Super-relation knowledge graph reasoning method and system based on structure perception hierarchical conformal prediction

Abstract

This invention relates to the field of intelligent question answering technology, and particularly to a super-relational knowledge graph reasoning method and system based on structure-aware hierarchical conformal prediction. The method utilizes a training set to train a structured representation learning model and obtains a candidate entity set and corresponding candidate entity scalar score vectors for training queries. An automatic temperature scaling calibration of the base probability distribution is introduced. Queries are dynamically routed to different structure groups based on the number of qualifiers, and group-specific conformal thresholds are learned. Test queries are routed to corresponding subsets based on the qualifier density of the test queries, and the conformal thresholds of the subsets are used to obtain a candidate entity prediction set to evaluate the accuracy of the structured representation model's knowledge reasoning. Models whose accuracy meets preset conditions are used as target models for knowledge reasoning and deployed to the intelligent question answering system. This invention enables refined and adaptive uncertainty graph modeling, facilitating the construction of a reliable reasoning intelligent question answering system with both powerful representational capabilities and rigorous statistical guarantees.

Description

Technical Field

[0001] This invention relates to the field of intelligent question answering technology, and in particular to a super-relationship knowledge graph reasoning method and system based on structure-aware hierarchical conformal prediction. Background Technology

[0002] Knowledge graphs (KGs), as structured carriers of human knowledge, have become the cornerstone of modern artificial intelligence systems. However, knowledge in the real world is often intricate and complex, and traditional binary relation triples are insufficient to contain fine-grained information such as the time, place, and conditions under which facts occur. Hyper-Relational KGs (HKGs) have emerged to address this, achieving precise characterization of complex facts by attaching a series of key-value pair "qualifiers" to the main triples. For example, in intelligent question-answering systems, the fact (Geoffrey Hinton, works at, Google) is often accompanied by qualifiers such as {(role, Vice President), (time, 2013-2023)}, thereby eliminating semantic ambiguity.

[0003] In recent years, significant progress has been made in representation learning research targeting HKG. Structured coding models, such as MAYPL, have achieved state-of-the-art (SOTA) ranking performance on link prediction tasks by explicitly modeling the interaction between the main triples and qualifiers. However, existing HKG inference methods generally suffer from a critical, overlooked flaw: the models focus only on "prediction accuracy" and lack rigorous quantification of "prediction trustworthiness." In challenging intelligent question-answering scenarios such as inductive reasoning, dot product scoring models often exhibit a severe tendency of "overconfidence," assigning extremely high confidence levels to incorrect predictions. This unreliable confidence estimation hinders the practical application of HKG in high-risk intelligent question-answering fields such as healthcare and finance.

[0004] Introducing conformal prediction (CP) is an effective way to solve this problem, as it can generate a prediction set with statistical coverage guarantees. However, directly applying standard CP to HKG faces a unique challenge—structural heterogeneity. In HKG, the semantic ambiguity of a query is negatively correlated with its qualifier density: queries with rich qualifiers are clearly targeted (low uncertainty), while queries lacking qualifiers are ambiguous (high uncertainty). This makes it impossible to achieve the "tailored" management of HKG and to provide reliable guarantees in practical high-risk intelligent question answering applications. Summary of the Invention

[0005] To address the problem that existing models often exhibit a serious tendency of "overconfidence" when facing complex inductive reasoning scenarios, making it difficult to meet the reliability requirements of trustworthy AI in high-risk intelligent question answering fields, this invention provides a super-relationship knowledge graph reasoning method and system based on structure-aware hierarchical conformal prediction. By deeply mining the inherent structural information of HKG, it achieves refined and adaptive uncertainty modeling, thereby constructing a trustworthy reasoning intelligent question answering system with both powerful representation capabilities and rigorous statistical guarantees, meeting the application needs of high-risk scenarios such as healthcare and finance.

[0006] According to the design scheme provided by this invention, on the one hand, a super-relationship knowledge graph reasoning method based on structure-aware hierarchical conformal prediction is provided, comprising:

[0007] Training and test sets are obtained based on the target domain hyper-relation knowledge graph. The hyper-relation knowledge graph is obtained by extracting hyper-relation facts from a given target domain document. The training set contains question-answer training query pairs consisting of some known entities and their relationships. The training set contains question-answer test query pairs consisting of entities that did not appear in the training phase. The hyper-relation facts consist of entity triples and a set of qualifiers.

[0008] The structured representation learning model is trained using the training set and the candidate entity set and corresponding candidate entity scalar score vectors for the training query are obtained. The optimal temperature parameter is obtained using the calibration set and the probability distribution of the candidate entity scalar score is calibrated using the optimal temperature parameter. The calibration set and the training set are independent but have the same data distribution.

[0009] The calibration set is divided into several non-overlapping subsets based on the qualifier density of the training queries in the training set, so as to split the given query into each subset, and the conformal threshold of each subset is calculated based on the calibrated probability distribution.

[0010] Extract the qualifier density of test queries in the test set, route the test queries to the corresponding subsets based on the qualifier density, and call the conformal threshold of the subsets to obtain the candidate entity prediction set. Evaluate the accuracy of knowledge reasoning of the structured representation model based on the candidate entity prediction set of the test queries, and use the model whose accuracy meets the preset conditions as the target model for knowledge reasoning and deploy it to the intelligent question answering system.

[0011] As part of the super-relational knowledge graph reasoning method based on structure-aware hierarchical conformal prediction of this invention, further, a training set and a test set are obtained based on the target domain super-relational knowledge graph, including:

[0012] Extract entity triples and qualifier sets from multi-source heterogeneous target domain documents, and construct a hyperrelationship knowledge graph for the target domain;

[0013] The hyperrelationship knowledge graph is divided into training graph and test graph according to preset conditions, and training set and test set are constructed based on the training graph and test graph. The preset conditions are entity type and / or timestamp.

[0014] As a super-relational knowledge graph reasoning method based on structure-aware hierarchical conformal prediction of the present invention, the structured representation learning model adopts the MAYPL encoder to capture the local topological structure between super-relational facts and entities through the MAYPL encoder and the message passing mechanism, and obtain the candidate entity set and the corresponding candidate entity variable score vector for a given query.

[0015] As a super-relational knowledge graph reasoning method based on structure-aware hierarchical conformal prediction in this invention, it further utilizes a MAYPL encoder to obtain a candidate entity set and corresponding candidate entity variable score vectors for a given query, including:

[0016] For a given query in the training / test set, aggregate the context and qualifier set attribute information of the main triples to construct a query vector;

[0017] For a given query, extract the neighborhood subgraph of a known entity, and obtain all entities and relationships that are directly or indirectly related to the known entity, and form a query candidate pool.

