An AI-based intelligent information extraction method and system

Abstract

The application discloses an AI-based intelligent information extraction method and system, and relates to the technical field of intelligent information extraction.The method comprises the following steps: establishing evidence mapping of each field candidate in a field candidate set and an original text evidence fragment by a semantic evidence boundary discrimination model, recording the support range, semantic connection path and derivation level of the original text evidence fragment on the field candidate, and obtaining field evidence boundary representation; dividing the field candidate according to the support result of the evidence gap item on the supplementary text fragment, and outputting the evidence boundary extraction result.The application establishes the mapping of the field candidate and the original text evidence fragment by the semantic evidence boundary discrimination model, records the support range, semantic connection path and derivation level, positions the evidence gap item to the specific field, original text position and to-be-supplemented semantic role, generates a field supplement object for review, and improves the traceability, structural integrity and output stability of the extraction result.

Description

Technical Field

[0001] This invention relates to the field of intelligent information extraction technology, and in particular to an AI-based intelligent information extraction method and system. Background Technology

[0002] With the development of natural language processing, deep semantic understanding, and artificial intelligence models, information extraction technology has been widely applied in scenarios such as text element recognition, entity relationship parsing, automatic field filling, and structured content generation. Information extraction for business fields is gradually shifting from keyword matching to semantic fragment recognition, context analysis, and multi-field collaborative processing.

[0003] In the process of extracting target field structures, related technologies usually focus on whether the field content is identified and filled in. However, they lack a continuous description of the support range, semantic connection path and deduction level between the field content and the original evidence fragments. This results in unclear evidence boundaries for field candidates, making it difficult to accurately identify direct support, indirect support and insufficient evidence, and affecting the reliability of subsequent evidence supplementation and structured verification. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an AI-based intelligent information extraction method to solve the problems of unclear evidence support boundaries and difficulty in verifying derivation levels in target field extraction.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an AI-based intelligent information extraction method, which includes: acquiring the text to be extracted and the target field structure; performing text cleaning, semantic slicing, context boundary marking and semantic role recognition on the text to be extracted to obtain semantic fragment data, and loading a semantic evidence boundary discrimination model; extracting semantic elements from the semantic fragment data; filling in the semantic elements according to the field name, semantics and position in the target field structure; merging fields and marking empty spaces to obtain a field candidate set; A semantic evidence boundary discrimination model is used to establish an evidence mapping between each field candidate in the field candidate set and the original text evidence fragment. The support range, semantic connection path, and derivation level of the original text evidence fragment for the field candidate are recorded to obtain the field evidence boundary representation. Based on the field evidence boundary representation, the consistency of the support range, semantic connection path, and derivation level in the original text evidence fragment for each field candidate is checked to identify the evidence gap items corresponding to the field candidate and generate an evidence supplementation list. Supplementary text fragments are retrieved based on the evidence supplementation list. The supplementary text fragments and the field evidence boundary representation are input into the semantic evidence boundary discrimination model for evidence mapping review and structured verification. Based on the support results of the supplementary text fragments for the evidence gap items, the field candidates are divided and the evidence boundary extraction results are output.

[0007] As a preferred embodiment of the AI-based intelligent information extraction method of the present invention, the steps of obtaining the text to be extracted and the target field structure, performing text cleaning, semantic slicing, context boundary labeling and semantic role recognition on the text to be extracted to obtain semantic fragment data, and loading the semantic evidence boundary discrimination model are as follows: The extracted text is processed by text normalization and content filtering to obtain cleaned text content; The cleaned text content is semantically sliced, contextual boundary labeled, and semantic role identified to obtain semantic fragment data, and the semantic evidence boundary discrimination model is loaded.

[0008] As a preferred embodiment of the AI-based intelligent information extraction method of the present invention, the semantic evidence boundary discrimination model is specifically constructed as follows: A semantic evidence boundary discrimination model is constructed based on a field fragment alignment layer, an evidence support determination layer, and an evidence boundary attribution layer. The field fragment alignment layer extracts the fragment source location, context boundary, semantic role, field name, field semantics and field location from the semantic fragment data, and aligns the source location, context boundary and field affiliation between field candidates and candidate evidence fragments to form field fragment alignment information. The evidence support determination layer, based on field fragment alignment information, determines the support status of the core semantic element coverage relationship, context boundary correspondence relationship and semantic connection path between candidate evidence fragments and field candidates, and determines whether the candidate evidence fragments provide direct support, indirect support or no support to the field candidates, thus forming a field support determination list. The evidence boundary attribution layer, based on the field support judgment list, performs boundary merging, level marking, and evidence gap attribution judgment on the support range, semantic connection path, and deduction level corresponding to each field candidate, forming the evidence boundary attribution content.

[0009] As a preferred embodiment of the AI-based intelligent information extraction method of the present invention, the specific steps for obtaining the field candidate set are as follows: Extract semantic elements from semantic fragment data, and perform semantic element identification, fragment source location and context location correspondence to obtain a set of semantic elements; The semantic element set is matched with the target field structure by field name matching, field semantic correspondence, and field position filling to obtain field filling information; By filling in information in the fields, the content of similar fields is merged and organized, empty target fields are marked, and candidate fields that have been filled and marked are organized to obtain a set of candidate fields.

[0010] As a preferred embodiment of the AI-based intelligent information extraction method of the present invention, the specific steps for obtaining the field evidence boundary representation are as follows: The field candidate set and semantic fragment data are input into the semantic evidence boundary discrimination model. The field candidates are split into candidates, correspond to the original text position, and extracted from the candidate evidence fragments to obtain the content corresponding to the field evidence. The semantic evidence boundary discrimination model is used to perform semantic support verification and context boundary matching between the content corresponding to the field evidence and the candidate evidence fragments. It determines whether the candidate evidence fragments support the field candidate directly, indirectly, or not. The semantic connection paths are organized according to the support status to obtain the field evidence mapping set. By using the field evidence mapping set, the semantic evidence boundary discrimination model performs boundary merging, hierarchical labeling, and result reorganization on the support range, semantic connection path, and inference level of each field candidate, thereby obtaining the field evidence boundary representation.

[0011] As a preferred embodiment of the AI-based intelligent information extraction method of the present invention, the specific steps for obtaining the field evidence mapping set are as follows: Based on the content corresponding to the field evidence, the field candidates, candidate evidence fragments and context boundaries are extracted and organized, and the semantic evidence boundary discrimination model is used to identify the coverage of the candidate evidence fragments on the core semantic elements of the field candidates, so as to obtain semantic support verification information. Based on the semantic support verification information, the context boundary of the candidate evidence fragment is matched with the fragment source position of the field candidate through the semantic evidence boundary discrimination model, and the initial semantic connection path from the candidate evidence fragment to the field candidate is sorted out to obtain the context boundary matching result. Based on semantic support verification information and context boundary matching results, candidate evidence fragments that cover core semantic elements and are in the same context boundary are judged as direct support, candidate evidence fragments that cover core semantic elements and are in continuous context boundaries through the initial semantic connection path are judged as indirect support, and any of the following situations are judged as not support: not covering core semantic elements, discontinuous context boundaries, and interrupted initial semantic connection path. The semantic connection path is then organized according to the judgment results to obtain a set of field evidence mappings.

[0012] As a preferred embodiment of the AI-based intelligent information extraction method of the present invention, the following steps are taken: Based on the field evidence boundary representation, the consistency of the supporting range, semantic connection path and inference level in the original text evidence fragments of each field candidate is checked to identify the evidence gap items corresponding to the field candidate and generate an evidence supplementation list: Based on the field evidence boundary representation, the supporting range, semantic connection path and derivation level corresponding to each field candidate are organized to obtain the field evidence verification content; Based on the content of the field evidence verification, the consistency of the supporting relationship, semantic connection relationship and inference hierarchy relationship between each field candidate and the original text evidence fragment is verified to obtain field evidence gap information; By using field evidence gap information, the candidate field identifier, gap type, original text location, semantic role to be supplemented, and direction of supplementation are reorganized to generate an evidence supplementation list.

[0013] As a preferred embodiment of the AI-based intelligent information extraction method of the present invention, the specific steps for outputting the evidence boundary extraction result are as follows: Based on the evidence supplementation list and the field evidence boundary representation, the evidence gap items are checked against the supporting scope, semantic connection path and deduction level. They are classified into four types of supplementation: direct evidence missing, insufficient supporting scope, semantic connection interruption and deduction level exceeding the scope of deduction, and field supplementation objects are generated. Based on the field supplementation object, extract the supplementary text fragments corresponding to the field candidate, the semantic role to be supplemented, and the original text position, and combine them with the surrounding text content of the original text evidence fragments to perform semantic slicing and supporting relationship marking to obtain the supplementary evidence fragments; By supplementing evidence fragments and representing field evidence boundaries, the candidate fields are subjected to evidence mapping verification, derivation level update and structured verification, and the evidence boundary extraction results are output.

