A local culture history space-time evidence three-state output method and system

Abstract

The application provides a local cultural history space-time evidence three-state output method and system, relates to the field of artificial intelligence and natural language processing technology, the method is to receive the evidence demand set extracted by user query, generate the first interception state label according to the knowledge base boundary perception degree; based on the first interception state label, user query and candidate evidence, the historical space-time conflict degree is obtained by fusion, and the second interception state label is generated according to the historical space-time conflict degree; and according to the missing information recoverability index, the shunt state label containing the clarification request state and the hard gap refusal state is generated, the three labels are comprehensively obtained, and the loss parameter is obtained; based on the evidence support degree and the loss parameter, and combined with the shunt state label, the structured three-state result containing the decision label and the routing track is output. The application solves the problems that the existing local culture field question and answer method is easy to produce error determination answer, fact paradox filtering ability is poor, interactive mode is rigid and inflexible, and decision process is difficult to trace back.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing, and in particular to a method and system for outputting three-state evidence of local cultural history in time and space. Background Technology

[0002] With the development of natural language processing technology and large language models, retrieval-enhanced generation (RAG)-based question answering systems have been widely applied in various vertical fields. However, when facing application scenarios with strong spatiotemporal attributes and complex entity relationships, such as local culture and historical changes, existing question answering systems and information retrieval methods still have obvious limitations.

[0003] First, existing systems generally lack a precise awareness of the boundaries of their own knowledge base capabilities. When a user's question exceeds the range of evidence types, modalities, or timeframes covered by the underlying database, the system often relies on the internal illusions of the large model to forcibly piece together an answer, resulting in an extremely high false positive rate, failing to achieve the secure gating of knowing what is known and not knowing what is unknown.

[0004] Secondly, when retrieving and verifying candidate evidence, most existing technologies rely solely on superficial text vector similarity or keyword matching. Local cultural resources have a rich historical heritage; the same location may have different ancient names in different dynasties, as seen in place name evolution, and the same person may have different relationship networks at different times. Based solely on literal similarity, the system cannot identify the historical and spatiotemporal logical conflicts hidden in the evidence and the issue across time spans, geographical changes, and complex interpersonal relationships. This easily leads to seemingly relevant but actually misattributed historical materials being output as the final evidence.

[0005] Furthermore, the interaction patterns of existing question-and-answer systems are too simplistic and rigid. When faced with insufficient user query conditions or missing key information, traditional systems typically adopt a one-size-fits-all, rigid rejection strategy, or ignore the missing conditions and output a one-sided answer. The system fails to make a deep judgment on the nature of the missing information, and cannot distinguish between common sense omissions that users can easily fill in and hard gaps that the system's underlying data simply does not support, thus missing the opportunity to salvage the task flow through clarification and follow-up questions.

[0006] Finally, in the final scoring and output decision-making stage, existing solutions mostly adopt a simple linear weighted accumulation mechanism. The initial similarity of evidence often directly dominates the output result, lacking a pre-emptive risk reduction mechanism with veto power. At the same time, its black-box output also lacks a record of the judgment process, making it difficult to audit and review the system decision. Summary of the Invention

[0007] To address the aforementioned shortcomings in existing technologies, this invention provides a three-state output method and system for local cultural historical spatiotemporal evidence. This solves the problems of existing question-and-answer systems in the field of local culture, which lack awareness of system capability boundaries, in-depth examination of historical spatiotemporal logical conflicts, and dynamic assessment of the recoverability of missing information. As a result, these systems are prone to generating incorrect answers, have poor ability to filter factual fallacies, have rigid and inflexible interaction methods, and have difficulty in tracing the decision-making process.

[0008] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a method for three-state output of local cultural historical spatiotemporal evidence, comprising: S1: Receive the set of evidence requirements extracted by the user query, calculate the coverage degree using the pre-built evidence base capability profile and the set of evidence requirements, obtain the knowledge base boundary awareness, and generate the first interception status label based on the knowledge base boundary awareness. S2: Based on the first interception status label, user query and pre-acquired candidate evidence, when the first interception status label is not intercepted, the distance between time intervals, the distance of place name evolution path and the distance of entity relationship path are calculated by using the local cultural historical spatiotemporal topology, and the historical spatiotemporal conflict degree is obtained by fusion, and the second interception status label is generated according to the historical spatiotemporal conflict degree. S3: Based on the second interception status label, the entity slot queried by the user, and the preset task completion conditions, the state machine matching algorithm is used to calculate the recoverability index of missing information, and a diversion status label including the clarification request status and the hard gap rejection status is generated according to the recoverability index of missing information. S4: Based on the first interception status label, the second interception status label, and the diversion status label, when none of them trigger the interception and hard gap rejection status, extract the knowledge base boundary awareness, historical spatiotemporal conflict degree, and pre-acquired knowledge gap degree generated in the preceding steps as loss parameters. S5: Based on the pre-acquired evidence support and attenuation parameters, the attenuation calculation is performed using a multiplicative attenuation model to obtain the definitive answer score. Combined with the diversion status label, the structured three-state results containing decision labels and routing trajectories are output to complete the three-state output of local cultural historical spatiotemporal evidence.

[0009] Further, S1 includes: Based on the evidence requirement set extracted from user queries, a pre-built evidence base capability profile containing evidence type support matrix, field completeness vector, modal capability identifier, and timeliness capability identifier is used to calculate the uncovered ratio of the evidence requirement set, obtain the knowledge base boundary awareness, and generate the first interception status label based on the knowledge base boundary awareness.

[0010] Further, S2 includes: Based on the first interception status label, the user query, and the pre-acquired candidate evidence, when the first interception status label is not intercepted, the local cultural history spatiotemporal topology structure containing ancient and modern place name nodes and their evolutionary relationship edges, entity nodes and their person relationship edges is used to obtain the place name evolutionary path distance through shortest path analysis between nodes, obtain the entity relationship path distance through entity relationship path comparison processing, and combine it with time interval distance comparison processing to obtain the historical spatiotemporal conflict degree, and generate the second interception status label based on the historical spatiotemporal conflict degree.