[0018] For each candidate entity in the candidate pool, obtain its co-occurrence frequency, connection pattern, and position role in the neighborhood subgraph and initialize the entity vector. The co-occurrence frequency is used to characterize how many super-relation facts the candidate entity appears in, the connection pattern is used to characterize the relation type / entity type connected by the candidate entity, and the position role is used to characterize the position / component in the set of qualifiers of the candidate entity in the principal triple.

[0019] Multi-level message passing is performed on the neighborhood subgraph to obtain context-aware candidate entity representation vectors by fusing higher-order neighborhood information;

[0020] Calculate the semantic similarity between the query vector and the representation vector of each candidate entity, sort them by score, and obtain the candidate entity set.

[0021] As part of the super-relational knowledge graph reasoning method based on structure-aware hierarchical conformal prediction of this invention, the method further includes obtaining optimal temperature parameters using a calibration set and calibrating the probability distribution of candidate entity scalar scores using these optimal temperature parameters, comprising:

[0022] On the calibration set, the optimal temperature parameter is automatically optimized by minimizing the negative log-likelihood.

[0023] The candidate entity scalar score vector is scaled using the optimal temperature parameter and a Softmax function is applied to obtain the calibrated candidate entity scalar score probability distribution.

[0024] As a super-relational knowledge graph reasoning method based on structure-aware hierarchical conformal prediction of this invention, further comprising dividing the calibration set into several non-overlapping subsets according to the qualifier density of the training queries in the training set, and further including:

[0025] Samples with a qualifier density greater than a threshold are assigned to the last subset, where the threshold is the same as the number of subsets.

[0026] As a super-relational knowledge graph reasoning method based on structure-aware hierarchical conformal prediction of this invention, further, the conformal threshold of each subset is calculated based on the calibrated probability distribution, including:

[0027] For each sample in each subset, a non-consistency score is calculated based on the calibrated probability distribution, which is described using the minimum cumulative probability quality required to include the true answer;

[0028] Based on the inconsistent scores of the samples, the empirical quantiles of the scores within each subset are obtained, and the conformal threshold specific to that subset is obtained based on the empirical quantiles.

[0029] Furthermore, this invention also provides a super-relationship knowledge graph reasoning system based on structure-aware hierarchical conformal prediction, comprising: a sample acquisition module, a model training module, a structure-aware module, and a model deployment module, wherein...

[0030] The sample acquisition module is used to acquire training and test sets based on the target domain hyperrelation knowledge graph. The hyperrelation knowledge graph is obtained by extracting hyperrelation facts from a given target domain document. The training set contains question-answer training query pairs samples consisting of some known entities and their relationships. The training set contains question-answer test query pairs samples consisting of entities that did not appear during the training phase. The hyperrelation facts consist of entity triples and a set of qualifiers.

[0031] The model training module is used to train the structured representation learning model using the training set and obtain the candidate entity set and corresponding candidate entity scalar score vectors for the training query. It uses the calibration set to obtain the optimal temperature parameters and uses the optimal temperature parameters to calibrate the probability distribution of the candidate entity scalar scores. The calibration set and the training set are independent but have the same data distribution.

[0032] The structure-aware module is used to divide the calibration set into several non-overlapping subsets based on the qualifier density of the training queries in the training set, so as to divide a given query into subsets and calculate the conformal threshold of each subset based on the calibrated probability distribution.

[0033] The model deployment module is used to extract the qualifier density of test queries in the test set, route the test queries to the corresponding subsets based on the qualifier density, and call the conformal threshold of the subsets to obtain the candidate entity prediction set. Based on the candidate entity prediction set of the test queries, the accuracy of the structured representation model knowledge reasoning is evaluated, and the model whose accuracy meets the preset conditions is used as the knowledge reasoning target model and deployed to the intelligent question answering system.

[0034] The beneficial effects of this invention are:

[0035] This invention introduces Auto-Temperature Scaling to calibrate the underlying probability distribution, then dynamically routes queries to different structure groups based on the number of qualifiers, and learns group-specific conformal thresholds. This endows the model with adaptive truncation capabilities, generating extremely compact prediction sets for semantically explicit queries while maintaining a wide search range for semantically ambiguous queries to ensure no missed detections. Furthermore, extensive experiments on 12 highly challenging inductive datasets containing varying proportions of new relations demonstrate that the proposed ReliableHKG solution, while strictly meeting user-preset coverage (e.g., 95%), significantly reduces the global average prediction set size by over 30% (with redundancy reduction reaching as high as 80% on specific query groups), outperforming traditional global calibration baselines. This lays a solid foundation for the reliable deployment of HKG in high-risk scenarios such as healthcare and finance. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the super-relationship knowledge graph reasoning process based on structure-aware hierarchical conformal prediction in the embodiment.

[0037] Figure 2 This is an example of the complexity and uncertainty of the HKG query structure in the embodiment;

[0038] Figure 3 This is a schematic diagram of the ReliableHKG hyperrelationship knowledge graph reasoning algorithm in the embodiment.

[0039] Figure 4 This example illustrates a comparison of the inductive link prediction performance of ReliableHKG and the baseline model on the dataset.

[0040] Figure 5 This example illustrates a comparison of the average prediction set size for different structural groupings of the WK-100 and FB-100 datasets.

[0041] Figure 6 The cumulative probability distribution of queries A and B and the truncation process of different conformal mechanisms are visualized in the example. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.

[0043] Traditional knowledge graphs (KGs) are typically represented as directed relation graphs, where facts exist in the form of binary relation triples, i.e. However, this simplified representation cannot capture the multidimensional context of facts in the real world (such as time, place, conditions, etc.). Hyperrelational knowledge graphs (HKGs) overcome this limitation by introducing "qualifiers".

[0044] In HKG, a fundamental fact involves more than just a primary triplet. It also comes with a qualifier set. Specifically, the set of qualifiers is represented as ,in This represents the number of quantifiers. Each quantifier is a key-value pair, where... This is called a qualifier relationship (i.e., attribute category, such as "occurrence time"). This is called a qualifier entity (i.e., a specific attribute value, such as "2008"). Therefore, a complete hyperrelational fact can be formalized as a tuple. The entire hyperrelationship knowledge graph can be defined as:

[0045]

[0046] in It is an entity set. and These are the master relation set and the qualifier relation set, This represents the power set.