[0014] As a preferred embodiment of the AI-based intelligent information extraction method of the present invention, the specific steps for generating the field supplementation object are as follows: Based on the evidence supplementation list, candidate field identifiers are extracted, the source of the gap is located, the semantic role to be supplemented is extracted, and the original text position is organized to obtain gap organization information; Based on the gap information and field evidence boundary representation, the evidence gap items are checked against the supporting scope, semantic connection path and inference level. The gap type is determined according to the check results. If there is no corresponding original text evidence fragment, it is determined as direct evidence missing. If the semantic role to be supplemented exceeds the supporting scope, it is determined as insufficient supporting scope. If there is a break in the semantic connection path, it is determined as semantic connection interruption. If the semantic role to be supplemented is not supported by evidence within the record level, it is determined as inference level exceeding the limit. The gap type determination result is obtained. Based on the gap type determination results, the missing direct evidence is matched with the direct evidence retrieval object of the same field, the insufficient support range is matched with the adjacent context evidence retrieval object, the interrupted semantic connection is matched with the intermediate semantic role evidence retrieval object, and the inference level exceeds the limit is matched with the near source evidence retrieval object. The field candidate, the semantic role to be supplemented, the original text position and the corresponding evidence retrieval object are reorganized into the field supplementation object.

[0015] Secondly, this invention provides an AI-based intelligent information extraction system, comprising: a text normalization module, used to acquire the text to be extracted and the target field structure, perform text cleaning, semantic slicing, context boundary marking, and semantic role recognition on the text to be extracted to obtain semantic fragment data, and load a semantic evidence boundary discrimination model; a field candidate module, used to extract semantic elements from the semantic fragment data, fill in the semantic elements according to the field name, semantics, and position in the target field structure, merge fields, and mark empty spaces to obtain a field candidate set; and an evidence discrimination module, used to establish an evidence mapping between each field candidate in the field candidate set and the original text evidence fragment through the semantic evidence boundary discrimination model. The system records the support range, semantic connection path, and derivation level of the original text evidence fragments for the field candidates, thus obtaining the field evidence boundary representation. The supplementary evidence module is used to perform consistency checks on the support range, semantic connection path, and derivation level of the original text evidence fragments for each field candidate based on the field evidence boundary representation, identify the evidence gap items corresponding to the field candidates, and generate an evidence supplement list. The extraction and output module is used to retrieve supplementary text fragments based on the evidence supplement list, input the supplementary text fragments and field evidence boundary representations into the semantic evidence boundary discrimination model for evidence mapping review and structured verification, divide the field candidates according to the support results of the supplementary text fragments for the evidence gap items, and output the evidence boundary extraction results.

[0016] The beneficial effects of this invention are as follows: by establishing a mapping between field candidates and original text evidence fragments through a semantic evidence boundary discrimination model, the support range, semantic connection path and derivation level are recorded, so that the evidence gap item is located to a specific field, original text position and semantic role to be supplemented, and field supplementation objects are generated for review, thereby improving the traceability, structural integrity and output stability of the extraction results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an AI-based intelligent information extraction method.

[0019] Figure 2 This is a flowchart of semantic fragments and field candidates.

[0020] Figure 3 This is a flowchart for boundary determination of field evidence mapping.

[0021] Figure 4 This is a flowchart for filling evidence gaps and extracting boundaries. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an AI-based intelligent information extraction method, including the following steps: S1: Obtain the text to be extracted and the target field structure, perform text cleaning, semantic slicing, context boundary marking and semantic role recognition on the text to be extracted to obtain semantic fragment data, and load the semantic evidence boundary discrimination model; the semantic evidence boundary discrimination model is constructed based on the field fragment alignment layer, the evidence support judgment layer and the evidence boundary attribution layer.

[0026] S1.1: Field Fragment Alignment Layer: Extracts the source location, context boundary, semantic role, field name, field semantics, and field location from the semantic fragment data. Aligns the source location, context boundary, and field affiliation between field candidates and candidate evidence fragments to form field fragment alignment information. S1.2: Evidence Support Judgment Layer. Based on field fragment alignment information, it performs support status judgment on the core semantic element coverage relationship, context boundary correspondence relationship and semantic connection path between candidate evidence fragments and field candidates. It determines whether the candidate evidence fragments provide direct support, indirect support or no support to the field candidates, and forms a field support judgment list. S1.3: Evidence Boundary Attribution Layer. Based on the field support judgment list, the support range, semantic connection path and deduction level corresponding to each field candidate are merged, level marked and evidence gap attribution is judged to form the evidence boundary attribution content.

[0027] The training process of the semantic evidence boundary discrimination model is as follows: a training sample set is established using historical text to be extracted, current text samples to be extracted, manually verified field extraction records, original evidence fragment records, and target field structures as training data sources. Historical text to be extracted includes business texts that have completed field extraction and retained the original text positions. Current text samples to be extracted include similar texts with the same or similar target field structures. Manually verified field extraction records include field names, field semantics, field positions, field value content, fragment source positions, and empty space markers. Original evidence fragment records include the start and end positions of original characters supporting field candidates, context boundaries, semantic roles, semantic connection paths, inference levels, and evidence gap attribution.

[0028] The training samples are hierarchically organized according to field type, text source, support status, and gap type, so that each target field includes directly supported samples, indirectly supported samples, unsupported samples, samples with missing direct evidence, samples with insufficient support scope, samples with broken semantic connections, and samples with inference levels exceeding the limit. The training sample set includes 12,000 field candidate samples, which are divided into training set, validation set, and test set in an 8:1:1 ratio. Field candidate samples generated from the same original text are assigned to the same set to avoid the same text appearing in the training set, validation set, or test set at the same time.

[0029] Manual annotation is performed on the candidate set of fields and semantic fragment data. The annotation specifications include: marking candidate evidence fragments that can cover the content pointed to by the field name, the content pointed to by the field semantics, and the content from which the field location originates within the same context boundary as direct support; marking candidate evidence fragments that need to be connected by adjacent context boundaries and intermediate semantic roles to cover the above core semantic elements as indirect support; marking candidate evidence fragments that do not cover the core semantic elements, have discontinuous context boundaries, or have an interrupted initial semantic connection path as unsupported; marking cases where field candidates do not correspond to original text evidence fragments as missing direct evidence; marking cases where the semantic roles to be supplemented exceed the range of already merged support as insufficient support range; marking cases where there are breaks in the semantic connection path as semantic connection interrupted; and marking cases where the semantic roles to be supplemented do not obtain support from original text evidence fragments within the record inference level as inference level exceeded.

[0030] During training, the field candidate set, semantic fragment data, and training annotation file are input into the semantic evidence boundary discrimination model. The field fragment alignment layer learns the source position alignment, context boundary alignment, and field attribution alignment between field candidates and candidate evidence fragments. The evidence support determination layer learns the correspondence between core semantic element coverage, context boundary correspondence, initial semantic connection path, and support status. The evidence boundary attribution layer learns the correspondence between support range, semantic connection path, derivation level, and evidence gap attribution.

[0031] The loss function includes field fragment alignment loss, evidence support state classification loss, semantic connection path annotation loss, inference level classification loss, and evidence gap attribution classification loss. The total training loss is formed by merging the above losses. The merged total training loss is used to back-update the parameters of the field fragment alignment layer, evidence support determination layer, and evidence boundary attribution layer. The initial learning rate is set to 2×10^-5, the batch size is set to 16, and the number of training rounds is set to 20. Training stops when there is no improvement on the validation set for 3 consecutive rounds. After each training round, the field fragment alignment accuracy, support state determination accuracy, evidence boundary attribution accuracy, and evidence gap identification recall are calculated on the validation set. When all validation metrics reach above 0.90 and the validation set loss no longer decreases, the model is considered to have met the training convergence condition.

[0032] After training, the semantic evidence boundary discrimination model is independently validated using a test set. The test content includes validation of the correspondence between the original text positions of candidate fields and candidate evidence fragments, validation of the distinction between direct and indirect support, validation of the exclusion of unsupported samples, validation of the merging of support ranges, validation of the integrity of semantic connection paths, validation of derivation level labeling, and validation of evidence gap types. When the field fragment alignment accuracy, support status determination accuracy, evidence boundary attribution accuracy, and evidence gap identification recall rate in the test set all reach above 0.88, the model is considered validated. After validation, the following files are generated: position label file corresponding to the field fragment alignment layer, semantic role label file corresponding to the evidence support determination layer, support status label file corresponding to the evidence boundary attribution layer, semantic connection path label file, derivation level label file, evidence gap attribution label file, and model parameter file. These files are used for loading the semantic evidence boundary discrimination model, evidence mapping, gap identification, and supplementary review.

[0033] S1.4: Perform text normalization and content filtering on the extracted text to obtain cleaned text content.

[0034] Furthermore, the main text, title text, table text, and footnote text in the text to be extracted are read, and the source position of each text is preserved according to the order of appearance in the text to be extracted. Then, the read text is subjected to character encoding unification, full-width and half-width unification, punctuation format unification, continuous space merging, abnormal line break merging, and paragraph sentence breaking correction to obtain a well-formatted text with consistent format.

