A chapter-aware information extraction method and system for structured long text

Abstract

The application discloses a chapter perception information extraction method and system for structured long text in the technical field of natural language processing, knowledge extraction and knowledge organization, and contains the following steps: S1, obtaining original structured long text; S2, performing hierarchical identification on a text block sequence and constructing path information; S3, combining chapter titles, path information and local context, performing chapter attribution determination on each text block and mapping into a standard chapter type; S4, according to the standard chapter type, dynamically generating chapter constraint configurations for each text block; S5, under the chapter constraint configurations, performing information extraction on each text block; S6, performing intra-chapter aggregation, global aggregation and standardization processing on local extraction results; and S7, constructing standard entity projection relationships based on the structured extraction results. The system comprises functional modules corresponding to the above steps. The application can reduce the context burden of long text extraction, reduce cross-chapter semantic confusion and mis-extraction, and improve extraction stability.

Description

Technical Field

[0001] This invention relates to a chapter-aware information extraction method and system for structured long texts in the fields of natural language processing, knowledge extraction, and knowledge organization. Background Technology

[0002] Information extraction technology aims to automatically identify entities, relationships, attributes, or events from natural language text and convert them into structured representations to support applications such as knowledge graph construction, knowledge base updates, retrieval enhancement, intelligent question answering, and decision support. Most existing information extraction methods are designed for sentence-level text, short paragraph text, or weakly structured text. They typically input the entire text into a unified extractor to perform entity recognition, relationship extraction, or triple extraction.

[0003] However, in textbooks, technical manuals, industry standards, professional explanatory materials, and other structured long text scenarios, existing technologies still have the following prominent problems:

[0004] (i) Structured long texts typically have a multi-level hierarchical structure, such as book, chapter, section, subsection, and main text, with a large text span and lengthy context. If an entire chapter or paragraph is directly input into the extraction model, the model needs to make unified inferences within an excessively long context, which can easily increase the inference burden and cause local facts to be interfered with by distant and irrelevant content, thereby affecting the stability and accuracy of the extraction results.

[0005] (ii) Different chapters typically carry different types of knowledge representations. Different chapters in the same data often correspond to different knowledge topics such as etiology, epidemiology, clinical symptoms, pathological changes, diagnosis or prevention. Existing methods usually lack explicit modeling mechanisms for chapter semantics, which can easily lead to relationship drift and entity boundary drift between chapters, resulting in cross-chapter mis-extraction, missed extraction, and unbalanced extraction granularity.

[0006] (iii) Existing methods typically perform extraction on a globally unified set of entity types and relation types, lacking constraint configuration and result normalization mechanisms for specific chapter content. This results in a large number of candidate types unrelated to the current chapter entering the reasoning process. At the same time, it is difficult to uniformly map abbreviations, aliases, and variant expressions in the extraction results, thereby affecting the controllability, standardization, and subsequent database utilization of the structured results. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for extracting chapter-aware information from structured long texts. Through a linkage mechanism of text block division, hierarchical recognition, chapter affiliation determination, standardized chapter type mapping, chapter constraint configuration, chapter-aware extraction, result aggregation, and standardized entity projection, the original long text is converted into a controllable, stable, and standardized structured knowledge result.

[0008] To achieve the above objectives, this invention provides a method for extracting chapter-aware information from structured long texts, comprising the following steps: S1. Obtaining the original structured long text and dividing it into text blocks to obtain a text block sequence; S2. Performing hierarchical recognition on the text block sequence and constructing path information to obtain the hierarchical labels and path information corresponding to each text block; S3. Combining the chapter title, path information, and local context, determining the chapter affiliation of each text block and mapping it to a standardized chapter type; S4. Dynamically generating chapter constraint configurations for each text block according to the standardized chapter type, wherein the chapter constraint configuration includes at least an entity type space, a relation type space, and a set of rule restrictions; S5. Under the chapter constraint configuration, performing information extraction on each text block and performing chapter consistency screening on the original candidate extraction results to obtain local extraction results; S6. Performing intra-chapter aggregation, global aggregation, and normalization processing on the local extraction results to obtain structured extraction results; S7. Constructing standardized entity projection relationships based on the structured extraction results and using the standardized entity projection relationships to normalize the entities in the extraction results to obtain standardized structured results.

[0009] As a further improvement to the present invention, S1 is specifically described below.

[0010] S1.1 Perform pre-cleaning on the original structured long text. Pre-cleaning includes removing meaningless blank lines, normalizing whitespace characters, identifying heading numbering patterns, removing redundant parentheses in headings, and preserving one or more of paragraph boundaries and sequential indexes.

[0011] S1.2. Based on one or more of the following: title style, numbering pattern, paragraph boundary, line break rule, punctuation pattern, and length threshold, the original structured long text is segmented into blocks.

[0012] S1.3 Divide the text fragments that meet the title determination conditions into title-type text blocks, divide the text fragments that carry the semantic content of the body text into body-type text blocks, and retain at least one of the sequential index, position index or source document offset of each text block.

[0013] S1.4. Temporarily store the body text blocks that are subordinate to the adjacent title blocks as attachable objects to form a sequence of text blocks for subsequent hierarchical identification, path construction and chapter affiliation determination.

[0014] As a further improvement to the present invention, S2 is specifically described below.

[0015] S2.1, Let the sequence of text blocks obtained after partitioning by S1 be... ,in Indicates the first I. Text blocks; for each text block, extract its structural feature representation:

[0016] ;

[0017] in, Indicates the characteristics of the title pattern. Indicates the structural characteristics of the numbering system. This indicates the continuity of local context. Represents local semantic features;

[0018] S2.2, Let the set of candidate level labels be: ;

[0019] in, Includes one or more hierarchical tags such as book level, document level, chapter level, section level, subsection level, and body text level; for text blocks and candidate level tags Define the hierarchical recognition scoring function as follows:

[0020] ;

[0021] In the formula, Represents a text block With candidate level labels Title pattern matching degree; This indicates consistency in the numbering structure; Indicates local context continuity; This indicates the strength of local semantic support for hierarchical tags; , , and These are the corresponding weighting coefficients;

[0022] Based on this scoring function, text blocks can be determined using the following formula. Hierarchical tags:

[0023] ;

[0024] That is, the candidate level label that maximizes the level recognition scoring function is selected as the text block. Hierarchical tags;

[0025] The confidence level for the rule-driven decision is defined as follows:

[0026] ;

[0027] when When, the result is determined directly by the rule; when At that time, the large language model is invoked for auxiliary determination, that is:

[0028] ;

[0029] In the formula, This represents the confidence threshold for hierarchical determination. This represents the auxiliary level decision function of the large language model. Indicates text block The associated local context; through this collaborative determination method, the efficiency of the rule system in structural pattern recognition and the flexible supplementary ability of the large language model in semantic ambiguity scenarios can be taken into account.