[0011] Furthermore, the expression for the historical spatiotemporal conflict degree is: ; ; ; in, Indicates the degree of historical spatiotemporal conflict. This represents an exponential function with base e. This represents the preset time weighting coefficient. This represents the preset place name weight coefficient. This represents the preset relational weight coefficients. This represents the time interval distance, specifically the overlap penalty value between the query time and the evidence time interval obtained based on the time axis mapping comparison. Indicates the distance along the path of the place name's evolution. Indicates the path distance of entity relationships. This represents the shortest path length in the local cultural and historical spatiotemporal topology between the ancient and modern place name nodes associated with the user query and the place name nodes associated with the candidate evidence, obtained based on the graph search algorithm. This represents the maximum graph diameter constant of the preset topology. This represents the similarity between the query-corresponding entity relationship path obtained based on graph feature embedding comparison and the evidence-corresponding entity relationship path.

[0012] Further, S3 includes: Based on the second interception status label, the entity slot queried by the user, and the preset task completion conditions, the slot comparison calculation is performed using state machine matching logic that includes recoverable missing information identifiers and hard gap identifiers to obtain the recoverability index of missing information. Slots with missing real-time data and no applicable modality are determined to have hard gap identifiers. Based on the recoverability index of missing information, the diversion status label including the clarification request status and the hard gap rejection status is generated.

[0013] Further, S5 includes: Based on the pre-acquired evidence support and the attenuation parameter, the attenuation calculation is performed using the multiplicative attenuation model to obtain the determined answer score. Combined with the diversion status label, the structured three-state result containing the decision label and the routing trajectory is output to complete the three-state output of local cultural historical spatiotemporal evidence. The routing trajectory is log data generated by recording the data processing link, which includes pre-gating trigger conditions, various scores, and the final output state.

[0014] Furthermore, the expression for the multiplication loss model is: ; in, This indicates that the answer has been confirmed and a score has been awarded. This represents the initial support score of the candidate evidence obtained based on the semantic matching model or slot coverage statistical algorithm for the user query. Indicates the knowledge base boundary awareness. Indicates the degree of historical spatiotemporal conflict. It represents the knowledge gap degree, which is the gap penalty factor obtained by quantitative mapping based on the recoverability index of missing information.

[0015] This invention provides a three-state output system for spatiotemporal evidence of local cultural history, comprising: The first tag generation module is used to receive the evidence requirement set extracted by the user query, calculate the coverage degree using the pre-built evidence base capability profile and the evidence requirement set, obtain the knowledge base boundary awareness, and generate the first interception status tag based on the knowledge base boundary awareness. The second tag generation module is used to calculate the time interval distance, place name evolution path distance and entity relationship path distance based on the first interception status tag, user query and pre-acquired candidate evidence when the first interception status tag is not intercepted, using the local cultural historical spatiotemporal topology to obtain the historical spatiotemporal conflict degree, and generate the second interception status tag based on the historical spatiotemporal conflict degree. The diversion label generation module is used to calculate the recoverability index of missing information based on the second interception status label, the entity slot queried by the user and the preset task completion conditions, using a state machine matching algorithm, and to generate diversion status labels including clarification request status and hard gap rejection status based on the recoverability index of missing information. The parameter acquisition module is used to extract the knowledge base boundary awareness, historical spatiotemporal conflict degree, and pre-acquired knowledge gap degree generated in the preceding steps as loss parameters based on the first interception state label, the second interception state label, and the diversion state label, when none of them have triggered the interception or hard gap rejection state. The output module is used to calculate the score of the definitive answer based on the pre-acquired evidence support and the attenuation parameter using the multiplicative attenuation model. It also combines the diversion status label to output a structured three-state result containing the decision label and the routing trajectory, thus completing the three-state output of local cultural historical spatiotemporal evidence.

[0016] The beneficial effects of this invention are as follows: This invention provides a three-state output method for local cultural history spatiotemporal evidence. By pre-constructing a capability profile of the evidence base containing multiple dimensions and calculating the coverage of user needs in the pre-process stage, this method can trigger the first interception state in advance when the knowledge base capabilities do not match, completely blocking the model illusion caused by exceeding the scope, and greatly reducing the probability of incorrectly determining the answer.

[0017] Breaking the limitations of pure text similarity, it places candidate evidence and user queries within the spatiotemporal topology of local culture and history. By calculating the network path distance of time intervals, the evolution of place names from ancient to modern times, and entity relationships, it measures the degree of historical spatiotemporal conflict, accurately identifying hidden chronological misalignments and geographical factual errors, thus ensuring the rigor of cultural dissemination.

[0018] Instead of abruptly refusing to answer, the system calculates the recoverability index of missing information based on task completion conditions. Through state machine matching logic, it finely divides missing slots into clarification request states and hard gap refusal states. This allows the system to proactively guide users to supplement information when necessary, greatly improving the coherence of multi-turn dialogues and the problem-solving rate.

[0019] In the final score calculation stage, a multiplicative attenuation model is introduced, using the aforementioned boundary awareness, conflict degree, and gap degree as penalty attenuation terms to prevent unreliable evidence from crossing the security threshold. At the same time, a structured result containing the routing trajectory is output, which fully records the entire process from gating interception to state transition, providing a reliable audit basis for manual review and system optimization. Attached Figure Description

[0020] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is a schematic diagram of a three-state output system for spatiotemporal evidence of local cultural history, as shown in some embodiments of this specification. Figure 2 This is an exemplary flowchart of a three-state output method for local cultural historical spatiotemporal evidence, as shown in some embodiments of this specification. Detailed Implementation

[0021] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0022] Example 1 Figure 1 This is a schematic diagram of a three-state output system for spatiotemporal evidence of local cultural history, as shown in some embodiments of this specification.