[0047] In link prediction tasks, a query is defined as an incomplete fact that a certain entity (usually a tail entity) is missing. For tail entity prediction, the query is represented as: To quantify the structural complexity and information richness of a query, the qualifier density of the query is defined as a set. The cardinality of is denoted as . This metric plays a crucial role in this article: The larger the value, the stronger the contextual constraints of the query, and the more certain the semantic space of the corresponding true answer (i.e., the lower the ambiguity).

[0048] Hyper-Relational Knowledge Graphs (HKGs) significantly enrich the expressive power and contextual precision of factual knowledge by associating principal triples with qualifiers in the form of key-value pairs. While recent structured representation learning models have achieved state-of-the-art ranking performance on link prediction tasks, they generally lack uncertainty quantification of prediction results. This leads to a severe tendency for models to exhibit overconfidence when facing complex inductive reasoning scenarios, making it difficult to meet the reliability requirements of trustworthy AI in high-risk domains. Furthermore, traditional conformal prediction methods typically employ globally uniform calibration thresholds, failing to adapt to the inherent structural heterogeneity in HKGs—that is, the semantic ambiguity of queries is often negatively correlated with their qualifier density.

[0049] like Figure 2 As shown on the left, query A contains only the core triple (Albert Einstein, nationality, ?), without any qualifiers. This simple query structure has highly ambiguous semantics—Einstein's nationality could be German, Swiss, or American (as he held multiple nationalities throughout his life)—resulting in a wide answer space and thus high uncertainty in the model's prediction. Conversely, Figure 2 Query B on the right not only contains the core triple (Marie Curie, received, Nobel Prize), but also includes two key qualifiers: (award category, Physics Prize) and (year of award, 1903). These qualifiers significantly narrow down the range of answers, making the query highly targeted and able to pinpoint a unique answer (such as the 1903 Nobel Prize in Physics), thus resulting in lower uncertainty in the model's predictions. From query A to query B, the qualifier density gradually increases, while the prediction uncertainty decreases accordingly—a typical manifestation of the structural heterogeneity of HKG. Query A has a simple structure, no qualifiers, a large answer space, and high uncertainty; query B has a complex structure, contains multiple qualifiers, a small answer space, and low uncertainty. Qualifier density is negatively correlated with uncertainty. (a) Simple queries: containing only principal triples (e.g., (?, occupation, professor)), no qualifiers, large answer space (high uncertainty); (b) Complex queries: containing principal triples and multiple qualifiers (e.g., (?, award, Nobel Prize), (year of award, 2021), (field of award, physics)), small answer space (low uncertainty). The comparison illustrates the negative correlation between qualifier density and uncertainty.

[0050] If the standard CP's globally unified threshold is used, then... Figure 2 The two types of queries presented present a dilemma: for the fuzzy query A, the threshold may be too aggressive, resulting in an overly small prediction set (missing correct answers such as Germany, Switzerland, or the United States); for the specific query B, the threshold may be too conservative, resulting in an overly large prediction set (including irrelevant entities such as Nobel Prizes from other award categories in 1903 or other years). This "one-size-fits-all" calibration approach ignores the inherent structural heterogeneity of HKG, fails to achieve tailored management, and is ill-suited to provide reliable assurance in high-risk applications.

[0051] Therefore, in the embodiments of the present invention, such as Figure 1 As shown, a super-relationship knowledge graph reasoning method based on structure-aware hierarchical conformal prediction is provided, including:

[0052] S101. Obtain training and test sets based on the target domain hyperrelation knowledge graph. The hyperrelation knowledge graph is obtained by extracting hyperrelation facts from a given target domain document. The training set contains question-answer training query pairs consisting of some known entities and their relationships. The training set contains question-answer test query pairs consisting of entities that did not appear during the training phase. The hyperrelation facts consist of entity triples and a set of qualifiers.

[0053] This example focuses on the most challenging inductive reasoning scenario. In a traditional transductive setting, the training graph and the test graph share the same set of entities (i.e., In the inductive setting, the model can directly learn the embedding representation of a specific entity. However, in the inductive setting, the model learns from the training set... Optimizations were performed on the existing test set, but on a test set containing entirely new entities. An evaluation was conducted on the above (i.e.) (Or only partially overlapping). This means the model must completely abandon the memory of entity IDs and instead learn general subgraph structural patterns. Since the test entities are unseen, inductive reasoning inherently contains extremely high uncertainty.

[0054] Specifically, entity triples and qualifier sets can be extracted from multi-source heterogeneous target domain documents, and a super-relationship knowledge graph of the target domain can be constructed. The super-relationship knowledge graph is divided into training graphs and test graphs according to preset conditions, and training sets and test sets are constructed based on the training graphs and test graphs. The preset conditions are entity type and / or timestamp.

[0055] This can be achieved by continuously collecting multi-source heterogeneous texts containing structured data, unstructured data, and real-time dynamic data in the medical or financial fields, and using natural language processing technology to extract entity, relationship, and attribute-related data from the multi-source heterogeneous texts, so as to obtain a super-relationship knowledge graph in the corresponding target domain through graph construction.

[0056] The main symbols involved in this scheme and their physical meanings are shown in Table 1.

[0057] Table 1: Explanation of Core Symbols

[0058] symbol describe Traditional knowledge graphs From the head entity Tail entity ,relation The triplet Quantifier set This represents the number of quantifiers. Each quantifier is a key-value pair, where... This is called a qualifier relationship (i.e., attribute category, such as "occurrence time"). This is called a qualifier entity (i.e., a specific attribute value, such as "2008"). A complete hyperrelational fact Hyperrelationship knowledge graph. Among them... It is an entity set. and These are the master relation set and the qualifier relation set, This represents the power set. In link prediction tasks, a query is defined as an incomplete fact that a certain entity (usually a tail entity) is missing. To quantify the structural complexity and information richness of a query, the qualifier density of the query is defined as a set. cardinality ， ， ， Training entity set, test entity set, training set, test set Standard link prediction scoring function Each query Output a prediction set that statistically satisfies the user-preset edge coverage rate. It is a real tail entity. The user-tolerable error rate Expected size of the prediction set A heuristic scoring function is used to evaluate candidate answers. With query The degree of "mismatch". A calibration set that is independent of but identically distributed with the training set. Inconsistency scores for each real question-answer pair in the calibration set The empirical quantile of the inconsistency score for each real question-answer pair in the calibration set. A brand new test query Prediction set

[0059] Standard link prediction tasks aim to learn a scoring function. It outputs Top-1 or Top-K results by sorting the candidate entity set. However, this point estimation cannot inform the user of the magnitude of the predicted risk.

