Artificial intelligence-based method and system for analyzing content of a document

Abstract

This invention discloses an artificial intelligence-based method and system for document content analysis, belonging to the field of artificial intelligence analysis technology. The method includes separating structured and unstructured text information from the original corpus of target documents; dynamically constructing an analysis network framework containing multiple hierarchical processing nodes and a global semantic understanding engine; the global semantic understanding engine scanning the document analysis object to generate a node layout strategy; activating hierarchical processing nodes on the core logical path; iteratively capturing text features to form a preliminary semantic parsing tree; tracing back the semantic contribution weights from the terminal nodes to determine key semantic segments and importing them into the global semantic understanding engine for refined reasoning, context bridging, and logical verification; extracting the marked semantic relationship links to construct a semantic association graph; and outputting the analysis results after verifying cyclic convergence. This method can improve the accuracy and reliability of document content analysis and is suitable for in-depth analysis of various types of documents.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence analysis technology, specifically a method and system for document content analysis based on artificial intelligence. Background Technology

[0002] In the field of document content analysis, artificial intelligence technology has been widely applied to the extraction and parsing of text information. The conventional implementation process typically involves separating structured and unstructured text information from the original target document corpus, then capturing features of the separated text information through pre-defined fixed processing nodes, subsequently constructing semantic parsing structures, and finally outputting the document content analysis results. This AI-based document analysis approach is currently the main technical path for achieving in-depth and efficient processing of document information, and is suitable for batch parsing and information extraction scenarios of various types of documents.

[0003] Conventional AI-based document content analysis techniques have significant shortcomings. Their processing node layout is fixed and cannot be dynamically adjusted according to the actual textual characteristics and semantic complexity of the document being analyzed. This results in key textual information on the core logical path not being processed in a targeted manner, and the depth and accuracy of textual feature capture are limited. At the same time, the optimization of semantic parsing results lacks specificity. It fails to identify and focus on the key parts affecting the quality of semantic parsing, and only performs simple logical checks on the overall semantic parsing results. This easily leads to problems such as semantic logic disconnect and unclear semantic relationships, making it difficult to form an accurate and complete semantic relationship map. As a result, it cannot meet the needs of precise document content analysis. These problems need to be solved through technological improvements. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art; To this end, this invention proposes a document content analysis method based on artificial intelligence, including: Structured and unstructured text information are separated from the original corpus of the target document to form a document analysis object. An analysis network framework containing multiple levels of processing nodes and a global semantic understanding engine is dynamically constructed, and the global semantic understanding engine is triggered to perform a holistic scan of the document analysis object to generate a node layout strategy. Based on the node layout strategy, the hierarchical processing nodes located on the core logical path in the analysis network framework are activated, so that the activated hierarchical processing nodes perform iterative deep feature capture on the input text information to form a preliminary semantic parsing tree. Starting from the terminal node of the initial semantic parsing tree, the semantic contribution weights are traced backward to identify several key semantic segments that play a dominant role in the formation of the semantic parsing tree. The key semantic fragments are imported into the global semantic understanding engine, which then enters the refined reasoning mode to perform context bridging and logical verification on the key semantic fragments. After the global semantic understanding engine completes the refined reasoning, the semantic relationship links marked by the global semantic understanding engine in the semantic parsing tree are extracted to construct the semantic association graph of the target document; After the semantic association graph stabilizes, the analysis result verification loop is started. When the analysis result verification loop reaches the preset convergence condition, the content analysis result is output.

[0005] Furthermore, the global semantic understanding engine is triggered to perform a holistic scan of the document analysis object, generating a node layout strategy, including: Structured and unstructured text information are separated from the original corpus of the target document to form the document analysis object; Based on the text attribute features of the document analysis object, an analysis network framework containing multiple levels of processing nodes is dynamically constructed, and a global semantic understanding engine is configured for the analysis network framework. The structured text information and the unstructured text information are input into the analysis network framework, triggering the global semantic understanding engine to perform a holistic scan of the document analysis object and generate a semantic scene description; Based on the semantic scenario description, the topology of multiple processing nodes in the analysis network framework is reconstructed to generate a node layout strategy that is compatible with the document analysis object. The process of separating structured and unstructured text information from the original corpus of the target documents to form the document analysis object includes: A grammatical structure scan is performed on the original corpus of the target document to identify and extract sentence clusters with complete subject-verb-object structures, and the sentence clusters are marked as structured text information. Fuzzy semantic capture is performed on the original corpus of the target document. After filtering out the structured text information, the remaining text components constitute an unstructured text information set. The phrases, terms and incomplete sentences in the unstructured text information set are clustered to form unstructured text information. Add grammatical logic tags to the structured text information and add semantic ambiguity tags to the unstructured text information; Structured text information carrying grammatical logic tags is bound to unstructured text information carrying semantic ambiguity tags, together forming the document analysis object.

[0006] Furthermore, based on the text attribute features of the document analysis object, an analysis network framework containing multiple levels of processing nodes is dynamically constructed, and a global semantic understanding engine is configured for the analysis network framework, including: The syntactic complexity of structured text information and the semantic entropy of unstructured text information in the document analysis object are analyzed, and the network architecture complexity index is calculated based on the syntactic complexity and the semantic entropy. Based on the network architecture complexity index, a corresponding number of processing nodes with different functions are selected from a preset node type library. The functions of the processing nodes include syntactic parsing, entity recognition, relation extraction, sentiment judgment, and topic clustering. Based on the dependency relationships between different text information segments in the document analysis object, the selected multiple processing nodes are arranged into a hierarchical structure with upstream and downstream data flow directions to form an initial analysis network framework. A reasoning module with state memory is injected into the top layer of the initial analysis network framework. The reasoning module is defined as a global semantic understanding engine, and a bidirectional communication link is established between the global semantic understanding engine and each processing node in the analysis network framework.