[0035] Remove interfering text located in headers, footers, page numbers, tables of contents, duplicate declarations, and formatting instructions from the regularized text. Retain text content containing field names, semantic terms, and position identifiers of the target field structure. For text content that matches the target field structure, simultaneously retain the adjacent previous and next statements of the paragraph to ensure that the field semantics have a contextual source during processing. Reorder the retained text content according to the source position, merge paragraphs, and remove blank lines to obtain cleaned text content.

[0036] S1.5: Perform semantic slicing, context boundary labeling, and semantic role recognition on the cleaned text content to obtain semantic fragment data, and complete the loading of the semantic evidence boundary discrimination model.

[0037] Furthermore, the paragraph number, sentence number, character start and end positions, and punctuation separation positions of the cleaned text content are read. The cleaned text content is divided into basic sentence segments according to paragraph line breaks and sentence-end punctuation. The basic sentence segments are then subjected to word segmentation, part-of-speech tagging, named entity recognition, and dependency parsing, retaining the character positions corresponding to entity words, numerical words, time words, action words, attribute words, and negation words.

[0038] The basic sentence segments are semantically mapped to the field names, field semantics, and field positions in the target field structure. Basic sentence segments containing the same field semantics, the same entity object, the same event action, or the same limiting condition are merged into candidate semantic fragments, and the source position and character boundaries of the fragments are recorded. Taking the candidate semantic fragments as the center, the context expansion range is determined according to the text type. One to three basic sentence segments are read forward and backward in continuous text. Long sentence texts are first split into clauses according to punctuation and parallel predicates before expansion. Table texts use the same row title, the same column title, merged cells, and footnotes as associated basic sentence segments.

[0039] When cross-paragraph texts have entity objects, time conditions, conditional constraints, negative constraints, or event action succession relationships in adjacent paragraphs, they are included in the expansion scope. During the expansion process, if the current base sentence segment and the candidate semantic segment have any of the above-mentioned related content, they are included in the context scope. Expansion stops when the related content is not valid or the maximum expansion scope is reached. The context boundary is marked, and the context boundary is used for subsequent identification of direct support, indirect support, semantic connection paths, and evidence gaps.

[0040] Semantic role annotation is performed on candidate semantic fragments that have completed context boundary marking. This involves identifying field fill values, subject objects, action relationships, attribute constraints, time conditions, condition ranges, negation constraints, and source locations, forming semantic fragment data that includes fragment number, start and end positions of original characters, context boundaries, semantic roles, field correspondences, and fragment source locations. The parameter file, vocabulary file, and field label file are read sequentially according to the field fragment alignment layer, evidence support judgment layer, and evidence boundary attribution layer of the semantic evidence boundary discrimination model. The field names, field semantics, and field positions in the target field structure are written into the input configuration of the semantic evidence boundary discrimination model. The format of the field labels, character positions, and context boundaries of the semantic fragment data is validated to obtain the semantic fragment data and complete the loading of the semantic evidence boundary discrimination model.

[0041] It should be noted that the location tag file is used to identify the source location and context boundary of the fragment, the semantic role tag file is used to identify the semantic role correspondence between the field candidate and the candidate evidence fragment, and the support status tag file is used to identify the direct support, indirect support and non-support status. The three types of files together define the input recognition caliber of the semantic evidence boundary discrimination model.

[0042] S2: Extract semantic elements from semantic fragment data, fill in the semantic elements according to the field name, semantics and position in the target field structure, merge the fields and mark empty spaces to obtain a field candidate set.

[0043] S2.1: Extract semantic elements from semantic fragment data, and perform semantic element identification, fragment source location and context location correspondence to obtain a set of semantic elements.

[0044] Furthermore, the semantic fragment data is read segment by segment according to the fragment number, and the text content, start character position, end character position, paragraph position, and context boundary of each semantic fragment are extracted. Each semantic fragment is then segmented, labeled with parts of speech, named entity recognition, dependency parsing, and semantic role recognition are performed to identify words related to field names, entity objects, attribute words, numerical words, time words, location words, action words, and relation trigger words. The identified words are then merged with the semantic role recognition results to form semantic elements.

[0045] Based on the starting character position, ending character position, and paragraph position recorded in the semantic fragment data, the fragment source position is written for each semantic element. Then, based on the context boundary, semantic role, and dependency syntax relationship, the context position correspondence is established for semantic elements belonging to the same field expression within the same context boundary. Finally, deduplication, merging, and sorting are performed according to the semantic element content, semantic role, fragment source position, and context position correspondence to obtain a set of semantic elements.

[0046] S2.2: Perform field name matching, field semantic correspondence, and field position filling on the semantic element set and the target field structure to obtain field filling information.

[0047] Furthermore, the field names, semantics, positions, and filling order in the target field structure are read to form a sequence of fields to be filled. Then, the semantic elements, semantic roles, fragment source positions, and context positions in the semantic element set are read from front to back according to the fragment source position. After the text content of the semantic elements is processed by character normalization, it is compared with the field names in the sequence of fields to be filled item by item. Semantic elements with completely identical field names establish a field name matching relationship. Semantic elements with incompletely identical field names but with a field name inclusion relationship continue to be checked in conjunction with the semantic roles and the object types in the field semantics. After the check is consistent, a field name matching relationship is established.

[0048] After completing the field name matching, the context boundaries, semantic roles, and predicate content connecting the semantic elements are checked against the corresponding field semantics in the target field structure. When the semantic element's object, attribute category, and contextual role are consistent with the corresponding field semantics, a field semantic correspondence is established, and the corresponding semantic element is filled into the corresponding field position according to its position in the target field structure. When multiple semantic elements exist in the same field position, they are merged according to the order of their source positions, while simultaneously retaining the field value fragment list, source position list, and context boundary list. When the same semantic element corresponds to multiple field positions, the field position that first satisfies both the field name matching and field semantic correspondence is filled in according to the field filling order in the target field structure.

[0049] Fields in the target field structure that have not been filled with semantic elements retain empty slot markers and record field names, field semantics, field positions, field value fragment lists, source position lists, context boundary lists, and empty slot markers to obtain field filling information. When performing field evidence mapping, original text evidence fragments are extracted item by item from multi-source field candidates according to the source position list, and evidence mapping results are generated separately before field-level merging.

[0050] S2.3: Based on the information entered into the fields, merge and organize the content of similar fields, mark the empty spaces of the target fields that have not been filled, and organize the candidate fields of the target fields that have been filled and marked as empty to obtain a set of candidate fields.

[0051] Furthermore, the field name, field semantics, field position, semantic element content, fragment source position, and context position in the field information are read item by item. First, semantic element content with the same field name, the same field semantics, and the same field position is grouped into the same field merging group. Then, semantic element content with the same field name, the same field semantics, and the same field position is merged into the corresponding field merging group according to the field position order. For the semantic element content in each field merging group, the character format is unified, duplicate content is eliminated, and the source position is sorted. For completely duplicate semantic element content, the one with the fragment source position closest to the target field position is retained. Semantic element content with different content but pointing to the same field semantics is merged according to the original text appearance order to obtain the merged field content of the same type.

[0052] After merging the content of similar fields, the field names, semantics, and positions in the target field structure are checked item by item. Target fields that do not correspond to semantic elements in the field filling information are marked as empty fields, and the field name, semantics, position, and unmatched semantic element status of the empty fields are recorded to obtain empty mark content. The merged content of similar fields and empty mark content are written into the target field structure. Candidates are sorted according to field name, semantics, position, content, fragment source position, context position, and empty mark status. Filled field candidates are generated for the target fields that have been filled, and field candidates to be filled are generated for the target fields marked as empty. The field candidate set is then obtained.

[0053] S3: By using the semantic evidence boundary discrimination model, establish an evidence mapping between each field candidate in the field candidate set and the original text evidence fragment, record the support range, semantic connection path and derivation level of the original text evidence fragment for the field candidate, and obtain the field evidence boundary representation.

[0054] S3.1: Input the field candidate set and semantic fragment data into the semantic evidence boundary discrimination model, perform candidate splitting, original text position correspondence and candidate evidence fragment extraction on the field candidates to obtain the content corresponding to the field evidence.

[0055] Furthermore, after inputting the field candidate set and semantic fragment data into the semantic evidence boundary discrimination model, the semantic evidence boundary discrimination model reads the field candidates one by one according to the field candidate identifier, and splits the field candidates into field name, field semantics, field value content, empty space marker, fragment source location and context boundary, forming candidate split content.

[0056] The semantic evidence boundary discrimination model locates the semantic fragments corresponding to the field candidates in the semantic fragment data based on the source position of the fragments in the candidate split content. It writes the start and end character positions, paragraph positions, and context boundaries of the semantic fragments into the field candidates to complete the original text position correspondence. The semantic evidence boundary discrimination model then uses the field name, field semantics, and field value content as comparison objects to search for semantic fragments containing the same entity, action relationship, numerical content, time condition, and range limitation within the same context boundary. It extracts the hit semantic fragments and semantic fragments that are continuous with the hit semantic fragments and are used to supplement the subject, predicate, object, time, condition, and range as candidate evidence fragments. The field candidate identifier, candidate split content, original text position correspondence, and candidate evidence fragments are organized according to the same field candidate to obtain the corresponding content of the field evidence.