[0060] S102. Train the structured representation learning model using the training set and obtain the candidate entity set and corresponding candidate entity scalar score vectors for the training query. Obtain the optimal temperature parameter using the calibration set and calibrate the probability distribution of the candidate entity scalar score using the optimal temperature parameter. The calibration set and the training set are independent but have the same data distribution.

[0061] Credible reasoning does not require the model to provide a single, definitive answer, but rather requires the model to provide a consistent answer for each query. Output a prediction set. This set must statistically satisfy the user-preset marginal coverage:

[0062]

[0063] in, It is a real tail entity. The user's tolerable error rate (e.g.) (Corresponding to a 95% confidence level). Under the premise of strictly satisfying the above inequalities, the optimization objective is to minimize the expected size of the prediction set. The smaller the set, the stronger the model's ability to eliminate incorrect options, and the higher the information efficiency it provides.

[0064] S103. Based on the qualifier density of the training query in the training set, divide the calibration set into several non-overlapping subsets to split the given query into subsets, and calculate the conformal threshold of each subset based on the calibrated probability distribution.

[0065] Conformal prediction (CP) is a statistical framework that provides a strict guarantee of finite sample coverage without making assumptions about the data distribution. Its core relies on the assumption that data points satisfy the exchangeability assumption between the calibration and test sets. The standard split-CP process includes the following key steps:

[0066] 1. Non-conformity Measure (NCM): First, a heuristic scoring function is needed. Used to measure candidate answers With query The degree of "mismatch". A higher score indicates a greater degree of mismatch. The less likely it is to be the true answer.

[0067] 2. Score Calibration: Given a calibration set that is independent of but identically distributed with the training set. Calculate the inconsistency score for each real question-answer pair in the calibration set. .

[0068] 3. Empirical quantile calculation: To ensure... Coverage, calculate the empirical quantiles of these scores. :

[0069]

[0070] 4. Prediction Set Construction: For a brand new test query Its prediction set is defined as all inconsistent scores less than or equal to The set of entities:

[0071]

[0072] In classification and knowledge graph completion tasks, a widely used NCM (Non-Conformance Model) algorithm is the Adaptive Prediction Sets (APS) algorithm based on cumulative probability. However, standard CP implicitly assumes that all input queries... They are homogeneous in terms of uncertainty, therefore a globally uniform threshold is used. In HKG, the query structure (i.e., qualifier density) This leads to significant heterogeneity. Using a single global threshold to simultaneously adapt to both fuzzy queries (requiring a large set) and specific queries (requiring a small set) inevitably results in a decrease in overall prediction efficiency.

[0073] S104. Extract the qualifier density of the test queries in the test set, route the test queries to the corresponding subsets based on the qualifier density of the test queries, and call the conformal threshold of the subsets to obtain the candidate entity prediction set. Evaluate the accuracy of the knowledge reasoning of the structured representation model based on the candidate entity prediction set of the test queries, and use the model whose accuracy meets the preset conditions as the knowledge reasoning target model and deploy it to the intelligent question answering system.

[0074] Inductive reasoning inherently contains extremely high uncertainty, and the inherent structural heterogeneity of hyperrelational knowledge graphs (HKG) leads to a significant decrease in the efficiency of standard conformal prediction methods. To achieve reliable predictions in complex scenarios facing high uncertainty, such as... Figure 2 The ReliableHKG framework is shown. Unlike traditional knowledge graph completion methods that only focus on the absolute order of nodes, ReliableHKG constructs an end-to-end system from structural feature extraction to probability calibration and then to hierarchical adaptive truncation. The framework consists of three tightly coupled stages: (1) Structured encoding: using the MAYPL backbone network to aggregate main triples and qualifiers The information outputs the original score vector. (2) Distribution calibration: By minimizing the negative log-likelihood (NLL) through automatic temperature scaling, the overconfidence tendency of the model is suppressed, and the unbounded scores are transformed into a smooth and calibrated probability distribution. (3) Structure-aware hierarchical reasoning: using dynamic routers to extract the qualifier density of queries. The query will be routed to the corresponding structure group. This is achieved by applying a group-specific conformal threshold. The system can generate a final prediction set that strictly meets the coverage requirements and is adaptive in size. .

[0075] The framework's operation can be naturally decoupled into three core steps. First, the framework needs to obtain the topological representation of the graph, extract subgraph patterns independent of entity IDs using a structured encoder, and output the original score vector for entity ranking. Second, to eliminate confidence distortion caused by high-dimensional dot products, a distribution calibration layer is introduced to map extreme scores to a probability distribution that conforms to the true accuracy, providing a reliable underlying statistical foundation for subsequent conformal prediction. Finally, and this is the core innovation of the framework, the structure-aware hierarchical reasoning mechanism. Since the qualifier density of different queries directly determines their semantic ambiguity, a "Dynamic Router" is used to divert queries to different structure groups and independently calculate group-specific conformal thresholds. This mechanism truly achieves "personalized instruction," minimizing the expected size of the prediction set while strictly ensuring global coverage.

[0076] In inductive link prediction, the model must be able to generalize to unseen entities. In this embodiment, MAYPL is selected as the basic encoder. Because it relies purely on message passing to capture the local topology between facts and entities, without remembering the global entity embedding.

[0077] Specifically, the MAYPL encoder is used to obtain the candidate entity set and the corresponding candidate entity variable score vector for a given query, which can be designed to include:

[0078] For a given query in the training / test set, aggregate the context and qualifier set attribute information of the main triples to construct a query vector;

[0079] For a given query, extract the neighborhood subgraph of a known entity, and obtain all entities and relationships that are directly or indirectly related to the known entity, and form a query candidate pool.

[0080] For each candidate entity in the candidate pool, obtain its co-occurrence frequency, connection pattern, and position role in the neighborhood subgraph and initialize the entity vector. The co-occurrence frequency is used to characterize how many super-relation facts the candidate entity appears in, the connection pattern is used to characterize the relation type / entity type connected by the candidate entity, and the position role is used to characterize the position / component in the set of qualifiers of the candidate entity in the principal triple.

[0081] Multi-level message passing is performed on the neighborhood subgraph to obtain context-aware candidate entity representation vectors by fusing higher-order neighborhood information;

[0082] Calculate the semantic similarity between the query vector and the representation vector of each candidate entity, sort them by score, and obtain the candidate entity set.

[0083] Given a query MAYPL aggregates the context and qualifier set of the principal triple. Attribute information for candidate entity sets. any entity in The model outputs a scalar score (Logit):

[0084]

[0085] The scores of all candidate entities are denoted as vectors. .Although The relative ranking can be directly used for Top-K truncation, but point estimation is prone to failure when faced with high uncertainty in inductive scenarios. To apply conformal prediction, a reasonable non-consistency measure (NCM) needs to be constructed. Its non-probabilistic nature hinders this process.