[0007] Furthermore, the structured text information and the unstructured text information are input into the analysis network framework, triggering the global semantic understanding engine to perform a holistic scan of the document analysis object, generating a semantic scene description including: The entry node of the analysis network framework receives the structured text information and the unstructured text information, and routes different text information segments to the corresponding hierarchical processing nodes for processing according to the predetermined data distribution rules. When the global semantic understanding engine detects that all processing nodes at all levels have been started and have generated intermediate outputs through the bidirectional communication link, the global semantic understanding engine enters the overall scanning state. In the overall scanning state, the global semantic understanding engine collects the intermediate outputs of each level of processing nodes in a time-slice polling manner, and splices the intermediate outputs according to their logical positions in the analysis network framework to generate a global snapshot of the current processing state of the document analysis object. The global semantic understanding engine performs pattern recognition on the global snapshot, extracts recurring semantic patterns, potential logical conflict areas, and information density distribution features from the snapshot, and synthesizes the semantic scene description based on the recurring semantic patterns, potential logical conflict areas, and information density distribution features.

[0008] Furthermore, based on the node layout strategy, the hierarchical processing nodes located on the core logical path in the analysis network framework are activated, enabling the activated hierarchical processing nodes to perform iterative deep feature capture on the input text information, forming a preliminary semantic parsing tree, including: The analysis network framework closes or suspends hierarchical processing nodes located on non-core logical paths according to the node layout strategy. The hierarchical processing nodes on the activated core logic path enter the iterative working state, where the upstream node passes the processed feature vector to the downstream node, and the downstream node, while receiving the feature vector, sends its own generated feedback signal back to the upstream node. After multiple rounds of iterative transmission and feedback, the hierarchical processing nodes located in the final output layer will generate a set of semantic tags with hierarchical relationships. The semantic tags are automatically organized into a tree structure according to their dependencies in the generation process, forming the preliminary semantic parsing tree. The root node of the preliminary semantic parsing tree corresponds to the core topic of the document analysis object, the branch nodes correspond to the sub-topics at each level, and the leaf nodes correspond to specific semantic features or entities.

[0009] Furthermore, by tracing back the semantic contribution weights from the end nodes of the preliminary semantic parse tree, several key semantic segments that play a dominant role in the formation of the semantic parse tree are identified, including: A semantic contribution weight value is calculated for each node in the initial semantic parsing tree. The semantic contribution weight value is derived based on the number of times the node is activated in the analysis network framework, the frequency with which its output is referenced by downstream nodes, and the attention paid to it by the global semantic understanding engine. Starting from all leaf nodes of the initial semantic parsing tree, backtrack upwards along the parent node link, and accumulate the semantic contribution weight values ​​of the path nodes to form the semantic contribution path weight from each leaf node to the root node. Several paths with semantic contribution path weights greater than a preset threshold are selected, and the original text information segments corresponding to the nodes traversed by the paths with semantic contribution path weights greater than the preset threshold are located and extracted from the document analysis object. The extracted original text information segments are merged and deduplicated to obtain the aforementioned key semantic segments.

[0010] Furthermore, the key semantic fragments are imported into the global semantic understanding engine, driving the engine into a refined reasoning mode to perform context bridging and logical verification on the key semantic fragments, including: The global semantic understanding engine switches to refined inference mode. In this mode, the global semantic understanding engine pauses receiving real-time data from the processing nodes of the analysis network framework and focuses on processing the key semantic segments. The global semantic understanding engine places each key semantic segment in its original context and the overall semantic scene description for bidirectional verification, detecting whether the key semantic segment has semantic contradictions or logical breaks in the local and overall contexts. For detected semantic contradictions, the global semantic understanding engine calls a preset knowledge base or reasoning rules to generate correction suggestions or connection segments that enable semantic self-consistency. For detected logical breaks, the global semantic understanding engine attempts to insert inferential logical relationship links between the key semantic segments, or adjust the relative importance order of the key semantic segments to conform to the semantic scene description.

[0011] Furthermore, after the global semantic understanding engine completes the refined reasoning, the semantic relationship links marked by the global semantic understanding engine in the semantic parsing tree are extracted to construct a semantic association graph of the target document, including: The global semantic understanding engine uses the correction suggestions, connecting segments, logical relationship links, and importance ranking information generated in the refined reasoning mode as new semantic relationship link tags, and feeds them back to update the preliminary semantic parsing tree to form an enhanced semantic parsing tree. From the enhanced semantic parsing tree, extract the direct or indirect semantic relationship links between all nodes. The types of semantic relationship links include, but are not limited to, causal, parallel, progressive, inclusive, referential, and refutation. Using semantic concepts or entities as vertices and the semantic relationship links as edges, draw a network graph and attach attribute information inherited from the enhanced semantic parsing tree to the vertices and edges; The generated network graph is subjected to loop removal and pruning operations to eliminate redundant connections and ensure the clarity and structure of the semantic association graph.

[0012] Furthermore, after the semantic association graph stabilizes, a verification loop for the analysis results is initiated. When the verification loop reaches a preset convergence condition, the content analysis results are output, including: After the semantic association graph stabilizes, an analysis result verification loop is started. The analysis result verification loop is used to perform multiple rounds of alignment and calibration between the semantic association graph and the preset knowledge structure template. When the analysis result verification loop reaches the preset convergence condition, the final semantic association graph is output as the content analysis result of the target document; After the semantic association graph stabilizes, a verification loop for analysis results is initiated. This verification loop is used to perform multiple rounds of alignment and calibration between the semantic association graph and a preset knowledge structure template, including: Multiple knowledge structure templates for different fields are preset, and each knowledge structure template defines the standard concept hierarchy and relationship types in the corresponding field; The semantic association graph is sequentially matched with each knowledge structure template to calculate the matching degree between the topological structure of the semantic association graph and each knowledge structure template. The knowledge structure template with the highest matching degree is selected as the alignment benchmark, and the vertices and edges in the semantic association graph are mapped to the concepts and relationships corresponding to the knowledge structure template. For parts that cannot be successfully mapped, a calibration sub-process is initiated. The calibration sub-process, based on the constraints of the knowledge structure template, fine-tunes, merges, or splits the corresponding local structure in the semantic association graph. After completing one round of alignment and calibration, the matching degree is recalculated. If the improvement in matching degree does not reach the preset convergence condition, the alignment and calibration process with the knowledge structure template is repeated based on the calibrated semantic association graph until the matching degree change tends to stabilize and the preset convergence condition is reached.

