An archival auxiliary compilation and content reconstruction method and system based on generative artificial intelligence

Abstract

The present application relates to a kind of based on generative artificial intelligence archive auxiliary compilation and content reconstruction method and system, comprising: extracting archive multidimensional semantic features, the semantic constraint of the relationship of time, theme and event is automatically mined and formalized by historical sample and field ontology, and the dimension contract expression set that can be parsed by machine is generated;Based on finite state machine, integrate the expression into content generation real-time checking process, quickly determine semantic compliance and capture default signal;After detecting default, local context is automatically frozen and deep quantitative analysis semantic conflict, generate the operational semantic snapshot input lightweight reconciliation model, directional correction strategy is outputted after joint modeling, and conflict content is accurately automatically reconstructed;Contract closed loop is formed by combining user adoption feedback and text coherence score, drive semantic constraint parameter adaptive evolution.This method improves the multidimensional semantic consistency and intelligent reconciliation capability of automatic archive content generation.

Description

Technical Field

[0001] This invention relates to the field of generative artificial intelligence archival content reconstruction technology, and in particular to a method and system for archival assisted compilation and content reconstruction based on generative artificial intelligence. Background Technology

[0002] Generative AI-driven archival content reconstruction technology has made significant progress in recent years in areas such as intelligent archival compilation and research, automatic text generation, and historical knowledge-assisted expression. Current mainstream solutions generally rely on large-scale pre-trained language models (such as the GPT series and knowledge-enhanced generative models), combined with structured knowledge such as domain knowledge graphs and event process ontology. Through the analysis and modeling of original archival data, they achieve content compilation, style transformation, and multi-dimensional narrative thread reorganization to meet diverse user needs. Industry trends show continuously strengthening integration capabilities such as large-scale model intelligent generation, semantic parsing and structuring, and multi-channel content reconstruction, promoting the deep utilization of digitized archives and the implementation of intelligent language services.

[0003] However, existing technologies generally employ global regeneration or knowledge graph-based reasoning methods to address the semantic consistency issues between chapter modules across time, themes, and events. On the one hand, large models often rely on cross-segment memory and centralized attention mechanisms to maintain the overall chapter logic. However, this often weakens the fine-grained control over local consistency due to increased input length or theme transitions, resulting in "semantic breaks" such as event timeline breaks, theme abrupt changes, and distorted character relationships within the archive content. On the other hand, while knowledge graph reasoning or multi-dimensional entity relationship reconstruction technologies can provide constraint rules and content verification capabilities to some extent, they generally face practical bottlenecks such as high barriers to knowledge base construction and maintenance, large overhead of global reasoning, and high consumption of software and hardware resources (e.g., high complexity of steps such as graph search and feature weighted aggregation).

[0004] Therefore, the industry urgently needs a lightweight, pluggable, and real-time method for dynamically maintaining semantic consistency during multi-dimensional content reconstruction without relying on high-cost global computation and knowledge graph reasoning. Summary of the Invention

[0005] This application provides a method and system for archival research and content reconstruction based on generative artificial intelligence, aiming to solve the problems mentioned in the background art, such as broken event timelines, abrupt themes, and distorted relationships between people in archival content.

[0006] This application provides a method for archival-assisted compilation and content reconstruction based on generative artificial intelligence, specifically including: S1: For multi-source heterogeneous archival data in archival auxiliary compilation scenarios, extract multi-dimensional semantic features of timeline, theme and relationship between people and events, and construct a set of dimensional contract expressions; S2: Compile the set of dimensional contract expressions into a lightweight state machine form of contract validator, perform online structured mapping processing on the raw text word stream output in real time by the decoder in the file content generation process to obtain a structured semantic summary, input the structured semantic summary into the contract validator, and output the dimensional compliance judgment result. S3: Identify breach signals based on the compliance determination results of the aforementioned dimensions. If a breach signal is detected, immediately freeze the local context window, capture and generate a local semantic snapshot marked with the breach dimension identifier and conflict type label. S4: Using a reconciliation proxy model with fewer than a preset threshold of parameters, the local semantic snapshot and the corresponding dimensional contract expression are used as joint inputs to perform minimal intervention correction calculations to generate a directional semantic reconciliation strategy that includes transitional explanatory phrase insertion instructions and time node fine-tuning parameters. S5: Perform partial reconstruction of the current archive content according to the directional semantic reconciliation strategy, dynamically adjust the logical connection relationship of the breach dimension to output the intermediate result of multi-dimensional content reconstruction, and complete the compilation and content reconstruction of the archive based on the intermediate result of multi-dimensional content reconstruction. S6: Collect user adoption rate feedback signals and downstream coherence scoring data of the intermediate results of the multi-dimensional content reconstruction, and construct a contract feedback closed-loop dataset that includes reconciliation operation records and effect evaluation indicators. S7: Dynamically update the dimensional contract expression set based on the contract feedback closed-loop dataset, generate a new version of the dimensional contract expression set that evolves adaptively, and replace the original configuration.

[0007] This application also proposes a generative artificial intelligence-based system for archival compilation and content reconstruction, used to execute a generative artificial intelligence-based method for archival compilation and content reconstruction, including: Feature extraction module: used to extract multi-dimensional semantic features of timeline, theme and relationship between people and events from multi-source heterogeneous archival data in archival auxiliary compilation and research scenarios, and to construct a set of dimensional contract expressions; Contract validator: It is formed by compiling the set of dimensional contract expressions into a lightweight state machine. It is used to receive the structured semantic summary of the currently generated fragment as input after each step of the decoder output in the content generation process, perform millisecond-level matching verification, and output the dimensional compliance judgment result. The default detection module is used to identify default signals based on the compliance judgment results of the aforementioned dimensions. If a default signal is detected, the local context window is immediately frozen, and a local semantic snapshot marked with the default dimension identifier and conflict type label is captured and generated. Harmony Proxy Model: A proxy model with fewer than a preset threshold of parameters, used to take the local semantic snapshot and the corresponding dimensional contract expression as joint input, perform minimal intervention correction calculation, and generate a directional semantic harmonization strategy that includes transitional explanatory phrase insertion instructions and time node fine-tuning parameters; The local reconstruction module is used to perform local reconstruction operations on the current archive content according to the directional semantic reconciliation strategy, dynamically adjust the logical connection relationship of the breach dimension, output the intermediate results of multi-dimensional content reconstruction, and complete the compilation and content reconstruction of the archive based on the intermediate results of multi-dimensional content reconstruction.

[0008] This application provides a method for archival research and content reconstruction based on generative artificial intelligence, which has the following beneficial effects: This solution introduces a lightweight semantic coordination mechanism driven by contracts, achieving for the first time an innovative path that ensures multi-dimensional logical consistency without global reconstruction. Specifically, it extracts executable symbolic contract expressions from high-quality historical samples and domain ontologies, compiles them into a low-latency state machine embedded in the generation process, and completes millisecond-level local verification without decoupling the backbone model. This effectively avoids the resource waste and response lag issues caused by the repeated calls to large-scale language models for correction in traditional methods, significantly improving system operating efficiency and deployment flexibility. It is particularly suitable for real-time content-assisted generation needs in long texts and highly structured scenarios such as archives. Attached Figure Description

[0009] Figure 1 This is the main flowchart of a method for archival research and content reconstruction based on generative artificial intelligence.

[0010] Figure 2 This is a sub-flowchart of a generative artificial intelligence-based method for archival research and content reconstruction.

[0011] Figure 3 This is another sub-flowchart of a method for archival-assisted compilation and content reconstruction based on generative artificial intelligence. Detailed Implementation

[0012] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0013] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0014] like Figure 1 As shown, this application provides a method for archival-assisted compilation and content reconstruction based on generative artificial intelligence, specifically including: S1: For multi-source heterogeneous archival data in archival auxiliary compilation scenarios, extract multi-dimensional semantic features of timeline, theme and relationship between people and events, and construct a set of dimensional contract expressions; S2: Compile the set of dimensional contract expressions into a lightweight state machine form of contract validator, perform online structured mapping processing on the raw text word stream output in real time by the decoder in the file content generation process to obtain a structured semantic summary, input the structured semantic summary into the contract validator, and output the dimensional compliance judgment result. S3: Identify breach signals based on the compliance determination results of the aforementioned dimensions. If a breach signal is detected, immediately freeze the local context window, capture and generate a local semantic snapshot marked with the breach dimension identifier and conflict type label. S4: Using a reconciliation proxy model with fewer than a preset threshold of parameters, the local semantic snapshot and the corresponding dimensional contract expression are used as joint inputs to perform minimal intervention correction calculations to generate a directional semantic reconciliation strategy that includes transitional explanatory phrase insertion instructions and time node fine-tuning parameters. S5: Perform partial reconstruction of the current archive content according to the directional semantic reconciliation strategy, dynamically adjust the logical connection relationship of the breach dimension to output the intermediate result of multi-dimensional content reconstruction, and complete the compilation and content reconstruction of the archive based on the intermediate result of multi-dimensional content reconstruction. S6: Collect user adoption rate feedback signals and downstream coherence scoring data of the intermediate results of the multi-dimensional content reconstruction, and construct a contract feedback closed-loop dataset that includes reconciliation operation records and effect evaluation indicators. S7: Dynamically update the dimensional contract expression set based on the contract feedback closed-loop dataset, generate a new version of the dimensional contract expression set that evolves adaptively, and replace the original configuration.

[0015] like Figure 2 As shown, step S1 involves extracting multi-dimensional semantic features of timelines, themes, and relationships between individuals and events from multi-source heterogeneous archival data in archival-assisted compilation scenarios, and constructing a set of dimensional contract expressions. Specifically, this includes: S1.1: Preprocess the multi-source heterogeneous archival data to generate a standardized archival text sequence. Use named entity recognition and dependency parsing techniques to extract the original timestamp entities, original topic tags, and original person / event triples from the standardized archival text sequence, thereby obtaining a multi-dimensional set of original features containing basic semantic units.

