context types were specified based on the result of a Wikidata profiling procedure identified relevant contextual predicates. These were then used to specify the context mappings.

The Wikidata graph data structure represents statements of claims associated with qualifiers and references, but does not establish how users should retrieve, interpret, and use them. For this PoC, a set of Competency Questions was elaborated to generate queries with and without context, and their respective answers were obtained.

We agree that the quality of statements/claims is crucial in DSS. Still, different from them, in our proposed architecture for DSS the Trust Layer should deal with potentially inconsistent, conflicting, and contradictory data using context information. CoaKG, under the DOWA, represents different perspectives that can be applied to different decision-making scenarios. The Knowledge Layer does not resolve apparent conflicts among the data, as the Trust Layer should handle this according to the user and task requirements. Conflict resolution in the Knowledge Layer can introduce bias and restrict the user’s decision-making power.

KGs have been used in various tasks, and some recent works about recommendation systems (RS) incorporate KGs to alleviate the item cold-start and sparsity problems. However, KG-based tasks, applications, and state-of-the-art KGR methods need to consider that real-world KGs inevitably end up having a large amount of untrustworthy information [16]. The authors developed a trustworthiness-aware KG representation method called TrustRec.

TrustRec leverages internal structures of KGs to calculate a trustworthiness estimator, which gives the degree of triples’ certainty and integrates it into noise-tolerant KGR and item representations for RS. Similarly, the Trust Layer, included in the proposed architecture, can compute a trust indicator based on the trust policies of the decision-making task to rank the contextualized answers retrieved from the Knowledge Layer. This computation may consider both the KG instances and the context mappings.

Regarding exploratory search applications, it is crucial to consider that users often need to become more familiar with the KG structure, actual contents, and the domain of the information they are exploring. Solutions to address these challenges have been proposed based on approximate methods, query expansion, and query refinement techniques.

TriniT is a KG exploratory querying system [41] that addresses vocabulary mismatch problems using rules for query relaxation. Users are often unfamiliar with the KG structure and the labels of entities, classes, and relations. This phenomenon makes it difficult for users to formulate proper graph queries. Their relaxation rules translate the original triple pattern into a semantic similar one that can be answered. Besides, TriniT also treats KG incompleteness at query time using Open Information Extraction (OIE) tools to extract triples from external sources. The eXtend KG (XKG) comprises the original KG and new triples associated with their provenance context. Although the new triples have lower confidence than triples from the original KG, both contribute to the final answer.

The CoaKG Query Engine also considers that users may be unfamiliar with the KG structure and the information needed. Graph queries generated by the user or system interface or the Trust Layer may need to be completed regarding context. The original graph query K is evaluated in relation to entity types, concepts, predicates, and qualifiers against the context mappings C and can be expanded to retrieve context information. Co-domain Algebra for context values can also modify the graph query when the graph pattern is disconnected to extract implied relations. KG incompleteness regarding contextual information is addressed through semantic rules I and Co-domain Algebra, inferring implicit context.

Another approach found in the literature involves interactive graph query expansion over KGs [24]. Sample query results are used as input to generate the top k-most relevant expansions. The original query input is complemented with edges using their Bag-of-Labels Model for Graph definition inspired by how language models expand keyword queries. Two unsupervised methods (Labels and pseudo-relevance feedback) rank query expansions based on estimated relevance.

Unlike this work, CoaKG Query Engine expands a graph query K with context information, adding qualifiers for claims and predicates for entities based on the context mappings the KG Engineer added to the KG. All Contextualized Answers A are retrieved and sent to the Trust Layer, which can apply trust policies and rules using context information to rank or filter the contextualized claims and entities according to user and task context of interest.