[0013] Furthermore, the present invention also includes an artificial intelligence-based document content analysis system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the artificial intelligence-based document content analysis method described above.

[0014] Compared with the prior art, the beneficial effects of the present invention are: A dynamically constructed analysis network framework, comprising multiple hierarchical processing nodes and a global semantic understanding engine, is employed. The global semantic understanding engine performs a holistic scan of the document analysis object, composed of both structured and unstructured text information, and generates a node layout strategy. Based on this strategy, hierarchical processing nodes along the core logical paths of the analysis network framework are activated. These activated nodes then iteratively capture deep features from the input text information, forming a preliminary semantic parsing tree. This approach dynamically adjusts the node layout according to the actual characteristics of the document analysis object, avoiding the computational waste caused by fixed node layouts. It allows processing nodes to precisely focus on core text information, and iterative deep feature capture gradually uncovers the deeper semantics of the text, avoiding superficial or incomplete feature capture. This results in a semantic parsing tree that more closely matches the original semantics of the document, improving the accuracy of semantic analysis.

[0015] Starting from the terminal node of the initial semantic parse tree, the semantic contribution weights are traced backward to identify key semantic segments that play a dominant role in the formation of the semantic parse tree. These key semantic segments are then imported separately into the global semantic understanding engine, driving the engine into a refined reasoning mode to specifically perform context bridging and logical verification on the key semantic segments. This approach can accurately locate the core parts affecting the semantic parsing results, avoiding the problem of key information being obscured due to the equal treatment of all semantic segments in conventional techniques. Context bridging and logical verification can fill logical gaps between semantic segments, correct semantic deviations, and make the logic of key semantic segments more coherent and the semantics more accurate. This, in turn, makes the constructed semantic association graph more logical and complete, improving the reliability of the document content analysis results. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of the document content analysis method based on artificial intelligence described in this invention. Figure 2 A flowchart for generating a semantic scene description; Figure 3 A diagram illustrating the key semantic segment extraction process; Figure 4 To verify the cyclic convergence analysis graph of the semantic association graph; Figure 5 A radar chart for multi-dimensional adaptability analysis of semantic association graphs and knowledge templates from different domains. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1This AI-based document content analysis method starts with the original corpus of the target document, separating structured and unstructured text information to form a unified document analysis object. It dynamically constructs an analysis network framework containing multiple hierarchical processing nodes and a global semantic understanding engine. This triggers the global semantic understanding engine to perform a holistic scan of the document analysis object and generate a node layout strategy accordingly. Based on this strategy, the analysis network framework activates hierarchical processing nodes along the core logical path, iteratively capturing deep features of the text information to form a preliminary semantic parsing tree. By tracing the semantic contribution weights backward from the end nodes of this tree, several key semantic fragments that play a dominant role in the formation of the tree structure can be identified. These key semantic fragments are imported into the global semantic understanding engine, driving it into a refined inference mode to complete context bridging and logical verification. Subsequently, the labeled semantic relationship links are extracted from the refined inference-updated semantic parsing tree to construct a semantic association graph of the target document. After the graph stabilizes, an analysis result verification loop is initiated. When the loop reaches a preset convergence condition, the final content analysis result is output.

[0019] In one embodiment of the present invention, a grammatical structure scan is performed on the original corpus of the target document to identify and extract sentence clusters with complete subject-verb-object structures. For example, in the full text of an academic paper, the system can locate all declarative and concluding sentences that conform to the standard grammatical structure and mark these sentence clusters as structured text information. Fuzzy semantic capture is then performed on the same original corpus. After filtering out the parts that have been marked as structured text information, the remaining text components constitute a set of unstructured text information. This set may include core phrases from the paper abstract, a list of technical terms from the methodology section, figure and table titles, and citations from the literature review. The phrases, terms, and incomplete sentences in the unstructured text information set are clustered based on semantic similarity to form several groups of unstructured text information.

[0020] Structured text information is affixed with grammatical logic tags, which indicate the functional role of sentences within a discourse, such as "background statement," "method definition," "data presentation," or "conclusion summary." Unstructured text information is affixed with semantic ambiguity tags, which quantify the clarity of meaning of phrases or terms when taken out of context. The structured text information with grammatical logic tags and the unstructured text information with semantic ambiguity tags are then bound together to form a document analysis object, which is a composite data volume containing multiple attribute tags.

[0021] In some embodiments, an analysis network framework containing multiple levels of processing nodes is dynamically constructed based on the textual attribute characteristics of the document analysis object. The textual attribute characteristics of the document analysis object are jointly characterized by the syntactic complexity of structured text information and the semantic entropy value of unstructured text information. A network architecture complexity index is calculated based on the syntactic complexity and semantic entropy value, which determines the scale and node configuration of the analysis network framework. Based on the network architecture complexity index, a corresponding number of processing nodes with different functions are selected from a preset node type library. The functions of the processing nodes include syntactic parsing, entity recognition, relation extraction, sentiment judgment, and topic clustering. A global semantic understanding engine is configured for the analysis network framework, injected into the top layer of the analysis network framework, and establishes a bidirectional communication link with all its processing nodes.