[0086] In conformal prediction based on adaptive prediction sets (APS), inconsistency scores often depend on cumulative probabilities. If we directly apply the original Logits... Applying the standard Softmax function can easily lead to an "overconfident" probability distribution in the model due to the numerical explosion of high-dimensional dot products—that is, the maximum value approaches 1, and the rest approach 0. This distorted distribution can cause the CP algorithm to prematurely truncate the accumulation process, resulting in severe under-coverage.

[0087] To address this issue, this embodiment introduces temperature scaling as a probabilistic calibration layer to automatically optimize the optimal temperature parameter on the calibration set by minimizing the negative log-likelihood. The optimal temperature parameter is then used to scale the candidate entity scalar score vector, and a softmax function is applied to obtain the calibrated candidate entity scalar score probability distribution.

[0088] Specifically, the calibrated probability distribution can be defined as follows:

[0089]

[0090] in It is a temperature parameter. To obtain the probability of "honesty" (i.e., the prediction confidence level aligning with the true accuracy), in a separate calibration set... The optimal temperature is automatically found by minimizing the negative log-likelihood (NLL). :

[0091]

[0092] go through After scaling, confidence overflow caused by high-frequency errors is effectively suppressed. This provides a smooth and reliable probability measure, laying a solid mathematical foundation for subsequent precise hierarchical conformal inference.

[0093] Qualifier density This determines the degree of ambiguity in the query. The standard global CP method uses a uniform quantile threshold for all queries. This not only penalizes information-rich, specific queries (forcing them to output redundant answers), but also fails to accurately characterize the uncertainty boundaries of fuzzy queries. Therefore, in this embodiment, the structure-aware hierarchical strategy can be summarized as follows:

[0094] Step 1, Structural Stratification: Based on the structural features of the query (i.e., qualifier density), the calibration set is grouped... Divided into Non-overlapping subsets (structure groups):

[0095]

[0096] For long-tail samples with too many qualifiers (i.e.) To ensure the statistical sample size, they are uniformly classified into Group.

[0097] Step 2, Group-wise Threshold Learning: In each structural group... Internally, using calibrated probabilities Construct inconsistent scores. For the first score within a group... The non-consistency score of each sample Defined as containing true answers Minimum cumulative probability mass required:

[0098]

[0099] Then, each group was calculated independently. The empirical quantiles of the inner fractions yield the group-specific conformal threshold. :

[0100]

[0101] The elegance of this mechanism lies in its adaptability: for semantically explicit groups (high... Model predictions are usually quite concentrated. Smaller, learned Naturally, the requirements are more stringent; while for fuzzy groups ( ), It will automatically loosen to accommodate long-tail uncertainty.

[0102] Step 3, Adaptive Inference: In the inductive testing phase, for entirely new test queries... The framework first acts as a "Dynamic Router," extracting its qualifier density. The system then routes the query to the corresponding structure group rule and invokes the dedicated threshold. Construct the final prediction set:

[0103]

[0104] in, It is the th in descending order Highly probable. Through this structure-aware mechanism, ReliableHKG strictly adheres to global... While implementing the coverage theorem, it also achieved Minimize.

[0105] Furthermore, based on the above method, this embodiment of the invention also provides a super-relationship knowledge graph reasoning system based on structure-aware hierarchical conformal prediction, comprising: a sample acquisition module, a model training module, a structure-aware module, and a model deployment module, wherein,

[0106] The sample acquisition module is used to acquire training and test sets based on the target domain hyperrelation knowledge graph. The hyperrelation knowledge graph is obtained by extracting hyperrelation facts from a given target domain document. The training set contains question-answer training query pairs samples consisting of some known entities and their relationships. The training set contains question-answer test query pairs samples consisting of entities that did not appear during the training phase. The hyperrelation facts consist of entity triples and a set of qualifiers.

[0107] The model training module is used to train the structured representation learning model using the training set and obtain the candidate entity set and corresponding candidate entity scalar score vectors for the training query. It uses the calibration set to obtain the optimal temperature parameters and uses the optimal temperature parameters to calibrate the probability distribution of the candidate entity scalar scores. The calibration set and the training set are independent but have the same data distribution.

[0108] The structure-aware module is used to divide the calibration set into several non-overlapping subsets based on the qualifier density of the training queries in the training set, so as to divide a given query into subsets and calculate the conformal threshold of each subset based on the calibrated probability distribution.

[0109] The model deployment module is used to extract the qualifier density of test queries in the test set, route the test queries to the corresponding subsets based on the qualifier density, and call the conformal threshold of the subsets to obtain the candidate entity prediction set. Based on the candidate entity prediction set of the test queries, the accuracy of the structured representation model knowledge reasoning is evaluated, and the model whose accuracy meets the preset conditions is used as the knowledge reasoning target model and deployed to the intelligent question answering system.

[0110] To verify the effectiveness of this solution, the following explanation is based on experimental data:

[0111] The experimental design revolves around the following three core research questions (RQs):

[0112] RQ1 (Efficiency and Robustness): When faced with demanding inductive scenarios involving a large number of unseen entities and new relationships, can ReliableHKG outperform existing baselines and maintain efficient predictions while providing strict coverage guarantees?

[0113] RQ2 (Mechanism Adaptability): Can the "structure-aware hierarchical calibration" mechanism in this case correctly capture the correlation between query structure (qualifier density) and uncertainty, thereby achieving adaptive truncation and improving information efficiency?

[0114] RQ3 (Component Necessity): Is the automatic temperature scaling module indispensable for correcting the "overconfident" distribution of the dot product model?

[0115] To comprehensively evaluate the model's generalization and uncertainty quantification capabilities under different distribution shifts, extensive experiments were conducted on 12 inductive hyperrelation datasets: NL-25 / 50 / 75 / 100, WK-25 / 50 / 75 / 100, and FB-25 / 50 / 75 / 100.

[0116] NL, WK, and FB correspond to knowledge sources in different domains. The suffix numbers (25, 50, 75, 100) represent the proportion of new relations in the test set. For example, WK-100 means that all relations in the test graph did not appear during the training phase. This setting greatly increases the epistemic uncertainty and is an ideal touchstone for testing trustworthy AI.

[0117] The algorithm ReliableHKG proposed in this case is comprehensively compared with the current state-of-the-art methods in this field, as follows:

[0118] Standard baselines: Includes all standard hyperrelation inference baselines (such as StarE, RAM, etc.) introduced in the HyRel paper.

[0119] State-of-the-art inductive reasoning models: HyRel and MAYPL.