[0016] This process receives raw data from diverse, heterogeneous archives, including OCR text from digitized scans, structured database records, and unstructured log files. The raw data undergoes a unified character set encoding conversion to remove invisible control characters and garbled symbols. Text cleaning is performed to eliminate residual HTML tags and special formatting marks. A named entity recognition method based on a BERT pre-trained model is used to semantically annotate standardized archive text sequences, extracting entity fragments with time attributes as raw timestamp entities and keyword combinations representing core issues as raw topic tags. Dependency parsing techniques are used to analyze the internal grammatical structure of sentences, identifying subject-verb-object relationships and modifiers, constructing a person-action-object triple structure to form a raw set of person-event triples. The extracted raw timestamp entities, raw topic tags, and raw person-event triples are spatiotemporally aligned and indexed to generate a multi-dimensional raw feature set containing basic semantic units. Through these processing methods, unstructured archive data is transformed into machine-readable structured semantic elements, achieving a precise mapping from raw text to multi-dimensional semantic features, providing a high-fidelity data foundation for subsequent dimensional contract modeling.

[0017] For example, taking a document document on enterprise reform from 1980 to 1990 as an example, firstly, noisy characters generated by OCR are removed. Then, the BERT-BiLSTM-CRF model is used to identify entities such as "March 1985" and "factory director responsibility system". Dependency parsing is used to extract triples (Zhang Moumou, implementation, reform plan). An entity confidence threshold of 0.85 is set to filter out low-quality extraction results. The final output is a raw feature set containing 50 time nodes, 20 topic tags, and 100 event triples, ensuring a semantic unit coverage rate of over 95%, significantly improving the data completeness for subsequent contract verification.

[0018] S1.2: Based on the multi-dimensional original feature set, perform semantic alignment processing to generate structured semantic graph nodes. Use domain ontology mapping rules to convert the original timestamp entities into normalized time node objects, map the original topic tags into standard topic concept vectors, and reorganize the original person event triples into logical event relationship edges, thereby outputting a multi-dimensional structured semantic feature set with topological association characteristics.

[0019] The system receives a multi-dimensional set of raw features from S1.1, including raw timestamp entities, raw topic tags, and raw person / event triples. It performs standardization and cleaning on the raw timestamp entities, using regular expressions to match non-standard date formats and convert them into ISO 8601 standard format time strings. Based on the time granularity definition in the domain ontology, it parses the standardized time strings into structured time node objects, each containing absolute time value, relative time offset, and time confidence fields. It performs vector space mapping on the raw topic tags, using a pre-trained domain-specific BERT model to extract the semantic embedding vectors of the tags. It compares the raw tags with prototype vectors in the standard topic concept library using cosine similarity calculation, selecting the standard concept with the highest similarity as the mapping target to generate standard topic concept vectors. It performs logical recombination on the raw person / event triples, identifying the role attributes of subject, predicate, and object based on dependency parsing results. Finally, it combines event pattern templates from the domain ontology to aggregate the scattered triples into logical event relationship edges with causal or temporal associations. A multi-dimensional semantic graph node is constructed, mapping normalized time node objects, standard topic concept vectors, and logical event relationship edges to a unified graph structure coordinate system. The semantic association strength between nodes is quantified using an edge weighting method, forming a multi-dimensional structured semantic feature set with topological association characteristics. Through the above semantic alignment processing, the unstructured raw features from the previous step are transformed into standardized structured semantic graph node data, achieving semantic unification of multi-source heterogeneous archival data in the time, topic, and event dimensions, providing a consistent data foundation for subsequent contract mining.

[0020] S1.3: Pattern mining is performed on high-quality historical archive samples to generate a typical semantic constraint pattern library. The frequent subgraph mining method is applied to statistically analyze the time interval distribution of adjacent event nodes and the transition probability matrix of event type sequences from multi-dimensional structured semantic features. This allows us to extract time span tolerance interval parameters and event evolution topology order rules that reflect the experience of domain experts.

[0021] It receives a multi-dimensional structured semantic feature set that has undergone semantic alignment processing. This dataset contains normalized time node objects, standard topic concept vectors, and logical event relationship edges, which serve as the input basis for pattern mining.

[0022] A directed graph model of event evolution based on high-quality historical archive samples is constructed, mapping each independent archive entry to a graph node. Node attributes include timestamp, topic tag and event type, and directed edges between nodes represent the chronological order and causal relationship.

[0023] The frequent subgraph mining method is applied to traverse and search the directed graph model of event evolution. The minimum support threshold is set to 0.05 to filter low-frequency noise patterns and identify local topological subgraphs that recur in multiple high-quality archives.

[0024] For the identified frequent subgraphs, the time interval data between adjacent event nodes are statistically analyzed, the mean and standard deviation of the time intervals are calculated, a time span distribution model is established, and the time span tolerance interval parameter reflecting the regular narrative rhythm of the domain is extracted.

[0025] Construct a state transition matrix for the event type sequence, calculate the conditional probabilities from the initial event type to the development, climax and conclusion event types, determine the optimal state transition path based on the maximum likelihood estimation principle, and extract the event evolution topological order rules.

[0026] The time span tolerance interval parameters and event evolution topology order rules are encapsulated in a structured manner to generate a typical semantic constraint pattern library, which contains two types of metadata: numerical constraints and logical constraints.

[0027] By using pattern mining, the structured semantic features from the previous step are transformed into a library of typical semantic constraint patterns that reflect the experience of domain experts. This enables the automated extraction of quantifiable semantic rules from unstructured archival data, providing data support for the subsequent construction of contract expressions.

[0028] S1.4: Based on the typical semantic constraint pattern library, perform symbolic encoding operations to generate dimensional contract expression templates. Use formal language to define syntax to instantiate the time span tolerance interval parameters into numerical comparison assertions and the event evolution topological order rules into state transition assertions, thereby constructing an initial dimensional contract expression prototype that can be directly parsed by the machine.

[0029] It receives time span tolerance interval parameters and event evolution topology order rules from a library of typical semantic constraint patterns as raw input data for symbolic encoding.

[0030] A contract expression template framework is constructed using a formal language definition syntax, establishing a standard structural paradigm for numerical comparison assertions and state transition assertions.

[0031] For the timeline dimension, the time span tolerance interval parameter is mapped to a numerical comparison operator to construct the validity verification logic for the timestamp difference between adjacent event nodes.

[0032] For the dimensions of theme and people / events, the event evolution topology order rules are mapped to the state transition matrix of a finite state machine, and the legal event type sequence paths are defined.

[0033] By using predicate logic, topic drift constraints are instantiated into semantic distance threshold assertions to ensure that the cosine similarity of topic vectors of adjacent events is higher than a preset lower limit.

[0034] The numerical comparison assertions and state transition assertions mentioned above are encapsulated into structured objects that can be directly parsed by the machine, generating the prototype of the initial dimension contract expression.

[0035] By using symbolic encoding, the typical semantic constraint patterns extracted in the previous step are transformed into initial dimensional contract expression prototypes with clear logical judgment capabilities, realizing the transformation from statistical regularities to executable code rules, and providing standardized input for subsequent lightweight verification.

[0036] S1.5: Perform integrity verification and encapsulation on the initial dimension contract expression prototypes to generate a set of dimension contract expressions. The logical consistency verification engine checks for constraint conflicts within each initial dimension contract expression prototype and merges similar terms. Finally, the output is a set of dimension contract expressions containing the constraints of the timestamp difference range between adjacent event nodes and the topological order rules of the event type sequence.

[0037] Step S2: The set of dimensional contract expressions is compiled into a lightweight state machine-based contract validator. Online structured mapping processing is performed on the raw text lexical stream output in real-time by the decoder during the file content generation process to obtain a structured semantic summary. This structured semantic summary is then input into the contract validator, which outputs the dimensional compliance determination result. Specifically, this includes: S2.1: Based on the set of dimensional contract expressions constructed in the preceding steps, the dimensional contract expressions containing timestamp difference interval constraints and event type sequence topological order rules are compiled using the finite state automaton construction method to generate a lightweight state machine form of contract verifier with multi-state transition logic.

[0038] Receive the set of dimensional contract expressions output from step S1. This set contains formal definitions of timestamp difference interval constraints and event type sequence topological order rules. Parse the state node definitions and transition conditions in the dimensional contract expressions, extracting the initial state identifier, termination state identifier, and intermediate state set. Construct the state space of a finite state automaton, mapping each semantic constraint stage to an independent state node, ensuring that each state node has a unique integer index identifier. Define directed transition edges between states according to the event type sequence topological order rules, associating each transition edge with a trigger condition predicate, including time interval value range verification logic and topic tag matching logic. Introduce error state nodes to capture abnormal input sequences that do not meet any preset transition conditions, ensuring the state machine has complete fault tolerance. Compile the state transition functions, converting the formal contract logic into a deterministic state transition matrix, optimizing matrix sparsity to reduce memory usage. Generate a lightweight state machine-style contract validator, which includes a state register, transition logic engine, and output interface, supporting millisecond-level concurrent verification requests. By constructing a finite state automaton, the dimensional contract expression from the previous step is transformed into a lightweight state machine-based contract validator with multi-state transition logic, enabling executable compilation of semantic constraints and providing a low-latency computational foundation for subsequent real-time verification.

[0039] S2.2: For the raw text word stream output in real time by the decoder in the content generation process, online structured mapping is performed using domain entity extraction methods and relation triple parsing to extract a structured semantic summary containing event node identifiers, time attribute values ​​and topic classification tags.

