Intelligent document review method and system based on large model fusion causal reasoning

Abstract

The invention relates to the technical field of intelligent document review, in particular to an intelligent document review method and system based on large model fusion causal reasoning, and the method comprises the steps: collecting documents, extracting formats and null lines, recognizing roles and actions, retrieving nouns after verbs, processing semantic repetition, outputting a table, checking causal connections, comparing action sequences, screening and joining, ranking verbs, and pulling through actions. And outputting causal fragments, checking consistency, separating objects, calibrating behaviors, outputting chain fracture, splitting adverbs, sorting relationships, and outputting conflict paths and review paths. According to the method and the device, by extracting the paragraph formats, roles and actions of the documents, the meticulousness of review is improved. And analyzing a causal conjunction and an action relationship, and identifying a causal relationship and a semantic conflict. The method comprises the steps of calibrating an action sequence, identifying missing parts, sorting multiple actions and role changes, optimizing an examination path, reducing manual intervention and improving efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent document review technology, and in particular to an intelligent document review method and system based on large model fusion causal reasoning. Background Technology

[0002] The field of intelligent document review technology refers to the automatic analysis, classification, review, and processing of documents by combining natural language processing, machine learning, and deep learning technologies. Its core includes text structure parsing, information extraction, semantic understanding, and compliance review. With the development of digitalization, intelligent document review is widely used in fields such as law, finance, and contract management, aiming to improve processing efficiency and reduce the risk of errors.

[0003] Traditional intelligent document review methods and systems based on large-scale models and causal reasoning combine large-scale models with causal reasoning to conduct intelligent review by analyzing document content and its inherent relationships. Large-scale models extract document features, and causal reasoning is used to infer causal relationships in the text to achieve information verification and automated review. Unlike traditional review methods based on rule matching or statistical models, large-scale model-based solutions can effectively capture deep semantic and causal relationships, improving the accuracy and intelligence of the review process.

[0004] Existing technologies rely on large models and causal reasoning for text analysis, but they cannot effectively capture subtle semantic changes and multiple causal relationships in complex documents. Traditional methods have limitations in identifying roles and actions, making it difficult to accurately locate potential relationships and behavioral repetitions in complex texts, leading to omissions or misjudgments in review results. Furthermore, existing technologies cannot flexibly adjust the text context when processing documents with multiple roles and actions, resulting in incomplete review results in complex structured or unstructured texts. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide an intelligent document review method and system based on large-model fusion causal reasoning. The technical solution is as follows:

[0006] The intelligent document review method based on large-scale model fusion causal reasoning includes the following steps:

[0007] S1: Based on all paragraphs of the collected document, extract the first line format and blank line spacing of each paragraph in turn, extract the sentences containing words with roles and their corresponding actions one by one, search for the first object noun in the phrase containing action verbs, use the role correspondence table to process semantic repetition, merge the results according to paragraph position, and output the clause behavior positioning table.

[0008] S2: Based on the aforementioned clause behavior positioning table, sequentially retrieve paragraphs containing causal connectors, compare the order of appearance of action words, screen whether roles and actions are continuously connected, group the segments with no more than three lines between verb phrases, connect continuous text actions with corresponding positions, and output causal relationship segment sequence groups.

[0009] S3: Based on the causal relationship fragment sequence group, extract the action word and role combination of each segment, check the consistency of role usage between the first and last sentences, separate the action expression of words without connected objects, calibrate the relative positions of the same behavior multiple times but with different objects, confirm the missing parts according to the paragraph order, and output the expression chain break record set;

[0010] S4: Extract the record set of broken expression chains, extract repeated action expression statements of the same role, call the list of role behaviors according to the distance between the beginning and end of the text, match the action with the root words of the aforementioned terms, split the collocation of negative adverbs and action words, sort out the subject-object structure connection relationship, and the result is a list of semantic conflict paths;

[0011] S5: Based on the semantic conflict path list and the expression chain break record set, separate the action repetition and role change segments, organize the sentences containing result words at the intersection of role and verb, arrange the verb order of multi-action expression segments, filter out unconnected subject or adverb segments, and the result is the intelligent review task path set.