[0120] It is worth emphasizing that the algorithm ReliableHKG in this case is a model-agnostic general-purpose reliable inference framework. To explore the performance ceiling of this framework, MAYPL, currently the most powerful ranking algorithm, was selected as the underlying structured feature extractor in the experiments. A target error rate was set. (That is, the target coverage rate is 95%).

[0121] To comprehensively and scientifically evaluate model performance, the experimental evaluation system is divided into two orthogonal dimensions: basic ranking ability and reliable reasoning ability.

[0122] 1) Basic Ranking Metrics: Used to evaluate the model's relative ranking accuracy for candidate entities (i.e., point estimation ability). These include:

[0123] MRR (Mean Reciprocal Rank): The average of the reciprocals of the correct answer's rank across all test queries. A higher MRR indicates better overall ranking quality from the model.

[0124] Hit10 and Hit1: These represent the proportion of correct answers appearing in the top 10 and number 1 of the model's predictions, respectively. Following the latest trends in knowledge graph representation, Hit10 and Hit1 are used instead of the earlier Hit@10 and Hit@1 expressions.

[0125] 2) Reliability & Efficiency Metrics: These are the core evaluation dimensions used to assess the framework's performance in providing statistical security guarantees. They include:

[0126] Actual Coverage: The proportion of queries that contain the true answer in the prediction set "C(q)". Ideally, it should be strictly greater than or equal to the target confidence level (e.g., 95%).

[0127] Average Set Size: The average number of entities contained in the candidate set output by the model while meeting coverage requirements. The smaller this value, the more redundant options the model eliminates, and the higher the information efficiency.

[0128] 1. Basic sorting performance evaluation (answer RQ1 - sorting dimension)

[0129] To objectively address the premise regarding model validity in RQ1, a comprehensive comparison was made between ReliableHKG and the state-of-the-art inductive reasoning methods in the existing literature, including HyRel [Ref] and all its referenced baseline models, as well as the latest MAYPL model. Figure 4 The basic ranking performance of each model on 12 inductive datasets is presented, with baseline data directly cited from existing literature to ensure the fairness of the comparison.

[0130] Figure 4 In the table, the number appended to each dataset indicates the proportion of new relations. The symbol "-" indicates a model reproduced by the proposed solution; to ensure fair comparison, the remaining baselines directly report the best results from the original paper. Bold and underlined values ​​represent the best and second-best results, respectively. For example... Figure 4 As shown, thanks to its powerful structured encoding module, ReliableHKG demonstrates dominant ranking performance in demanding inductive scenarios. Compared to state-of-the-art baseline models (such as HyRel and MAYPL), the framework achieves a significant performance leap when faced with a large number of new relations, specifically in the following two aspects:

[0131] 1) Robust generalization in extreme, unseen environments: In extreme scenarios where 100% of the test graph relationships are entirely new (i.e., datasets with the -100 suffix), traditional models are prone to failure, while ReliableHKG's subgraph pattern capture capability has withstood the test. Particularly on the structurally complex FB-100 dataset, the framework's MRR metric jumped dramatically from 0.282 on HyRel to 0.341 (an absolute improvement of 0.059); and on the Hit1 metric, it improved from 0.188 to 0.236 (an absolute improvement of 0.048). In the field of knowledge graph inductive reasoning, this leap in absolute accuracy exceeding 0.04 is a landmark achievement.

[0132] 2) Complete dominance in local precision and global recall: Even in scenarios containing partially known relationships (such as the -75, -50, and -25 series), ReliableHKG maintains a crushing advantage. For example, on NL-50, compared to the strong baseline MAYPL (MRR: 0.343), our solution achieved an MRR of 0.357; and on the extremely difficult WK-75 dataset, the framework not only broke through to 0.206 in local precision (Hit1) (surpassing HyRel's 0.187), but its global recall metric (Hit10) also soared to 0.432 (an absolute improvement of 0.043 compared to HyRel). These quantitative data irrefutably demonstrate that the structured message passing mechanism upon which our solution relies has superior performance in capturing topological patterns of unseen entities.

[0133] Despite ReliableHKG's superior performance on ranking metrics, traditional evaluation paradigms mask the cognitive uncertainty inherent in inductive reasoning. For example, on WK-100, even with state-of-the-art MRR of 0.112 and Hit1 of 0.061 achieved through architectural innovation, from a physics application perspective, this means that the model's absolute highest score prediction (Top-1) still has nearly a 93.9% (1 - 0.061) probability of being completely wrong.

[0134] This extremely low absolute prediction accuracy highlights the fatal flaws of forcibly performing "deterministic point estimation" in an open-world context. Neither HyRel nor MAYPL, traditional baseline models, can answer the core question of "how reliable is this prediction?" This profound experimental insight completely refutes the illusion of solving the risks of inductive reasoning simply by improving the model architecture, thus forcing the introduction of conformal predictions in subsequent mitigation efforts and a systematic analysis of the reliability and statistical safety boundaries of ReliableHKG.

[0135] 2. Trustworthy Reasoning Performance and Structure-Aware Analysis (Answering RQ1 - Trustworthy Dimension & RQ2)

[0136] This study aims to answer the RQ1-credibility dimension and explore the adaptability of the structure-aware mechanism in RQ2, examining the framework's performance under a 95% target coverage constraint. Since traditional ranking baselines (such as HyRel) cannot provide coverage guarantees, the following internal comparison baselines were constructed:

[0137] 1) MAYPL (Top-10): The traditional fixed-length truncation strategy (forces the output of the top 10 entities with the highest scores).

[0138] 2) Global APS: Standard conformal prediction baseline based on global threshold.

[0139] The experimental results are shown in Table 2.

[0140] Table 2: Comparison of Trusted Reasoning Efficiency in Inductive Scenarios (Target Coverage) 95%)

[0141] Observing the data in Table 2, it is clear that the traditional MAYPL (Top-10) strategy suffered a catastrophic security failure, with an actual coverage of only 12.1% on FB-100. This once again confirms that simply improving sorting accuracy cannot meet the reliability requirements of high-risk scenarios. In contrast, ReliableHKG robustly maintained a security baseline of nearly 95% on all datasets, providing conclusive data to demonstrate the absolute effectiveness and robustness of the framework in dealing with extreme uncertainty, and fully answering RQ1.

[0142] To further understand the underlying mechanism of ReliableHKG's efficiency improvement and to accurately answer RQ2, test queries were divided into three structural groups based on their "qualifier density (|Q|)". The truncation performance of each group was tracked on an extreme dataset. The experimental results are shown in Table 3. Figure 5 As shown.