[0040] The system receives the raw text word stream output in real time from the decoder during the content generation process and uses it as the initial input data object for online structured mapping processing. A pre-trained domain entity extraction method is used to perform a sliding window scan on the raw text word stream to identify and locate key semantic segments containing time, people, and events. For the identified time semantic segments, regular expression matching and date normalization rules are applied to extract standardized time attribute values, eliminating ambiguity in relative time expressions. For person and event semantic segments, dependency parsing techniques are used to parse the subject-verb-object structure and construct preliminary person-action-object relation triples. Based on the domain ontology library, entity linking and disambiguation processing are performed on the preliminary relation triples, mapping non-standardized entities to unique event node identifiers. The contextual semantics of the currently generated segment are vector-encoded using a topic classification model, and the cosine similarity between the vector and the preset standard topic concept vector is calculated to determine the topic classification label with the highest confidence. The extracted event node identifiers, time attribute values, and topic classification labels are integrated and serialized and encapsulated according to a predefined data structure protocol. The above-mentioned online structured mapping process transforms the unstructured raw text word stream into a structured semantic summary containing event node identifiers, time attribute values, and topic classification labels. This enables real-time quantification and standardization of the semantic features of the archive content, providing accurate input for subsequent millisecond-level matching verification of the lightweight state machine.

[0041] S2.3: The structured semantic summary is injected as an input signal into the lightweight state machine form of the contract verifier, and parallel matching verification calculation is performed based on predefined state transition conditions to generate a Boolean state tag sequence representing whether the currently generated fragment satisfies the semantic constraints of each dimension.

[0042] The structured semantic summary is parsed into a state machine input vector, and the timestamp value, topic tag ID, and relationship edge type are extracted as key attribute fields for state transition.

[0043] Invoke a lightweight state machine-based contract validator, load pre-compiled timeline-dimensional contract rules, and construct numerical comparison logic based on the time difference between the current time node and the previous time node.

[0044] A parallel computing architecture is adopted, and state verification threads for the topic dimension and the person / event dimension are started simultaneously. The semantic similarity threshold between the current topic tag and the historical topic sequence, as well as the topological connectivity constraints of the person relationship edge, are compared respectively.

[0045] For the timeline dimension, perform interval matching operation to determine whether the timestamp difference between adjacent event nodes falls within the preset tolerance range and generate a time compliance boolean flag.

[0046] For the topic dimension, calculate the cosine distance between the current topic vector and the center of the context topic cluster. If the distance exceeds the drift mutation threshold, it is determined to be a topic violation, and a topic compliance boolean flag is generated.

[0047] For the character event dimension, check whether the newly generated character relationship triplet has a logical conflict with the established character relationship graph. If there is a contradictory relationship pointing, it is judged as a relationship breach, and a relationship compliance Boolean flag is generated.

[0048] The Boolean status markers of each dimension are aggregated to form a status marker sequence containing three dimensions: time, topic, and person. Each element represents the compliance status of the corresponding dimension at the current generation step.

[0049] By using a parallel matching and verification process, the structured semantic summary from the previous step is transformed into a multi-dimensional Boolean state label sequence, achieving the expected technical effect of real-time monitoring of multi-dimensional semantic constraints and preliminary identification of defaults at the millisecond level.

[0050] For example, suppose the currently generated archive fragment contains the event "Victory in the War of Resistance Against Japan in 1945". The structured semantic summary extracts the timestamp 1945, the topic tag "End of the War", and the relationship "Japanese Surrender". The state machine engine loads the timeline contract, setting the minimum interval between adjacent major historical events to 1 year. The previous event is "The Battle of Henan-Hunan-Guangxi in 1944", with a time difference of 1 year, satisfying the constraint of greater than or equal to 1 year, so the time dimension is marked as true. In the topic dimension, the cosine distance between the topic vectors of "End of the War" and the preceding "Battle in Progress" is 0.85, which is less than the set mutation threshold of 0.9, so the topic dimension is marked as true. In the relationship dimension, "Japanese Surrender" conforms to the logical edge type of "Aggressor-Defeat", with no conflict, so the relationship dimension is marked as true. The final output state label sequence is [True, True, True], indicating that the current fragment conforms to the semantic contract in all dimensions, and no reconciliation mechanism needs to be triggered.

[0051] S2.4: Based on the continuous change trend of the Boolean state label sequence, the multi-dimensional verification results are fused and discriminated using time-series logic aggregation rules to output a dimension compliance judgment result containing specific default dimension indexes and conflict type labels.

[0052] Based on a Boolean state tag sequence, which contains real-time compliance binary flags for three dimensions: timeline, topic, and person / event, a sliding window temporal analysis is performed on the Boolean state tag sequence. The window length is set to the combination of the tags of the currently generated segment and its N previous historical segments to capture the continuous evolution trend of semantic state rather than the instantaneous state of isolated points. A temporal logic aggregation method is used to calculate the state reversal frequency and duration of continuous violations within the sliding window for each dimension, distinguishing between occasional noise interference and systematic semantic drift. For the timeline dimension, if the timestamp difference of multiple consecutive windows is detected as a violation, it is judged as a 'temporal logic break'; if it is only a single jump followed by recovery, it is marked as 'local temporal jitter' and ignored. For the topic dimension, the cosine similarity change rate of the topic tag vector within the window period is calculated. When the change rate exceeds a preset mutation threshold of 2 and is accompanied by a persistently false compliance flag, it is judged as a 'topic drift mutation'. For the person / event dimension, a topological connectivity verification matrix is ​​constructed for event triples. If the matrix rank undergoes unexpected dimensionality reduction and the corresponding compliance flag is false, it is determined as a 'person / event relationship break'. The judgment results of the above dimensions are mapped to a predefined conflict type label table, generating a structured data tuple containing a breach dimension index, conflict severity level, and specific conflict type label. Through multi-dimensional temporal logic aggregation processing, discrete Boolean labels are transformed into semantically interpretable dimensional compliance judgment results, achieving accurate location and classification of multi-dimensional semantic conflicts, and providing precise triggering basis for subsequent local semantic snapshot capture.

[0053] Step S3: Identify breach signals based on the dimensional compliance determination results. If a breach signal is detected, immediately freeze the local context window, capture and generate a local semantic snapshot marked with the breach dimension identifier and conflict type label. Specifically, this includes: S3.1: Perform logic level parsing on the dimensional compliance determination result to extract the high-level valid default signal, and map the default signal to a specific default dimension identifier and conflict type label, thereby generating a default event object with metadata tags.

[0054] The system receives the dimensional compliance judgment result output from S2.4. This data packet contains a Boolean-type tag sequence representing the verification status of each semantic dimension and a preliminary conflict index. The Boolean-type tag sequence undergoes logic level parsing processing. A bitmask scanning method is used to traverse the status bits of each dimension, identifying violation signals with a logic high level value, and separating them from the continuous data stream as independent abnormal event identifiers. The extracted high-level violation signals are mapped to a predefined semantic conflict dictionary. The specific violation dimension identifier is determined based on the dimension position index of the signal, such as a timeline dimension, topic dimension, or interpersonal relationship dimension. Based on the conflict type encoding rules and the entity attributes in the structured semantic summary of the currently generated fragment, the specific nature of the conflict is determined. For example, exceeding the timestamp difference limit is mapped to "time jump" encoding, and excessive cosine distance of the topic vector is mapped to "topic drift" encoding. A violation event object with metadata tags is constructed, encapsulating the violation dimension identifier, conflict type label, trigger time timestamp, and corresponding original lexical index into a structured data unit. Through the above logical parsing and mapping process, the Boolean judgment result of the previous step is transformed into a default event object with a clear semantic orientation, realizing the rapid location and classification of multi-dimensional semantic conflicts, and providing precise target guidance for subsequent local context freezing.

[0055] S3.2: Based on the default dimension identifier in the default event object, perform a read lock operation on the decoder status register in the current generation process to freeze the local context window data containing the vectors of the three preceding event nodes and the vectors of the two subsequent prediction nodes, thereby generating a local context data block in a static holding state.

[0056] Receive the default event object output from step S3.1. This object contains a high-level active default signal, a specific default dimension identifier, and a conflict type label. Parse the default dimension identifier in the default event object to determine the semantic dimension range that needs to be frozen, such as the timeline dimension or the interpersonal relationship dimension. Based on the default dimension identifier, locate the address space of the state register inside the decoder in the content generation process and identify the memory region storing the currently generated context vector sequence. Perform an atomic read lock operation on the decoder state register, prohibiting the write process from modifying this memory region, ensuring the static consistency of the context data during snapshot capture. Read the locked state register data and extract the high-dimensional embedding vectors containing the three preceding historical event nodes. These vectors represent the stable semantic background of the archived content. Simultaneously, extract the latent semantic vectors of the two subsequent prediction nodes. These vectors represent the next generation direction inferred by the model based on the current context. Concatenate the extracted preceding historical event node vectors and subsequent prediction node vectors in chronological order to construct a local context window data matrix. Add a read-only marker to this local context window data matrix to generate a statically held local context data block, preventing data drift due to continued decoder operation. By using the above-mentioned read locking and data extraction processing methods, the breach signal in the previous step is transformed into a local context data block containing semantic information from the preceding and following texts. This achieves precise freezing and isolation of key contextual information around the point of semantic conflict, providing a stable data foundation for subsequent multi-dimensional feature decoupling.

[0057] S3.3: Using predefined semantic structure extraction operators, multi-dimensional feature decoupling processing is performed on the local context data block to separate the timeline dimension temporal feature sequence, the topic dimension semantic cluster distribution matrix, and the person-event relationship dimension topological connection graph, thereby generating a multimodal semantic feature component set.