[0022] It is understandable that inputting structured and unstructured text information into the analysis network framework triggers the global semantic understanding engine to perform a holistic scan of the document analysis object. The entry node of the analysis network framework receives both types of text information and routes text segments with different tags to the corresponding hierarchical processing nodes according to predefined data distribution rules. When the global semantic understanding engine detects through a bidirectional communication link that all hierarchical processing nodes have started and generated intermediate outputs, the global semantic understanding engine enters the overall scanning state. In the overall scanning state, the global semantic understanding engine collects the intermediate outputs of each hierarchical processing node in a time-slice polling manner, and splices the intermediate outputs according to their logical positions in the analysis network framework to generate a global snapshot of the current processing state of the document analysis object. The global semantic understanding engine performs pattern recognition on the global snapshot, extracting recurring semantic patterns, potential logical conflict areas, and information density distribution features. Based on the recurring semantic patterns, potential logical conflict areas, and information density distribution features, it synthesizes a semantic scene description, which is a high-level summary of the core content and logical skeleton of the document.

[0023] In practice, based on the semantic scene description, the topology of multiple processing nodes in the analysis network framework is reconstructed to generate a node layout strategy adapted to the literature analysis object. The topology reconstruction adjusts the connection weights and data flow directions between processing nodes based on the logical conflict areas and information density distribution identified in the semantic scene description. The node layout strategy clarifies the core logical paths that need to be activated in the subsequent in-depth analysis stage and the execution order of nodes on those paths.

[0024] In one embodiment of the present invention, see [reference] Figure 2This study analyzes the syntactic complexity of structured text information and the semantic entropy of unstructured text information in the literature analysis object. Syntactic complexity can be comprehensively quantified by indicators such as average sentence length, clause nesting depth, and conjunction usage frequency. Semantic entropy measures the degree of semantic confusion and uncertainty in the unstructured text information set. A network architecture complexity index is calculated based on syntactic complexity and semantic entropy, which determines the scale and type configuration of processing nodes required for the constructed analysis network framework. Based on the network architecture complexity index, a corresponding number of processing nodes with different functions are selected from a pre-defined node type library. The functions of these processing nodes include syntactic parsing, entity recognition, relation extraction, sentiment judgment, and topic clustering. The node type library predefines the algorithm implementation and interface specifications for each processing node. According to the dependencies between different text information segments in the literature analysis object, the selected processing nodes are arranged into a hierarchical structure with upstream and downstream data flows, forming the initial analysis network framework. The dependencies are determined by the relationship between the syntactic logic tags and semantic ambiguity tags carried by the text information segments.

[0025] In some embodiments, a reasoning module with state memory is injected at the top layer of the initial analysis network framework. This reasoning module is defined as a global semantic understanding engine, and a bidirectional communication link is established between the global semantic understanding engine and each processing node in the analysis network framework. See also... Figure 2 The analysis process involves the entry node of the network framework receiving structured and unstructured text information and routing different text segments to corresponding processing nodes according to predetermined data distribution rules. For example, structured text information marked as "method definition" is preferentially sent to the relation extraction node, while terms with high semantic ambiguity are clustered and sent to the topic clustering node. When the global semantic understanding engine detects through the bidirectional communication link that all processing nodes at all levels have started and generated intermediate outputs, the global semantic understanding engine enters the overall scanning state.

[0026] Understandably, during the overall scanning process, the global semantic understanding engine collects intermediate outputs from each processing node in a time-slice polling manner. These intermediate outputs include syntactic parse tree fragments, entity lists, preliminary relation triples, sentiment polarity values, and topic distribution vectors. The global semantic understanding engine then concatenates these intermediate outputs according to their logical positions within the analysis network framework, generating a global snapshot of the current processing state of the document analysis object. This global snapshot is a multi-dimensional data structure reflecting the intermediate processing state. The global semantic understanding engine performs pattern recognition on the global snapshot, extracting recurring semantic patterns, potential logical conflict areas, and information density distribution features. Based on these recurring semantic patterns, potential logical conflict areas, and information density distribution features, it synthesizes a semantic scene description.

[0027] In practical implementation, the network architecture complexity index The calculation can be quantitatively evaluated based on the following relationship:

[0028] in: This represents an index indicating network architecture complexity. This represents the calculated syntactic complexity of structured text information. The semantic entropy value represents the unstructured text information, with α and β being the configuration weight coefficients for the calculated syntactic complexity and the semantic entropy value, respectively. When generating semantic scene descriptions, the global semantic understanding engine integrates all extracted features. Potential logical conflict areas may manifest as attribute contradictions in the outputs of the same entity at different processing nodes, while information density distribution features reveal the spatial clustering of core arguments and supporting content in the global snapshot.

[0029] In one embodiment of the present invention, the analysis network framework shuts down or suspends hierarchical processing nodes located on non-core logical paths according to a node layout strategy. This node layout strategy is generated by a global semantic understanding engine based on a semantic scene description, defining the core and non-core logical paths for information processing. Hierarchical processing nodes on activated core logical paths enter an iterative working state, where upstream nodes pass processed feature vectors to downstream nodes. Downstream nodes, while receiving feature vectors, simultaneously send feedback signals back to upstream nodes. The feature vectors contain deep semantic encoding of the text, and the feedback signals are used to adjust the feature extraction granularity of the upstream nodes. After multiple rounds of iterative transmission and feedback, the hierarchical processing nodes at the final output layer generate a set of semantic tags with hierarchical relationships. These semantic tags represent semantic concepts of different granularities, ranging from specific features to abstract topics. The semantic tags are automatically organized into a tree structure based on their dependencies during generation, forming a preliminary semantic parsing tree. The root node of the preliminary semantic parsing tree corresponds to the core topic of the document analysis object, branch nodes correspond to sub-topics at various levels, and leaf nodes correspond to specific semantic features or entities.