Concerning knowledge context representation and manipulation of triple-based KGs, Homola et al., 2010 [21] proposed contextualized knowledge repository (CKR). A CKR is a pair K = <D, C> where C represents the set of contexts, and D is a Description Logic KB. CKR adopts the theoretical perspective of context as a multidimensional space and the metaphor of context as a box that contains triples associated with the same context. The CoaKG also contemplates explicit contextual information associated with claims, represented as triples, besides entities and concepts. Unlike CKR, CoaKG adopts a multi-layer graph data structure that enables in-depth exploration of context information navigating through n-layers. Additionally, implicit context can be inferred using rules and Co-domain Algebra at query time by CoaKG Query Engine.

In this section, we review our results and contributions, and point out future work.

Considering the flow of information for decision-making presented (Figure 2) and the opportunity to use KGs as knowledge repositories, some relevant issues related to the context needed to be addressed:

• $\mathrm{■}$

  How can each element type be differentiated if the knowledge repository consists of entities and claims with respective meta-information regarding the context?

• $\mathrm{■}$

  How to identify the different types of context represented in the knowledge graph?

• $\mathrm{■}$

  How to identify the lack of information about a relevant context for decision-making?

• $\mathrm{■}$

  If queries do not specify context, how can they be modified to retrieve the relevant context?

• $\mathrm{■}$

  If context information is implicit and not materialized in the repository, how can it be inferred using the retrieved information?

Our research answered theses questions, preparing and querying KGs to support context-aware exploration in decision-making scenarios. We designed a framework for CoaKG4DSS composed of a Decision Layer, a Trust Layer, and a Knowledge Layer. Our observation of how KGs operate in practice led to the formulation of DOWA, hitherto implicit in crowd-sourced KGs such as Wikidata. DOWA asserts that the mere presence of a statement in a KG does not warrant its truthfulness.

Overall, the research considered some premises: (i) Graph queries can be contextually incomplete; (ii) KGs can be contextually incomplete in a general sense and also regarding contextual information; (iii) implicit context can be extracted from KGs based on semantic rules and context values operations; (iv) CoaKG should operate under DOWA; and (v) contextual information, absent or already available in KGs, should be explicitly identified to enable a more informed decision by users regarding trust.

A graph query K, issued so that the information retrieved at least partially satisfies an information need, may be complete or incomplete with respect to context specification. In the case of incomplete queries, the proposed CoaKG Query Engine uses the context mappings C to expand the graph query K to retrieve all Contextualized Answers A from H’.

This PoC is a small-scale approach used to demonstrate the feasibility of an idea. Our PoC aims to demonstrate that the Knowledge Layer is capable of providing Contextualized Answers to the Trust Layer, enabling the construction of DSSs based on CoaKGs. Since we cannot rely on a knowledge expert, we chose a domain that is relatively easy to understand, explore, and develop CQs. We selected a subset of Wikidata, restricted to the Geopolitical History domain, using a method of filtering nodes, claims, and qualifiers to generate a standard KG represented in the multilayer model. Then, we applied our KG Engineering steps using this standard KG as input to identify the explicit context, built the mappings using the proposed schema, and added these mappings to the KG itself, transforming it into a CoaKG. We also developed two examples of rules to infer the implicit context. The CQs were manually converted into graph queries that were completed by simulating the operation of the proposed CoaKG Query Engine. The scripts and graph queries developed, as well as the datasets and contextualized responses generated throughout the PoC are publicly available on GitHub, allowing reproducibility.

We acknowledge that our PoC currently focuses on Wikidata and does not explicitly evaluate CoaKG against other existing KGs. Given the scope of this work, we frame this aspect as future work.

Although the POC only used a subset of Wikidata, we claimed that any KG, crowd-sourced or constructed by integrating several different information sources, should operate under DOWA and be consumed through a Trust Layer. However, the Trust Layer was not detailed for the PoC, being indicated as future work. We are currently evaluating the use of the Reference Ontology of Trust (ROT) [4] to specify trust policies. ROT can answer questions about the nature of trust and its influential factors. Trust policies can be used to identify trustworthiness evidence that the decision maker agent (trustor) would like to receive to trust Contextualized Answers A (trustee). It could also be used by the Trust Layer to generate trust calibration signals that transform Contextualized Answers A into Trusted Information A’.