[0143] Table 3: Comparison of information efficiency under different qualifier densities in the extreme inductive datasets (WK-100, FB-100) (target coverage ≥ 95%)

[0144] Figure 5 In this paper, we compare the average prediction set size of different structural groups on the WK-100 and FB-100 datasets. Traditional Global APS is constrained by the global threshold, resulting in a large candidate set in all groups; while ReliableHKG exhibits significant asymmetric adaptive truncation capability, achieving up to 80% redundancy reduction in semantically clear specific query groups (Group≥2), greatly improving the information efficiency of graph inference.

[0145] Combining the detailed data in Table 3 with the intuitive visual comparison in Figure 5, an asymmetric truncation phenomenon can be clearly observed. Global APS, due to its "one-size-fits-all" globally uniform threshold, is "hijacked" by a large number of high-uncertainty samples in Group 0, resulting in its inability to handle semantically clear specific queries (Group 0). This also outputs an extremely redundant set. This "punishing the high achievers" strategy leads to extremely poor global information efficiency.

[0146] In contrast, ReliableHKG successfully achieves precise routing tailored to individual needs. For fuzzy queries without any contextual constraints (Group 0), the framework honestly provides a slightly larger set to handle extremely high cognitive uncertainty; however, once the query possesses structural information gain from qualifiers (Group 1 and Group ≥2), the dynamic router astutely captures this collapse of ambiguity and automatically applies extremely stringent group-specific thresholds. This structure-aware, fine-grained tuning allows the framework to reduce the size of specific query groups by nearly 80% without compromising 95% actual coverage within the group, ultimately driving a reduction of over 30% in the global average set size. This empirical result perfectly answers RQ2, conclusively demonstrating the superior adaptability of the core mechanism presented in this paper.

[0147] 3. Ablation experiment: The effect of temperature scaling (Answer RQ3)

[0148] To directly and thoroughly answer the question in RQ3 regarding the necessity of the automatic temperature scaling module, a strip test was performed on this component. This compared the results without using temperature scaling (forced setting). ) and temperature automatically optimized using the framework ( The actual coverage of the system changes when the system is in a certain state, and the experimental results are shown in Table 4.

[0149] Table 4: Ablation Experiment Results of Temperature Scaling Module

[0150] The ablation results in Table 4 empirically reveal the inherent statistical flaws of the dot product model. On all tested datasets, the temperature parameter driving the optimal calibration of the model is significantly greater than 1 (e.g., as high as 1.58 on FB-100). This key value directly exposes a serious "overconfidence" bias in the original graph neural network—that is, the model is extremely inclined to assign extreme spurious probabilities close to 1.0 to incorrect head entities.

[0151] The disastrous consequences of removing this module further confirm its irreplaceable nature. Once temperature scaling (i.e.) is removed... This distorted underlying probability distribution propagates upwards along the system pipeline, leading to severe misjudgments in the calculation of inconsistency scores. When predicting cumulative probabilities conformally, the system mistakenly assumes sufficient probability quality has been included and prematurely triggers truncation, ultimately causing the actual coverage to plummet below the 95% tolerance threshold (e.g., a sharp drop to 81.4% on FB-100). Therefore, the automatic temperature scaling module is not merely an edge trimming, but a crucial component that forcibly transforms unbounded feature similarity into "honest" probabilities strictly aligned with the true error. As the underlying foundation supporting subsequent accurate hierarchical truncation, it perfectly answers RQ3.

[0152] 4. Qualitative Case Analysis

[0153] To intuitively reveal the working mechanism of ReliableHKG at the micro level, representative query pairs were extracted from the test set for qualitative case analysis. These query pairs target the same primary triple, but are presented as an ambiguous state without qualifiers and a concrete state containing rich qualifiers, respectively. Table 5 compares the truncation behavior of different methods when processing this set of queries.

[0154] Table 5: Qualitative Case Analysis Regarding the Residence of Entity "Albert Einstein"

[0155]

[0156] Experimental results are as follows Figure 6 As shown, the cumulative probability distributions of queries A and B, as well as the truncation process of different conformal mechanisms, are visualized. The left side (a) shows the smooth probability distribution of highly ambiguous queries. Global APS misses true values ​​due to insufficient thresholds (red stars), while ReliableHKG's high threshold ensures safe coverage. The right side (b) shows the steep distribution of low-ambiguity queries. ReliableHKG accurately truncates Top-1 by lowering the threshold, effectively removing the gray redundant areas generated by Global APS.

[0157] Through Table 5 and Figure 6 As shown, this set of cases profoundly reveals the necessity of solving the problem of structural heterogeneity.

[0158] When faced with query A (fuzzy query), due to the lack of time constraints, all the cities where Einstein lived are plausible, and the cumulative probability curve output by the model is extremely flat (Figure 6(a)). At this point, the fixed truncation strategy (Top-10) and the Global APS strategy using a globally uniform threshold are compared. The search was stopped too early, resulting in the true value "Princeton" being missed when it was ranked 15th. ReliableHKG, however, identified the high uncertainty and raised the threshold to 0.99, successfully including the true value, demonstrating a safety bottom line of "better to be broad than to miss a value".

[0159] However, when the query evolves into query B (the specific query), the qualifier... The significant collapse of the understanding space greatly increases the probability of "Princeton" being the first to fall. Figure 6 (b)). At this point, the inherent flaws of Global APS are fully exposed: constrained by a rigid global threshold, it still "inertiafully" accumulates the subsequent 5 obviously conflicting irrelevant options (the gray shaded area in the figure) even when the truth value has been hit with extremely high confidence. In contrast, ReliableHKG's dynamic router decisively calls the option targeted at... stringent thresholds ( This directly reduced the prediction set to 1. The case clearly demonstrates that structure-aware hierarchical structures endow the model with a powerful reasoning ability, much like a human expert, to adaptively and asymmetrically adjust its cognitive breadth based on the amount of known contextual clues.

[0160] The experimental results demonstrate that the proposed solution successfully overcomes the efficiency bottleneck caused by a globally uniform threshold by automatically calibrating the underlying probabilities through temperature scaling and utilizing qualifier density for dynamic routing and group-specific threshold learning. In extensive evaluations across 12 inductive datasets encompassing multiple domains and varying proportions of new relations, the framework robustly provides 95% statistical coverage even under extreme conditions where basic ranking capabilities face significant challenges. Furthermore, it endows the system with asymmetric adaptive truncation capabilities tailored to specific needs, significantly reducing redundant candidate options without sacrificing security, thus providing a robust safety net for graph-based question answering and decision support in high-risk fields such as healthcare and finance.