[0058] The system receives the statically held local context data block output from step S3.2. This block contains frozen state information for the vectors of the three preceding event nodes and the vectors of the two subsequent prediction nodes. It then calls a predefined semantic structure extraction operator library to perform temporal feature decoupling processing along the timeline dimension. The system extracts the standardized timestamp attributes of each event node from the local context data block to construct a one-dimensional time series vector. It calculates the time interval difference between adjacent nodes to generate a time span feature sequence, used to characterize the rhythm density of event evolution. Finally, it calculates the standard deviation and skewness of the time intervals using a sliding window statistical method to quantify the uniformity of the time distribution, forming a timeline-dimensional temporal feature sequence.

[0059] The semantic cluster distribution matrix is ​​constructed for the topic dimension. Topic classification labels and their corresponding confidence scores are extracted from all text fragments within the local context window. These labels are mapped to a pre-trained topic embedding space to obtain high-dimensional topic vector representations. Weighted aggregation of topic vectors within the window is performed, and the centroid coordinates of the topics are calculated. A topic similarity matrix is ​​constructed, where each element represents the cosine similarity between topic vectors of any two nodes. A spectral clustering algorithm is used to softly partition the topic vectors, generating a topic membership distribution matrix that characterizes the concentration and dispersion of topic concepts within the current local window, thus forming the topic-dimensional semantic cluster distribution matrix.

[0060] A topological connection graph reconstruction process is performed for the character-event relationship dimension. The character-action-object triplet information in the local context data block is parsed. A directed weighted graph model is constructed, where nodes represent entities (characters or events) and edges represent semantic relationship types. Edge weights are assigned based on the semantic strength of the relationship type; for example, causal relationships have higher weights than parallel relationships. The adjacency matrix and Laplacian matrix of the graph are calculated, and topological features are extracted, including node degree centrality, clustering coefficient, and shortest path length. Connected components and isolated nodes in the graph are identified, potential relationship breakpoints are marked, and a topological connection graph for the character-event relationship dimension is generated. The generated timeline-dimensional temporal feature sequence, topic-dimensional semantic cluster distribution matrix, and person-event relationship-dimensional topological connection graph are structurally encapsulated. A unified data interface format is implemented, and dimension identifiers and timestamp index metadata are added. A multimodal feature alignment mechanism ensures the synchronization of features across dimensions on the timeline. The final output is a multimodal semantic feature component set containing three-dimensional semantic feature components, providing standardized input data for subsequent cross-dimensional alignment deviation calculations.

[0061] By using a multi-dimensional feature decoupling process, the local context data block from the previous step is transformed into a structured feature representation with three orthogonal dimensions: time, topic, and relationship. This enables refined separation of semantic conflict sources and significantly improves the targeting and accuracy of subsequent reconciliation strategies.

[0062] For example, the local context data block contains 5 event nodes. Timestamps [1945, 1946, 1949, 1978, 1980] are extracted from the timeline dimension, and the calculated intervals are [1, 3, 29, 2] years with a standard deviation of 12.8, identifying a significant jump from 1949 to 1978. After extracting vectors from the topic dimension, a cosine similarity matrix is ​​calculated, revealing that the similarity between nodes 3 and 4 is only 0.35, below the threshold of 0.6, and is marked as topic drift. A graph is constructed from the relationship dimension, detecting that node 4 has no edge connection to its preceding and following nodes, with a clustering coefficient of 0, and is marked as a relationship break. The final output is a multimodal feature set containing a time interval sequence, a topic similarity matrix, and a sparse relationship graph, accurately quantifying the specific forms of semantic inconsistency.

[0063] S3.4: Select the corresponding structural consistency verification rule according to the conflict type label, perform cross-dimensional alignment deviation calculation on the multimodal semantic feature component set, so as to quantify the time jump limit, topic drift mutation amplitude or character relationship break index, thereby generating a semantic conflict quantification descriptor containing specific numerical indicators.

[0064] The system receives a set of multimodal semantic feature components output from S3.3. This set includes a timeline-dimensional sequence of temporal features, a topic-dimensional semantic cluster distribution matrix, and a topological connection graph representing the relationship between people and events. It parses conflict type labels and routes them to the corresponding structural consistency verification rule engine based on the encoded values, loading a predefined deviation calculation method. For time jump exceeding limits scenarios, it extracts the timestamp values ​​of adjacent event nodes in the temporal feature sequence and calculates the deviation between the actual time interval and the tolerance range specified in the contract. It uses an absolute difference function to quantify the degree of semantic breakage in the time dimension, generating a time jump exceeding limit index. For topic drift and mutation scenarios, it calculates the cosine distance between the current topic vector and the center vector of the context topic cluster, combining this with the topic change rate within a sliding window to assess the dispersion of topic semantics. It uses a weighted Euclidean distance formula to fuse local mutation amplitude and global drift trend, generating a topic drift and mutation amplitude index. In the scenario of broken relationships between people, it traverses the edge weights of entity nodes in the topological connection graph to detect the existence of isolated nodes or contradictory relationship edges. It calculates the change rate of connected components and the proportion of logically conflicting edges in the relationship graph, quantifying the degree of structural damage to the relationship between people and events, and generating a relationship breakage index. The quantitative indicators of each dimension mentioned above are standardized and mapped to a unified dimensional range to construct a semantic conflict quantitative descriptor containing specific numerical indicators. By using a cross-dimensional alignment deviation calculation method, the multimodal features of the previous step are transformed into a measurable semantic conflict quantitative descriptor, achieving accurate quantification of multi-dimensional semantic inconsistencies and providing a precise basis for correction of the subsequent reconciliation proxy model.

[0065] S3.5: The default dimension identifier, conflict type label, local context data block and semantic conflict quantification descriptor are structurally encapsulated to construct a local semantic snapshot with a unified data interface, thereby outputting a standardized correction input data package that can be directly called by the reconciliation proxy model.

[0066] Step S4: Using a reconciliation proxy model with fewer than a preset threshold of parameters, the local semantic snapshot and the corresponding dimensional contract expression are used as joint inputs to perform minimal intervention correction calculations to generate a directional semantic reconciliation strategy that includes transitional explanatory phrase insertion instructions and time node fine-tuning parameters. Specifically, this includes: S4.1: Perform feature decoupling processing on local semantic snapshots, and use multimodal semantic feature component extraction operators to separate the timeline dimension temporal feature sequence, topic dimension semantic cluster distribution matrix and person-event relationship dimension topological connection graph, so as to generate a standardized multidimensional semantic conflict quantification descriptor set as the benchmark data for model input.

[0067] The system receives a local semantic snapshot output from S3.5. This snapshot contains a breach dimension identifier, a conflict type label, a frozen local context data block, and a preliminary semantic conflict quantization descriptor. It parses the multidimensional feature tensor in the local context data block and uses a predefined multimodal semantic decoupling operator to separate the hybrid embedding vector into a temporal feature sequence (timeline dimension), a semantic cluster distribution matrix (topic dimension), and a topological connectivity graph (person-event relationship dimension). For the timeline dimension, it extracts the timestamp difference sequence of adjacent event nodes to construct a one-dimensional temporal feature vector reflecting the chronological order and interval density of events. For the topic dimension, it calculates the cosine similarity matrix between the current generated fragment topic vector and the centroids of historical context topics, generating a semantic cluster distribution heatmap representing topic stability. For the person-event relationship dimension, it extracts the co-occurrence frequency and relationship path length between key entities based on the graph adjacency matrix, constructing a sparse topological connectivity graph to characterize the tightness of the person interaction network. Based on the semantic conflict quantization descriptor generated in S3.4, it performs weighted normalization on the features of each dimension to eliminate numerical biases caused by different units. Z-score normalization is used to transform the temporal feature sequence, ensuring that the distribution of time interval data has a mean of 0 and a standard deviation of 1. Softmax normalization is performed on the semantic cluster distribution matrix to transform it into a probability distribution form, highlighting the weights of the main thematic components. Laplacian normalization is applied to the sparse topological connectivity graph to enhance the numerical stability of the graph structure features. The processed feature vectors of each dimension are concatenated into a unified set of multi-dimensional semantic conflict quantification descriptors, serving as the baseline input data for the reconciliation surrogate model. Through multimodal semantic feature component extraction and normalization, the local semantic snapshot from the previous step is transformed into a structured input tensor that the model can understand, achieving a refined representation of multi-dimensional semantic conflicts and providing a high-fidelity data foundation for subsequent minimal intervention correction.

[0068] S4.2: Perform joint encoding operations based on the multidimensional semantic conflict quantization description subset and the dimensional contract expression set, and use a sparse attention mechanism to construct a context-aware input vector. Map the default dimension identifier, conflict type label and numerical constraint rules into a high-dimensional latent space representation to generate a model inference input tensor with semantic alignment characteristics.

[0069] S4.3: Using a harmonic surrogate model with fewer than a preset threshold of parameters, perform forward propagation calculation on the model inference input tensor. Identify semantic breakpoints and predict the optimal repair path through an internally deployed minimal intervention correction neural network layer to output the original correction decision vector containing the potential logical buffer position index and the time axis offset probability distribution.

[0070] The system receives a model inference input tensor with semantic alignment characteristics generated by S4.2. This tensor integrates multidimensional features of local semantic snapshots with constraint rules of dimensional contracts. The model inference input tensor is fed into a harmonic proxy model with fewer than a preset threshold parameters, initiating the forward propagation computation process. Using an internally deployed minimal intervention correction neural network layer, the system performs deep analysis of the high-dimensional latent space representation in the input tensor, capturing implicit dependencies between timeline, topic, and person / event dimensions through a multi-head attention mechanism. Semantic breakpoints in the input data are identified, locating the keyword meta-positions or entity nodes causing logical inconsistencies. Based on the identified semantic breakpoints and the topological order rules defined in the dimensional contract expression, the optimal repair path is predicted. A probability distribution estimation method is used to calculate the latent logical buffer position index for inserting transitional explanatory phrases at the breakpoints. Simultaneously, for conflicts in the timeline dimension, a regression analysis head is used to predict the probability distribution of time axis offset, generating preliminary estimates of time node fine-tuning parameters. The latent logical buffer position index and the time axis offset probability distribution are concatenated as vectors to form an original correction decision vector containing multi-level correction information. By employing a minimal intervention approach to correct the nonlinear mapping of the neural network layers, the output correction decision vector is ensured to meet both semantic coherence requirements and the hard constraints of the dimensional contract. Through the aforementioned forward propagation computation and prediction processing, the model inference input tensor is transformed into an original correction decision vector containing the latent logical buffer position index and the temporal axis offset probability distribution. This achieves accurate quantification and repair path planning for multi-dimensional semantic conflicts, providing a structured data foundation for subsequent discretization mapping and instruction generation.