[0057] S3.2: Based on the content corresponding to the field evidence, extract and organize the field candidates, candidate evidence fragments and context boundaries, and identify the coverage of the candidate evidence fragments to the core semantic elements of the field candidates through the semantic evidence boundary discrimination model to obtain semantic support verification information.

[0058] Furthermore, based on the content corresponding to the field evidence, the candidate field identifier, field name, field semantics, field position, candidate field text, candidate evidence fragment, fragment source position, and context boundary are read. The candidate evidence fragments corresponding to the same field candidate are arranged according to the fragment source position. The candidate evidence fragments that cross the context boundary are trimmed according to the marked context boundary, and the fragments that still contain the semantic role related to the field candidate after trimming are retained as candidate evidence fragments to be verified.

[0059] Read the field name, field semantics, and field position in the target field structure, and then combine the semantic role, dependency relationship, and original character position of the semantic fragment in which the field candidate is located to determine the main element, attribute element, numerical element, time element, limiting condition element, and relation element in sequence. When multiple elements of the same type appear, prioritize the element that has a direct dependency relationship with the field name, is located on the same context boundary, and has the closest source position to the field candidate. The remaining elements are retained as supplementary elements.

[0060] The core semantic elements are compared with the candidate evidence fragments to be verified. The coverage relationship includes consistency of original characters, consistency of normalized expression, consistency of synonyms registered in the target field structure, consistency of abbreviation and full name, and consistency of hierarchical concepts. If the correspondence relationship is not met, it is not considered as coverage.

[0061] The semantic connection path starts with the core semantic elements in the candidate evidence fragment and ends with the field candidate. It connects the subject, predicate, object, limiting condition, and referential item in the order of context boundaries. When an adjacent node lacks any of the contents of dependency relationship, referential relationship, or context continuity relationship, it is marked as a path breakpoint. The derivation level is determined according to the number of intermediate semantic roles traversed by the semantic connection path. The first level of derivation is marked as the core semantic element directly covered within the same context boundary. The second level of derivation is marked as the connection through one intermediate semantic role. The levels of derivation through multiple intermediate semantic roles are marked layer by layer according to the connection order. The core semantic element, coverage relationship, semantic connection path, path breakpoint, and derivation level are written into the semantic support verification information.

[0062] The semantic evidence boundary discrimination model determines whether the candidate evidence fragment to be verified covers the main elements and attribute elements of the field candidate. Then, it determines whether the numerical elements, time elements, limiting condition elements, and relational elements form continuous semantic support within the same context boundary. Core semantic elements that are not directly expressed but are taken over by the referents within the context boundary are marked as having been taken over. Core semantic elements that are missing, negated, or inconsistent with the semantics of the field candidate are marked as missing. The semantic support verification information is obtained by organizing the field candidate identifier, candidate evidence fragment identifier, covered core semantic elements, uncovered core semantic elements, taking over mark, missing mark, and corresponding context boundary.

[0063] S3.3: Based on the semantic support verification information, the context boundary of the candidate evidence fragment is matched with the fragment source position of the field candidate through the semantic evidence boundary discrimination model, and the initial semantic connection path from the candidate evidence fragment to the field candidate is organized to obtain the context boundary matching result.

[0064] Furthermore, the candidate field identifier, the source position of the candidate field fragment, the field semantics, the semantic role corresponding to the candidate field, the source position of the candidate evidence fragment, the start position of the candidate evidence fragment, the end position of the candidate evidence fragment, and the context boundary of the candidate evidence fragment are read from the content corresponding to the field evidence. The semantic evidence boundary discrimination model uses the candidate field identifier as the basis to arrange the candidate evidence fragments under the same candidate field according to the source position of the candidate evidence fragment, and matches the context boundary into which the source position of the candidate field fragment falls with the context boundary of the candidate evidence fragment one by one. If the candidate evidence fragment and the candidate field fall into the same context boundary, a same boundary matching mark is written.

[0065] Candidate evidence fragments and field candidates fall into the consecutive context boundaries of the cleaned text content. If the semantic support verification information records that the candidate evidence fragment covers the core semantic elements of the field candidate, a continuous matching mark is written. If there is any situation where there is no matching context boundary between the candidate evidence fragment and the field candidate, or where the semantic support verification information does not record the coverage relationship of the core semantic elements, a boundary interruption mark is written. After writing the same boundary matching mark and the continuous matching mark, the semantic evidence boundary discrimination model takes the semantic role that covers the core semantic elements in the candidate evidence fragment as the starting point and the semantic role corresponding to the field candidate as the ending point. According to the arrangement order of the context boundaries in the cleaned text content, it sequentially connects the semantic roles in the context boundary where the starting point is located, the connecting semantic roles in the adjacent context boundaries, and the ending semantic role to form an initial semantic connection path. The field candidate, candidate evidence fragment, context boundary matching mark, fragment source position correspondence, and initial semantic connection path are written into the context boundary matching result.

[0066] S3.4: Based on semantic support verification information and context boundary matching results, candidate evidence fragments that cover core semantic elements and are in the same context boundary are judged as direct support, candidate evidence fragments that cover core semantic elements and are in continuous context boundaries through the initial semantic connection path are judged as indirect support, and any of the following situations are judged as not support: not covering core semantic elements, discontinuous context boundaries, and interrupted initial semantic connection path. The semantic connection path is then organized according to the judgment results to obtain the field evidence mapping set.

[0067] Furthermore, the core semantic element coverage relationship between field candidates and candidate evidence fragments is read from the semantic support verification information. The core semantic element coverage relationship is formed by the fact that the content pointed to by the field name, the content pointed to by the field semantics, and the content of the source of the field location are all covered by the candidate evidence fragments. The same boundary matching mark, continuous matching mark, boundary interruption mark, fragment source location correspondence and initial semantic connection path are read from the context boundary matching results, and the field candidates and candidate evidence fragments are judged item by item.

[0068] When a candidate evidence fragment has a core semantic element coverage relationship and the context boundary matching result is written to the same boundary matching tag, the support status of the candidate evidence fragment to the field candidate is recorded as direct support, and the text range between the starting position of the candidate evidence fragment and the source position of the field candidate fragment is recorded as the support range.

[0069] The original text location of candidate evidence fragments, the source location of candidate field fragments, and the same boundary matching markers are organized into direct semantic connection paths. When candidate evidence fragments have a core semantic element coverage relationship, the context boundary matching results are written into continuous matching markers, and the initial semantic connection path includes a starting semantic role, a connecting semantic role, and an ending semantic role, the support status of candidate evidence fragments for candidate fields is recorded as indirect support.

[0070] The text range covered by the initial semantic connection path is recorded as the support range, and the initial semantic connection path is retained as an indirect semantic connection path. When the semantic support verification information does not record the coverage relationship of core semantic elements, the context boundary matching result is written into the boundary interruption mark, or the initial semantic connection path lacks any of the starting semantic role, connecting semantic role, and ending semantic role, the support status of the candidate evidence fragment to the field candidate is recorded as unsupported, and no semantic connection path is generated. The field evidence mapping set is obtained by merging according to the correspondence between the field candidate, candidate evidence fragment, support status, support range, semantic connection path, and context boundary matching mark.

[0071] The formula for determining the supporting status of evidence fragments is: ; in, This represents the field candidate index in the field candidate set, used to uniquely identify a field candidate. This indicates the candidate evidence fragment number corresponding to the candidate for the same field, used to uniquely identify the candidate evidence fragment used for supporting relationship verification with the candidate for the field. Indicates the first The candidate evidence fragments for the first The supporting status formed by the candidate fields Indicates direct support. Indicates indirect support. This indicates that it is not supported.

[0072] This represents the core semantic element coverage flag, with a value of 0 or 1. The candidate evidence fragments have already covered the first... When the core semantic elements required for each field candidate are =1, otherwise =0, the core semantic element refers to the semantic role content in the field candidate determined by the content pointed to by the field name, the content pointed to by the field semantics, and the content from which the field position is located. It is used to determine whether the candidate evidence fragment can support the meaning of the field candidate.

[0073] Indicates the first Candidates for the 1st field and the 2nd field The number of context boundaries between candidate evidence fragments, measured in units of the number of boundaries. =0 indicates that the field candidate and the candidate evidence fragment are on the same context boundary, enabling the determination of same boundary support. =1 indicates that the field candidate and the candidate evidence fragment are in a continuous context boundary, which can be used to make indirect support judgment under continuous boundary.

[0074] Indicates the first From the first candidate evidence fragment to the second The number of semantic role nodes in the initial semantic connection path between candidates for each field, measured in terms of the number of nodes. ≥2 indicates that the initial semantic connection path at least includes the starting semantic role corresponding to the candidate evidence fragment and the ending semantic role corresponding to the field candidate, and can form a direct connection. ≥3 indicates that the initial semantic connection path includes a connecting semantic role in addition to the starting semantic role and the ending semantic role, and can form an indirect connection.

[0075] This indicates the remaining situations that do not meet the direct support conditions and the indirect support conditions, including situations where the core semantic elements are not covered, the context boundaries are discontinuous, the initial semantic connection path nodes are insufficient, or the semantic connection path is interrupted.