[0030] S2.3 After determining the hierarchical labels, attach the body text block to the nearest valid title node to construct the chapter tree structure and build path information for the body text block:

[0031] ;

[0032] in, Represents a text block Path information, Indicates the first Layer title node, Indicates the path level; path information not only indicates the structural position of the text block in the document tree, but also provides structural priors for subsequent chapter affiliation determination.

[0033] As a further improvement to the present invention, the specific content of S3 is as follows.

[0034] S3.1, Let the set of candidate chapter types be... The standard chapter type set is as follows For text blocks and candidate chapter types The chapter attribution scoring function is defined as follows:

[0035] ;

[0036] In the formula, This indicates the title semantic matching score. Indicates the path structure consistency score. This indicates that the local context supports the score; , and These correspond to the weighting coefficients. This yields the text block. Initial chapter assignment results:

[0037] ;

[0038] S3.2 Since the same chapter in a real document may have different naming conventions, it is necessary to further map the initial chapter assignment results to the standardized chapter space and define a chapter semantic representation function: ;

[0039] in, This represents a chapter semantic representation function constructed based on the initial chapter affiliation result, path information, and local context;

[0040] S3.3, The standard chapter type mapping satisfies: ;

[0041] In the formula, Represents a text block Standard chapter types, This represents the semantic similarity function of chapters. Through this mapping, chapter descriptions with different writing styles but the same or similar semantics can be uniformly mapped to the same standardized chapter type, thereby ensuring the consistency of the chapter semantic space and providing stable input for the next step of chapter constraint configuration generation.

[0042] As a further improvement of the present invention, the specific content of S4 is as follows:

[0043] S4.1, For text blocks Corresponding standard chapter type Configure its chapter constraints: ;

[0044] in, This indicates the allowed entity type space for this chapter. This indicates the space of allowed relation types in this chapter. This represents the set of rules and restrictions corresponding to this chapter;

[0045] S4.2, The local candidate space of the current text block is defined as follows: ;

[0046] This indicates that for text blocks Subsequent information extraction is not performed unconstrained within a globally unified entity type and relation type space, but rather within a local candidate space of chapters induced by the standardized chapter types. This effectively reduces irrelevant candidates from entering the extraction process. This represents the space of entity type and relation type combinations induced by chapter constraints, and is not a complete triple space;

[0047] Set of rules and restrictions Including but not limited to: candidate entity expression length limit, tail entity phrase form limit, parallel item splitting rule, title fragment filtering rule, granularity control rule and noise filtering rule;

[0048] Chapter constraint configuration is generated by loading synonym chapter merging rules, default extraction rules, chapter customization rules, and chapter constraint overriding rules, so that chapter structure information and extraction control information form a one-to-one correspondence.

[0049] As a further improvement to the present invention, the specific content of S5 is as follows.

[0050] Set text block The original candidate extraction result set is To ensure that the extraction results are consistent with the semantics of the current chapter, candidate results are defined. In standardizing chapter types The validity determination function is as follows:

[0051] ;

[0052] Then text block The set of local extraction results satisfies:

[0053] ;

[0054] In the formula, Represents a text block The set of local extraction results;

[0055] Through the chapter consistency screening process, only candidate results that meet the current specification chapter type constraints are retained, thereby suppressing cross-chapter mis-sampling and the mixing of irrelevant relationships;

[0056] Chapter consistency screening includes at least one or more of the following criteria: entity type validity, relation type validity, head and tail entity combination validity, tail entity phrase form validity, and chapter semantic consistency.

[0057] As a further improvement of the present invention, when post-processing the local extraction results in S6, one or more of the following operations are included: chapter-level aggregation, global aggregation, bracket alias extraction, alias merging, pronoun resolution, relational direction correction, deduplication, parallel tail entity splitting, title fragment filtering, and noise filtering; wherein, chapter-level aggregation is used to summarize the knowledge expressed in different text blocks under the same standardized chapter type, and standardization processing is used to eliminate the formal differences caused by different text blocks, different expression styles, and different alias forms;

[0058] Let the set of local extraction results for all text blocks be . n represents the number of text blocks participating in the aggregation, and n is a positive integer greater than or equal to 1; First, the local extraction results of multiple text blocks with the same chapter affiliation or the same standard chapter type are merged to obtain the chapter-level aggregation result; Through intra-chapter aggregation, knowledge expressed by different text blocks under the same chapter can be summarized;

[0059] Based on chapter-level aggregation, a global aggregation is performed on the results of the entire document to form a candidate structured result set. Then to Performs lightweight normalization of bracket, punctuation, connector, and alias expressions, and removes duplicate results.

[0060] As a further improvement of the present invention, the normalized entity projection relationship in S7 is obtained by constructing an entity mapping table from the structured extraction results. The entity mapping table records the normalized entity name, entity type, and alias set, and performs normalized mapping on the original entity representation in the extraction results accordingly; the normalized mapping satisfies:

[0061] Let the original entity be represented as The standard entity projection function is The solid projection relationship can then be expressed as: ;

[0062] In the formula, Represents the original entity expression The corresponding canonical entity;

[0063] Construct the entity mapping table: ;

[0064] In the formula, Representing a canonical entity entity type, Representation and specification entities The corresponding set of aliases; the entity mapping table can be automatically constructed from the extraction results, or it can be supplemented and improved by combining manual thesaurus or external knowledge;

[0065] After performing normalized entity projection on the entities in the result, normalized triples are obtained:

[0066] ;

[0067] In the formula, Indicates the head entity. Indicates a relationship. This represents the tail entity; the output can be either a chapter-level structured result or a full-text-level structured result, and can be directly used in scenarios such as knowledge graph database entry, entity fusion, and entity disambiguation.

[0068] To achieve the above objectives, the present invention also provides a chapter-aware information extraction system for structured long texts, including a text block segmentation module, a hierarchy recognition and path construction module, a chapter affiliation determination and standardized chapter type mapping module, a chapter constraint configuration module, an information extraction and chapter consistency screening module, a result aggregation and standardization processing module, and a standardized entity projection module.

[0069] The text block segmentation module is used to segment the original structured long text into text blocks to obtain a text block sequence.

[0070] The hierarchy recognition and path construction module is used to perform hierarchy recognition on the text block sequence and construct path information to obtain the hierarchy label and path information corresponding to each text block.

[0071] The chapter attribution determination and standardized chapter type mapping module is used to determine the chapter attribution of each text block by combining the chapter title, path information and local context, and map it to a standardized chapter type;

[0072] The chapter constraint configuration module is used to generate chapter constraint configurations for each text block according to the standard chapter type. The chapter constraint configuration includes at least an entity type space, a relation type space, and a set of rule constraints.