[0071] For example, the reconciliation surrogate model has 50M parameters and a preset threshold of 60M, employing a Transformer encoder architecture. The input tensor has a dimension of 768, containing embedded representations of the preceding three event nodes and the subsequent two prediction nodes. The minimal intervention correction neural network layer within the model contains four layers of self-attention modules. When a time jump exceeding the limit conflict is detected, the attention mechanism focuses on two event nodes with significantly different timestamps. The model predicts that there is an optimal logical buffer point at index position 152, outputting the potential logical buffer position index as

[152] . Simultaneously, the regression analysis head outputs the time axis offset probability distribution, which, after softmax normalization, determines the optimal time fine-tuning parameter to be +5 years to smooth the huge span from 1949 to 1980. The final output of the original correction decision vector significantly improves the accuracy and interpretability of the correction strategy.

[0072] S4.4: Perform discretization mapping on the original correction decision vector, apply a natural language generation decoder to convert the potential logical buffer position index into specific transitional explanation phrase text fragments, and use regression analysis to calculate the time axis offset probability distribution into precise time node fine-tuning parameter values ​​to generate a structured local correction instruction candidate set.

[0073] The system receives the original correction decision vector output from S4.3, which contains latent logical buffer position indices and time axis offset probability distribution data. Discretization mapping is performed on the latent logical buffer position indices in the original correction decision vector. A pre-trained semantic cohesion lexicon is used to retrieve transitional phrase templates that match the default dimension identifier and conflict type label. The semantic similarity distance between the semantic features of the preceding event and the semantic features of the subsequent predicted event within the local context window is calculated and used as the weighting criterion for selecting specific transitional explanatory phrases. A natural language generation decoder is applied to instantiate the selected transitional phrase templates into specific text fragments, ensuring that the generated text maintains grammatical coherence with the preceding and following text, forming the initial logical buffer layer text data.

[0074] Based on the probability distribution of time axis offsets in the original revised decision vector, a weighted regression analysis method is used to calculate the precise time node fine-tuning parameter values. A time offset calculation model is constructed, using the high-confidence intervals in the probability distribution as weights to perform a weighted average calculation on the candidate time offsets. The final time node fine-tuning parameters are calculated using the following formula: in, Fine-tune the parameter values ​​for the final time point. For the first The probability weights of each candidate time offset For the first A number of candidate time offset values ​​were calculated. This calculation process eliminates noise interference in the probability distribution, ensuring the statistical significance and logical rationality of the time adjustment values.

[0075] The generated transitional explanatory phrase text fragments and the calculated time node fine-tuning parameter values ​​are structurally encapsulated. A data structure for the local correction instruction candidate set is established, including the insertion position index, the text content to be inserted, the target timestamp, and the adjustment range fields. The instruction candidate set is then standardized to conform to the input interface specifications required by step S5. Through the above discretization mapping and regression analysis, the correction decision vector abstracted in the previous step is transformed into a concrete and executable local correction instruction candidate set. This achieves a precise transformation from the quantitative description of semantic conflict to specific text and time correction operations, significantly improving the operability and execution efficiency of semantic reconciliation.

[0076] S4.5: Based on the structured local correction instruction candidate set, perform consistency verification and encapsulation operations, use policy optimization methods to select the optimal instruction combination that meets the constraints of the dimensional contract expression, and finally generate a directional semantic reconciliation policy containing a definite transitional explanation phrase insertion instruction and precise time node fine-tuning parameters as the final output of this step.

[0077] Step S5: Perform a partial reconstruction operation on the current archive content according to the directional semantic reconciliation strategy, dynamically adjust the logical connection relationships of the breach dimensions to output multi-dimensional content reconstruction intermediate results, and complete the archive compilation and content reconstruction based on the multi-dimensional content reconstruction intermediate results. Specifically, this includes: S5.1: Obtain the directional semantic reconciliation strategy containing the transitional explanatory phrase insertion instruction and the time node fine-tuning parameter, parse the transitional explanatory phrase insertion instruction in the directional semantic reconciliation strategy to extract the target insertion position index and the text segment to be inserted, and parse the time node fine-tuning parameter to calculate the time axis offset, and generate a structured local correction instruction set.

[0078] The system receives the directional semantic reconciliation strategy data packet output from step S4. This data packet contains transitional explanatory phrase insertion instructions and time node fine-tuning parameters. A structured parsing operation is performed on the transitional explanatory phrase insertion instructions, using JSON path lookup technology to extract the target insertion position index field and the text fragment to be inserted field, transforming the unstructured instruction stream into addressable text operation metadata. Based on the target insertion position index, the preceding boundary of the semantic breakpoint is located in the character-level coordinate system of the original archive content, determining the physical embedding anchor point of the transitional explanatory phrase. Numerical calculation processing is performed on the time node fine-tuning parameters, reading the time axis offset probability distribution vector and applying a weighted average method to calculate the final timestamp correction value. The time axis offset is calculated, and the obtained time axis offset is algebraically added to the timestamp of the original event node to generate calibrated absolute time coordinates. The extracted target insertion position index, the text fragment to be inserted, and the calibrated absolute time coordinates are integrated to construct a structured local correction instruction set containing position mapping relationships and time correction values. Through the above analysis and calculation methods, the abstract reconciliation strategy generated in the previous step is transformed into a structured local correction instruction set with precise execution coordinates and numerical parameters, realizing a lossless mapping from the decision space to the execution space, and providing standardized input for subsequent text slicing and fusion operations.

[0079] S5.2: Receive the original text sequence of the current file content and the local context window marked with the default dimension identifier, slice the original text sequence based on the target insertion position index in the structured local correction instruction set, separate the preceding stable semantic segment, the conflict semantic segment to be reconstructed and the subsequent predicted semantic segment, and generate a three-segment semantic segmentation data block.

[0080] The system receives the original text sequence of the current archive content and local context window data marked with breach dimension identifiers as input sources for semantic slicing processing. It parses the structured local correction instruction set and extracts the defined target insertion position index, which points to the specific character offset of the semantic break in the original text sequence. Based on the target insertion position index, it performs a bidirectional truncation operation on the original text sequence, dividing the sequence into a pre-sequence part and a post-sequence part. From the pre-sequence part, it extracts a text block containing the first three complete event nodes, defining it as a pre-sequence stable semantic fragment, ensuring the logical consistency of the timeline and thematic dimension within this fragment. From the post-sequence part, it extracts the text block to be processed immediately adjacent to the break point, defining it as a conflict semantic fragment to be reconstructed, containing the anomalous semantic units that lead to contract breach. It retains the remaining predictive text content in the post-sequence part, defining it as a subsequent predictive semantic fragment, used to maintain the coherence expectation of the archive content. It adds metadata tags to the pre-sequence stable semantic fragment, the conflict semantic fragment to be reconstructed, and the subsequent predictive semantic fragment, identifying their role attributes in the three-segment structure. It encapsulates the three text fragments with metadata tags into a unified data structure, generating a three-segment semantic segmentation data block. Through the above slicing and encapsulation processing methods, the continuous raw text stream is transformed into structured three-segment semantic segmentation data blocks, which realizes the precise isolation of the defaulting local area and the protection of the non-defaulting area, providing a clear operational boundary for subsequent targeted semantic reconciliation.

[0081] S5.3: Using the conflicting semantic segments to be reconstructed in the three-segment semantic segmentation data block, and combining the text segments to be inserted in the structured local correction instruction set, perform text fusion operation to embed the transitional explanation phrases into the connection between the preceding stable semantic segments and the conflicting semantic segments to be reconstructed, thereby generating a preliminary connected text stream with a logical buffer layer.

[0082] The system receives conflicting semantic segments to be reconstructed from a three-segment semantic segmentation data block and text segments to be inserted from a structured local correction instruction set. Boundary feature extraction is performed on the conflicting semantic segments to be reconstructed, and a dependency parser is used to identify the initial subject and predicate verb of the segment, determining the grammatical dependencies at the semantic breakpoint. The semantic attributes of the text segment to be inserted are parsed, and its cosine similarity with the preceding and following text in the topic vector space is calculated to ensure contextual compatibility of transition phrases. A dynamic text fusion matrix is ​​constructed to map the text segment to be inserted into an embedding vector sequence, and its attention weight distribution at the starting position of the conflicting semantic segment to be reconstructed is calculated. Based on the attention weight distribution, vector interpolation is performed to generate a smooth semantic transition layer, eliminating semantic abrupt changes caused by hard splicing. A dynamic alignment method based on edit distance is used to adjust the word order structure of the text segment to be inserted and the conflicting semantic segment to be reconstructed, ensuring grammatical coherence. The aligned text segment to be inserted is embedded at the connection point between the preceding stable semantic segment and the conflicting semantic segment to be reconstructed, forming a preliminary connected text flow with a logical buffer layer. By using text fusion processing, the discrete text fragments from the previous step are transformed into preliminarily connected text stream data with semantic continuity, achieving the expected technical effect of logical repair and smooth transition of local semantic breaks in multi-dimensional content reconstruction.