[0023] In some embodiments, the local cultural history spatiotemporal evidence three-state output system may include a first tag generation module, used to receive a set of evidence requirements extracted by a user query, calculate the coverage using a pre-built evidence base capability profile and the set of evidence requirements, obtain the knowledge base boundary awareness, and generate a first interception status tag based on the knowledge base boundary awareness; a second tag generation module, used to calculate the time interval distance, place name evolution path distance, and entity relationship path distance using the local cultural history spatiotemporal topology when the first interception status tag is not intercepted, based on the first interception status tag, the user query, and pre-acquired candidate evidence, and fused to obtain the historical spatiotemporal conflict degree, and generate a second interception status tag based on the historical spatiotemporal conflict degree; a diversion tag generation module, used to calculate the time interval distance, place name evolution path distance, and entity relationship path distance using the local cultural history spatiotemporal topology when the first interception status tag is not intercepted, and generate a second interception status tag based on the historical spatiotemporal conflict degree; and a diversion tag generation module, used to calculate the second interception status tag, the user query, and the first interception status tag based on the second interception status tag and the first interception status tag. The system uses entity slots and preset task completion conditions, employs a state machine matching algorithm to calculate the recoverability index of missing information, and generates triage status labels including clarification request status and hard gap rejection status based on the recoverability index. A parameter acquisition module, based on the first interception status label, the second interception status label, and the triage status label, extracts the knowledge base boundary awareness, historical spatiotemporal conflict degree, and pre-acquired knowledge gap degree generated in the preceding steps as loss parameters when no interception or hard gap rejection status is triggered. An output module, based on the pre-acquired evidence support degree and loss parameters, uses a multiplicative loss model to calculate the attenuation score of the definitive answer, and combines the triage status labels to output a structured three-state result including decision labels and routing trajectories, completing the three-state output of local cultural historical spatiotemporal evidence.

[0024] In some embodiments, a three-state output system for local cultural historical spatiotemporal evidence can be used to execute a three-state output method for local cultural historical spatiotemporal evidence, including: S1: receiving a set of evidence requirements extracted by a user query, calculating the coverage degree using a pre-built evidence database capability profile and the set of evidence requirements, obtaining the knowledge base boundary awareness, and generating a first interception status label based on the knowledge base boundary awareness; S2: based on the first interception status label, the user query, and pre-acquired candidate evidence, when the first interception status label is not intercepted, calculating the time interval distance, place name evolution path distance, and entity relationship path distance using the local cultural historical spatiotemporal topology, fusing them to obtain the historical spatiotemporal conflict degree, and generating a second interception status label based on the historical spatiotemporal conflict degree; S3: based on the second interception status label, The system uses a state machine matching algorithm to calculate the recoverability index of missing information based on the entity slot queried by the user and the preset task completion conditions. Based on the recoverability index of missing information, it generates a diversion status label that includes the clarification request status and the hard gap rejection status. S4: Based on the first interception status label, the second interception status label, and the diversion status label, when no interception or hard gap rejection status is triggered, it extracts the knowledge base boundary awareness, historical spatiotemporal conflict degree, and pre-acquired knowledge gap degree generated in the previous steps as the loss parameters. S5: Based on the pre-acquired evidence support degree and the loss parameters, it uses a multiplicative loss model to calculate the attenuation and obtain the definite answer score. Combined with the diversion status label, it outputs a structured three-state result that includes the decision label and the routing trajectory, thus completing the three-state output of local cultural historical spatiotemporal evidence.

[0025] In some embodiments of this specification, the processor utilizes a three-state output system for local cultural and historical spatiotemporal evidence to execute a three-state output method for local cultural and historical spatiotemporal evidence. In this way, the complex process of reviewing unstructured cultural historical materials can be transformed into an automated pipeline operation. While ensuring high concurrency and low latency operation of the local cultural Q&A service, this significantly improves the utilization efficiency of hardware computing resources and the stability of decision-making processes.

[0026] Example 2 Figure 2 This is an exemplary flowchart illustrating a three-state output method for local cultural historical spatiotemporal evidence, as shown in some embodiments of this specification. Figure 2 As shown, the process includes the following steps. In some embodiments, the process may be executed by a processor.

[0027] S1: Receive the evidence requirement set extracted by the user query, calculate the coverage degree using the pre-built evidence base capability profile and the evidence requirement set, obtain the knowledge base boundary awareness, and generate the first interception status label based on the knowledge base boundary awareness.

[0028] A user query is a natural language question text submitted by a user, containing the user's intent and specific constraints. Entity slots are key information points with specific semantic attributes extracted from the user query. For example, a user query may include the original string of statements input by the user and intent category labels, while entity slots may include specific data such as dynasty slots, person slots, location slots, and event slots.

[0029] In some embodiments, the processor can perform natural language preprocessing on the received raw user voice or text data to convert it into a standard text string, and then use a sequence labeling model or word segmentation tool to perform entity recognition and part-of-speech tagging on the standard text string to extract key fields representing specific semantics, thereby obtaining the user query and its corresponding entity slot data.

[0030] The evidence requirements set is a list of capabilities and information that the system needs to retrieve from the local cultural evidence database in order to answer user queries. For example, the evidence requirements set may include specific types of evidence necessary to answer the current question, such as biographical records and local chronicles; required fields to be covered, such as birth and death years and changes in official positions; desired output modality, such as plain text or with accompanying images; and specific data such as the required timeliness.

[0031] In some embodiments, the processor can perform semantic parsing and intent expansion on the extracted entity slots and intent classification tags, compare them with a preset task requirement template library, decompose the user's problem into multiple sub-dimensions of information retrieval instructions and capability matching requirements, and structure and summarize these requirements to obtain evidence requirement set data for subsequent matching and determination.

[0032] An evidence repository capability profile is a data set that provides a structured and quantitative representation of the overall resource reserves and supply capacity of the current local cultural and historical spatiotemporal evidence repository. For example, an evidence repository capability profile may include specific data such as an evidence type support matrix representing the degree of support for different types of evidence, a field completeness vector representing the missing or complete status of various evidence fields, a modal capability identifier representing whether the system can provide text, images, and audio / video, and a timeliness capability identifier representing the system's data update frequency and timeliness status.