[0073] The information extraction and chapter consistency filtering module is used to perform information extraction under chapter constraint configuration and perform chapter consistency filtering on the original candidate extraction results to obtain local extraction results.

[0074] The result aggregation and normalization module is used to perform intra-chapter aggregation, global aggregation, and normalization on local extraction results to obtain structured extraction results.

[0075] The normalized entity projection module is used to construct normalized entity projection relationships based on the structured extraction results, and to normalize the entities in the extraction results to obtain standardized structured results.

[0076] As a further improvement of the present invention, the chapter constraint configuration generated by the chapter constraint configuration module includes an entity type space, a relation type space, and a set of rule restrictions corresponding to the standard chapter type; the information extraction and chapter consistency screening module performs information extraction within the local candidate space defined by the chapter constraint configuration and performs chapter consistency screening on the candidate extraction results.

[0077] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0078] 1. By dividing text into blocks, identifying levels, and constructing paths, the problem of extracting the whole structured long text is transformed into an extraction problem on multiple local processing units, thereby reducing the context burden caused by direct processing of long text and improving the efficiency and stability of information extraction in long text scenarios.

[0079] 2. By standardizing chapter type mapping and chapter constraint configuration, information extraction is performed in different chapters under the corresponding entity type space, relation type space and rule constraints, thereby reducing cross-chapter erroneous extraction, omission extraction and relation drift, and improving the consistency and controllability of extraction results with chapter semantics.

[0080] 3. By performing intra-chapter aggregation, global aggregation, normalization processing, and normalized entity projection on the extracted results, the abbreviations, aliases, and variant spellings are uniformly merged, thereby improving the standardization and stability of the extracted results and their effectiveness in subsequent knowledge graph database entry, entity fusion, and entity disambiguation.

[0081] 4. Through a collaborative mechanism involving text block segmentation, hierarchy recognition, chapter affiliation determination, standardized chapter type mapping, chapter constraint configuration, chapter-aware extraction, result aggregation, and standardized entity projection, the original long text is transformed into a controllable, stable, and standardized structured knowledge result. Compared with existing technologies, this invention effectively reduces the burden of long text processing, reduces cross-chapter erroneous extraction, and significantly improves the stability and usability of the extraction results. Attached Figure Description

[0082] Figure 1 This is a flowchart of the present invention.

[0083] Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0084] The present invention will be further described below with reference to the accompanying drawings:

[0085] like Figure 1 The method for extracting chapter-aware information from structured long texts, as shown, includes the following steps:

[0086] S1. Obtain the original structured long text and divide the original structured long text into text blocks to obtain a text block sequence;

[0087] S1.1 Perform pre-cleaning on the original structured long text. Pre-cleaning includes removing meaningless blank lines, normalizing whitespace characters, identifying heading numbering patterns, removing redundant parentheses in headings, and preserving one or more of paragraph boundaries and sequential indexes.

[0088] S1.2. Based on one or more of the following: title style, numbering pattern, paragraph boundary, line break rule, punctuation pattern, and length threshold, the original structured long text is segmented into blocks.

[0089] S1.3 Divide the text fragments that meet the title determination conditions into title-type text blocks, divide the text fragments that carry the semantic content of the body text into body-type text blocks, and retain at least one of the sequential index, position index or source document offset of each text block.

[0090] S1.4. Temporarily store the body text blocks that are subordinate to the adjacent title blocks as attachable objects to form a sequence of text blocks for subsequent hierarchical identification, path construction and chapter affiliation determination.

[0091] S2. Perform hierarchical recognition on the text block sequence and construct path information to obtain the hierarchical labels and path information corresponding to each text block;

[0092] S2.1, Let the sequence of text blocks obtained after partitioning by S1 be... ,in Indicates the first I. Text blocks; for each text block, extract its structural feature representation:

[0093] ;

[0094] in, Indicates the characteristics of the title pattern. Indicates the structural characteristics of the numbering system. This indicates the continuity of local context. Represents local semantic features;

[0095] S2.2, Let the set of candidate level labels be: ;

[0096] in, Includes one or more hierarchical tags such as book level, document level, chapter level, section level, subsection level, and body text level; for text blocks and candidate level tags Define the hierarchical recognition scoring function as follows:

[0097] ;

[0098] In the formula, Represents a text block With candidate level labels Title pattern matching degree; This indicates consistency in the numbering structure; Indicates local context continuity; This indicates the strength of local semantic support for hierarchical tags; , , and These are the corresponding weighting coefficients;

[0099] Based on this scoring function, text blocks can be determined using the following formula. Hierarchical tags:

[0100] ;

[0101] That is, the candidate level label that maximizes the level recognition scoring function is selected as the text block. The hierarchical labels; the confidence level for the rule-driven decision is defined as:

[0102] ;

[0103] when When, the result is determined directly by the rule; when At that time, the large language model is invoked for auxiliary determination, that is:

[0104] ;

[0105] In the formula, This represents the confidence threshold for hierarchical determination. This represents the auxiliary level decision function of the large language model. Indicates text block The associated local context; through this collaborative determination method, the efficiency of the rule system in structural pattern recognition and the flexible supplementary ability of the large language model in semantic ambiguity scenarios can be taken into account.

[0106] S2.3 After determining the hierarchical labels, attach the body text block to the nearest valid title node to construct the chapter tree structure and build path information for the body text block:

[0107] ;

[0108] in, Represents a text block Path information, Indicates the first Layer title node, Indicates the path level; path information not only indicates the structural position of the text block in the document tree, but also provides structural priors for subsequent chapter affiliation determination.

[0109] S3 combines chapter titles, path information, and local context to determine the chapter affiliation of each text block and maps it to a standard chapter type;

[0110] S3.1, Let the set of candidate chapter types be... The standard chapter type set is as follows For text blocks and candidate chapter types The chapter attribution scoring function is defined as follows:

[0111] ;

[0112] In the formula, This indicates the title semantic matching score. Indicates the consistency score of the path structure. This indicates that the local context supports the score; , and These correspond to weighting coefficients; thus, the text block is obtained. Initial chapter assignment results:

[0113] ;

[0114] S3.2 Since the same chapter in a real document may have different naming conventions, it is necessary to further map the initial chapter assignment results to the standardized chapter space and define a chapter semantic representation function: ;

[0115] in, This represents a chapter semantic representation function constructed based on the initial chapter affiliation result, path information, and local context;

[0116] S3.3, The standard chapter type mapping satisfies: ;

[0117] In the formula, Represents a text block Standard chapter types, This represents the semantic similarity function of chapters. Through this mapping, chapter descriptions with different writing styles but the same or similar semantics can be uniformly mapped to the same standardized chapter type, thereby ensuring the consistency of the chapter semantic space and providing stable input for the next step of chapter constraint configuration generation.