[0083] S5.4: For event nodes involving the timeline dimension in the preliminary connected text stream with the logical buffer layer, the timestamp attribute of the event node is recalibrated using the time axis offset in the structured local correction instruction set to correct the time logic discontinuity caused by semantic jumps and generate a time-logically consistent time-corrected text stream.

[0084] The system receives the initial concatenated text stream with a logical buffer layer and the timeline offset from the structured local correction instruction set, and locates the event nodes in the text stream that need to be calibrated along the timeline dimension. It extracts the original timestamp attribute values ​​of each event node and constructs an initial time-series vector sequence containing node identifiers and original time values.

[0085] Based on the time axis offset parameter, a linear translation transformation is performed on each time node in the initial time-series vector sequence. The corrected timestamp is determined using the following formula: in, This is the corrected timestamp. This is the original timestamp. This is the time axis offset.

[0086] The system traverses the initially connected text stream, identifying contextual semantic units associated with the corrected time nodes, including verb tense markers, time adverbial clauses, and causal conjunctions. It detects tense inconsistencies or logical inversions caused by timestamp changes and generates a list of tense conflict markers.

[0087] For the list of tense conflict markers, a predefined tense mapping rule library is invoked to adjust the affected verb forms to match the tense forms of the corrected timestamps. Relative time references in time adverbs are updated synchronously to ensure that indicator words such as "before" and "after" remain logically consistent with the new time series.

[0088] The corrected timestamps, adjusted tense markers, and updated semantic units are reassembled to replace the corresponding segments in the original text stream. A local syntactic integrity check is performed to eliminate syntactic breaks caused by the replacement, generating a temporally consistent, time-corrected text stream.

[0089] By using the above-mentioned linear translation and temporal synchronization processing methods, the initial connecting text stream from the previous step is transformed into a temporally corrected text stream with rigorous temporal logic and consistent temporality, thereby achieving precise repair of semantic breaks in the temporal dimension.

[0090] S5.5: Integrate the preceding stable semantic segments, the temporally consistent time-corrected text stream, and the subsequent predicted semantic segments, perform semantic coherence verification to ensure that the non-default dimension semantic structure has not been distorted, and finally synthesize and output the multi-dimensional content reconstruction intermediate result with semantic consistency preservation characteristics.

[0091] The system receives stable semantic segments, a temporally consistent corrected text stream, and subsequent predicted semantic segments as input data. Boundary alignment is performed on the three-segment semantic segmentation data blocks, extracting the end entity vector of the preceding segment and the beginning entity vector of the subsequent segment to construct a local semantic cohesion matrix. The semantic distance between adjacent entities in the cohesion matrix is ​​calculated using the cosine similarity method to generate an initial coherence score benchmark. For the temporally consistent corrected text stream, dependency parsing is applied to analyze the syntactic dependencies between the corrected event nodes, identifying potential subject-verb-object structure breakpoints. The identified syntactic dependencies are mapped to a domain ontology library, verifying role consistency in the person-event relationship dimension and generating a role conflict marker sequence. If a role conflict marker is detected, a local referencing resolution mechanism is initiated, correcting pronoun references based on entity co-occurrence frequencies within the context window and outputting the corrected referencing chain data. The stable semantic segments, the referencing-resolved temporally corrected text stream, and the subsequent predicted semantic segments are integrated, and a linear concatenation operation is performed to generate a preliminary reconstructed text sequence. A global semantic scan is performed on the initially reconstructed text sequence, and a sliding window mechanism is used to calculate the KL divergence of topic distribution between every two adjacent sentences. Then, a semantic distortion index is calculated. If the semantic distortion index exceeds a preset threshold of 0.15, distortion is determined to have occurred in the non-breach dimension, triggering a backtracking correction instruction and readjusting the weight coefficients of transitional explanatory phrases. If the semantic distortion index is below the preset threshold, the semantic structure of the non-breach dimension is confirmed to be intact. The initially reconstructed text sequence confirmed to be distortion-free undergoes format standardization processing, unifying punctuation and paragraph spacing, generating a multi-dimensional content reconstruction intermediate result with semantic consistency preservation characteristics. Through the above semantic coherence verification and distortion suppression methods, scattered semantic fragments are transformed into logically rigorous and structurally complete intermediate results, achieving dynamic maintenance of semantic consistency and zero-interference protection of the non-breach dimension during the multi-dimensional content reconstruction process.

[0092] like Figure 3 As shown, step S6 involves collecting user adoption rate feedback signals and downstream coherence scoring data from the intermediate results of the multi-dimensional content reconstruction to construct a contract feedback closed-loop dataset containing reconciliation operation records and effect evaluation indicators. Specifically, this includes: S6.1: Perform user interaction behavior monitoring and processing on the intermediate results of multi-dimensional content reconstruction, and use the tracking log analysis method to extract the user's adoption operation sequence and modification rejection instructions on the archive compilation interface, thereby generating a user adoption rate feedback signal vector that represents the user's subjective recognition.

[0093] The system receives intermediate results of multi-dimensional content reconstruction with semantic consistency preservation from the S5 output and uses them as a data source for monitoring user interaction behavior. An embedded listener is deployed on the front-end interface of the archival compilation system to capture real-time user interaction events, including clicks for acceptance, partial modification, complete deletion, and rejection of generation. The captured raw interaction logs are timestamped and associated with session IDs to construct an operation sequence dataset based on user session dimensions. A sliding window method is used to segment the operation sequence, extracting an independent subset of interaction behaviors corresponding to each reconstruction segment. Acceptance weight coefficients and rejection penalty coefficients are defined: direct acceptance operations are assigned a high positive weight; modification operations are assigned a partial acceptance weight based on the modification magnitude; and rejection operations are assigned a negative weight. A weighted statistical method is used to calculate the user's subjective acceptance score for each reconstruction segment. The subjective acceptance scores of each segment are arranged chronologically to construct a user acceptance rate feedback signal vector representing user subjective acceptance. Through the above processing, the reconstruction results of the previous step are transformed into a user acceptance rate feedback signal vector, achieving a quantitative assessment of the subjective quality of archival content and providing a basis for real user preferences for subsequent contract parameter optimization.

[0094] For example, the adoption weight coefficient is set to 1.0, the weight of modification operations decreases linearly from 0.8 to 0.2 as the edit distance increases, and the weight of rejection operations is -0.5. A user was detected performing direct adoption, minor modification (edit distance 5), and rejection operations on three reconstructed fragments within a 5-minute session. The calculated scores for each fragment were 1.0, 0.7, and -0.5, respectively. The generated user adoption rate feedback signal vector is [1.0, 0.7, -0.5]. This vector accurately reflects the user's acceptance of different semantic reconciliation strategies, significantly improving the granularity and accuracy of the feedback data.

[0095] S6.2: Based on the intermediate results of multi-dimensional content reconstruction corresponding to the user adoption rate feedback signal vector, the natural language processing consistency verification method is applied to calculate the logical coherence score and narrative fluency index between the generated text and the original archive semantic graph, thereby generating a downstream coherence score data set that represents the objective quality of the content.

[0096] The system receives the intermediate text sequence of multi-dimensional content reconstruction results output by S5 and the corresponding original archive semantic graph node data as input benchmarks for objective quality assessment.

[0097] The intermediate reconstruction results are processed by syntactic dependency parsing and reference resolution to construct a local semantic dependency tree of the generated text, extract the subject-verb-object core triples and modification relation edges, and form the semantic structure vector of the generation end.

[0098] The semantic structure vector generated by the graph isomorphism matching method is used to align the semantic structure vector of the generated end with the semantic graph of the original archive. The structural similarity scores of the two in terms of node coverage and edge connectivity consistency are calculated to quantify the basic indicators of logical coherence.

[0099] The reconstructed text is segmented into narrative units based on a sliding window mechanism. The rate of change of semantic cosine distance between adjacent narrative units is calculated, semantic abrupt change points are identified, and a smoothness index is calculated to generate a preliminary narrative fluency score.

[0100] A timeline constraint verification module is introduced to detect whether the timestamp sequence of event nodes in the generated text satisfies monotonically increasing or reasonable backtracking logic, and calculates the number of time logic violations and the average offset as a penalty factor for logical coherence.

[0101] A comprehensive score for downstream coherence is calculated using a weighted fusion formula, taking into account structural similarity, narrative smoothness index, and temporal logic penalty factor. in, The overall score is based on downstream continuity. Score the structural similarity. The narrative fluency index, This is the time logic penalty value. , , These are the preset weighting coefficients.

[0102] The calculated downstream coherence comprehensive score and narrative fluency sub-indicators are encapsulated into a standardized downstream coherence score data set and associated with the current reconstruction batch ID.

[0103] By using the above-mentioned natural language processing consistency verification method, the intermediate results of the previous step of reconstruction are transformed into quantitative logical coherence scores and narrative fluency indicators, thereby realizing an automated and accurate assessment of the objective quality of the archive content and providing reliable data support for subsequent contract parameter optimization.

[0104] S6.3: Extract metadata from historical reconciliation operation records that trigger the directional semantic reconciliation strategy. Use time-series association mapping technology to structurally bind specific transitional explanation phrase insertion instructions, time node fine-tuning parameters, and corresponding default dimension identifiers and conflict type labels, thereby generating a reconciliation operation record dataset containing the complete correction trajectory.

[0105] Receive the raw log stream of historical reconciliation operation records that triggered the targeted semantic reconciliation strategy, and use the regular expression parsing engine to extract the timestamp, session ID and operation type fields from each record to generate a preliminary structured metadata index table.

[0106] Perform time-series alignment processing on the structured metadata index table, reassemble discrete operation logs into continuous operation sequences based on session IDs, identify the start point and reconciliation end point of each semantic conflict, and construct a time-series association mapping window.

[0107] Within the temporal correlation mapping window, the specific text content, insertion position offset, and context window identifier of the transitional explanation phrase insertion instruction are extracted and encoded into a natural language correction feature vector.