[0076] It should be noted that the evidence fragment support state determination formula uses core semantic elements to cover the markers. Number of context boundaries and the number of semantic role nodes As a basis for judgment, when =1、 =0 and When ≥2, candidate evidence fragments and candidate fields are on the same context boundary and form a direct semantic connection, which is determined as direct support. =1、 =1 and When the number of candidate evidence fragments is ≥3, the candidate evidence fragments and candidate fields are in a continuous context boundary and form a semantic connection through the connecting semantic role, which is judged as indirect support. The remaining cases are judged as not support.

[0077] This is a flag indicating whether the value is 0 or 1. and All are count values. The evidence fragments supporting the state determination formula do not add up the values ​​of objects with different dimensions, and the dimensions remain consistent.

[0078] The determination of direct support, indirect support, and no support is based on the semantic support verification information and the context boundary matching results. The same boundary matching mark is used to indicate that the candidate evidence fragment and the field candidate are on the same context boundary, and direct support is formed when the candidate evidence fragment covers the core semantic elements at the same time.

[0079] Continuous matching markers are used to indicate that candidate evidence fragments and field candidates are located at the context boundaries of consecutive contexts. They form indirect support when the initial semantic connection path includes a starting semantic role, a connecting semantic role, and an ending semantic role. Any of the following situations will result in no support: boundary interruption marker, core semantic element not covered, or the initial semantic connection path lacking necessary semantic roles.

[0080] S3.5: Through the field evidence mapping set, the semantic evidence boundary discrimination model performs boundary merging, hierarchical labeling and result reorganization on the support range, semantic connection path and inference level of each field candidate to obtain the field evidence boundary representation.

[0081] Furthermore, when the semantic evidence boundary discrimination model performs boundary merging, hierarchical labeling, and result reorganization on the support range, semantic connection path, and inference level of each field candidate, it first reads all candidate evidence fragments corresponding to the same field candidate in the field evidence mapping set, and extracts the fragment source position, context boundary, support status, initial semantic connection path, and start and end positions of the original text evidence fragment for each candidate evidence fragment.

[0082] Using field candidates as the merging object, the original text evidence fragments that directly support the field candidate are taken as the starting fragments of the support range. The original text evidence fragments that indirectly support the field candidate are included in the support range according to the initial semantic connection path. The candidate evidence fragments that do not support the field candidate are retained as content that does not participate in the support range merging. Then, according to the start and end positions of the original text evidence fragments and the context boundaries, the original text evidence fragments that directly support the field candidate and the included original text evidence fragments are merged. The original text evidence fragments that overlap, are connected end to end, or belong to the same context boundary are merged into the support range corresponding to the field candidate.

[0083] The semantic evidence boundary discrimination model determines the inference level based on the number of context boundary intervals between the candidate field and the original text evidence fragment, the completeness of the semantic connection path, and the number of intermediate semantic roles. When the candidate evidence fragment and the candidate field are located on the same context boundary and the candidate evidence fragment covers the core semantic elements of the candidate field, the corresponding support state is marked as direct support and marked as the first level of inference. When the candidate evidence fragment and the candidate field are located within a continuous context boundary and the semantic connection path can sequentially connect the starting core semantic elements, intermediate semantic roles, and the candidate field, the corresponding support state is marked as indirect support, and the inference level is marked layer by layer according to the connection order of the intermediate semantic roles. The level connected by one intermediate semantic role is marked as the second level of inference, and the level connected by multiple intermediate semantic roles continues to be marked according to the actual connection order.

[0084] The field candidate identifier, field name, field semantics, support scope, semantic connection path, derivation level, support status, and content not included in the support scope are reorganized according to the field candidate to obtain the field evidence boundary representation.

[0085] S4: Based on the field evidence boundary representation, the consistency of the supporting range, semantic connection path and inference level in the original text evidence fragments of each field candidate is checked, the evidence gap items corresponding to the field candidate are identified, and an evidence supplement list is generated.

[0086] S4.1: Based on the field evidence boundary representation, the supporting range, semantic connection path and derivation level corresponding to each field candidate are organized to obtain the field evidence verification content.

[0087] Furthermore, based on the field evidence boundary representation, the field name, field semantics, original text evidence fragment, original text position, supporting scope, semantic connection path and derivation level of each field candidate are read item by item according to the field candidate identifier, and the original text evidence fragments corresponding to the same field candidate are arranged in the order of text appearance.

[0088] Using the field candidate identifier as the basis for organization, the supporting scope is bound to the start position, end position and covered semantic roles of the original text evidence fragment, forming a scope correspondence between the field candidate and the original text evidence fragment. Then, the starting evidence fragment, intermediate semantic roles and field candidates in the semantic connection path are connected in the order of context, and the connection status of direct connection and indirect connection is marked between adjacent semantic roles, so that the semantic connection path reflects the support transmission process from the original text evidence fragment to the field candidate.

[0089] The derivation levels are mapped to semantic connection paths segment by segment. Field candidates that come directly from the original text evidence fragments are marked as direct levels, and field candidates that need to be passed through intermediate semantic roles are marked as corresponding transmission levels. Field candidate identifiers, field names, field semantics, original text evidence fragments, original text positions, supporting scope, semantic connection path order, and derivation levels are organized in the same verification unit to obtain field evidence verification content.

[0090] It should be noted that the field evidence verification content is used for the corresponding organization before the consistency verification of the field evidence boundary representation, and is not used to regenerate new derivation level names. The direct level is the first derivation corresponding to the direct support, and the corresponding transmission level is the second derivation and the derivation mark corresponding to the intermediate indirect support.

[0091] When organizing the field evidence verification content, the derivation levels already recorded in the field evidence boundary representation are written into the same verification unit according to the arrangement order of the semantic connection path, so that the field candidates, original evidence fragments, supporting scope, semantic connection path and derivation level maintain the same source and the same naming standard in the field evidence verification content.

[0092] S4.2: Based on the field evidence verification content, perform consistency verification on the supporting relationship, semantic connection relationship and inference hierarchy relationship between each field candidate and the original text evidence fragment to obtain field evidence gap information.

[0093] Furthermore, based on the content of the field evidence verification, the original text evidence fragments, supporting scope, semantic connection paths and derivation levels corresponding to the field candidates are read, with the field candidate identifier as the processing unit, and the verification order between the field candidates and the original text evidence fragments is established according to the field positions in the target field structure.

[0094] In the verification of supporting relationships, the field names and semantics in the field candidates are matched with the semantic elements in the original evidence fragments one by one. The coverage of the core semantic elements of the field candidates by the original evidence fragments is checked. Gaps that do not cover the core semantic elements are recorded as missing direct evidence, gaps that cover the position beyond the support range are recorded as insufficient support range, and uncovered semantic roles are marked as semantic roles to be supplemented.

[0095] In the semantic connection relationship verification, the connection between candidate evidence fragments, context boundaries, semantic roles and field candidates is checked along the semantic connection path. Gaps in the semantic connection path with breaks or discontinuous source positions are recorded as semantic connection interruptions. In the deduction level relationship verification, the deduction level recorded in the field evidence boundary representation is read, and it is checked whether the semantic role to be supplemented is supported by the original text evidence fragment within the record level. Gaps that are not supported by the original text evidence fragment within the record level are recorded as deduction level exceedance.

[0096] By merging and organizing the candidate field identifiers, gap types, original text positions, and semantic roles to be supplemented for fields with missing direct evidence, insufficient supporting scope, broken semantic connections, and inference levels exceeding the corresponding levels, we can obtain field evidence gap information.

[0097] It should be noted that exceeding the derivation level does not refer to a simple increase in the number of derivation levels. Rather, it means that the semantic role to be supplemented cannot obtain original text evidence fragments to support it within the derivation levels and semantic connection paths that have been recorded in the field evidence boundary representation. When verifying the derivation level relationship, first determine the semantic connection path segment that the semantic role to be supplemented should fall into, and then check whether there are original text evidence fragments supporting the semantic role to be supplemented within the derivation level corresponding to the semantic connection path segment. When the semantic role to be supplemented does not fall into the recorded semantic connection path segment, or when the semantic role to be supplemented needs to rely on an unrecorded intermediate semantic role to establish a supporting relationship with the field candidate, the corresponding gap is recorded as exceeding the derivation level.

[0098] S4.3: Based on the field evidence gap information, the candidate field identifier, gap type, original text position, semantic role to be supplemented, and supplementation direction are reorganized to generate an evidence supplementation list.

[0099] Furthermore, the field evidence gap information is grouped according to the field candidate identifier, and the field name, field semantics and field position corresponding to the field candidate identifier are read in each group to form the field candidate positioning content. Then, the gap type is extracted from the field evidence gap information, and the gap type is bound to the field candidate positioning content to determine one of the gap types among the following: lack of direct evidence, insufficient support scope, interrupted semantic connection and excessive inference level.