[0012] As a further aspect of the present invention, the clause behavior positioning table includes paragraph first-line format, blank line spacing, role words, action words, object nouns, semantic repetition, and paragraph position identifiers; the causal relationship segment sequence group includes causal connectors, action word order, role and action connection, verb phrase intervals, and continuous action text; the expression chain break record set includes role and action combination, first and last sentence role consistency, unconnected objects, inconsistent action objects, position calibration, and missing part identifiers; the semantic conflict path list includes repeated actions, role behavior list, root word matching, use of negative adverbs, and subject-object relationship analysis; the intelligent review task path set includes action repetition, role change, result word extraction, verb intersection identifiers, action arrangement order, and paragraphs without connected subject-adverb phrases.

[0013] As a further aspect of the present invention, the step of obtaining the clause behavior positioning table is as follows:

[0014] S101: Obtain the document paragraph content and simultaneously obtain the first-line format item and blank line spacing item of each paragraph. Call the paragraph content and format item to perform a check action on the paragraph position order. Record the corresponding role words and their action verb phrases that appear in the paragraph content and establish a sequence relationship with the paragraph number to generate the paragraph action structure.

[0015] S102: Based on the paragraph action structure, call the position of the action verb phrase to perform the search action on the subsequent word groups in sequence, confirm the first noun in the search path with the phrase execution object and record the sequence structure between the character words, action phrases and nouns, and generate the character action object chain;

[0016] S103: Based on the character action object chain, call the character correspondence table to perform a comparison action on the repeated expression items, merge the items that have been matched with the original paragraph position number and keep the original paragraph order, and obtain the clause behavior positioning table.

[0017] As a further aspect of the present invention, the step of obtaining the causal association fragment sequence group is as follows:

[0018] S201: Based on the clause behavior positioning table, detect the text fields involving causal conjunctions, perform a comparison action on the order of occurrence of action words in these text fields, perform a screening operation on the connection relationship between role words and action words in adjacent sentences and record the order difference items, and generate the action connection order.

[0019] S202: Call the action connection order to perform a detection action on the line order of each action word segment, group the segments with an interval of no more than three lines between adjacent action words and record the first and last line order of the segments, perform a judgment operation on the order of character words and action words in the grouped segments and establish a continuous structure to obtain the continuous action interval degree;

[0020] S203: Based on the continuous action interval degree, the sequence of the classified fragments and the original paragraph position number are combined, the combined structure is arranged according to the order of text appearance and the paragraph position of the causal connector is recorded, the arrangement result is established as a sequence group, and the causal association fragment sequence group is obtained.

[0021] As a further aspect of the present invention, the step of obtaining the expression chain break record set is as follows:

[0022] S301: Based on the causal association segment sequence group, extract the action words and role combinations of each segment, call the role words in the first and last sentences of the segment to perform a comparison action on their usage, record the role usages and corresponding action words that show differences after comparison as difference items and establish a corresponding relationship with the segment number to generate a role consistency order;

[0023] S302: Detect the position of the action word according to the role consistency degree, separate the expression fragments that are not followed by object words after the action word and record their separation point order, and at the same time, perform relative position judgment on the expressions that appear in different fragments and have different objects and record the relative position difference value to obtain the action object breakpoint degree.

[0024] S303: Call the action object breakpoint degree to perform calibration actions on the separated segments and relative position difference items in the order of text paragraphs, establish a sequence structure with the missing object words or missing role usage parts and the original paragraph order, and record the segment continuity change degree to establish a record set, thus obtaining the expression chain break record set.

[0025] As a further aspect of the present invention, the step of obtaining the semantic conflict path list is as follows:

[0026] S401: Based on the expression chain break record set, extract the expression statements involving repeated actions of the same character, call the statements in the text to perform a search action on the character behavior list, match the action roots in the expression statements with the behavior terms in the behavior list and record the matching differences, establish a correspondence between the differences and the statement position number, and generate action root matching groups;

[0027] S402: Based on the action root matching group, perform a detection action on the position of the negation adverb in the statement, split the negation adverb and the subsequent action words and record the split items, and at the same time perform a one-way judgment on the relationship between the role and the object in the statement connection structure and record the relationship direction item to obtain the role relationship structure group;

[0028] S403: Call the role relationship structure group to perform a combination action on the matching difference items and relationship direction items in the sentence fragment, arrange the combined structure in text order and record the conflict association items, establish the arranged content as a path list, and obtain the semantic conflict path list.