[0033] In some embodiments, the processor can periodically scan and statistically analyze all historical spatiotemporal evidence units in the local cultural historical spatiotemporal evidence database, perform aggregation calculations and normalization on the classification of each unit, the fill rate of the included fields, the mounting status of multimedia files, and the last update time, and encapsulate and vectorize the statistical results according to a preset dimensional structure to obtain evidence database capability profile data for representing system capabilities.

[0034] Knowledge base boundary awareness is a quantitative assessment metric used to measure the risk that a user's current query needs exceed the existing local cultural evidence base's ability to answer them. For example, knowledge base boundary awareness may include specific data such as the difference between the type of evidence required by the user and the system's support matrix, the mismatch ratio between the required key fields and the system's field completeness vector, and the risk assessment score for failure to meet modality or timeliness requirements.

[0035] In some embodiments, the processor can extract the set of evidence requirements corresponding to the user query, perform item-by-item cross-comparison and subtraction operations with the corresponding dimensions in the pre-built evidence base capability profile, calculate the proportion of requirement features not covered by the existing profile, and perform weighted summation of the uncovered proportions according to the preset weights of each dimension, thereby obtaining knowledge base boundary awareness data.

[0036] The first interception status label is a discrete status marker indicating whether the current task must be immediately terminated and the response rejected due to serious limitations, based on the system's knowledge base boundary awareness assessment. For example, the first interception status label may include "allow access" indicating full capability coverage, "risk observation" indicating partial capability deficiency but still possible to continue, or "strong rejection" indicating direct blocking due to exceeding the scope of core services.

[0037] In some embodiments, the processor can input the calculated knowledge base boundary awareness data into a preset threshold determiner and compare the awareness value with a set strong rejection threshold for the security service boundary. If the value exceeds the strong rejection threshold, a label representing blocking is generated; if it is lower than or equal to the threshold, a label representing allowing is generated, thereby obtaining first interception status label data for controlling the flow of the process.

[0038] In some embodiments, the processor may, based on the evidence requirement set extracted from the received user query, use a pre-built evidence base capability profile including an evidence type support matrix, a field completeness vector, a modal capability identifier, and a timeliness capability identifier to calculate the uncovered ratio of the evidence requirement set, obtain the knowledge base boundary awareness, and generate the first interception status label based on the knowledge base boundary awareness.

[0039] S2: Based on the first interception status label, user query and pre-acquired candidate evidence, when the first interception status label is not intercepted, the distance between time intervals, the distance of place name evolution path and the distance of entity relationship path are calculated by using the local cultural historical spatiotemporal topology, and the historical spatiotemporal conflict degree is obtained by fusion, and the second interception status label is generated according to the historical spatiotemporal conflict degree.

[0040] Candidate evidence is a set of historical documents, records, or factual fragments that the system initially filters from the local cultural and historical spatiotemporal evidence database based on user queries, and that may contain answers to the questions. For example, candidate evidence may include text paragraphs containing descriptions of relevant historical events, images of ancient books with geographic coordinates, structured attribute tables of biographical information, and unique identifiers of related local gazetteers.

[0041] In some embodiments, the processor can convert the user query text into a retrieval feature vector, use the feature vector to perform similarity search and recall in the inverted index or vector index of the local cultural history spatiotemporal evidence database, and sort and truncate the recall results based on the preliminary keyword matching degree and text similarity score, thereby selecting a combination of several top-ranked related records to obtain candidate evidence data.

[0042] Local cultural historical spatiotemporal topology is used to organize unstructured local cultural data into a complex graph-like data model with logical connections between time, space, and people. For example, local cultural historical spatiotemporal topology may include ancient and modern place name node data representing ancient administrative divisions and modern geographical locations, evolutionary relationship edge data representing the historical change path of place names, entity node data representing historical figures or organizations, and specific data such as person relationship edge data representing interactive relationships such as teachers and students, relatives, and colleagues.

[0043] In some embodiments, the processor can extract entity nouns with spatiotemporal and identity attributes and their co-occurrence dependencies at the sentence level by performing named entity recognition and relation extraction on massive amounts of original local documents. These entities are then mapped to nodes in a graph database, and the dependencies and temporal order between entities are transformed into associated edges connecting the nodes. Synonymous nodes are continuously merged to obtain complete spatiotemporal topological data of local cultural history.

[0044] Time interval distance is a numerical metric used to measure the degree of difference between the historical period specified in a user's query and the actual year recorded or occurred in the candidate evidence. For example, time interval distance can include specific data such as the non-overlapping span between the start and end years of the query requirement and the start and end years contained in the evidence, the absolute year difference between the center time points of the two, or the discrete penalty value caused by dynasty mismatch.

[0045] In some embodiments, the processor can parse the time constraint information in the user query and convert it into query interval segments on a standard time axis. At the same time, it can extract the time attributes corresponding to the candidate evidence to generate evidence interval segments, calculate the intersection length and union length of the two interval segments on the standard time axis, and perform quantization conversion and absolute value calculation on the non-overlapping or deviating parts to obtain time interval distance data.

[0046] The path distance of place name evolution is a structured distance value used to quantify the deviation and leap degree between the geographical name in a user query and the geographical names involved in candidate evidence on the historical evolution map. For example, the path distance of place name evolution may include specific data such as the minimum number of connections required between two place name nodes in the local cultural and historical spatiotemporal topology, the number of administrative level changes traversed in the history of place name changes, or the maximum value assigned when they cannot be connected.

[0047] In some embodiments, the processor can map and match the place names extracted by the user query and the place names associated with the candidate evidence to the corresponding ancient and modern place name nodes in the pre-constructed local cultural and historical spatiotemporal topology, respectively, use a graph traversal algorithm to calculate the shortest number of steps or the total weighted path length between the two nodes connected by the evolutionary relationship edge, and normalize the length based on the maximum diameter of the whole graph to obtain the place name evolution path distance data.

[0048] Entity relationship path distance is an evaluation parameter used to reflect the degree of inconsistency between the expected relationship state between entities in a user's query and the actual entity relationship network supported by candidate evidence. For example, entity relationship path distance can include specific data such as the difference between the expected relationship path features between the two people mentioned in the query and the relationship edge sequence actually reflected by the evidence, missing markers of connected paths of relevant nodes in the topology, or the inverse of the similarity of relationship attributes.