[0118] S4. Based on the standard chapter type, dynamically generate chapter constraint configurations for each text block. The chapter constraint configuration includes at least an entity type space, a relation type space, and a set of rule constraints.

[0119] S4.1, For text blocks Corresponding standard chapter type Configure its chapter constraints: ;

[0120] in, This indicates the allowed entity type space for this chapter. This indicates the space of allowed relation types in this chapter. This represents the set of rules and restrictions corresponding to this chapter;

[0121] S4.2, The local candidate space of the current text block is defined as follows: ;

[0122] This indicates that for text blocks Subsequent information extraction is not performed unconstrained within a globally unified entity type and relation type space, but rather within a local candidate space of chapters induced by the standardized chapter types. This effectively reduces irrelevant candidates from entering the extraction process. This represents the space of entity type and relation type combinations induced by chapter constraints, and is not a complete triple space;

[0123] Set of rules and restrictions Including but not limited to: candidate entity expression length limit, tail entity phrase form limit, parallel item splitting rule, title fragment filtering rule, granularity control rule and noise filtering rule;

[0124] Chapter constraint configuration is generated by loading synonym chapter merging rules, default extraction rules, chapter customization rules, and chapter constraint overriding rules, so that chapter structure information and extraction control information form a one-to-one correspondence.

[0125] S5, under the chapter constraint configuration, performs information extraction on each text block and performs chapter consistency filtering on the original candidate extraction results to obtain local extraction results;

[0126] Set text block The original candidate extraction result set is To ensure that the extraction results are consistent with the semantics of the current chapter, candidate results are defined. In standardizing chapter types The validity determination function is as follows:

[0127] ;

[0128] Then text block The set of local extraction results satisfies:

[0129] ;

[0130] In the formula, Represents a text block The set of local extraction results;

[0131] Through the chapter consistency screening process, only candidate results that meet the current specification chapter type constraints are retained, thereby suppressing cross-chapter mis-sampling and the mixing of irrelevant relationships;

[0132] Chapter consistency screening includes at least one or more of the following criteria: entity type validity, relation type validity, head and tail entity combination validity, tail entity phrase form validity, and chapter semantic consistency.

[0133] S6 performs intra-chapter aggregation, global aggregation, and normalization on the local extraction results to obtain structured extraction results;

[0134] When post-processing the results of local extraction, one or more of the following operations are included: chapter-level aggregation, global aggregation, bracket alias extraction, alias merging, pronoun resolution, relational direction correction, deduplication, parallel tail entity splitting, title fragment filtering, and noise filtering. Among them, chapter-level aggregation is used to summarize the knowledge expressed in different text blocks under the same standardized chapter type, and standardization processing is used to eliminate the formal differences caused by different text blocks, different expression styles, and different alias forms.

[0135] Let the set of local extraction results for all text blocks be . n represents the number of text blocks participating in the aggregation, and n is a positive integer greater than or equal to 1; First, the local extraction results of multiple text blocks with the same chapter affiliation or the same standard chapter type are merged to obtain the chapter-level aggregation result; Through intra-chapter aggregation, knowledge expressed by different text blocks under the same chapter can be summarized;

[0136] Based on chapter-level aggregation, a global aggregation is performed on the results of the entire document to form a candidate structured result set. Then to Performs lightweight normalization of bracket, punctuation, connector, and alias expressions, and removes duplicate results.

[0137] S7. Based on the structured extraction results, construct the normalized entity projection relationship, and use the normalized entity projection relationship to normalize the entities in the extraction results to obtain the standardized structured results.

[0138] The standardized entity projection relationship is obtained by constructing an entity mapping table from the structured extraction results. The entity mapping table records the standardized entity name, entity type, and alias set, and is used to standardize the original entity representations in the extraction results. The standardized mapping satisfies the following:

[0139] Let the original entity be represented as The standard entity projection function is The solid projection relationship can then be expressed as: ;

[0140] In the formula, Represents the original entity expression The corresponding canonical entity;

[0141] Construct the entity mapping table: ;

[0142] In the formula, Representing a canonical entity entity type, Representation and specification entities The corresponding set of aliases; the entity mapping table can be automatically constructed from the extraction results, or it can be supplemented and improved by combining manual thesaurus or external knowledge;

[0143] After performing normalized entity projection on the entities in the result, normalized triples are obtained:

[0144] ;

[0145] In the formula, Indicates the head entity. Indicates a relationship. This represents the tail entity; the output can be either a chapter-level structured result or a full-text-level structured result, and can be directly used in scenarios such as knowledge graph database entry, entity fusion, and entity disambiguation.

[0146] like Figure 2 The system shown is a chapter-aware information extraction system for structured long texts, including a text block segmentation module, a hierarchy recognition and path construction module, a chapter affiliation determination and standardized chapter type mapping module, a chapter constraint configuration module, an information extraction and chapter consistency filtering module, a result aggregation and standardization processing module, and a standardized entity projection module.

[0147] The system comprises the following modules: a text block segmentation module for segmenting the original structured long text into text blocks to obtain a text block sequence; a hierarchy recognition and path construction module for performing hierarchy recognition on the text block sequence and constructing path information to obtain the hierarchy label and path information corresponding to each text block; a chapter attribution determination and standardized chapter type mapping module for determining the chapter attribution of each text block by combining the chapter title, path information, and local context, and mapping it to a standardized chapter type; a chapter constraint configuration module for generating chapter constraint configurations for each text block based on the standardized chapter type, wherein the chapter constraint configuration includes at least an entity type space, a relation type space, and a set of rule restrictions; an information extraction and chapter consistency filtering module for performing information extraction under the chapter constraint configuration and performing chapter consistency filtering on the original candidate extraction results to obtain local extraction results; a result aggregation and normalization module for performing intra-chapter aggregation, global aggregation, and normalization processing on the local extraction results to obtain structured extraction results; and a standardized entity projection module for constructing standardized entity projection relationships based on the structured extraction results and normalizing the entities in the extraction results to obtain standardized structured results.

[0148] The chapter constraint configuration generated by the chapter constraint configuration module includes an entity type space, a relation type space, and a set of rule restrictions corresponding to the standard chapter type; the information extraction and chapter consistency filtering module performs information extraction within the local candidate space defined by the chapter constraint configuration and performs chapter consistency filtering on the candidate extraction results.

[0149] The core of this invention lies in: first, dividing the original structured long text into text blocks and identifying their hierarchical levels to construct a sequence of text blocks with hierarchical labels and path information; then, transforming the hierarchical identification results into chapter semantic constraints, and further into entity type space, relation type space, and rule restrictions; finally, completing information extraction, result aggregation, and normalized entity projection under chapter-aware conditions. Through this method, the original global extraction problem for the entire long text is transformed into a local constraint extraction problem for different chapter contents.