[0029] As a further aspect of the present invention, the step of obtaining the intelligent review task path set is as follows:

[0030] S501: Based on the semantic conflict path list and expression chain break record set, take relevant content, monitor the action repetition and role change segments, perform sorting action on the sentences containing result words in these segments according to the role position and verb intersection, establish a correspondence between the sorted sentence structure and the original text line order, and generate role verb intersection sequence;

[0031] S502: Call the character verb cross sequence to perform judgment actions on expression segments containing multiple actions, arrange the order of verbs according to line position and record the arrangement order item, and at the same time perform screening processing on paragraphs where there is no subject or adverb connecting verb actions and record the screening item to obtain action connection sorting group;

[0032] S503: Based on the action connection sorting group, perform combined actions on the sorted segments and the filtered items, build the structure of the combined content segment by segment according to the text paragraph order, record the action and role arrangement changes, organize these structures into a task path set, and obtain the intelligent review task path set.

[0033] An intelligent document review system based on large-scale model fusion causal reasoning, the system comprising:

[0034] The collection and extraction module collects all paragraphs in the document, extracts the first line format and blank line spacing of each paragraph, extracts sentences with role words and their corresponding actions one by one, retrieves the first object noun for phrases containing action verbs, uses a role correspondence table to handle semantic repetition, merges the results according to the original paragraph position, and outputs a clause behavior positioning table.

[0035] The causal association module sequentially retrieves text fields involving causal connection words from the clause behavior positioning table, compares the order of occurrence of action words, performs screening operations to determine whether the roles and actions in the text are connected continuously, groups segments with no more than three lines between verb phrases, connects continuous text actions with corresponding positions, and outputs a causal association segment sequence group.

[0036] The consistency check module extracts the action words and role combinations of each segment based on the associated segment sequence group, checks the consistency of role usage between the first and last sentences, separates the expression of the action without the object word, performs relative position calibration for the same behavior that appears multiple times in the text but with different objects, confirms the missing parts in the order of the text segments in which they appear, and outputs a record set of expression chain breaks.

[0037] The semantic conflict analysis module extracts the expression statements involving repeated actions of the same character from the expression chain break record set, calls the character behavior list according to the distance between them in the text, performs root matching between the expression action and the above behavior terms, performs splitting of the position of the negative adverb in the sentence and the collocation of the behavior word, performs unidirectional sorting of the subject-object relationship in the sentence connection structure, and outputs a list of semantic conflict paths.

[0038] The intelligent review module extracts relevant content from the semantic conflict path list and the expression chain break record set, separates the segments involving repeated actions and role changes, organizes the sentences containing result words according to their intersection with the role and verb, arranges the expression segments containing multiple actions according to the order of the verbs, and filters the segments where the verb actions are not connected by subjects or adverbs, and outputs the intelligent review task path set.

[0039] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0040] This invention improves the accuracy and intelligence of document review by accurately identifying document paragraph structure, roles, and actions. Extracting paragraph format, role-related words, and actions enhances the detail of text parsing, providing a solid foundation for subsequent causal relationship identification. By comparing causal connectors, actions, and role order in the text, the system can accurately capture potential causal relationships and semantic conflicts, ensuring the integrity and consistency of information. When processing the relationship between actions and objects, the accuracy of review is further improved by calibrating the action order and identifying missing parts. Furthermore, the organization and filtering of multiple actions and role changes optimizes the review path, reduces the need for manual intervention, and improves review efficiency. Attached Figure Description

[0041] Figure 1 This is a flowchart of the method of the present invention;

[0042] Figure 2 This is a flowchart illustrating the process of obtaining the behavior positioning table for the clauses of this invention.

[0043] Figure 3 This is a flowchart illustrating the process of obtaining causal-related fragment sequence groups according to the present invention.