[0030] In some embodiments, a semantic contribution weight value is calculated for each node in the initial semantic parsing tree. This weight value is derived based on the number of times the node is activated within the analysis network framework, the frequency with which its output is referenced by downstream nodes, and the global semantic understanding engine's attention to the node. The global semantic understanding engine's attention to a node can be measured by the frequency with which it accesses the node's intermediate output during the overall scanning and polling process. Starting from all leaf nodes of the initial semantic parsing tree, the process traces back along the parent node's link, accumulating the semantic contribution weight values ​​of the nodes along the path to form the semantic contribution path weight from each leaf node to the root node. Several paths with semantic contribution path weights greater than a preset threshold are selected. This preset threshold is dynamically set based on the overall information content and analysis depth requirements of the document analysis object. The original text information segments corresponding to the nodes traversed by the paths with semantic contribution path weights greater than the preset threshold are located and extracted from the document analysis object. The location process is completed based on the mapping relationship between nodes and text segments in the document analysis object.

[0031] It is understandable that the node semantic contribution weight value The calculation can be comprehensively evaluated based on the following relationship:

[0032] in: The semantic contribution weight of a node is represented by A, the number of times the node is activated in the analysis network framework, R, the frequency of the node's output being referenced by downstream nodes, and G, the quantification of the global semantic understanding engine's attention to the node. γ, δ, and ε are the configuration coefficients for the corresponding statistical values. The process of backtracking from all leaf nodes of the initial semantic parse tree is essentially evaluating the contribution strength of each path back from a specific semantic feature (leaf node) to the core topic (root node) to the construction of the entire semantic parse tree. The original text content corresponding to the nodes on paths with high contribution strength is considered the backbone material supporting the core logic.

[0033] In practice, the extracted original text segments undergo merging and deduplication. Merging connects different text fragments belonging to the same semantic cluster or describing the same core concept, while deduplication eliminates completely repetitive or highly overlapping text content, resulting in several key semantic fragments. These key semantic fragments are selected and integrated sets of continuous or discontinuous text that play a dominant role in forming the initial semantic parsing tree. These fragments are then submitted to the global semantic understanding engine for further refinement and reasoning. The calculation of semantic contribution path weights and the extraction of key semantic fragments enable the reverse location and extraction of human-readable original text fragments crucial for understanding the core content of the document from the machine-generated tree-like semantic structure.

[0034] See Figure 3 This is a diagram analyzing the key semantic fragment extraction process, visually illustrating the trends in the number of semantic fragments and extraction accuracy across the four stages from the original text to the final extraction. The number of semantic fragments shows a continuous decreasing trend, gradually decreasing from 120 to 28, reflecting the gradual convergence of screening, merging, and deduplication. The extraction accuracy shows a continuous increasing trend, rising from 0.65 to 0.95, indicating that the recognition accuracy of core semantic fragments continuously improves as redundant information is removed. The original text stage contains complete corpora but has high noise levels, resulting in the lowest accuracy and forming the basis for subsequent processing. The initial screening stage uses semantic contribution weighting to initially remove low-contribution fragments, reducing the number of fragments and significantly improving accuracy. The merging and deduplication stage merges semantic clusters and eliminates duplicate content from high-contribution fragments, further compressing redundancy and significantly improving accuracy. The final extraction stage yields highly condensed core semantic fragments with peak accuracy, which can be directly used for subsequent refined inference.

[0035] In one embodiment of the present invention, the global semantic understanding engine switches to a refined inference mode. In this mode, the global semantic understanding engine suspends receiving real-time data from the processing nodes of the analysis network framework and instead focuses on processing several key semantic segments extracted from the initial semantic parsing tree. The global semantic understanding engine performs bidirectional verification on each key semantic segment by placing it in its original context and the overall semantic scene description, detecting whether there are semantic contradictions or logical breaks in the key semantic segments in the local and overall contexts. The original context refers to the preceding and following paragraphs of the key semantic segment in the original corpus of the target document, and the overall semantic scene description is generated by the global semantic understanding engine in the early overall scanning stage. For detected semantic contradictions, the global semantic understanding engine calls a preset knowledge base or inference rules to generate correction suggestions or connecting segments that enable semantic self-consistency. The preset knowledge base includes domain common sense and logical rules. For detected logical breaks, the global semantic understanding engine attempts to insert inferential logical relationship links between key semantic segments or adjust the relative importance order of key semantic segments to make them conform to the semantic scene description.

[0036] In some embodiments, the global semantic understanding engine uses the correction suggestions, connecting segments, logical relationship links, and importance ranking information generated in the refined reasoning mode as new semantic relationship link markers, feeding them back to and updating the initial semantic parse tree to form an enhanced semantic parse tree. From the enhanced semantic parse tree, direct or indirect semantic relationship links between all nodes are extracted; the types of semantic relationship links are defined and marked during the refined reasoning process. A network graph is drawn with semantic concepts or entities as vertices and semantic relationship links as edges, and attribute information inherited from the enhanced semantic parse tree is attached to the vertices and edges. Vertex attributes include concept name, type, and source text location; edge attributes include relationship type, confidence level, and reasoning basis. The drawn network graph undergoes delooping and pruning operations to eliminate redundant connections. Delooping addresses circular references, and pruning removes edges with confidence levels below a threshold or those contributing little to the overall structure.