[0150] To more clearly illustrate the overall method of this invention, it can be defined as a composite mapping from the original structured long text to the standardized structured result. Let the original structured long text be... The final output of the standardized structured result is Then we have: ;

[0151] in, This represents the chapter-aware information extraction mapping. This represents the standardized, structured result after chapter constraints, post-processing, and entity projection.

[0152] Mapping It can be decomposed into: ;

[0153] in, This represents the structure recognition mapping corresponding to text block segmentation, hierarchy recognition, and path construction. This indicates the determination of chapter affiliation and the mapping of standardized chapter types; This indicates the chapter constraint configuration mapping; This represents the chapter-aware information extraction mapping; This represents the aggregation, normalization, and normalized entity projection mapping of results.

[0154] In the rule-driven and large language model-assisted collaborative implementation, hierarchical identification, chapter affiliation determination, and standardized chapter mapping all adopt a rule-driven and large language model-assisted collaborative mechanism. Let the object to be determined be... The comprehensive decision function can then be expressed as:

[0155] ;

[0156] in, This represents the output of the rules system. This indicates the auxiliary output of the large language model. This indicates the trigger threshold. The reason for adopting this collaborative mechanism is that the rule system has a strong ability to determine structural patterns, numbering patterns, and fixed title templates, while the large language model has a strong ability to supplement variant titles, weakly structured text, and cases of local semantic ambiguity. The collaboration between the two can balance efficiency, robustness, and engineering economy.

[0157] This invention is applicable to structured long texts across various domains. For new application domains, only the following aspects need to be adjusted: the set of standardized chapter types, the corresponding entity type space and relation type space in the chapter constraint configuration, the chapter rule restrictions, and the implementation method of the standardized entity space. Thus, the main framework of this invention remains unchanged, while the chapter semantic space and constraint configuration can be locally adapted according to domain differences, thereby achieving good cross-domain transferability.

[0158] The following explanation uses the chapter on "Infectious Bursal Disease" from "Avian Pathology" as the input object.

[0159] The original text of this chapter includes a disease overview and several subsections such as "History and Distribution," "Etiology," "Epidemiology," "Clinical Symptoms," "Pathological Changes," "Diagnosis," and "Prevention and Control." The overview states that "Infectious bursal disease is a serious, immunosuppressive, and highly contagious infectious disease that severely harms chicks. Its pathogen is infectious bursal disease virus infection. The disease is characterized by high morbidity and short course. Typical symptoms include diarrhea, shivering, and extreme weakness. Characteristic lesions include hemorrhage and edema of the bursa of Fabricius, kidney swelling, and hemorrhage in the leg and pectoral muscles." This is followed by detailed discussions in each subsection.

[0160] Step 1: Obtaining the original structured long text and dividing it into text blocks.

[0161] Step 1.1: Obtain the original structured long text.

[0162] The original text of the "Infectious Bursal Disease" chapter in the textbook *Avian Pathology* was obtained. This text contains both a disease overview and multiple subsections with clearly defined headings, making it a typical structured long text. The chapter's organization begins with a general definition of the disease, the pathogen, typical symptoms, and lesions, followed by in-depth descriptions in different subsections. Therefore, it is suitable as a target for extracting chapter-based information.

[0163] Step 1.2: Divide the original structured long text into text blocks.

[0164] The text blocks are identified as headings, with short headings such as "Infectious Bursal Disease," "History and Distribution," "Etiology," "Epidemiology," "Clinical Symptoms," "Pathological Changes," "Diagnosis," and "Prevention and Control." The body paragraphs within each section are identified as body text blocks. This process converts the original long text into a sequential sequence of text blocks.

[0165] In this embodiment, disease names and subsection titles are typically short, clearly structured, and lack complete sentence punctuation, thus they can be identified as title-type text blocks. Content such as "All chicken breeds can be affected, with chickens aged 3-6 weeks being most susceptible" and "Agar diffusion test, fluorescent antibody technology, double-antibody sandwich ELISA, and RT-PCR can be used to diagnose this disease" are identified as body text blocks. After block segmentation, each text block retains its sequential index and relative position for subsequent hierarchical recovery and path construction.

[0166] Step 2: Hierarchical identification and path construction.

[0167] Step 2.1, perform hierarchical recognition.

[0168] The system extracts heading patterns, numbering structures, local context, and semantic features from heading and body text blocks to determine the hierarchical labels of each text block. Since "Infectious Bursal Disease" is a chapter title, and "History and Distribution," "Etiology," and "Epidemiology" are subsection titles, while the paragraphs following each subsection belong to the body text, the system can restore a stable hierarchical structure through heading templates and contextual continuity.

[0169] Step 2.2: Perform text linking and path construction.

[0170] After determining the hierarchy, the text following the "Pathogen" heading is attached to the "Pathogen" node, and the text following the "Epidemiology" heading is attached to the "Epidemiology" node. The same applies to other sections. Thus, each text block obtains a clear path information indicating its position in the chapter tree.

[0171] For example, a text containing "IBDV has two serotypes, namely serotype I and serotype II" is classified under "Infectious Bursal Disease / Etiology" in the path; a text containing "All breeds of chickens can be affected, and chickens aged 3 to 6 weeks are most susceptible to this disease" is classified under "Infectious Bursal Disease / Epidemiology" in the path.

[0172] Step 3: Determine chapter affiliation and standardize chapter type mapping.

[0173] Step 3.1: Perform the initial chapter affiliation determination.

[0174] Based on title information, path information, and local context, each text block is categorized into candidate chapter types such as "History and Distribution," "Etiology," "Epidemiology," "Clinical Symptoms," "Pathological Changes," "Diagnosis," and "Prevention and Control." For example, text blocks describing serotypes and viral biological characteristics are prioritized for inclusion in the Pathology chapter; text blocks describing susceptible age, hosts, and transmission routes are prioritized for inclusion in the Epidemiology chapter; and text blocks describing detection techniques and identification targets are prioritized for inclusion in the Diagnosis chapter.

[0175] Step 3.2: Perform the specification chapter type mapping.

[0176] For synonymous, alternative, and variant titles that may appear in different textbooks or chapters, the system further categorizes them into a unified, standardized chapter type. Through this step, subsequent rule configuration no longer relies on the original literal form of the title, but rather on stable chapter semantic tags.

[0177] In this embodiment, the system ultimately outputs a clear, standardized chapter type for each text block. For example, the text block containing "IBDV is particularly heat-resistant" is mapped to the etiology standardized chapter type; the text block containing "Sick chickens and carrier chickens are the main sources of infection for this disease" is mapped to the epidemiology standardized chapter type; and the text block containing "Agar diffusion test, double-antibody sandwich ELISA, and RT-PCR can be used for the diagnosis of this disease" is mapped to the diagnostic standardized chapter type.