[0044] Figure 4 This is a flowchart illustrating the process of obtaining the expression chain break record set in this invention.

[0045] Figure 5 This is a flowchart illustrating the process of obtaining the semantic conflict path list in this invention.

[0046] Figure 6 This is a flowchart illustrating the process of obtaining the intelligent review task path set in this invention. Detailed Implementation

[0047] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0048] refer to Figures 1 to 6 The intelligent document review method based on large-scale model fusion causal reasoning includes the following steps:

[0049] S1: Collect all paragraphs in the document, extract the first line format and blank line spacing of each paragraph in turn, extract the sentences with role words and their corresponding actions one by one, retrieve the first object noun of the phrase containing action verbs, use the role correspondence table to process semantic repetition, merge the results according to the original paragraph position, and output the clause behavior positioning table.

[0050] S2: Based on the clause behavior positioning table, retrieve the text fields involving causal connection words in sequence, compare the order of occurrence of action words in the text, perform screening operation on whether the roles and actions in the text are connected continuously, classify the segments with no more than three lines between verb phrases, connect continuous text actions with corresponding positions, and output the causal relationship segment sequence group.

[0051] S3: Extract action words and role combinations from each segment based on causal relationship segment sequence groups, check the consistency of role usage between the first and last sentences, separate the expression of the action without the object word, perform relative position calibration for the same behavior that appears multiple times in the text but with different objects, confirm the missing parts in the order of the text segments in which they appear, and output the expression chain break record set.

[0052] S4: Extract expressions involving repeated actions of the same character from the record set of broken expression chains, call the character behavior list according to the distance between them in the text, perform root matching between the expression action and the above behavior terms, split the position of the negative adverb in the sentence and the collocation of the behavior words, perform unidirectional sorting of the subject-object relationship in the sentence connection structure, and output a list of semantic conflict paths.

[0053] S5: Based on the semantic conflict path list and the expression chain break record set, relevant content is extracted to separate the segments involving repeated actions and role changes. Sentences containing result words are organized according to their intersection with the role and verb. Expression segments containing multiple actions are arranged according to the order of the verbs. Segments without connected subjects or adverbs between verb actions are screened and processed to output the intelligent review task path set.

[0054] The clause behavior positioning table includes paragraph first-line format, blank line spacing, role words, action words, object nouns, semantic repetition, and paragraph position markers; the causal relationship segment sequence group includes causal connectors, action word order, role and action connection, verb phrase intervals, and continuous action text; the expression chain break record set includes role and action combination, first and last sentence role consistency, unconnected objects, inconsistent action objects, position calibration, and missing part markers; the semantic conflict path list includes repeated actions, role behavior list, root word matching, use of negative adverbs, and subject-object relationship analysis; the intelligent review task path set includes action repetition, role change, result word extraction, verb intersection markers, action arrangement order, and paragraphs without connected subject adverbs.

[0055] Please see Figure 2 The steps to obtain the clause behavior positioning table are as follows:

[0056] S101: Obtain the document paragraph content and simultaneously obtain the first-line format item and blank line spacing item of each paragraph. Call the paragraph content and format item to perform a check action on the paragraph position order. Record the corresponding role words and their action verb phrases that appear in the paragraph content and establish a sequence relationship with the paragraph number to generate the paragraph action structure.

[0057] The document is divided into consecutive paragraphs according to the display order and numbered. The number of indentation characters at the first line of each paragraph is read character by character and converted into an integer value. The number of blank lines before and after each paragraph is counted line by line and recorded as independent integer values. These two values ​​are used as basic parameters for paragraph formatting. The text of each paragraph, starting from the first sentence, is then split word by word in sequence. After splitting, recognition actions are performed sequentially. Each word is compared word by word with an established character title list to determine if it belongs to a character term. When characters are completely identical or match the alias list in the title list, it is confirmed as a character term, and its position index in the paragraph is recorded. The starting point of the verb phrase is determined by reading the position of the verb after the character term. Then, character continuity is judged from the verb to the subsequent consecutive words to confirm whether an action phrase is formed. The phrase is recorded in its original character form. Simultaneously, multiple character-action association extraction actions are performed within the paragraph. Character words appearing in different orders are numbered according to their chronological order of appearance, creating a one-to-one correspondence between character numbers and action phrase numbers. During this process, if multiple character words appear in the same paragraph, they are arranged in ascending order based on their starting position in the text. By rereading the paragraph's first-line formatting parameters and blank line parameters, the paragraph formatting information and character action information are combined to form a paragraph action structure example. For instance, when "Li Mou organizes materials" appears in a paragraph, it constitutes an action structure, forming a unit chain with a fixed order of paragraph number, character word, and action phrase. To ensure the consistency of the paragraph sequence, all structures are arranged in ascending order of number. After reading multiple paragraphs, a complete paragraph action structure sequence is formed.