[0049] In some embodiments, the processor can obtain the association path between two sets of entities in the local cultural history spatiotemporal topology by extracting the entities and expected relationships in the user query and the corresponding entities and actual relationships in the candidate evidence, and extract their topological feature vectors. By calculating the edit distance or feature difference value between these two sets of topological feature vectors or relationship attribute label sequences, the processor can obtain entity relationship path distance data.

[0050] Historical spatiotemporal conflict degree is a comprehensive quantitative indicator that measures whether candidate evidence contradicts or contains errors in objective historical facts such as time chronology, geographical evolution, and interpersonal relationships with the user's query intent. For example, historical spatiotemporal conflict degree can include severe conflict penalty scores caused by time misalignment, geographical deviation scores caused by place name errors, and comprehensive deviation numerical data synthesized from time interval distance, place name evolution path distance, and entity relationship path distance.

[0051] In some embodiments, the processor can extract the time interval distance, place name evolution path distance, and entity relationship path distance obtained from previous calculations, assign preset empirical weight coefficients to these three distances according to local cultural characteristics, multiply these three distances by their corresponding weights and sum them, and input the summation result into a function with a nonlinear decay term for smoothing and boundary constraints, thereby obtaining historical spatiotemporal conflict degree data that reflects the overall factual risk.

[0052] The second interception status label is a status control flag used by the system to determine whether the output of evidence must be blocked due to serious violations of historical facts, based on the calculation results of historical spatiotemporal conflict degree. For example, the second interception status label may include "passed identification" flag data indicating good spatiotemporal consistency and allowing use, "weak conflict warning" flag data indicating slight deviations and requiring caution, or "conflict interception" flag data indicating serious spatiotemporal logical errors and direct rejection.

[0053] In some embodiments, the processor can receive the historical spatiotemporal conflict degree data calculated in the previous step, input its value into the logic judgment module, and compare it with the conflict interception threshold preset by the system based on the tolerance bottom line of historical objective facts. If the conflict degree value is greater than or equal to the interception threshold, tag data for forcibly blocking the flow is generated; otherwise, tag data allowing the process to continue to be passed downstream is generated, thereby obtaining the second interception status tag data.

[0054] In some embodiments, the processor can, based on the first interception status label, the user query, and the pre-acquired candidate evidence, when the first interception status label is not intercepted, utilize the local cultural historical spatiotemporal topology structure containing ancient and modern place name nodes and their evolutionary relationship edges, entity nodes and their person relationship edges, obtain the place name evolutionary path distance through shortest path analysis between nodes, obtain the entity relationship path distance through entity relationship path comparison processing, and combine it with time interval distance comparison processing to obtain the historical spatiotemporal conflict degree, and generate the second interception status label based on the historical spatiotemporal conflict degree.

[0055] In some embodiments, the expression for the degree of historical spatiotemporal conflict is: ; ; ; in, Indicates the degree of historical spatiotemporal conflict. This represents an exponential function with base e. This represents the preset time weighting coefficient. This represents the preset place name weight coefficient. This represents the preset relational weight coefficients. This represents the time interval distance, specifically the overlap penalty value between the query time and the evidence time interval obtained based on the time axis mapping comparison. Indicates the distance along the path of the place name's evolution. Indicates the path distance of entity relationships. This represents the shortest path length in the local cultural and historical spatiotemporal topology between the ancient and modern place name nodes associated with the user query and the place name nodes associated with the candidate evidence, obtained based on the graph search algorithm. This represents the maximum graph diameter constant of the preset topology. This represents the similarity between the query-corresponding entity relationship path obtained based on graph feature embedding comparison and the evidence-corresponding entity relationship path.

[0056] S3: Based on the second interception status label, the entity slot queried by the user, and the preset task completion conditions, the state machine matching algorithm is used to calculate the recoverability index of missing information, and a diversion status label including the clarification request status and the hard gap rejection status is generated according to the recoverability index of missing information.

[0057] Task completion conditions are structured standards and sets of information rules that the system sets to determine whether a user's question has been fully answered. For example, task completion conditions may include a list of key fields that must be filled in to answer a question about a specific local culture, such as the time and location of historical events, the minimum amount of supporting evidence that must be provided, and conditional validation expressions that require the answer logic to form a closed loop.

[0058] In some embodiments, the processor can extract the standard constraint framework corresponding to the type of intent from a pre-set business rule template library by parsing the task intent classification in the user query, and extract and encapsulate the necessary slot types, pre- and post-dependencies, and data type requirements in combination with the current system's output quality baseline requirements, transforming them into a set of rules that can be verified by automated programs, thereby obtaining task completion condition data.

[0059] The recoverability metric for missing information is a quantitative assessment parameter used to evaluate whether, given insufficient information in a current user question, the user can potentially fill in the missing information through subsequent interactive dialogue. For example, the recoverability metric may include a recovery probability score measuring whether the missing slot is of common-sense nature, statistical data on the success rate of guiding questions in the system's history, or classification coding data indicating whether the missing content is recoverable manually or due to a fundamental flaw in the underlying database.

[0060] In some embodiments, the processor can extract the set of entity slots queried by the user and the set of required slots required by the task completion conditions, perform a difference operation to find the specific set of missing slots, and then perform matching analysis on these missing slots with a preset system support capability dictionary to evaluate whether they belong to the current knowledge base supply scope. Based on the difficulty of the matching results, a score is assigned and calculated to obtain the recoverability index data of the missing information.

[0061] In one implementation, before calculating the recoverability index of missing information, the system needs to construct a set of task constraint architectures based on user intent. The expression for the set of task constraint architectures is: ; in, This represents a pre-built set of task constraint architectures. Indicates the type of task requirement. This represents the set of critical entity slots that must be included. This represents the set of non-critical entity slots used as supplementary reference. This represents the set of recoverable slots that the system determines can be supplemented by subsequent user interactions, and is used to assign the aforementioned recoverable missing identifier. This represents the set of hard gap slots that the system determines cannot be filled due to missing objective data, and is used to assign the aforementioned hard gap identifier. This represents the aforementioned task completion condition verification logic expression. In actual processing, the processor maps the task requirement type to the above subsets by reading the configuration file, thereby constructing the task constraint architecture set used to support the state machine matching logic.