[0178] Step 4: Generate chapter constraint configuration.

[0179] Step 4.1: Generate chapter constraint configurations based on the specified chapter types.

[0180] When the main text block belongs to the "Pathogen" section, the system prioritizes opening entity types such as pathogen, serotype, and biological characteristics, as well as relationship types such as "pathogen is," "classified as," "belongs to serotype," and "possesses biological characteristics." When the main text block belongs to the "Epidemiology" section, the system prioritizes opening entity types such as host, transmission route, susceptibility factors, and morbidity / mortality rate, as well as relationship types such as "susceptible to," "transmission route is," "source of infection is," and "morbidity / mortality rate is." When the main text block belongs to the "Diagnosis" section, the system prioritizes opening diagnostic methods and identification targets, as well as relationship types such as "used for diagnosis" and "used for identification." When the main text block belongs to the "Prevention and Control" section, the system prioritizes opening relationships related to prevention and control measures, disinfection methods, and immunization methods.

[0181] Step 4.2: Load chapter rule restrictions.

[0182] In this embodiment, the rules and restrictions are further specified: in pathological changes, the phrase "location + lesion" is prioritized for output, such as "bursal hemorrhage" and "kidney enlargement"; in epidemiology, "chicks, chicks, turkeys" are identified as hosts, and "3-6 weeks old" is identified as a susceptible factor; in etiology, pathogens, serotypes, and biological characteristics with discriminative significance are allowed to be extracted; at the same time, molecular biological noise content such as RNA, DNA, VP numbers, and sequence sites is prohibited from being output. Through these constraints, the system completes local candidate space pruning before entering the extraction step.

[0183] Step 5: Perform information extraction and chapter consistency screening under chapter constraints.

[0184] Step 5.1, perform extraction under the Pathogens section.

[0185] In the "Etiology" section, the system extracts the following information from statements such as "The pathogen of this disease is infectious bursal virus," "IBDV has two serotypes, namely serotype I and serotype II," "IBDV has no hemagglutination properties," "IBDV is particularly heat-resistant," and "IBDV is resistant to acids and alkalis":

[0186] Results included: "Infectious bursal disease - pathogen is - infectious bursal virus"; "Infectious bursal disease - pathogen is - IBDV"; "IBDV - belongs to serotype - serotype I"; "IBDV - belongs to serotype - serotype II"; "IBDV - has biological characteristics - no hemagglutination properties"; "IBDV - has biological characteristics - heat resistant"; "IBDV - has biological characteristics - acid and alkali resistant", etc.

[0187] Because the current chapter is limited to etiology, the system will not misattribute these results to clinical symptoms or prevention and treatment outcomes.

[0188] Step 5.2, perform the sampling under the Epidemiology section.

[0189] In the "Epidemiology" section, the system extracts the following from statements such as "All breeds of chickens can be affected, including turkeys," "Chickens aged 3-6 weeks are most susceptible to this disease," "Sick and carrier chickens are the main sources of infection," "Infection routes include the digestive tract, respiratory tract, and conjunctiva," and "Morality rates in flocks can be as high as 70% or more."

[0190] Results included: "Infectious bursal disease (IBD) - susceptible to - chickens"; "Infectious bursal disease (IBD) - susceptible to - turkeys"; "Infectious bursal disease (IBD) - susceptible to - 3-6 weeks of age"; "Infectious bursal disease (IBD) - source of infection is - sick chickens"; "Infectious bursal disease (IBD) - source of infection is - carrier chickens"; "Infectious bursal disease (IBD) - transmission route is - digestive tract"; "Infectious bursal disease (IBD) - transmission route is - respiratory tract"; "Infectious bursal disease (IBD) - transmission route is - conjunctiva"; "Infectious bursal disease (IBD) - morbidity / mortality rate is - over 70%", etc.

[0191] Step 5.3: Perform extraction under the section on clinical symptoms and pathological changes.

[0192] In the main text related to disease overview, clinical symptoms, and pathological changes, the system extracted the following from descriptions such as "typical symptoms include diarrhea, tremors, and extreme weakness," "bursal hemorrhage, edema, kidney swelling, and hemorrhage in the leg and pectoral muscles," "transverse hemorrhage at the junction of proventriculus and gizzard," and "kidney enlargement and urate deposition":

[0193] Results included: “Infectious bursal disease – manifested as – diarrhea”; “Infectious bursal disease – manifested as – tremors”; “Infectious bursal disease – manifested as – extreme weakness”; “Infectious bursal disease – pathological change – hemorrhage of the bursa of Fabricius”; “Infectious bursal disease – pathological change – kidney swelling”; “Infectious bursal disease – necropsy reveals – linear and punctate hemorrhages in the pectoral and hamstring muscles”; “Infectious bursal disease – necropsy reveals – transverse hemorrhages at the junction of the proventriculus and gizzard”; “Infectious bursal disease – necropsy reveals – kidney enlargement and urate deposition”.

[0194] The pathological changes extracted here follow the phrase rule of "location + lesion" to avoid outputting isolated location terms such as "leg muscle" or "chest muscle".

[0195] Step 5.4, perform extraction under the Diagnostics section.

[0196] In the "Diagnosis" section, the system extracts the following from statements such as "Agar diffusion test, fluorescent antibody technology, double antibody sandwich ELISA, and virus neutralization test can be used to diagnose this disease," "RT-PCR in molecular biology techniques can be used to diagnose this disease," and "Note that it should be differentiated from renal infectious bronchitis."

[0197] Results including: "Agar diffusion test - for diagnosis of infectious bursal disease"; "Fluorescent antibody technology - for diagnosis of infectious bursal disease"; "Double antibody sandwich ELISA - for diagnosis of infectious bursal disease"; "RT-PCR - for diagnosis of infectious bursal disease"; "Infectious bursal disease - for differential diagnosis of renal infectious bronchitis", etc.

[0198] Since the current chapter's constraint configuration only allows entity types and relationships such as diagnostic methods and identification objects, content such as hosts and symptoms occasionally mentioned in this chapter will not be mixed into the diagnostic results.

[0199] Step 5.5, perform extraction under the prevention section.

[0200] In the "Prevention and Control" section, the system prioritizes extracting general and actionable prevention and control measures, such as: "Live vaccine immunization—for the prevention of infectious bursal disease"; "Oil adjuvant inactivated vaccine immunization—for the prevention of infectious bursal disease"; "Multivalent vaccine immunization—for the prevention of infectious bursal disease"; "Formalin fumigation disinfection—for the prevention of infectious bursal disease"; "Phenolic disinfection—for the prevention of infectious bursal disease"; "Strong alkaline disinfectant solution—for the prevention of infectious bursal disease", etc.