[0037] It is understandable that during the context bridging process, the global semantic understanding engine may perform a logical consistency evaluation on the set of key semantic fragments, and its evaluation function can be expressed as:

[0038] in: This represents the overall logical consistency score of a set of key semantic fragments within a given context. Indicates the number of key semantic segments, This represents the i-th key semantic segment. A function that represents the overall semantic scene description. Used to compute a single key semantic fragment With semantic scene description The semantic consistency score between them. The operation performed by the global semantic understanding engine in the refined reasoning mode is essentially to align and refine the preliminary semantic structure (preliminary semantic parsing tree) automatically generated by the machine with the logical consistency requirements that humans can understand. Logical verification ensures that the relationships derived from key semantic fragments do not conflict with the overall argument of the document. In specific implementation, the type of semantic relationship link is the basis for constructing the semantic association graph, see Table 1.

[0039] Table 1: Examples of Semantic Relationship Link Types

[0040] After completing the loop removal and pruning operations, the resulting network graph is the semantic association graph of the target document. The semantic association graph, with its graphical structure, reveals the complex network of relationships between core concepts and entities within the document, serving as a structured representation of the content analysis results. The process of drawing the network graph transforms the hierarchical parent-child relationships in the enhanced semantic parse tree into many-to-many connections in the network graph, thus enabling a richer expression of the cross-connections between concepts.

[0041] See Figure 4 This is a convergence analysis diagram of the semantic association graph verification loop, showing the changing trends of the number of nodes, edges, and knowledge template matching degree during the iteration process. The number of nodes gradually increased in the first 5 iterations, stabilizing at 63 after the 5th iteration, indicating that the core set of semantic concepts has been determined and no new nodes will be added subsequently. The number of edges continuously increased in the first 7 iterations, stabilizing at 160 after the 7th iteration, representing that the connection structure of semantic relationships has converged, with no new or removed edges. The matching degree continuously increased throughout, gradually rising from 60% to 95%, with the growth rate gradually slowing down, reflecting the convergence characteristics of the verification loop; the closer to the final structure, the more gradual the improvement in matching degree. During the graph construction period of iterations 1-5, the number of nodes and edges increased rapidly, and the matching degree improved simultaneously, gradually perfecting the core semantic concepts and relationships. During the structural stabilization period of iterations 5-7, the number of nodes reached stability before the number of edges, and the matching degree still showed a slight increase, indicating that the main focus was on optimizing the semantic connections between existing nodes. During the convergence period of iterations 7-10, the number of nodes and edges no longer changed, and the matching degree slowly approached its upper limit, indicating that the verification loop had reached the preset convergence condition.

[0042] In one embodiment of the present invention, after the semantic association graph stabilizes, an analysis result verification loop is initiated. This loop is used to perform multiple rounds of alignment and calibration between the semantic association graph and preset knowledge structure templates. Multiple preset knowledge structure templates from different domains are used. Each template defines a standard concept hierarchy and relation types within its corresponding domain. For example, a knowledge structure template for computer science might include standard concept hierarchies such as "algorithm," "data structure," and "complexity," as well as standard relation types such as "implementation," "superior to," and "contains." Templates for biology, however, define different concept hierarchies. The semantic association graph is sequentially matched with each knowledge structure template to calculate the matching degree between the topological structure of the semantic association graph and each template. The matching degree calculation considers the accuracy of vertex (concept) mapping and the coverage of edge (relation) mapping. The knowledge structure template with the highest matching degree is selected as the alignment benchmark. Vertices and edges in the semantic association graph are mapped to the corresponding concepts and relations in the knowledge structure template. This mapping process assigns standardized labels to the elements in the semantic association graph.

[0043] In some embodiments, for parts that cannot be successfully mapped, a calibration sub-process is initiated. This sub-process fine-tunes, merges, or splits the corresponding local structure in the semantic association graph according to the constraints of the knowledge structure template. Fine-tuning may adjust vertex attributes or edge types to conform to the template definition; merging groups multiple semantically similar vertices under the same standard concept in the template; and splitting breaks down an overly broad vertex into multiple more refined concepts within the template. After one round of alignment and calibration, the matching degree is recalculated. If the improvement in matching degree does not reach the preset convergence condition, the alignment and calibration process with the knowledge structure template is repeated based on the calibrated semantic association graph. The preset convergence condition is typically set as the improvement in matching degree being less than a threshold in multiple consecutive iterations, or reaching the maximum number of iterations, until the matching degree change stabilizes and the preset convergence condition is met.

[0044] Understandable, matching degree The calculation can be quantified using the following formula:

[0045] in: This indicates the degree of matching between the semantic association graph and a certain knowledge structure template. This represents the total number of vertices in the semantic association graph. This represents the number of vertices that can be successfully mapped to the standard concepts of this knowledge structure template. This represents the total number of edges in the semantic association graph. This represents the number of edges that can be successfully mapped to the standard relation type of this knowledge structure template. and These are the weighting coefficients for vertex matching contribution and edge matching contribution, respectively. When the calibration sub-process is started, the system adjusts the structure of the semantic association graph according to the standardization requirements of the knowledge structure template. For example, if the semantic association graph contains vertex A "deep learning model" and vertex B "neural network" and the two are closely related, but the knowledge structure template only contains the standard concept "artificial neural network model", the calibration sub-process may merge vertex A and vertex B and uniformly map them to the standard concept "artificial neural network model".

[0046] In practice, when the analysis result verification loop reaches the preset convergence condition, the final semantic association graph is output as the content analysis result of the target document. The final semantic association graph is a relationship network with standardized labels and a clear structure, after being aligned and calibrated with the domain's standard knowledge structure. The analysis result verification loop process ensures that the semantic association graph automatically extracted from a specific document is compatible with a broader and more standardized domain knowledge system, thereby improving the universality and interpretability of the analysis results. The iterative process of multiple rounds of alignment and calibration continuously optimizes the structure of the semantic association graph until it reaches a stable high-match state with the most relevant domain knowledge structure template. At this point, the output graph is considered a reliable result of the structured understanding of the document content.