[0058] S102: Based on the paragraph action structure, call the position of the action verb phrase to execute the search action on the subsequent phrases in sequence, confirm the first noun in the search path with the phrase execution object and record the sequence structure between the character words, action phrases and nouns, and generate the character action object chain;

[0059] Based on the paragraph action structure, each action phrase is read sequentially. The starting character number of the action phrase in the paragraph text is used as the retrieval starting point. By continuously scanning the text, the retrieval action is performed word by word starting from the position after the verb phrase. Each subsequent word is added to the candidate list in the order of appearance. During the retrieval process, a noun determination action is performed on each candidate word. By comparing the candidate word character by character with the noun list entries, when the character sequence matches any noun in the list, it is determined as a noun item and its appearance position is recorded. When multiple noun items appear consecutively after the same verb phrase, they are recorded in order of appearance, and a position comparison action is performed to determine which noun item is closest to the action phrase. That is, the index size of their character starting positions is compared, and the one with the smallest index is the target word for the action. The remaining nouns are treated as non-target record items. The process involves exclusion. If additional modifiers such as "number table" or "application record table" appear during this process, the two consecutive words are treated as a whole noun and matched as a whole to ensure the completeness and reliability of the object recognition process. The identified role words, action phrases, and object nouns are combined into a role action object chain according to a fixed sequence. For example, in "Zhang submitted an application form for review", after the action phrase "submit", the nouns "application form" and "review form" are identified. By comparing their respective positions, "application form" is determined to be the closest to the verb phrase, making it an object noun, thus forming the object chain "Zhang - submit - application form". The corresponding paragraph number is then assigned to it. Finally, the object chains generated in all paragraphs are arranged and output in ascending order of paragraph number to form a complete role action object chain sequence.

[0060] S103: Based on the character action object chain, call the character correspondence table to perform a comparison action on the repeated expression items, merge the items that have been matched with the original paragraph position number and keep the original paragraph order, and obtain the clause behavior positioning table.

[0061] Based on the character action object chain, all character words are extracted sequentially and compared character by character. Character-by-character alignment is used to determine if character names are identical. If the number of characters is identical and each character is completely identical, they are considered the same character. When the same character appears in multiple object chains, its corresponding paragraph number is recorded. These paragraph numbers are arranged in natural number order to form a sequence of paragraph numbers corresponding to the character. After obtaining the paragraph number sequence, the corresponding character's action phrases and object nouns are extracted from the object chain. The action phrases are then arranged according to their order of appearance in the text, forming an action record sequence in ascending order of paragraph number. For action phrases of the same character... The process involves comparing the verbs to determine if the action phrases in different paragraphs are completely identical. If the verbs are identical but the following words are different, they are identified as two independent phrases and recorded separately. If the action phrases are completely identical but the object words are different, they are recorded as two object chain units and assigned to the same role entry. During this process, character comparison is performed again for each object word to eliminate duplicate object records. All object chains of the same role are arranged in ascending order according to the original paragraph number to maintain consistency with the document order. The role words, the sorted action phrase sequence, the corresponding object word sequence, and the paragraph number sequence are combined into a clause behavior positioning table.