[0100] The original text positions in the field evidence gap information are extracted and refined into the positions of the original text evidence fragments, context boundaries, and semantic connection path breakpoints. The original text positions are then mapped to gap types. Based on the gap types, semantic roles to be supplemented are extracted. Semantic roles corresponding to missing direct evidence are identified as core semantic elements of the field. Semantic roles corresponding to insufficient support range are identified as semantic elements outside the support range. Semantic roles corresponding to interrupted semantic connections are identified as intermediate semantic roles on both sides of the path breakpoint. Semantic roles corresponding to inference levels exceeding the corresponding levels are identified as near-source semantic roles.

[0101] Based on the semantic role to be supplemented and the original text position, a supplementary direction is generated. The candidate field identifier, gap type, original text position, semantic role to be supplemented, and supplementary direction are merged according to the same candidate field identifier and arranged in order of original text position to generate an evidence supplementary list.

[0102] S5: Based on the evidence supplementation list, retrieve supplementary text fragments, input the supplementary text fragments and field evidence boundary representations into the semantic evidence boundary discrimination model for evidence mapping review and structured verification, divide the field candidates according to the support results of the supplementary text fragments for the evidence gap items, and output the evidence boundary extraction results.

[0103] S5.1: Based on the evidence supplementation list, extract candidate field identifiers, locate the source of the gap, extract the semantic role to be supplemented, and organize the original text position for the evidence gap items to obtain gap organization information.

[0104] Furthermore, the evidence gap items in the evidence supplementation list are read one by one according to the record order of field candidate identifier and original text position. Field candidate identifiers are extracted from each evidence gap item, and the field candidate identifiers are matched with the corresponding field candidates in the field candidate set to determine the field candidate to which the evidence gap item belongs.

[0105] Based on the supporting scope, semantic connection path, and deduction level records corresponding to the evidence gap items in the evidence supplementation list, the specific source location of the evidence gap items is located where the original evidence fragments are insufficiently covered, the semantic connection path is broken, or the deduction level support is insufficient. The source location of the gap is completed. The semantic roles to be supplemented are extracted from the evidence gap items, and the semantic roles to be supplemented are matched with the field names, field semantics, and field positions in the target field structure to determine the semantic content of the fields that need to be supplemented.

[0106] Based on the original text position in the evidence gap item, the fragment source position, context boundary and field candidate position of the original text evidence fragment are sorted out so that the field candidate identifier, gap source, semantic role to be filled and original text position are corresponding under the same evidence gap item, and gap sorting information is obtained.

[0107] S5.2: Based on the gap information and field evidence boundary representation, the evidence gap items are matched with the supporting scope, semantic connection path and deduction level. The gap type is determined according to the matching results. If there is no corresponding original text evidence fragment, it is determined as direct evidence missing. If the semantic role to be supplemented exceeds the supporting scope, it is determined as insufficient supporting scope. If there is a break in the semantic connection path, it is determined as semantic connection interruption. If the semantic role to be supplemented is not supported by evidence within the record level, it is determined as deduction level exceeding the limit. The gap type determination result is obtained.

[0108] Furthermore, each evidence gap item is located according to the field candidate identifier. First, the field candidate identifier, gap source location, semantic role to be filled, and original text position are extracted from the gap information. Then, the original text evidence fragment, supporting range, semantic connection path, and deduction level corresponding to the same field candidate identifier are extracted from the field evidence boundary representation. The semantic role to be filled and the original text position are then placed into the supporting range, semantic connection path, and deduction level of the corresponding field candidate for verification.

[0109] During verification, it is determined whether the candidate field identifier and the original text position can correspond to any original text evidence fragment. If they cannot correspond to any original text evidence fragment, the evidence gap item is written into the direct evidence missing type. If they can correspond to the original text evidence fragment, it is further determined whether the semantic role to be supplemented is within the support range of the field evidence boundary representation record. If the semantic role to be supplemented is not covered by the support range, the evidence gap item is written into the support range insufficient type. If the semantic role to be supplemented is covered by the support range, it is further checked along the semantic connection path to see if the connection nodes between the original text evidence fragment, the semantic role to be supplemented, and the candidate field are continuous. If there are no connection nodes in the semantic connection path, or there is no context boundary correspondence between adjacent connection nodes, the evidence gap item is written into the semantic connection interruption type. If the semantic connection path is continuous, it is further checked whether the semantic role to be supplemented is supported by the original text evidence fragment within the deduction level of the field evidence boundary representation record. If the semantic role to be supplemented is not supported by evidence within the record level, the evidence gap item is written into the deduction level exceeded type.

[0110] Following the judgment order of missing direct evidence, insufficient supporting scope, broken semantic connection, and excessive inference level, only one gap type is written for each evidence gap item. The candidate field identifier, gap source location, semantic role to be supplemented, original text position, and written gap type are organized to obtain the gap type judgment result.

[0111] S5.3: Based on the gap type determination result, the missing direct evidence is matched with the direct evidence retrieval object of the same field, the insufficient support range is matched with the adjacent context evidence retrieval object, the interrupted semantic connection is matched with the intermediate semantic role evidence retrieval object, and the inference level exceeds the limit is matched with the near source evidence retrieval object. The field candidate, the semantic role to be supplemented, the original text position and the corresponding evidence retrieval object are reorganized into the field supplementation object.

[0112] Furthermore, based on the gap type determination results, each evidence gap item is read according to the order of the field candidate identifiers in the evidence supplementation list, and the field candidate, semantic role to be supplemented, original text position and gap type are extracted from the gap sorting information. The gap type is then correlated with the support range, semantic connection path and deduction level corresponding to the same field candidate in the field evidence boundary representation, so that each evidence gap item corresponds to a clear evidence boundary position.

[0113] When the gap type is missing direct evidence, the content to be obtained is limited by the field name, field semantics, field position, semantic role to be supplemented, and original text position of the field candidate. A direct evidence retrieval object of the same field is generated. The direct evidence retrieval object of the same field is used to obtain supplementary text fragments that belong to the same field name and field semantics as the field candidate and directly cover the semantic role to be supplemented.

[0114] When the gap type is insufficient support range, the starting and ending positions of the support range recorded in the field evidence boundary representation are used as the benchmark to lock the context boundary adjacent to the support range, and the consecutive occurrence positions of the semantic role to be supplemented in the adjacent context boundary are written into the adjacent context evidence retrieval object.

[0115] When the gap type is semantic connection interruption, the semantic role before the breakpoint, the semantic role after the breakpoint, and the corresponding fragment source position are read from the semantic connection path. The intermediate semantic role located between the semantic role before the breakpoint and the semantic role after the breakpoint, and the intermediate semantic role between the candidate connection field and the semantic role to be supplemented, is written into the intermediate semantic role evidence retrieval object.

[0116] When the gap type is "exceeding the derivation level", the original text evidence fragment that is adjacent to the field candidate in the semantic connection path and has formed a supporting state is used as the near-source benchmark fragment. The fragment source position, corresponding semantic role and semantic role to be supplemented of the near-source benchmark fragment are written into the near-source evidence retrieval object.

[0117] The candidate field identifier, candidate field, semantic role to be supplemented, original text position, gap type, corresponding evidence retrieval object, and evidence boundary position are merged according to the same candidate field identifier to generate a field supplementation object.

[0118] The formula for generating evidence gap types and field supplementation objects is as follows: ; ; in, Indicates field candidate The corresponding set of valid candidate evidence fragments consists of candidate evidence fragments whose supporting status is either direct or indirect. Indicates field candidate The set of core semantic elements that need to be covered This represents the set of core semantic elements that have been covered by valid candidate evidence fragments; both sides are sets of semantic roles. Indicates field candidate The corresponding initial set of semantic connection paths, This represents the set of path breakpoints in the initial set of semantic connection paths. Indicates field candidate The corresponding set of semantic roles to be supplemented. This represents the set of semantic roles to be supplemented that have already been supported by fragments of original text evidence within the derivation level of the record representing the field evidence boundary. Indicates field candidate The corresponding gap type, This indicates a lack of direct evidence. This indicates insufficient support range. This indicates a semantic connection interruption. This indicates that the derivation level has exceeded the limit. Indicates field candidate The corresponding position in the original text is measured in terms of character position number. Indication and gap type The corresponding evidence retrieval object, Indicates field candidate The corresponding field is appended to the object.

[0119] In the formulas for generating evidence gap types and field supplementation objects, set operations occur within the semantic role set or semantic connection path set. The original text position is only used as a position record field in the field supplementation object and is not numerically mixed with the number of semantic roles, thus maintaining consistent dimensions.

[0120] It should be noted that when a field candidate corresponds to a valid original text evidence fragment, all core semantic elements are covered, there are no breaks in the semantic connection path, and the semantic roles to be supplemented are all supported within the record deduction level, the field candidate does not meet any of the gap types among direct evidence missing, insufficient support scope, semantic connection interruption, and deduction level exceeding, and no field supplementation object is generated. It is only included in the evidence boundary extraction result as a field candidate with completed evidence support.