[0201] In this step, specific vaccine codes or model lists are not prioritized for retention, thus ensuring that the output is more suitable for knowledge graph database entry and standardized utilization.

[0202] Step 6: Aggregate and normalize the local extraction results.

[0203] Step 6.1: Perform intra-chapter aggregation.

[0204] After completing the local extraction of each section, the system first aggregates the results by chapter, and then performs a global aggregation of the entire text. Through this step, the etiological facts, epidemiological facts, diagnostic facts, and prevention and control facts scattered in different text blocks can be summarized into chapter-level structured results.

[0205] Step 6.2: Perform normalization processing.

[0206] Based on the aggregated results, the system further performs deduplication, lightweight normalization, parallel item splitting, relation direction correction, and alias merging. For example, if the same fact is repeatedly stated in different paragraphs, only one standard result is retained; for parallel tail entities, they are split into multiple independent triples; and for phrases with slightly different expressions, they are unified into standard phrase forms.

[0207] Step 7: Perform normal entity projection and build entity mapping table.

[0208] Step 7.1, perform standard entity projection.

[0209] Based on the original entity representations in the extraction results, the system constructs canonical entity projection relationships and entity mapping tables. For example, "infectious bursal disease," "Gumboro disease," and "infectious bursal virus" are uniformly projected to the same disease canonical entity, and "infectious bursal virus," "IBDV," and "IB-DV" are uniformly projected to the same pathogen canonical entity. Canonical entity projection can be performed on both head and tail entities to obtain standardized triplet outputs.

[0210] Step 7.2: Generate the entity mapping table.

[0211] The entity mapping table records the canonical entity name, entity type, and alias set, and outputs them as an independent result. Through this entity mapping table, subsequent knowledge graph import, entity fusion, entity disambiguation, and relaxed evaluation can all be processed within a unified entity space.

[0212] Step 7.3: Output standardized structured results.

[0213] After entity projection is completed, the system outputs chapter-level structured results, full-text-level structured results, and entity mapping tables for use in knowledge graph database entry, entity fusion, entity disambiguation, and lenient evaluation. This embodiment demonstrates that the present invention can transform the chapter structure of textbook-style long texts into extraction constraints and further output structured knowledge results suitable for standardized utilization in knowledge graphs.

[0214] This invention is not limited to the above embodiments. Based on the technical solutions disclosed herein, those skilled in the art can make some substitutions and modifications to some of the technical features without creative effort, and all such substitutions and modifications are within the protection scope of this invention.

Claims

1. A method for extracting chapter-aware information from structured long texts, characterized in that: Includes the following steps, S1. Obtain the original structured long text and divide the original structured long text into text blocks to obtain a text block sequence; S2. Perform hierarchical recognition on the text block sequence and construct path information to obtain the hierarchical labels and path information corresponding to each text block; S3. Combine the chapter title, path information and local context to determine the chapter affiliation of each text block and map it to the standard chapter type; S4. Based on the standard chapter type, dynamically generate chapter constraint configurations for each text block. The chapter constraint configurations should include at least the entity type space, the relation type space, and the set of rule constraints. S5. Under the chapter constraint configuration, perform information extraction on each text block and perform chapter consistency filtering on the original candidate extraction results to obtain local extraction results. S6. Perform intra-chapter aggregation, global aggregation, and normalization on the local extraction results to obtain structured extraction results; S7. Construct a normalized entity projection relationship based on the structured extraction results, and use the normalized entity projection relationship to normalize the entities in the extraction results to obtain a standardized structured result.

2. The method for extracting chapter-aware information from structured long texts according to claim 1, characterized in that: The specific content of S1 is as follows: S1.1 Perform pre-cleaning on the original structured long text. Pre-cleaning includes removing meaningless blank lines, normalizing whitespace characters, identifying heading numbering patterns, removing redundant parentheses in headings, and preserving one or more of paragraph boundaries and sequential indexes. S1.

2. Based on one or more of the following: title style, numbering pattern, paragraph boundary, line break rule, punctuation pattern, and length threshold, the original structured long text is segmented into blocks. S1.3 Divide the text fragments that meet the title determination conditions into title-type text blocks, divide the text fragments that carry the semantic content of the body text into body-type text blocks, and retain at least one of the sequential index, position index or source document offset of each text block. S1.

4. Temporarily store the body text blocks that are subordinate to the adjacent title blocks as attachable objects to form a sequence of text blocks for subsequent hierarchical identification, path construction and chapter affiliation determination.

3. The method for extracting chapter-aware information from structured long texts according to claim 2, characterized in that: The specific details of S2 are as follows. S2.1, Let the sequence of text blocks obtained after partitioning by S1 be... ,in Indicates the first I. Text blocks; for each text block, extract its structural feature representation: ； in, Indicates the characteristics of the title pattern. Indicates the structural characteristics of the numbering system. This indicates the continuity of local context. Represents local semantic features; S2.2, Let the set of candidate level labels be: ； in, Includes one or more hierarchical tags such as book level, document level, chapter level, section level, subsection level, and body text level; for text blocks and candidate level tags Define the hierarchical recognition scoring function as follows: ； In the formula, Represents a text block With candidate level labels Title pattern matching degree; This indicates consistency in the numbering structure; Indicates local context continuity; This indicates the strength of local semantic support for hierarchical tags; , , and These are the corresponding weighting coefficients; Based on this scoring function, text blocks can be determined using the following formula. Hierarchical tags: ； That is, the candidate level label that maximizes the level recognition scoring function is selected as the text block. Hierarchical tags; The confidence level for the rule-driven decision is defined as follows: ； when When, the result is determined directly by the rule; when At that time, the large language model is invoked for auxiliary determination, that is: ； In the formula, This represents the confidence threshold for hierarchical determination. This represents the auxiliary level decision function of the large language model. Indicates text block The associated local context; through this collaborative determination method, the efficiency of the rule system in structural pattern recognition and the flexible supplementary ability of the large language model in semantic ambiguity scenarios can be taken into account. S2.3 After determining the hierarchical labels, attach the body text block to the nearest valid title node to construct the chapter tree structure and build path information for the body text block: ； in, Represents a text block Path information, Indicates the first Layer title node, Indicates the path level; path information not only indicates the structural position of the text block in the document tree, but also provides structural priors for subsequent chapter affiliation determination.