[0047] See Figure 5 This is a radar chart analyzing the multi-dimensional adaptability of a semantic graph to knowledge templates from different domains, used to evaluate the graph's matching performance under templates from the computer science, biology, and linguistics domains. The computer science template showed the best overall adaptability, with the highest scores in structural similarity (0.90), edge mapping rate (0.86), and vertex mapping rate (0.85), indicating a high degree of consistency between the semantic graph's topological structure and the knowledge system of the computer science domain. The linguistics template showed the second best adaptability, matching the computer science template in concept coverage (0.88), but slightly lower in other dimensions, resulting in a balanced overall performance. The biology template showed the weakest adaptability, with scores significantly lower than the other two domains across all dimensions, indicating a low degree of matching between the semantic graph's concept system and the knowledge structure of the biological domain. Concept coverage was a relatively high score across all three domains, indicating that the graph extracted core concepts relatively completely, with the differences in adaptability mainly reflected in the structural and relational levels.

[0048] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A document content analysis method based on artificial intelligence, characterized in that, include: Structured and unstructured text information are separated from the original corpus of the target document to form a document analysis object. An analysis network framework containing multiple levels of processing nodes and a global semantic understanding engine is dynamically constructed, and the global semantic understanding engine is triggered to perform a holistic scan of the document analysis object to generate a node layout strategy. Based on the node layout strategy, the hierarchical processing nodes located on the core logical path in the analysis network framework are activated, so that the activated hierarchical processing nodes perform iterative deep feature capture on the input text information to form a preliminary semantic parsing tree. Starting from the terminal node of the initial semantic parsing tree, the semantic contribution weights are traced backward to identify several key semantic segments that play a dominant role in the formation of the semantic parsing tree. The key semantic fragments are imported into the global semantic understanding engine, which then enters the refined reasoning mode to perform context bridging and logical verification on the key semantic fragments. After the global semantic understanding engine completes the refined reasoning, the semantic relationship links marked by the global semantic understanding engine in the semantic parsing tree are extracted to construct the semantic association graph of the target document; After the semantic association graph stabilizes, the analysis result verification loop is started. When the analysis result verification loop reaches the preset convergence condition, the content analysis result is output.

2. The document content analysis method based on artificial intelligence as described in claim 1, characterized in that, The global semantic understanding engine is triggered to perform a holistic scan of the document analysis object, generating a node layout strategy, including: Structured and unstructured text information are separated from the original corpus of the target document to form the document analysis object; Based on the text attribute features of the document analysis object, an analysis network framework containing multiple levels of processing nodes is dynamically constructed, and a global semantic understanding engine is configured for the analysis network framework. The structured text information and the unstructured text information are input into the analysis network framework, triggering the global semantic understanding engine to perform a holistic scan of the document analysis object and generate a semantic scene description; Based on the semantic scenario description, the topology of multiple processing nodes in the analysis network framework is reconstructed to generate a node layout strategy that is compatible with the document analysis object. The process of separating structured and unstructured text information from the original corpus of the target documents to form the document analysis object includes: A grammatical structure scan is performed on the original corpus of the target document to identify and extract sentence clusters with complete subject-verb-object structures, and the sentence clusters are marked as structured text information. Fuzzy semantic capture is performed on the original corpus of the target document. After filtering out the structured text information, the remaining text components constitute an unstructured text information set. The phrases, terms and incomplete sentences in the unstructured text information set are clustered to form unstructured text information. Add grammatical logic tags to the structured text information and add semantic ambiguity tags to the unstructured text information; Structured text information carrying grammatical logic tags is bound to unstructured text information carrying semantic ambiguity tags, together forming the document analysis object.

3. The document content analysis method based on artificial intelligence as described in claim 2, characterized in that, Based on the text attribute features of the document analysis object, an analysis network framework containing multiple levels of processing nodes is dynamically constructed, and a global semantic understanding engine is configured for the analysis network framework, including: The syntactic complexity of structured text information and the semantic entropy of unstructured text information in the document analysis object are analyzed, and the network architecture complexity index is calculated based on the syntactic complexity and the semantic entropy. Based on the network architecture complexity index, a corresponding number of processing nodes with different functions are selected from a preset node type library. The functions of the processing nodes include syntactic parsing, entity recognition, relation extraction, sentiment judgment, and topic clustering. Based on the dependency relationships between different text information segments in the document analysis object, the selected multiple processing nodes are arranged into a hierarchical structure with upstream and downstream data flow directions to form an initial analysis network framework. A reasoning module with state memory is injected into the top layer of the initial analysis network framework. The reasoning module is defined as a global semantic understanding engine, and a bidirectional communication link is established between the global semantic understanding engine and each processing node in the analysis network framework.

4. The document content analysis method based on artificial intelligence as described in claim 3, characterized in that, The structured and unstructured text information are input into the analysis network framework, triggering the global semantic understanding engine to perform a holistic scan of the document analysis object and generate a semantic scene description including: The entry node of the analysis network framework receives the structured text information and the unstructured text information, and routes different text information segments to the corresponding hierarchical processing nodes for processing according to the predetermined data distribution rules. When the global semantic understanding engine detects that all processing nodes at all levels have been started and have generated intermediate outputs through the bidirectional communication link, the global semantic understanding engine enters the overall scanning state. In the overall scanning state, the global semantic understanding engine collects the intermediate outputs of each level of processing nodes in a time slice polling manner, and splices the intermediate outputs according to their logical positions in the analysis network framework to generate a global snapshot of the current processing state of the document analysis object. The global semantic understanding engine performs pattern recognition on the global snapshot, extracts recurring semantic patterns, potential logical conflict areas, and information density distribution features from the snapshot, and synthesizes the semantic scene description based on the recurring semantic patterns, potential logical conflict areas, and information density distribution features.