[0062] Triage status labels are specific action instructions used by the system to guide subsequent interactions when faced with queries lacking sufficient information. They contain detailed response strategies. For example, triage status labels may include clarification request status data that instructs the system to ask the user for clarification of missing elements, and hard gap rejection status data that indicates a direct failure due to missing core data that the user cannot supplement.

[0063] In some embodiments, the processor can obtain the recoverability index data of the missing information calculated in the preceding steps, and compare the index with the threshold of the value or category according to the built-in rule engine of the system. If the index meets the security judgment conditions that the user can supplement, action state data pointing to the interactive questioning logic is generated; if the index indicates that the information touches the blind spot of the system's objective capabilities, action state data pointing to termination and feedback of the failure reason is generated, thereby obtaining the final diversion status label data.

[0064] In some embodiments, the processor may perform slot comparison calculations based on the second interception status label, the entity slot queried by the user, and preset task completion conditions, using state machine matching logic that includes recoverable missing information identifiers and hard gap identifiers, to obtain the recoverability index of missing information, and determine slots with missing real-time data and no applicable modality as having hard gap identifiers, and generate the diversion status label including the clarification request status and the hard gap rejection status based on the recoverability index of missing information.

[0065] The state machine matching logic is a built-in dynamic rule-based judgment system and feature tag library used to control the flow of multi-turn dialogues and evidence retrieval. For example, the state machine matching logic may include rule matrix data that defines various dialogue state transition trigger conditions, recoverable missing information flag configuration data for marking commonly omitted user information, and feature dictionary data for marking hard gaps such as the absence of underlying interfaces that are absolutely irreparable errors.

[0066] In some embodiments, the processor can read the expert business rule configuration file in persistent storage during the system initialization phase, bind and map different categories of slot missing situations with the probability level of resolving the missing situation, establish a multi-dimensional state transition matrix and set the conditional expressions for each node jump, load these logical relationships and the corresponding identifier dictionary into the execution engine in memory, thereby obtaining state machine matching logic data that can be called at runtime.

[0067] S4: Based on the first interception status label, the second interception status label, and the diversion status label, when none of them trigger the interception or hard gap rejection status, extract the knowledge base boundary awareness, historical spatiotemporal conflict degree, and pre-acquired knowledge gap degree generated in the preceding steps as the loss parameters.

[0068] The interception and hard gap rejection status is a negative summary control signal triggered by the system when performing pre-gating judgment, indicating that the entire question-and-answer process must be terminated abnormally immediately. For example, the interception and hard gap rejection status may include specific data such as the first interception trigger record data caused by the knowledge base boundary awareness exceeding the limit, the second interception trigger record data caused by the historical spatiotemporal conflict exceeding the limit, and the hard gap rejection flag data caused by irreparable missing slots in slot matching.

[0069] In some embodiments, the processor can extract the first interception status label, the second interception status label, and the diversion status label output in the preceding execution flow and input them centrally into a composite judgment logic module. Whenever this module detects that any of the above labels contains a negative flag value indicating blocking, abort, or an unrecoverable error, it extracts the signal and combines it to output a global forced termination instruction, thereby obtaining the summarized interception and hard gap rejection status data.

[0070] The reduction parameter is a combination of variables used by the system to penalize and reduce the basic support level based on various potential risks identified in the early stages when generating the final answer confidence score. For example, the reduction parameter may include specific data sets such as knowledge base boundary awareness data that characterizes the degree of inadequacy in the system's capability coverage, historical spatiotemporal conflict data that characterizes the degree of contradiction between evidence and facts, and pre-acquired knowledge gap data that characterizes the degree of lack of required information elements.

[0071] In some embodiments, the processor can extract and read the knowledge base boundary awareness metric results and historical spatiotemporal conflict metric results that have been calculated and saved in the pre-detection step from the system's context memory or state transition channel, and statistically map the knowledge gap metric values ​​representing the degree of information incompleteness based on the difference set situation of slot matching. These three sets of metric values ​​are then encapsulated and combined in a unified floating-point format to obtain the loss parameter data used to calculate the penalty coefficient.

[0072] S5: Based on the pre-acquired evidence support and attenuation parameters, the attenuation calculation is performed using a multiplicative attenuation model to obtain the definitive answer score. Combined with the diversion status label, the structured three-state results containing decision labels and routing trajectories are output to complete the three-state output of local cultural historical spatiotemporal evidence.

[0073] Evidence support is an initial baseline score used to quantify the strength of a single or combined candidate piece of evidence in supporting a user's query in terms of surface text semantics and content slots. For example, evidence support may include specific data such as the percentage of literal overlap between candidate evidence and the user's query text, the cosine similarity score between high-dimensional semantic feature vectors, and the coverage ratio of entity fields contained in the candidate evidence to satisfy the necessary slots for the query.

[0074] In some embodiments, the processor can extract pre-selected candidate evidence text and user query text, convert them into dense vectors using a pre-trained language feature extraction model, and perform vector dot product or similarity operations. At the same time, the processor can use a field extraction module to compare the number of entities covered in the evidence with the total number of entities required by the query, and then linearly weight and fuse the similarity score and the entity coverage ratio score according to a certain ratio to obtain the initial evidence support data.

[0075] The multiplicative attenuation model is an algorithmic processing structure that uses a pre-assessment risk index to suppress and penalize the initial confidence score through a continuous multiplication and non-linear decreasing method. For example, the multiplicative attenuation model may include mathematical logic structure data in which the evidence support level is used as the base multiplier, and each multiplied by an attenuation factor representing the retention ratio, as well as smoothing processing procedure configuration data to prevent the value from reaching zero or underflowing after multiplication.