[0108] Simultaneously extract the original timestamp, corrected timestamp, and time axis offset values ​​from the time node fine-tuning parameters, and convert them into numerical time correction feature vectors to ensure the quantification and computability of time dimension data.

[0109] A hash mapping algorithm is used to bind the natural language modified feature vector and the numerical time modified feature vector with key values ​​through the default dimension identifier and conflict type label to establish a strong correlation index of multimodal modified features.

[0110] Perform integrity checks on the bound multimodal correction features, remove dirty data with missing key fields or abnormal formats, and ensure that each record contains a complete 'conflict-correction' causal chain.

[0111] The verified records are encapsulated into reconciliation operation objects in standard JSON format, including the default dimension, conflict type, correction instruction set, and time correction parameters, forming a structured reconciliation operation record dataset.

[0112] By using the aforementioned temporal correlation mapping and multimodal feature binding processing method, the discrete operation logs from the previous step are transformed into a structured harmonic operation record dataset with complete correction trajectories, thereby achieving traceability of the harmonic process and providing a basis for quantitative analysis.

[0113] S6.4: Based on the reconciliation operation record dataset, user adoption rate feedback signal vector, and downstream coherence scoring data set, perform multi-source heterogeneous data fusion and alignment operation, and use a standardized data encapsulation protocol to aggregate the correction trajectory, subjective recognition index, and objective quality index into a unified format effect evaluation index object, thereby generating a comprehensive effect evaluation index set with multi-dimensional evaluation dimensions.

[0114] The system receives a reconciliation operation record dataset, a user adoption rate feedback signal vector, and a downstream coherence scoring dataset as input sources for multi-source heterogeneous data fusion and alignment. It performs timestamp alignment on the correction trajectories in the reconciliation operation record dataset, extracting the specific time of each semantic reconciliation and its corresponding default dimension identifier, and constructs a time-series-based operation event index table. The user adoption rate feedback signal vector is mapped to the same time-series axis, and missing values ​​of discontinuous sampling points are filled using linear interpolation to generate a subjective acceptance time-series sequence with the same time granularity as the operation event index table. Normalization is performed on the downstream coherence scoring dataset to eliminate differences in different scoring dimensions, and the sliding window average of each scoring indicator and the operation event time is calculated to generate an objective quality time-series sequence. A unified data structure template is defined using a standardized data encapsulation protocol, including correction trajectory fields, subjective acceptance fields, and objective quality fields, ensuring semantic consistency of multi-source data at the logical level. Multi-source data association and matching operations are performed, using the operation event index table as the primary key, aggregating corresponding data units from the subjective acceptance time-series and the objective quality time-series into a unified format performance evaluation indicator object. The aggregated performance evaluation indexes undergo integrity checks to remove records with missing data or outliers, generating a comprehensive performance evaluation index set with multi-dimensional evaluation dimensions. Through standardized data encapsulation and multi-source temporal alignment, the discrete and heterogeneous data from the previous step is transformed into a unified, spatiotemporally aligned comprehensive performance evaluation index set, achieving the expected technical effect of quantitative evaluation of multi-dimensional content reconstruction effects.

[0115] S6.5: Perform persistent storage and index construction processing on the comprehensive effect evaluation index set, and use the distributed database write interface to archive the data units containing the correction trajectory, subjective recognition index and objective quality index in the order of timestamps, and finally output the contract feedback closed loop dataset for quantifying the collaborative boundary optimization requirements between dimensions.

[0116] The system receives the comprehensive performance evaluation index set generated by S6.4. This dataset includes correction trajectories, user adoption rate feedback signals, and downstream coherence scores. Data serialization preprocessing is performed on the comprehensive performance evaluation index set, using JSON-LD format to uniformly encapsulate unstructured text fragments and structured numerical indicators, ensuring the integrity of semantic metadata. A timestamp-based distributed index key-value pair is constructed, extracting the generation time from each record as the primary index field and the default dimension identifier as the secondary partition key, achieving hash sharding location of data in the storage cluster. The distributed database write interface is called to establish an asynchronous batch write channel, appending the serialized data units to the persistent storage layer in chronological order, utilizing a write-ahead log mechanism to ensure data atomicity and fault recovery capabilities. A cold / hot separation strategy is implemented for the incoming data, residing recently frequently accessed contract feedback data on high-speed cache nodes and migrating historical archived data to low-cost object storage, optimizing storage resource utilization. Through the above processing method, the comprehensive effect evaluation index of the previous step is transformed into a contract feedback closed-loop dataset with spatiotemporal indexing characteristics, which provides the data foundation support required for the continuous optimization of the dynamic maintenance mechanism of semantic consistency in the process of multi-dimensional content reconstruction.

[0117] Step S7: Dynamically update the dimensional contract expression set based on the contract feedback closed-loop dataset, generate a new version of the dimensional contract expression set that evolves adaptively, and replace the original configuration. Specifically, this includes: S7.1: Perform multi-dimensional feature extraction on the reconciliation operation records and effect evaluation indicators in the contract feedback closed-loop dataset to obtain a contract evolution feature vector set containing the time series sequence of user adoption rate and the downstream coherence score distribution vector.

[0118] The system reads the reconciliation operation records and comprehensive effect evaluation index set stored in the contract feedback closed-loop dataset, and extracts structured metadata containing default dimension identifiers, conflict type labels, and correction trajectories. Time series alignment processing is performed on the extracted metadata, mapping discrete reconciliation operation records to a continuous time axis based on the timestamps of user interaction behaviors, constructing a time-series operation event flow. For each default dimension, the trend of user adoption rate changes within a unit time window is statistically analyzed, and a sliding window averaging method is used to smooth noise interference, generating a time-series sequence of user adoption rate representing the dynamic evolution of user subjective acceptance. Simultaneously, the downstream coherence scoring dataset is acquired, and the logical coherence score distribution of each dimension under specific conflict types is calculated. The kernel density estimation method is used to fit the score probability density function, and the mean, variance, and skewness are extracted as statistical features to form a downstream coherence scoring distribution vector. The user adoption rate time-series sequence and the downstream coherence scoring distribution vector are mapped into a multi-dimensional space. Principal component analysis is used for dimensionality reduction to eliminate feature redundancy, retaining the principal component components with the highest contribution rate, and constructing a high-dimensional contract evolution feature vector set. Through the above processing method, the discrete feedback data of the previous step is transformed into a contract evolution feature vector set with temporal correlation and statistical distribution characteristics, thereby realizing the quantitative representation of the optimization requirements of the semantic consistency maintenance mechanism.

[0119] For example, a time window length of 24 hours and a sliding step of 1 hour are set. Harmonized records from the past 7 days are extracted from the database, containing 500 timeline-dimensional default events. The hourly user adoption rate is calculated. The raw data fluctuates significantly; after sliding window averaging, a smooth time-series adoption rate sequence is obtained, with values ​​ranging from 0.65 to 0.85. Kernel density estimation is performed on the logical coherence scores within the same time period, with a bandwidth parameter set to 0.1. The fitted distribution vector has a mean of 0.78 and a standard deviation of 0.05. These two feature vectors are concatenated and subjected to PCA dimensionality reduction, retaining the top three principal components with a cumulative contribution rate of 95%, ultimately outputting a 3D contract evolution feature vector. This vector accurately reflects the potential correlation between the current timeline contract tolerance threshold and user satisfaction and content quality, providing a high-precision input basis for subsequent parameter gradient calculations.

[0120] S7.2: Based on the contract evolution feature vector set, perform deviation mapping calculation, use gradient descent optimization method to derive the residual distance between the current tolerance threshold parameter and the theoretical optimal value, so as to generate a contract parameter update gradient including timestamp difference interval correction amount and event type sequence topology order adjustment coefficient.

[0121] The system receives a set of contract evolution feature vectors and extracts the user adoption rate time series and downstream coherence rating distribution vector as input variables for the optimization objective function. It constructs a gradient descent-based parameter optimization model and defines a loss function to quantify the deviation distance between the current tolerance threshold and the topology rule weights and the theoretical optimal values.

[0122] We use mean squared error as the basic metric, combine it with regularization to prevent overfitting, construct a loss function, calculate the partial derivative of the loss function, calculate the gradient vector, and determine the direction and step size of parameter updates.

[0123] By backpropagating the error using the chain rule, the gradient components of the time dimension tolerance threshold and the topic dimension topological weight are calculated respectively.

[0124] An adaptive moment estimation optimization algorithm is applied to adjust the learning rate, accelerate the convergence process, and avoid local minima traps.

[0125] Generate a contract parameter update gradient vector that includes timestamp difference interval correction and event type sequence topological order adjustment coefficients.

[0126] By using the gradient descent optimization method, the contract evolution characteristics are transformed into specific parameter correction gradients, thereby achieving adaptive and precise tuning of the dimensional contract expression.

[0127] For example, an initial timestamp difference tolerance threshold of 5 days and a topological order weight of 0.8 were set. 100 sets of feedback data were collected, with an actual coherence score of 0.85 and a predicted score of 0.75. Substituting these values ​​into the formula to calculate the loss function, backpropagation yielded a temporal tolerance gradient of -0.02 and a topological weight gradient of 0.05. Optimization methods were then applied to update the parameters, adjusting the tolerance threshold to 4.8 days and the weight to 0.85. Further verification showed a significant reduction in the temporal logical breakage rate of the archive content and a substantial improvement in the narrative fluency index.

[0128] S7.3: Use the contract parameters to update the gradient and perform a weighted iterative update operation on the tolerance threshold parameter and topology rule weight coefficient in the original set of dimensional contract expressions to generate a new set of dimensional contract expressions with adaptive evolution characteristics.

[0129] Receive the contract parameter update gradient generated by S7.2. This gradient includes the timestamp difference interval correction amount and the event type sequence topological order adjustment coefficient, which serves as the driving signal for the dynamic update of the dimensional contract expression set.