[0062] Please see Figure 3 The steps for obtaining causal association fragment sequence groups are as follows:

[0063] S201: Based on the clause behavior positioning table, detect the text fields that involve causal conjunctions, perform comparison actions on the order of occurrence of action words in these text fields, perform screening operations on the connection relationship between role words and action words in adjacent sentences and record the order difference items, and generate action connection order;

[0064] Based on the premise that the clause behavior location table detection requires identifying text paragraphs with causal connections, the process reads all paragraphs in the location table one by one during execution. Each paragraph is scanned character-wise to find causal connectors. The short sentences containing the scanned connectors are split into sentences before and after the connectors. The order of action words within the two sentences is then determined and their occurrence order is extracted. The position of the action word in the first sentence is recorded as parameter a1, and the position of the action word in the second sentence is recorded as parameter a2. A direct comparison is performed on the two integers to determine the difference in the order of actions in adjacent sentences. To ensure consistency, a word-by-word counting action is performed on the words within the short sentences. For example, in "Because Zhang failed to organize the record documents, the management personnel re-registered the content," "failed to organize" is determined as the action word in the first sentence, and its occurrence number is recorded as the 4th word. "Re-registered" is determined as the action word in the second sentence, and its occurrence number is recorded as the 3rd word. Then, a difference recording action is performed, and the two... Subtracting the sequence numbers yields a ranking difference value, which is saved as action connection difference information. Then, a positioning action is performed on the role words in the same paragraph. By matching words one by one, it is determined whether the role appears in the preceding or following sentence and its position index is recorded. For example, if "Zhang Mou" appears in the second position of the preceding sentence but not in the following sentence, it is marked as a role only in the preceding sentence and recorded along with the action position information. The same action is repeated for paragraphs with multiple roles. A connection screening action is performed on each "role-action word" combination. The connection strength is determined by judging the size of the action word ranking difference value. The action pair with the smaller absolute value of the difference is identified as the priority connection action. For example, if there is an action word sequence number 5 in the preceding sentence and an action word sequence number 7 in the following sentence in a paragraph, compared with another group of action words sequence number 12 in the following sentence, since the difference between 7 and 5 is 2, which is less than the difference between 12 and 5, the preceding action pair is recorded as the main connection action. Then, the connection actions detected in all paragraphs are arranged in order of paragraph number to form the action connection order.

[0065] S202: Call the action connection order to perform detection actions on the line order of each action word, group the segments with an interval of no more than three lines between adjacent action words and record the first and last line order of the segments, perform judgment operations on the order of character words and action words in the grouped segments and establish a continuous structure to obtain the continuous action interval degree;

[0066] Based on the premise of detecting the line order of action words according to the action sequence order, during the execution process, the line number of each action word position recorded in the sequence is read and converted into an integer. Then, all line numbers are arranged in ascending order, and an interval judgment is performed on adjacent line numbers. By comparing the difference between adjacent line numbers, when the difference is less than or equal to three lines, they are classified into the same segment, and the first and last line numbers of the segment are recorded. If multiple consecutive line numbers that meet the condition appear during the line number reading process, such as lines 8, 10, and 11, they are classified into the same segment in sequence, and the last line number is updated to the maximum line number 11. When the difference between line numbers exceeds three lines, the current segment ends and the next segment begins. After the above segments are established one by one, the order judgment action is performed on the action words and role words within the segment. By reading the character sequence positions of role words and action words in each line of text, each position is converted. The integers are used to determine the continuity of characters within a segment. For example, in line 8, "Zhang submitted materials," "Zhang" is recorded as position 1, and "submit" as position 3. In line 10, "Zhang organized the recorded content," "organize" is recorded as position 2. By determining whether the positions of two actions maintain continuous displacement in different sentence segments or whether there is a character matching relationship, they are connected into a continuous structure. If there are different characters within the segment, the characters are compared to confirm whether the characters are the same. If the characters are completely identical, they are classified as a continuous character chain. If they are inconsistent, they are marked separately but still recorded in the segment. When constructing the continuous structure, a sequence check is performed on each line. For example, if a character appears earlier in the segment the first time, and its position changes slightly in the next line but is still before the action word, such cases are recorded as valid continuous chains. The above continuous structures are arranged according to the segment line number order to form a continuous action interval.