[0076] In some embodiments, the processor can load and instantiate an algorithmic logic component for score calculation in memory. After receiving the evidence support level as input and the depreciation parameters consisting of multiple indicators, the processor can use an arithmetic processing unit within the logic component to convert the various risk values ​​into attenuation ratios in a pre-set order, perform continuous multiplication and accumulation calculation operations, and use upper and lower limit truncation mechanisms to constrain the calculation results, thereby obtaining the calculation results of the multiplication depreciation model.

[0077] The definitive answer score is a comprehensive evaluation score that the system is confident in providing a safe and reliable answer to the user, obtained after all pre-gating checks and multi-dimensional risk factor penalties. For example, the definitive answer score may include a floating-point numerical data in the range of 0 to 1, a standardized percentage confidence score, or discrete-level categorical numerical data that directly maps the strength of the system's response.

[0078] In some embodiments, the processor can invoke and execute the aforementioned multiplicative attenuation model logic within the processor. Starting with the extracted basic evidence support value, it sequentially performs multiple multiplication operations with the knowledge base boundary awareness attenuation factor, historical spatiotemporal conflict attenuation factor, and knowledge gap attenuation factor (all converted through subtraction). The final product is then rounded and subject to effective precision limitations to obtain the definitive answer score data used for controlling output decisions.

[0079] The structured three-state result is the final standard data output carrier after the system has processed the user query. It contains the system's action decisions and the details of the underlying decision-making process. For example, the structured three-state result may include decision label strings indicating whether the current task is to provide a normal answer, clarify to the user, or directly refuse to answer; specific text response content displayed to the user; and a routing trajectory data set recording the judgment status of each node in the entire process from query reception to gating interception and score reduction.

[0080] In some embodiments, the processor can aggregate the final state decision identifiers, called text templates, or generated content generated throughout the system execution process, and extract key event records such as the triggering reasons of each pre-gate control and intermediate scores arranged in chronological order from the system's internal operation log. These different types of data are then uniformly incorporated into a preset data serialization template, and assigned and nested according to the prescribed hierarchical field relationships to obtain a uniformly formatted structured three-state result data.

[0081] In some embodiments, the processor can calculate the determination answer score by using the multiplicative attenuation model based on the pre-acquired evidence support and the attenuation parameter, and combine the diversion status label to output the structured three-state result containing the decision label and the routing trajectory, thereby completing the three-state output of local cultural historical spatiotemporal evidence; wherein, the routing trajectory is log data generated by recording the data processing link, which includes pre-gating trigger conditions, various scores and the final output state.

[0082] In some embodiments, the expression for the multiplicative loss model is: ; in, This indicates that the answer has been confirmed and a score has been awarded. This represents the initial support score of the candidate evidence obtained based on the semantic matching model or slot coverage statistical algorithm for the user query. Indicates the knowledge base boundary awareness. Indicates the degree of historical spatiotemporal conflict. It represents the knowledge gap degree, which is the gap penalty factor obtained by quantitative mapping based on the recoverability index of missing information.

[0083] In one implementation, to ensure the rationality of the aforementioned interception gate threshold settings and reduce the probability of incorrect answers, a background verification stage is added after the system obtains the structured three-state result. Threshold calibration logic is introduced in the background verification stage, and the expression for its threshold calibration penalty loss function is: ; in, This represents the total global verification loss. This represents the basic evaluation loss value calculated based on the macro-average F1 score of the test sample set. This represents the false positive rate, which indicates the number of incorrect answers the system produces on the test set. This represents the penalty value for boundary leakage caused by the failure of knowledge base boundary awareness gating. This represents the penalty value for conflict leakage caused by the failure of historical spatiotemporal conflict gating. This represents the preset false positive penalty weighting coefficient. This represents the preset boundary leakage penalty weight coefficient. This represents the preset conflict leakage penalty weight coefficient. The processor uses gradient descent or grid search algorithms to minimize the total global verification loss, thereby obtaining the optimal decision threshold for the aforementioned knowledge base boundary awareness strong rejection condition and historical spatiotemporal conflict interception condition.

[0084] In some embodiments, to ensure the computational accuracy and convergence of the three-state output method for local cultural historical spatiotemporal evidence, the various constant parameters involved in the system have clearly defined value ranges and constraints. Specifically: for the constants in the aforementioned expression for historical spatiotemporal conflict degree, the time weight coefficient... Place name weight coefficient With relational weight coefficient The range of values ​​is Furthermore, in the preferred embodiment, the normalization constraint condition is satisfied: The maximum diagram diameter constant. It is a positive integer, and its value range is 1. This value is specifically obtained by dynamically determining the depth of the largest connected sublayer of the currently loaded local cultural and historical spatiotemporal topology, and is used to prevent the denominator from overflowing or becoming too small when normalizing the path distance of place name evolution.

[0085] For the constant in the aforementioned threshold calibration penalty loss function expression, the false positive penalty weight coefficient... Boundary leakage penalty weighting coefficient and conflict leakage penalty weighting coefficient All are real numbers greater than 0, and their value range is set to... Under the optimized parameter configuration, to reflect a zero-tolerance attitude towards inaccurate facts and services exceeding the scope of service, the values ​​of the three penalty weighting coefficients mentioned above are all set much larger than the basic assessment loss values. The implicit weight (i.e., 1.0) drives the system model to prioritize reducing the leakage rate of boundaries and facts.

[0086] It should be noted that different embodiments may produce different beneficial effects. In different embodiments, the beneficial effects may be any one or a combination of the above, or any other possible beneficial effects.