[0130] We perform weighted iterative calculations on the tolerance threshold parameters in the original set of dimensional contract expressions, and use the momentum-accelerated gradient descent method to linearly combine the current tolerance threshold with the gradient correction amount to generate a new version of the time span tolerance interval parameters.

[0131] Normalization and redistribution processing is performed on the weight coefficients of the topological order rules for the event type sequence. Based on the topological order adjustment coefficients, the probability weights of each event transition path are recalculated using the Softmax function to ensure that the sum of the weights of all possible transition paths is constant at 1, thereby generating a new version of the topological rule weight vector.

[0132] The calculated new time span tolerance interval parameters are instantiated as numerical comparison assertions, replacing the static threshold field in the original contract expression, thus forming a time-dimensional contract subset with dynamic adaptability.

[0133] The new topology rule weight vector is mapped to the conditional probability field of the state transition assertion, and the logical constraint strength of the character-event relationship dimension is updated to form a subset of relationship dimension contracts with evolutionary characteristics.

[0134] Perform a logical consistency merging operation on the updated time-dimensional contract subset and the relation-dimensional contract subset to eliminate internal constraint conflicts that may arise from independent parameter updates, and encapsulate and generate a new set of dimensional contract expressions with adaptive evolution characteristics.

[0135] By using weighted iterative updates and normalized redistribution, the parameter update gradient from the previous step is transformed into a new set of dimensional contract expressions with dynamic adaptability, thereby enabling continuous optimization and self-evolution of the semantic consistency maintenance mechanism during multi-dimensional content reconstruction.

[0136] S7.4: Load the new version of the dimensional contract expression set into the contract validator configuration register in the form of a lightweight state machine, and replace the original dimensional contract expression set configuration to complete the hot update of the parameters of the semantic consistency dynamic maintenance mechanism during the multi-dimensional content reconstruction process.

[0137] Receive the new set of dimensional contract expressions with adaptive evolution characteristics generated by S7.3. This set includes the updated timestamp difference interval tolerance threshold and the event type sequence topological order weight coefficient.

[0138] The new set of dimensional contract expressions is serialized and encoded, and the symbolic contract rules are converted into compact byte stream data using a binary protocol buffer format to reduce network transmission overhead and improve parsing efficiency.

[0139] Establish a secure communication channel with the contract verifier in the form of a lightweight state machine, and send the serialized byte stream data to the verifier's configuration register interface through atomic write operation instructions to ensure the atomicity and consistency of the configuration update process.

[0140] Version compatibility checks are performed inside the configuration register, comparing the syntax structure of the new contract expression with the instruction set version supported by the current state machine engine. If an incompatible item is detected, a rollback mechanism is triggered to retain the original configuration; otherwise, the loading process begins.

[0141] By using dynamic remapping technology, the numerical comparison assertions and state transition assertions in the new contract expression are compiled into the state transition table of a finite state automaton. The transition matrix and the set of accepting states of the state machine are updated in real time, and the underlying logic is hot-swapped.

[0142] Send a configuration activation signal to the decoder of the content generation module to activate the new state machine model instance, so that the subsequently generated word streams can be immediately matched and verified in milliseconds based on the updated semantic constraints.

[0143] By using the above processing method, the contract parameters optimized in the previous step are transformed into executable verifier configuration data, realizing uninterrupted hot updates of the semantic consistency dynamic maintenance mechanism during the multi-dimensional content reconstruction process, and ensuring that the continuous optimization effect takes effect immediately.

[0144] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0145] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0146] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for archival research and content reconstruction based on generative artificial intelligence, characterized in that, include: S1: For multi-source heterogeneous archival data in archival auxiliary compilation scenarios, extract multi-dimensional semantic features of timeline, theme and relationship between people and events, and construct a set of dimensional contract expressions; S2: Compile the set of dimensional contract expressions into a lightweight state machine form of contract validator, perform online structured mapping processing on the raw text word stream output in real time by the decoder in the file content generation process to obtain a structured semantic summary, input the structured semantic summary into the contract validator, and output the dimensional compliance judgment result. S3: Identify breach signals based on the compliance determination results of the aforementioned dimensions. If a breach signal is detected, immediately freeze the local context window, capture and generate a local semantic snapshot marked with the breach dimension identifier and conflict type label. S4: Using a reconciliation proxy model with fewer than a preset threshold of parameters, the local semantic snapshot and the corresponding dimensional contract expression are used as joint inputs to perform minimal intervention correction calculations to generate a directional semantic reconciliation strategy that includes transitional explanatory phrase insertion instructions and time node fine-tuning parameters. S5: Perform partial reconstruction of the current archive content according to the directional semantic reconciliation strategy, dynamically adjust the logical connection relationship of the breach dimension to output the intermediate result of multi-dimensional content reconstruction, and complete the compilation and content reconstruction of the archive based on the intermediate result of multi-dimensional content reconstruction.

2. The method for archival research and content reconstruction based on generative artificial intelligence according to claim 1, characterized in that, S5 is followed by: S6: Collect user adoption rate feedback signals and downstream coherence scoring data of the intermediate results of the multi-dimensional content reconstruction, and construct a contract feedback closed-loop dataset that includes reconciliation operation records and effect evaluation indicators. S7: Dynamically update the dimensional contract expression set based on the contract feedback closed-loop dataset, generate a new version of the dimensional contract expression set that evolves adaptively, and replace the original configuration.

3. The method for archival research and content reconstruction based on generative artificial intelligence according to claim 1, characterized in that, The local context window includes preceding event nodes and subsequent prediction nodes.

4. The method for archival research and content reconstruction based on generative artificial intelligence according to claim 1, characterized in that, The reconciliation proxy model adopts the Transformer encoder architecture.

5. The method for archival research and content reconstruction based on generative artificial intelligence according to claim 1, characterized in that, S2 specifically includes: Based on the set of dimensional contract expressions, the dimensional contract expressions are compiled to generate a contract checker in the form of a lightweight state machine. For the raw text stream output in real time by the decoder in the content generation process, online structured mapping processing is performed using relation triple parsing to extract structured semantic summaries; The structured semantic summary is injected as an input signal into the lightweight state machine-type contract verifier, and parallel matching verification calculations are performed based on predefined state transition conditions to generate a Boolean state tag sequence. Based on the continuous change trend of the Boolean state label sequence, the verification results of multiple dimensions are fused and discriminated using time-series logic aggregation rules to output a dimension compliance judgment result containing specific default dimension indexes and conflict type labels.

6. The method for archival research and content reconstruction based on generative artificial intelligence according to claim 5, characterized in that, The dimensional contract expression includes timestamp difference range constraints and event type sequence topological order rules.

7. The method for archival research and content reconstruction based on generative artificial intelligence according to claim 5, characterized in that, The structured semantic summary includes event node identifiers, time attribute values, and topic classification tags.

8. The method for archival research and content reconstruction based on generative artificial intelligence according to claim 3, characterized in that, S3 specifically includes: The dimensional compliance determination results are processed by logic level parsing to extract the high-level valid default signal, and the default signal is mapped to a specific default dimension identifier and conflict type label, thereby generating a default event object with metadata tags; Based on the default dimension identifier in the default event object, a read lock operation is performed on the decoder status register in the current generation process to freeze the local context window data containing the vectors of the three preceding event nodes and the vectors of the two subsequent prediction nodes, thereby generating a local context data block. Multidimensional feature decoupling is performed on local context data blocks to obtain a set of multimodal semantic feature components; Based on the conflict type label, select the corresponding structural consistency verification rule, perform cross-dimensional alignment deviation calculation on the multimodal semantic feature component set, and generate a semantic conflict quantification descriptor; The default dimension identifier, conflict type label, local context data block, and semantic conflict quantification descriptor are structurally encapsulated to construct a local semantic snapshot.

9. The method for archival research and content reconstruction based on generative artificial intelligence according to claim 8, characterized in that, The multi-dimensional feature decoupling process performed on the local context data block to obtain a set of multimodal semantic feature components includes: The local context data block is subjected to multi-dimensional feature decoupling processing to separate the timeline dimension temporal feature sequence, the topic dimension semantic cluster distribution matrix, and the person-event relationship dimension topological connection graph, thereby generating a multimodal semantic feature component set.

10. A system for archival research and content reconstruction based on generative artificial intelligence, characterized in that, The method for performing the generative artificial intelligence-based archival compilation and content reconstruction method as described in any one of claims 1-9 includes: Feature extraction module: used to extract multi-dimensional semantic features of timeline, theme and relationship between people and events from multi-source heterogeneous archival data in archival auxiliary compilation and research scenarios, and to construct a set of dimensional contract expressions; Contract validator: It is formed by compiling the set of dimensional contract expressions into a lightweight state machine. It is used to receive the structured semantic summary of the currently generated fragment as input after each step of the decoder output in the content generation process, perform millisecond-level matching verification, and output the dimensional compliance judgment result. The default detection module is used to identify default signals based on the compliance judgment results of the aforementioned dimensions. If a default signal is detected, the local context window is immediately frozen, and a local semantic snapshot marked with the default dimension identifier and conflict type label is captured and generated. Harmony Proxy Model: A proxy model with fewer than a preset threshold of parameters, used to take the local semantic snapshot and the corresponding dimensional contract expression as joint input, perform minimal intervention correction calculation, and generate a directional semantic harmonization strategy that includes transitional explanatory phrase insertion instructions and time node fine-tuning parameters; The local reconstruction module is used to perform local reconstruction operations on the current archive content according to the directional semantic reconciliation strategy, dynamically adjust the logical connection relationship of the breach dimension, output the intermediate results of multi-dimensional content reconstruction, and complete the compilation and content reconstruction of the archive based on the intermediate results of multi-dimensional content reconstruction.

Images