4. The method for extracting chapter-aware information from structured long texts according to claim 3, characterized in that: The specific details of S3 are as follows. S3.1, Let the set of candidate chapter types be... The standard chapter type set is as follows For text blocks and candidate chapter types The chapter attribution scoring function is defined as follows: ； In the formula, This indicates the title semantic matching score. Indicates the consistency score of the path structure. This indicates that the local context supports the score; , and These correspond to weighting coefficients; thus, the text block is obtained. Initial chapter assignment results: ； S3.2 Since the same chapter in a real document may have different naming conventions, it is necessary to further map the initial chapter assignment results to the standardized chapter space and define a chapter semantic representation function: ； in, This represents a chapter semantic representation function constructed based on the initial chapter affiliation result, path information, and local context; S3.3, The standard chapter type mapping satisfies: ； In the formula, Represents a text block Standard chapter types, This represents the semantic similarity function of chapters. Through this mapping, chapter descriptions with different writing styles but the same or similar semantics can be uniformly mapped to the same standardized chapter type, thereby ensuring the consistency of the chapter semantic space and providing stable input for the next step of chapter constraint configuration generation.

5. The method for extracting chapter-aware information from structured long texts according to claim 4, characterized in that: The specific content of S4 is as follows. S4.1, For text blocks Corresponding standard chapter type Configure its chapter constraints: ; in, This indicates the allowed entity type space for this chapter. This indicates the space of allowed relation types in this chapter. This represents the set of rules and restrictions corresponding to this chapter; S4.2, The local candidate space of the current text block is defined as follows: ； This indicates that for text blocks Subsequent information extraction is not performed unconstrained within a globally unified entity type and relation type space, but rather within a local candidate space of chapters induced by the standardized chapter types. This effectively reduces irrelevant candidates from entering the extraction process. This represents the space of entity type and relation type combinations induced by chapter constraints, and is not a complete triple space; Set of rules and restrictions Including but not limited to: candidate entity expression length limit, tail entity phrase form limit, parallel item splitting rule, title fragment filtering rule, granularity control rule and noise filtering rule; Chapter constraint configuration is generated by loading synonym chapter merging rules, default extraction rules, chapter customization rules, and chapter constraint overriding rules, so that chapter structure information and extraction control information form a one-to-one correspondence.

6. The method for extracting chapter-aware information from structured long texts according to claim 5, characterized in that: The specific details of S5 are as follows. Set text block The original candidate extraction result set is To ensure that the extraction results are consistent with the semantics of the current chapter, candidate results are defined. In standardizing chapter types The validity determination function is as follows: ； Then text block The set of local extraction results satisfies: ； In the formula, Represents a text block The set of local extraction results; Through the chapter consistency screening process, only candidate results that meet the current specification chapter type constraints are retained, thereby suppressing cross-chapter mis-sampling and the mixing of irrelevant relationships; Chapter consistency screening includes at least one or more of the following criteria: entity type validity, relation type validity, head and tail entity combination validity, tail entity phrase form validity, and chapter semantic consistency.

7. The method for extracting chapter-aware information from structured long texts according to claim 6, characterized in that: When post-processing the results of local extraction in S6, one or more of the following operations are included: chapter-level aggregation, global aggregation, bracket alias extraction, alias merging, pronoun resolution, relational direction correction, deduplication, parallel tail entity splitting, title fragment filtering, and noise filtering. Among them, chapter-level aggregation is used to summarize the knowledge expressed in different text blocks under the same standardized chapter type, and standardization processing is used to eliminate the formal differences caused by different text blocks, different expression styles, and different alias forms. Let the set of local extraction results for all text blocks be . n represents the number of text blocks participating in the aggregation, and n is a positive integer greater than or equal to 1; First, the local extraction results of multiple text blocks with the same chapter affiliation or the same standard chapter type are merged to obtain the chapter-level aggregation result; Through intra-chapter aggregation, knowledge expressed by different text blocks under the same chapter can be summarized; Based on chapter-level aggregation, a global aggregation is performed on the results of the entire document to form a candidate structured result set. Then to Performs lightweight normalization of bracket, punctuation, connector, and alias expressions, and removes duplicate results.

8. The method for extracting chapter-aware information from structured long texts according to claim 7, characterized in that: In S7, the normalized entity projection relationship is obtained by constructing an entity mapping table from the structured extraction results. The entity mapping table records the normalized entity name, entity type, and alias set, and uses this table to normalize the original entity representations in the extraction results. The normalized mapping satisfies the following: Let the original entity be represented as The standard entity projection function is The solid projection relationship can then be expressed as: ; In the formula, Represents the original entity expression The corresponding canonical entity; Construct the entity mapping table: ; In the formula, Representing a canonical entity entity type, Representation and specification entities The corresponding set of aliases; the entity mapping table can be automatically constructed from the extraction results, or it can be supplemented and improved by combining manual thesaurus or external knowledge; After performing normalized entity projection on the entities in the result, normalized triples are obtained: ； In the formula, Indicates the head entity. Indicates a relationship. This represents the tail entity; the output can be either a chapter-level structured result or a full-text-level structured result, and can be directly used in scenarios such as knowledge graph database entry, entity fusion, and entity disambiguation.

9. A chapter-aware information extraction system for structured long texts, characterized in that: The method for extracting chapter-aware information from structured long texts according to any one of claims 1-8 includes a text block segmentation module, a hierarchy recognition and path construction module, a chapter affiliation determination and standardized chapter type mapping module, a chapter constraint configuration module, an information extraction and chapter consistency screening module, a result aggregation and standardized processing module, and a standardized entity projection module. The text block segmentation module is used to segment the original structured long text into text blocks to obtain a text block sequence. The hierarchy recognition and path construction module is used to perform hierarchy recognition on the text block sequence and construct path information to obtain the hierarchy label and path information corresponding to each text block. The chapter attribution determination and standardized chapter type mapping module is used to determine the chapter attribution of each text block by combining the chapter title, path information and local context, and map it to a standardized chapter type; The chapter constraint configuration module is used to generate chapter constraint configurations for each text block according to the standard chapter type. The chapter constraint configuration includes at least an entity type space, a relation type space, and a set of rule constraints. The information extraction and chapter consistency filtering module is used to perform information extraction under chapter constraint configuration and perform chapter consistency filtering on the original candidate extraction results to obtain local extraction results. The result aggregation and normalization module is used to perform intra-chapter aggregation, global aggregation, and normalization on local extraction results to obtain structured extraction results. The normalized entity projection module is used to construct normalized entity projection relationships based on the structured extraction results, and to normalize the entities in the extraction results to obtain standardized structured results.

10. A chapter-aware information extraction system for structured long texts according to claim 9, characterized in that: The chapter constraint configuration generated by the chapter constraint configuration module includes an entity type space, a relation type space, and a set of rule restrictions corresponding to the standard chapter type; the information extraction and chapter consistency filtering module performs information extraction within the local candidate space defined by the chapter constraint configuration and performs chapter consistency filtering on the candidate extraction results.

Images