[0121] The evidence boundary position indicates the location where the evidence gap item occurs in the field evidence boundary representation. The evidence boundary position is jointly determined by the support range boundary position, the semantic connection path breakpoint position, and the corresponding position of the derivation level. The lack of direct evidence corresponds to the location where the original text evidence fragment is missing, the insufficient support range corresponds to the location outside the boundary of the support range, the semantic connection interruption corresponds to the breakpoint position in the semantic connection path, and the derivation level exceeds the derivation level position where the corresponding semantic role to be supplemented has not been supported. The evidence boundary position is written into the field supplementation object.

[0122] S5.4: Based on the field supplementation object, extract the supplementary text fragments corresponding to the field candidate, the semantic role to be supplemented, and the original text position, and combine them with the surrounding text content of the original text evidence fragments to perform semantic slicing and supporting relationship marking to obtain supplementary evidence fragments.

[0123] Furthermore, the corresponding evidence retrieval objects in the field candidates, semantic roles to be supplemented, original text positions, and field supplementation objects are read. The original text evidence fragments are first located by the original text position, and the semantic fragments adjacent to the original text evidence fragments are obtained as the surrounding text content of the original text evidence fragments.

[0124] The scope of supplementation is determined according to the corresponding evidence retrieval object in the supplementation object. The retrieval scope of supplementary text fragments is limited to the text to be extracted, supplementary text from the same business source, and attachment text that has been included in this extraction task. Evidence is not extracted from external materials from unregistered sources.

[0125] For direct evidence retrieval objects within the same field, the search retrieves text segments containing the semantic elements corresponding to the field candidate and the semantic roles to be supplemented. For adjacent contextual evidence retrieval objects, the search retrieves text segments that inherit the semantics of the field candidate within the three basic sentence segments before and after the context boundary of the original text evidence segment. For intermediate semantic role evidence retrieval objects, the search retrieves intermediate semantic segments that can connect the field candidate and the semantic roles to be supplemented. For near-source evidence retrieval objects, the search retrieves text segments that have the same source as the original text evidence segment and are closest in text order.

[0126] When determining supplementary text fragments, the shortest continuous text content that can form a complete subject, predicate, object, time condition, and limiting condition is extracted, using the start and end positions covering the semantic role to be supplemented as boundaries. The source position, context boundary, and field candidate identifier are recorded. When multiple candidate supplementary text fragments exist, they are sorted according to the completeness of the semantic role to be supplemented, the continuity of the semantic connection path, the proximity of the source position, and the simplicity of the text range. When the contents of multiple candidate supplementary text fragments conflict with each other, the text segment that is consistent with the source of the original evidence fragment, has a continuous context boundary, and has a complete semantic role is retained first, and the conflicting content is marked as content to be reviewed. When no supplementary text fragment is found, the original evidence gap item is retained, and no supplementary evidence fragment is generated.

[0127] The located text segment is merged with the surrounding text of the original evidence segment. Statements unrelated to the field candidate are deleted, and the shortest continuous text content containing the semantic role to be supplemented and establishing a semantic connection with the field candidate is retained to obtain the supplementary text segment. The supplementary text segment is then semantically sliced ​​according to sentence boundaries, paragraph boundaries, and semantic role boundaries to obtain the supplementary semantic segment.

[0128] The supplementary semantic fragments are checked against the field candidates, the semantic roles to be supplemented, and the original text positions respectively. When the supplementary semantic fragments cover the semantic roles to be supplemented and can be connected to the semantic connection path of the original text evidence fragments, a supporting relationship mark is written. The supplementary semantic fragments with field candidate marks, semantic roles to be supplemented, original text positions, and supporting relationship marks are organized into supplementary evidence fragments.

[0129] It should be noted that statements unrelated to field candidates refer to statements that do not contain content corresponding to field names, content corresponding to field semantics, or semantic roles to be supplemented, and cannot be connected to the semantic connection path of the original text evidence fragment. The shortest continuous text content is not arbitrarily truncated text, but continuous text content that is retained in the order of appearance in the original text, starting from the semantic fragment that first covers the semantic role to be supplemented and ending at the semantic fragment that connects to the semantic connection path of the original text evidence fragment. The continuous text content also retains the correspondence between field candidates, semantic roles to be supplemented, and positions in the original text.

[0130] S5.5: By supplementing evidence fragments and field evidence boundary representations, the field candidates are subjected to evidence mapping verification, derivation level update and structured verification, and the evidence boundary extraction results are output.

[0131] Furthermore, the candidate field identifiers, semantic roles to be supplemented, source locations of the fragments, context boundaries, and supporting relationship markers in the supplementary evidence fragments are read. The candidate field identifiers are matched with the candidate fields recorded in the field evidence boundary representation, the semantic roles to be supplemented are matched with the evidence gap items recorded in the field evidence boundary representation, and the source locations and context boundaries of the fragments are matched with the supporting range of the original evidence fragments.

[0132] The corresponding supplementary evidence fragments are input into the semantic evidence boundary discrimination model. The semantic evidence boundary discrimination model reviews the core semantic element coverage relationship, context boundary continuity relationship, and semantic connection path continuity relationship between the supplementary evidence fragments and the field candidates. When the supplementary evidence fragment covers the semantic role to be supplemented, the context boundary of the supplementary evidence fragment is continuous with the context boundary of the original evidence fragment, and the supplementary evidence fragment continues the semantic connection path, the evidence mapping relationship between the supplementary evidence fragments and the field candidates is added to the field evidence boundary representation.

[0133] Based on the derivation level corresponding to the original text evidence fragment in the field evidence boundary representation, the supplementary evidence fragment is connected to the interruption position and the position of the semantic role to be supplemented in the semantic connection path. The derivation level is re-marked according to the semantic continuation order between the original text evidence fragment, the supplementary evidence fragment, and the field candidate to obtain the updated field evidence boundary representation.

[0134] Based on the updated field evidence boundary representation, the field name, field semantics, field position, gap marker, support status, gap type, original evidence fragment, supplementary evidence fragment, support range, semantic connection path, and deduction level of the candidate fields are structurally verified. The candidate fields are then classified and organized. Field names, field semantics, and field positions are consistent, core semantic elements are all covered by original evidence fragments or supplementary evidence fragments, there are no breakpoints in the semantic connection path, and the deduction level is within the record range and there are no more evidence gaps. These are determined to be output field candidates.

[0135] If some supplementary evidence fragments have been obtained to support the field, but there are still source conflicts, unconfirmed support scope, unverified semantic connection paths, or unconfirmed derivation levels, the field is identified as a candidate field to be reviewed. If the target field has not been filled with effective semantic elements, or if no supplementary evidence fragments have been obtained after supplementing the evidence, the field is retained as a candidate field for vacancy. If the candidate evidence fragment is judged to be unsupported, or if there is semantic negation, evidence conflicts that cannot be merged, semantic connection paths that are interrupted and cannot be supplemented, or derivation levels that are exceeded and have no near-source evidence to support the field, the field is identified as a candidate field to be rejected for output.

[0136] The evidence is organized according to the correspondence between field candidates, classification categories, original evidence fragments, supplementary evidence fragments, supporting scope, semantic connection path, inference level, gap type, and empty space marker, and the results of evidence boundary extraction are output.

[0137] This embodiment also provides an AI-based intelligent information extraction system, including: a text normalization module, used to acquire the text to be extracted and the target field structure, perform text cleaning, semantic slicing, context boundary marking, and semantic role recognition on the text to be extracted to obtain semantic fragment data, and load a semantic evidence boundary discrimination model; a field candidate module, used to extract semantic elements from the semantic fragment data, fill in the semantic elements according to the field name, semantics, and position in the target field structure, merge fields, and mark empty spaces to obtain a field candidate set; and an evidence discrimination module, used to establish an evidence mapping between each field candidate in the field candidate set and the original text evidence fragment through the semantic evidence boundary discrimination model, and record... The original text evidence fragments support the field candidate's scope, semantic connection path, and derivation level to obtain the field evidence boundary representation. The supplementary evidence module is used to perform consistency checks on the support scope, semantic connection path, and derivation level of the original text evidence fragments for each field candidate based on the field evidence boundary representation, identify the evidence gap items corresponding to the field candidate, and generate an evidence supplement list. The extraction and output module is used to retrieve supplementary text fragments based on the evidence supplement list, input the supplementary text fragments and field evidence boundary representations into the semantic evidence boundary discrimination model for evidence mapping review and structured verification, divide the field candidate according to the support results of the supplementary text fragments for the evidence gap items, and output the evidence boundary extraction results.

[0138] In summary, this invention establishes a mapping between candidate fields and original text evidence fragments through a semantic evidence boundary discrimination model, records the support range, semantic connection path, and derivation level, locates the evidence gap item to a specific field, original text position, and semantic role to be supplemented, and generates field supplementation objects for verification, thereby improving the traceability, structural integrity, and output stability of the extraction results.