[0161] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0162] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0163] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0164] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0165] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A super-relational knowledge graph reasoning method based on structure-aware hierarchical conformal prediction, characterized in that, Include: Training and test sets are obtained based on the target domain hyper-relation knowledge graph. The hyper-relation knowledge graph is obtained by extracting hyper-relation facts from a given target domain document. The training set contains question-answer training query pairs consisting of some known entities and their relationships. The training set contains question-answer test query pairs consisting of entities that did not appear in the training phase. The hyper-relation facts consist of entity triples and a set of qualifiers. The structured representation learning model is trained using the training set and the candidate entity set and corresponding candidate entity scalar score vectors for the training query are obtained. The optimal temperature parameter is obtained using the calibration set and the probability distribution of the candidate entity scalar score is calibrated using the optimal temperature parameter. The calibration set and the training set are independent but have the same data distribution. The calibration set is divided into several non-overlapping subsets based on the qualifier density of the training queries in the training set, so as to split the given query into each subset, and the conformal threshold of each subset is calculated based on the calibrated probability distribution. Extract the qualifier density of test queries in the test set, route the test queries to the corresponding subsets based on the qualifier density, and call the conformal threshold of the subsets to obtain the candidate entity prediction set. Evaluate the accuracy of knowledge reasoning of the structured representation model based on the candidate entity prediction set of the test queries, and use the model whose accuracy meets the preset conditions as the target model for knowledge reasoning and deploy it to the intelligent question answering system.

2. The super-relationship knowledge graph reasoning method based on structure-aware hierarchical conformal prediction according to claim 1, characterized in that, Training and test sets are obtained based on the target domain hyper-relation knowledge graph, including: Extract entity triples and qualifier sets from multi-source heterogeneous target domain documents, and construct a hyperrelationship knowledge graph for the target domain; The hyperrelationship knowledge graph is divided into training graph and test graph according to preset conditions, and training set and test set are constructed based on the training graph and test graph. The preset conditions are entity type and / or timestamp.

3. The super-relational knowledge graph reasoning method based on structure-aware hierarchical conformal prediction according to claim 1, characterized in that, The structured representation learning model employs the MAYPL encoder to capture the local topological structure between hyperrelational facts and entities through a message passing mechanism, and to obtain a set of candidate entities and corresponding candidate entity variable score vectors for a given query.

4. The super-relationship knowledge graph reasoning method based on structure-aware hierarchical conformal prediction according to claim 3, characterized in that, The MAYPL encoder is used to obtain the candidate entity set and the corresponding candidate entity variable score vector for a given query, including: For a given query in the training / test set, aggregate the context and qualifier set attribute information of the main triples to construct a query vector; For a given query, extract the neighborhood subgraph of a known entity, and obtain all entities and relationships that are directly or indirectly related to the known entity, and form a query candidate pool. For each candidate entity in the candidate pool, obtain its co-occurrence frequency, connection pattern, and position role in the neighborhood subgraph and initialize the entity vector. The co-occurrence frequency is used to characterize how many super-relation facts the candidate entity appears in, the connection pattern is used to characterize the relation type / entity type connected by the candidate entity, and the position role is used to characterize the position / component in the set of qualifiers of the candidate entity in the principal triple. Multi-level message passing is performed on the neighborhood subgraph to obtain context-aware candidate entity representation vectors by fusing higher-order neighborhood information; Calculate the semantic similarity between the query vector and the representation vector of each candidate entity, sort them by score, and obtain the candidate entity set.

5. The super-relationship knowledge graph reasoning method based on structure-aware hierarchical conformal prediction according to claim 1, characterized in that, The optimal temperature parameters are obtained using a calibration set, and the probability distribution of the candidate entity scalar score is calibrated using these optimal temperature parameters. This includes: On the calibration set, the optimal temperature parameter is automatically optimized by minimizing the negative log-likelihood. The candidate entity scalar score vector is scaled using the optimal temperature parameter and a Softmax function is applied to obtain the calibrated candidate entity scalar score probability distribution.

6. The super-relational knowledge graph reasoning method based on structure-aware hierarchical conformal prediction according to claim 1, characterized in that, The calibration set is divided into several non-overlapping subsets based on the qualifier density of the training queries in the training set, which also include: Samples with a qualifier density greater than a threshold are assigned to the last subset, where the threshold is the same as the number of subsets.

7. The super-relationship knowledge graph reasoning method based on structure-aware hierarchical conformal prediction according to claim 1, characterized in that, The conformal threshold for each subset is calculated based on the calibrated probability distribution, including: For each sample in each subset, a non-consistency score is calculated based on the calibrated probability distribution, which is described using the minimum cumulative probability quality required to include the true answer; Based on the inconsistent scores of the samples, the empirical quantiles of the scores within each subset are obtained, and the conformal threshold specific to that subset is obtained based on the empirical quantiles.

8. A super-relational knowledge graph reasoning system based on structure-aware hierarchical conformal prediction, characterized in that, It includes: a sample acquisition module, a model training module, a structure-aware module, and a model deployment module. The sample acquisition module is used to acquire training and test sets based on the target domain hyperrelation knowledge graph. The hyperrelation knowledge graph is obtained by extracting hyperrelation facts from a given target domain document. The training set contains question-answer training query pairs samples consisting of some known entities and their relationships. The training set contains question-answer test query pairs samples consisting of entities that did not appear during the training phase. The hyperrelation facts consist of entity triples and a set of qualifiers. The model training module is used to train the structured representation learning model using the training set and obtain the candidate entity set and corresponding candidate entity scalar score vectors for the training query. It uses the calibration set to obtain the optimal temperature parameters and uses the optimal temperature parameters to calibrate the probability distribution of the candidate entity scalar scores. The calibration set and the training set are independent but have the same data distribution. The structure-aware module is used to divide the calibration set into several non-overlapping subsets based on the qualifier density of the training queries in the training set, so as to divide a given query into subsets and calculate the conformal threshold of each subset based on the calibrated probability distribution. The model deployment module is used to extract the qualifier density of test queries in the test set, route the test queries to the corresponding subsets based on the qualifier density, and call the conformal threshold of the subsets to obtain the candidate entity prediction set. Based on the candidate entity prediction set of the test queries, the accuracy of the structured representation model knowledge reasoning is evaluated, and the model whose accuracy meets the preset conditions is used as the knowledge reasoning target model and deployed to the intelligent question answering system.

9. An electronic device, characterized in that, include: At least one processor, and a memory coupled to said at least one processor; The memory stores a computer program that can be executed by the at least one processor to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, enables the implementation of the method as described in any one of claims 1 to 7.

Images