Claims

1. A method for outputting local cultural historical spatiotemporal evidence in three states, characterized in that, include: S1: Receive the set of evidence requirements extracted by the user query, calculate the coverage degree using the pre-built evidence base capability profile and the set of evidence requirements, obtain the knowledge base boundary awareness, and generate the first interception status label based on the knowledge base boundary awareness. S2: Based on the first interception status label, user query and pre-acquired candidate evidence, when the first interception status label is not intercepted, the distance between time intervals, the distance of place name evolution path and the distance of entity relationship path are calculated by using the local cultural historical spatiotemporal topology, and the historical spatiotemporal conflict degree is obtained by fusion, and the second interception status label is generated according to the historical spatiotemporal conflict degree. S3: Based on the second interception status label, the entity slot queried by the user, and the preset task completion conditions, the state machine matching algorithm is used to calculate the recoverability index of missing information, and a diversion status label including the clarification request status and the hard gap rejection status is generated according to the recoverability index of missing information. S4: Based on the first interception status label, the second interception status label, and the diversion status label, when none of them trigger the interception and hard gap rejection status, extract the knowledge base boundary awareness, historical spatiotemporal conflict degree, and pre-acquired knowledge gap degree generated in the preceding steps as loss parameters. S5: Based on the pre-acquired evidence support and attenuation parameters, the attenuation calculation is performed using a multiplicative attenuation model to obtain the definitive answer score. Combined with the diversion status label, the structured three-state results containing decision labels and routing trajectories are output to complete the three-state output of local cultural historical spatiotemporal evidence.

2. The method for outputting local cultural historical spatiotemporal evidence in three states according to claim 1, characterized in that, S1 includes: Based on the evidence requirement set extracted from user queries, a pre-built evidence base capability profile containing evidence type support matrix, field completeness vector, modal capability identifier, and timeliness capability identifier is used to calculate the uncovered ratio of the evidence requirement set, obtain the knowledge base boundary awareness, and generate the first interception status label based on the knowledge base boundary awareness.

3. The method for outputting local cultural historical spatiotemporal evidence in three states according to claim 1, characterized in that, S2 includes: Based on the first interception status label, the user query, and the pre-acquired candidate evidence, when the first interception status label is not intercepted, the local cultural history spatiotemporal topology structure containing ancient and modern place name nodes and their evolutionary relationship edges, entity nodes and their person relationship edges is used to obtain the place name evolutionary path distance through shortest path analysis between nodes, obtain the entity relationship path distance through entity relationship path comparison processing, and combine it with time interval distance comparison processing to obtain the historical spatiotemporal conflict degree, and generate the second interception status label based on the historical spatiotemporal conflict degree.

4. The method for outputting local cultural historical spatiotemporal evidence in three states according to claim 3, characterized in that, The expression for the degree of historical spatiotemporal conflict is: ； ； ； in, Indicates the degree of historical spatiotemporal conflict. This represents an exponential function with base e. This represents the preset time weighting coefficient. This represents the preset place name weight coefficient. This represents the preset relational weight coefficients. This represents the time interval distance, specifically the overlap penalty value between the query time and the evidence time interval obtained based on the time axis mapping comparison. Indicates the distance along the path of the place name's evolution. Indicates the path distance of entity relationships. This represents the shortest path length in the local cultural and historical spatiotemporal topology between the ancient and modern place name nodes associated with the user query and the place name nodes associated with the candidate evidence, obtained based on the graph search algorithm. This represents the maximum graph diameter constant of the preset topology. This represents the similarity between the query-corresponding entity relationship path obtained based on graph feature embedding comparison and the evidence-corresponding entity relationship path.

5. The method for outputting local cultural historical spatiotemporal evidence in three states according to claim 1, characterized in that, S3 includes: Based on the second interception status label, the entity slot queried by the user, and the preset task completion conditions, the slot comparison calculation is performed using state machine matching logic that includes recoverable missing information identifiers and hard gap identifiers to obtain the recoverability index of missing information. Slots with missing real-time data and no applicable modality are determined to have hard gap identifiers. Based on the recoverability index of missing information, the diversion status label including the clarification request status and the hard gap rejection status is generated.

6. The method for outputting local cultural historical spatiotemporal evidence in three states according to claim 1, characterized in that, S5 includes: Based on the pre-acquired evidence support and the attenuation parameter, the attenuation calculation is performed using the multiplicative attenuation model to obtain the determined answer score. Combined with the diversion status label, the structured three-state result containing the decision label and the routing trajectory is output to complete the three-state output of local cultural historical spatiotemporal evidence. The routing trajectory is log data generated by recording the data processing link, which includes pre-gating trigger conditions, various scores, and the final output state.

7. The method for outputting local cultural historical spatiotemporal evidence in three states according to claim 6, characterized in that, The expression for the multiplication loss model is: ； in, This indicates that the answer has been confirmed and a score has been awarded. This represents the initial support score of the candidate evidence obtained based on the semantic matching model or slot coverage statistical algorithm for the user query. Indicates the knowledge base boundary awareness. Indicates the degree of historical spatiotemporal conflict. It represents the knowledge gap degree, which is the gap penalty factor obtained by quantitative mapping based on the recoverability index of missing information.

8. A three-state output system for local cultural historical spatiotemporal evidence, used to execute the three-state output method for local cultural historical spatiotemporal evidence as described in any one of claims 1 to 7, characterized in that, include: The first tag generation module is used to receive the evidence requirement set extracted by the user query, calculate the coverage degree using the pre-built evidence base capability profile and the evidence requirement set, obtain the knowledge base boundary awareness, and generate the first interception status tag based on the knowledge base boundary awareness. The second tag generation module is used to calculate the time interval distance, place name evolution path distance and entity relationship path distance based on the first interception status tag, user query and pre-acquired candidate evidence when the first interception status tag is not intercepted, using the local cultural historical spatiotemporal topology to obtain the historical spatiotemporal conflict degree, and generate the second interception status tag based on the historical spatiotemporal conflict degree. The diversion label generation module is used to calculate the recoverability index of missing information based on the second interception status label, the entity slot queried by the user and the preset task completion conditions, using a state machine matching algorithm, and to generate diversion status labels including clarification request status and hard gap rejection status based on the recoverability index of missing information. The parameter acquisition module is used to extract the knowledge base boundary awareness, historical spatiotemporal conflict degree, and pre-acquired knowledge gap degree generated in the preceding steps as loss parameters based on the first interception state label, the second interception state label, and the diversion state label, when none of them have triggered the interception or hard gap rejection state. The output module is used to calculate the score of the definitive answer based on the pre-acquired evidence support and the attenuation parameter using the multiplicative attenuation model. It also combines the diversion status label to output a structured three-state result containing the decision label and the routing trajectory, thus completing the three-state output of local cultural historical spatiotemporal evidence.

Images