[0139] It should be noted that the above embodiments are only used to illustrate the technical solutions 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 solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-based intelligent information extraction method, characterized in that, include: The process involves obtaining the text to be extracted and the target field structure, performing text cleaning, semantic slicing, context boundary labeling, and semantic role recognition on the text to be extracted to obtain semantic fragment data, and then loading the semantic evidence boundary discrimination model. The specific construction process of the semantic evidence boundary discrimination model is as follows: A semantic evidence boundary discrimination model is constructed based on a field fragment alignment layer, an evidence support determination layer, and an evidence boundary attribution layer. The field fragment alignment layer extracts the fragment source location, context boundary, semantic role, field name, field semantics and field location from the semantic fragment data, and aligns the source location, context boundary and field affiliation between field candidates and candidate evidence fragments to form field fragment alignment information. The evidence support determination layer, based on field fragment alignment information, determines the support status of the core semantic element coverage relationship, context boundary correspondence relationship and semantic connection path between candidate evidence fragments and field candidates, and determines whether the candidate evidence fragments provide direct support, indirect support or no support to the field candidates, thus forming a field support determination list. The evidence boundary attribution layer, based on the field support judgment list, performs boundary merging, level marking and evidence gap attribution judgment on the support range, semantic connection path and derivation level corresponding to each field candidate, to form evidence boundary attribution content; Extract semantic elements from semantic fragment data, fill in the semantic elements according to the field name, semantics and position in the target field structure, merge the fields and mark empty spaces to obtain a field candidate set; By using a semantic evidence boundary discrimination model, each field candidate in the field candidate set is mapped to the original text evidence fragment, and the support range, semantic connection path and derivation level of the original text evidence fragment for the field candidate are recorded to obtain the field evidence boundary representation. Based on the field evidence boundary representation, the consistency of the supporting range, semantic connection path and derivation level in the original text evidence fragments of each field candidate is checked, the evidence gap items corresponding to the field candidate are identified, and an evidence supplementation list is generated. Based on the evidence supplementation list, supplementary text fragments are retrieved. The supplementary text fragments and field evidence boundary representations are input into the semantic evidence boundary discrimination model for evidence mapping verification and structured validation. Based on the support results of the supplementary text fragments for the evidence gap items, the field candidates are divided, and the evidence boundary extraction results are output. The specific steps are as follows: Based on the evidence supplementation list and the field evidence boundary representation, the evidence gap items are checked against the supporting scope, semantic connection path and deduction level. They are classified into four types of supplementation: direct evidence missing, insufficient supporting scope, semantic connection interruption and deduction level exceeding the scope of deduction, and field supplementation objects are generated. Based on the field supplementation object, extract the supplementary text fragments corresponding to the field candidate, the semantic role to be supplemented, and the original text position, and combine them with the surrounding text content of the original text evidence fragments to perform semantic slicing and supporting relationship marking to obtain the supplementary evidence fragments; By supplementing evidence fragments and representing field evidence boundaries, the candidate fields are subjected to evidence mapping verification, derivation level update and structured verification, and the evidence boundary extraction results are output.

2. The AI-based intelligent information extraction method as described in claim 1, characterized in that, The steps for obtaining the text to be extracted and the target field structure, performing text cleaning, semantic slicing, context boundary labeling, and semantic role recognition on the text to be extracted to obtain semantic fragment data, and loading the semantic evidence boundary discrimination model are as follows: The extracted text is processed by text normalization and content filtering to obtain cleaned text content; The cleaned text content is semantically sliced, contextual boundary labeled, and semantic role identified to obtain semantic fragment data, and the semantic evidence boundary discrimination model is loaded.

3. The AI-based intelligent information extraction method as described in claim 1, characterized in that, The specific steps for obtaining the candidate set of fields are as follows: Extract semantic elements from semantic fragment data, and perform semantic element identification, fragment source location and context location correspondence to obtain a set of semantic elements; The semantic element set is matched with the target field structure by field name matching, field semantic correspondence, and field position filling to obtain field filling information; By filling in information in the fields, the content of similar fields is merged and organized, empty target fields are marked, and candidate fields that have been filled and marked are organized to obtain a set of candidate fields.

4. The AI-based intelligent information extraction method as described in claim 1, characterized in that, The specific steps for obtaining the field evidence boundary representation are as follows: The field candidate set and semantic fragment data are input into the semantic evidence boundary discrimination model. The field candidates are split into candidates, correspond to the original text position, and extracted from the candidate evidence fragments to obtain the content corresponding to the field evidence. The semantic evidence boundary discrimination model is used to perform semantic support verification and context boundary matching between the content corresponding to the field evidence and the candidate evidence fragments. It determines whether the candidate evidence fragments support the field candidate directly, indirectly, or not. The semantic connection paths are organized according to the support status to obtain the field evidence mapping set. By using the field evidence mapping set, the semantic evidence boundary discrimination model performs boundary merging, hierarchical labeling, and result reorganization on the support range, semantic connection path, and inference level of each field candidate, thereby obtaining the field evidence boundary representation.

5. The AI-based intelligent information extraction method as described in claim 4, characterized in that, The specific steps for obtaining the field evidence mapping set are as follows: Based on the content corresponding to the field evidence, the field candidates, candidate evidence fragments and context boundaries are extracted and organized, and the semantic evidence boundary discrimination model is used to identify the coverage of the candidate evidence fragments on the core semantic elements of the field candidates, so as to obtain semantic support verification information. Based on the semantic support verification information, the context boundary of the candidate evidence fragment is matched with the fragment source position of the field candidate through the semantic evidence boundary discrimination model, and the initial semantic connection path from the candidate evidence fragment to the field candidate is sorted out to obtain the context boundary matching result. Based on semantic support verification information and context boundary matching results, candidate evidence fragments that cover core semantic elements and are in the same context boundary are judged as direct support, candidate evidence fragments that cover core semantic elements and are in continuous context boundaries through the initial semantic connection path are judged as indirect support, and any of the following situations are judged as not support: not covering core semantic elements, discontinuous context boundaries, and interrupted initial semantic connection path. The semantic connection path is then organized according to the judgment results to obtain a set of field evidence mappings.

6. The AI-based intelligent information extraction method as described in claim 1, characterized in that, The process based on field-based evidence boundary representation involves a consistency check of the supporting range, semantic connection path, and derivation level in the original text evidence fragments for each field candidate, identifying the evidence gap items corresponding to the field candidate, and generating an evidence supplementation list. The specific steps are as follows: Based on the field evidence boundary representation, the supporting range, semantic connection path and derivation level corresponding to each field candidate are organized to obtain the field evidence verification content; Based on the content of the field evidence verification, the consistency of the supporting relationship, semantic connection relationship and inference hierarchy relationship between each field candidate and the original text evidence fragment is verified to obtain field evidence gap information; By using field evidence gap information, the candidate field identifier, gap type, original text location, semantic role to be supplemented, and direction of supplementation are reorganized to generate an evidence supplementation list.

7. The AI-based intelligent information extraction method as described in claim 1, characterized in that, The specific steps for generating the field appending object are as follows: Based on the evidence supplementation list, candidate field identifiers are extracted, the source of the gap is located, the semantic role to be supplemented is extracted, and the original text position is organized to obtain gap organization information; Based on the gap information and field evidence boundary representation, the evidence gap items are checked against the supporting scope, semantic connection path and inference level. The gap type is determined according to the check results. If there is no corresponding original text evidence fragment, it is determined as direct evidence missing. If the semantic role to be supplemented exceeds the supporting scope, it is determined as insufficient supporting scope. If there is a break in the semantic connection path, it is determined as semantic connection interruption. If the semantic role to be supplemented is not supported by evidence within the record level, it is determined as inference level exceeding the limit. The gap type determination result is obtained. Based on the gap type determination results, the missing direct evidence is matched with the direct evidence retrieval object of the same field, the insufficient support range is matched with the adjacent context evidence retrieval object, the interrupted semantic connection is matched with the intermediate semantic role evidence retrieval object, and the inference level exceeds the limit is matched with the near source evidence retrieval object. The field candidate, the semantic role to be supplemented, the original text position and the corresponding evidence retrieval object are reorganized into the field supplementation object.

8. An AI-based intelligent information extraction system, based on the AI-based intelligent information extraction method according to any one of claims 1 to 7, characterized in that, include: The text preparation module is used to obtain the text to be extracted and the target field structure, perform text cleaning, semantic slicing, context boundary labeling and semantic role recognition on the text to be extracted, obtain semantic fragment data, and load the semantic evidence boundary discrimination model; The field candidate module is used to extract semantic elements from semantic fragment data, fill in the semantic elements according to the field name, semantics and position in the target field structure, merge the fields and mark empty spaces to obtain a field candidate set; The evidence discrimination module is used to establish an evidence mapping between each field candidate in the field candidate set and the original text evidence fragment through the semantic evidence boundary discrimination model, record the support range, semantic connection path and derivation level of the original text evidence fragment for the field candidate, and obtain the field evidence boundary representation. The supplementary evidence module is used to perform consistency checks on the supporting range, semantic connection path and inference level of the original text evidence fragments for each field candidate based on the field evidence boundary representation, identify the evidence gap items corresponding to the field candidate, and generate an evidence supplementary list; The extraction output module is used to retrieve supplementary text fragments based on the evidence supplementation list, input the supplementary text fragments and field evidence boundary representations into the semantic evidence boundary discrimination model for evidence mapping verification and structured validation, divide the field candidates according to the support results of the supplementary text fragments for the evidence gap items, and output the evidence boundary extraction results.

Images