5. The document content analysis method based on artificial intelligence as described in claim 4, characterized in that, Based on the node layout strategy, the hierarchical processing nodes located on the core logical path in the analysis network framework are activated, enabling the activated hierarchical processing nodes to iteratively capture deep features of the input text information, forming a preliminary semantic parsing tree, including: The analysis network framework closes or suspends hierarchical processing nodes located on non-core logical paths according to the node layout strategy. The hierarchical processing nodes on the activated core logic path enter the iterative working state, where the upstream node passes the processed feature vector to the downstream node, and the downstream node, while receiving the feature vector, sends its own generated feedback signal back to the upstream node. After multiple rounds of iterative transmission and feedback, the hierarchical processing nodes located in the final output layer will generate a set of semantic tags with hierarchical relationships. The semantic tags are automatically organized into a tree structure according to their dependencies in the generation process, forming the preliminary semantic parsing tree. The root node of the preliminary semantic parsing tree corresponds to the core topic of the document analysis object, the branch nodes correspond to the sub-topics at each level, and the leaf nodes correspond to specific semantic features or entities.

6. The document content analysis method based on artificial intelligence as described in claim 5, characterized in that, Starting from the terminal node of the initial semantic parse tree, the semantic contribution weights are traced backward to identify several key semantic segments that play a dominant role in the formation of the semantic parse tree, including: A semantic contribution weight value is calculated for each node in the initial semantic parsing tree. The semantic contribution weight value is derived based on the number of times the node is activated in the analysis network framework, the frequency with which its output is referenced by downstream nodes, and the attention paid to it by the global semantic understanding engine. Starting from all leaf nodes of the initial semantic parsing tree, backtrack upwards along the parent node link, and accumulate the semantic contribution weight values ​​of the path nodes to form the semantic contribution path weight from each leaf node to the root node. Several paths with semantic contribution path weights greater than a preset threshold are selected, and the original text information segments corresponding to the nodes traversed by the paths with semantic contribution path weights greater than the preset threshold are located and extracted from the document analysis object. The extracted original text information segments are merged and deduplicated to obtain the aforementioned key semantic segments.

7. The document content analysis method based on artificial intelligence as described in claim 6, characterized in that, The key semantic fragments are imported into the global semantic understanding engine, which then enters a refined reasoning mode to perform context bridging and logical verification on the key semantic fragments, including: The global semantic understanding engine switches to refined inference mode. In this mode, the global semantic understanding engine suspends receiving real-time data from the processing nodes of the analysis network framework and focuses on processing the key semantic segments. The global semantic understanding engine places each key semantic segment in its original context and the overall semantic scene description for bidirectional verification, detecting whether the key semantic segment has semantic contradictions or logical breaks in the local and overall contexts. For detected semantic contradictions, the global semantic understanding engine calls a preset knowledge base or reasoning rules to generate correction suggestions or connection segments that enable semantic self-consistency. For detected logical breaks, the global semantic understanding engine attempts to insert inferential logical relationship links between the key semantic segments, or adjust the relative importance order of the key semantic segments to conform to the semantic scene description.

8. The document content analysis method based on artificial intelligence as described in claim 7, characterized in that, After the global semantic understanding engine completes the refined reasoning, the semantic relationship links marked by the global semantic understanding engine in the semantic parsing tree are extracted to construct the semantic association graph of the target document, including: The global semantic understanding engine uses the correction suggestions, connecting segments, logical relationship links, and importance ranking information generated in the refined reasoning mode as new semantic relationship link tags, and feeds them back to update the preliminary semantic parsing tree to form an enhanced semantic parsing tree. From the enhanced semantic parsing tree, extract the direct or indirect semantic relationship links between all nodes. The types of semantic relationship links include, but are not limited to, causal, parallel, progressive, inclusive, referential, and refutation. Using semantic concepts or entities as vertices and the semantic relationship links as edges, draw a network graph and attach attribute information inherited from the enhanced semantic parsing tree to the vertices and edges; The generated network graph is subjected to loop removal and pruning operations to eliminate redundant connections and ensure the clarity and structure of the semantic association graph.

9. The document content analysis method based on artificial intelligence as described in claim 8, characterized in that, After the semantic association graph stabilizes, a verification loop for the analysis results is initiated. When the verification loop reaches a preset convergence condition, the content analysis results are output, including: After the semantic association graph stabilizes, an analysis result verification loop is started. The analysis result verification loop is used to perform multiple rounds of alignment and calibration between the semantic association graph and the preset knowledge structure template. When the analysis result verification loop reaches the preset convergence condition, the final semantic association graph is output as the content analysis result of the target document; After the semantic association graph stabilizes, a verification loop for analysis results is initiated. This verification loop is used to perform multiple rounds of alignment and calibration between the semantic association graph and a preset knowledge structure template, including: Multiple knowledge structure templates for different fields are preset, and each knowledge structure template defines the standard concept hierarchy and relationship types in the corresponding field; The semantic association graph is sequentially matched with each knowledge structure template to calculate the matching degree between the topological structure of the semantic association graph and each knowledge structure template. The knowledge structure template with the highest matching degree is selected as the alignment benchmark, and the vertices and edges in the semantic association graph are mapped to the concepts and relationships corresponding to the knowledge structure template. For parts that cannot be successfully mapped, a calibration sub-process is initiated. The calibration sub-process, based on the constraints of the knowledge structure template, fine-tunes, merges, or splits the corresponding local structure in the semantic association graph. After completing one round of alignment and calibration, the matching degree is recalculated. If the improvement in matching degree does not reach the preset convergence condition, the alignment and calibration process with the knowledge structure template is repeated based on the calibrated semantic association graph until the matching degree change tends to stabilize and the preset convergence condition is reached.

10. An artificial intelligence-based document content analysis system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the document content analysis method based on artificial intelligence as described in any one of claims 1 to 9.

Images