[0067] S203: Combine the sequenced fragments with the original paragraph position numbers according to the continuous action interval degree, arrange the combined structure according to the text appearance order and record the paragraph position of the causal connector, establish the arrangement result as a sequence group, and obtain the causal association fragment sequence group.

[0068] Based on the premise of combining the sequenced fragments with the original paragraph position numbers according to the continuous action interval, the first and last line numbers of each fragment in the continuous action interval are read during the execution process. By comparing the original paragraph numbers to which the line numbers belong, the paragraph range corresponding to the fragment is confirmed. All paragraph numbers within the range are arranged in ascending order to generate a segment number sequence. Then, the action words and role words recorded in the fragment are read one by one and bound to the segment number sequence, so that the segment number and the corresponding action chain and role chain form a combined structure. In this process, if a fragment corresponds to multiple segment numbers, for example, if the fragment involves lines 9 to 11, but the original text shows that line 9 belongs to segment 3, and lines 10 and 11 belong to segment 4, then the fragment is combined into the segment number sequence (3, 4). The system locates the line number of the causal connector by comparing the line number with the original paragraph number, and writes the paragraph number into the combined structure. When a segment contains multiple causal connectors, its line number is read and converted into a paragraph number, and then arranged into the sequence group according to the order of the line numbers, so that the positions of the causal relationship can be kept in a consistent order in the sequence. The combined structure of all segments is sorted by comparing the first line number of each segment to determine the order of the segments in the sequence group, and the segment with the smaller line number is placed first. For example, if segment A starts at line 5 and segment B starts at line 12, then segment A is arranged first, and then segment B is arranged. The sorted segments are arranged into a causal relationship segment sequence group in a uniform format and used as the output result.

[0069] Please see Figure 4 The steps for obtaining the expression chain break record set are as follows:

[0070] S301: Based on the causal relationship segment sequence group, extract the action words and role combinations of each segment, call the role words in the first and last sentences of the segment to perform comparison actions on their usage, record the role usages and corresponding action words that show differences after comparison as difference items and establish a corresponding relationship with the segment number to generate role consistency order;

[0071] Based on the initial requirement of extracting action words and character combinations from causal relationship sequence groups, the process involves reading each segment in the sequence group one by one. The correspondence between action words and character words within each segment is extracted from the first and last sentences. The character word in the first sentence is compared character by character to confirm its actual appearance, word length, and position index within the sentence. The same character comparison process is then performed on the character word in the last sentence to confirm its usage in the text. Simultaneously, the order of appearance of action words in the same sentence is read and converted into integer indices to determine the relationship between character and action. A comparison is performed on the character usage in the first and last sentences. This comparison involves checking the character sequence, the grammatical position of the character in the sentence, and the coherence between the character and action words. When any difference is found, the difference is recorded and a correspondence is established between the difference and the segment number. If a pronoun relationship is found during the comparison, for example, if the first sentence uses "Zhang Mou" and the last sentence uses "Qi," then the preceding segment content needs to be read and traced back to its source. The process checks whether pronouns correspond to the same role. When a character comparison fails to match the role in the segment record, the difference is marked as a role replacement and recorded together with the action word as a difference unit. In processing segments with multiple action words, the comparison steps need to be repeated for each action. For example, if a segment's first sentence is "Li submitted the document" and the last sentence is "Zhang continued to register," the two roles are compared separately to determine their differences. The difference item is then bound to the two action words, and the difference records are merged and stored as a segment difference record unit. Simultaneously, the direction of role change is determined by comparing the positional changes of the roles in the first and last sentences to determine the text path of the role transition. This role transition is then bound to the corresponding action word. If there is an overlap between multiple roles and multiple actions within a segment, each action chain needs to be compared in the order of appearance. For example, if the first sentence is "Wang organized the data, Li reviewed the content," and the last sentence is "Wang signed the record," the role differences are confirmed for the "organize-sign" and "review-sign" chains, and the identified differences are combined into a role consistency order according to the segment number from smallest to largest.

[0072] S302: Detect the position of the action word according to the consistency degree of the role, separate the expression fragments that are not followed by object words after the action word and record the separation point line order, and at the same time, perform relative position judgment on the expressions that appear in different fragments and have different objects and record the relative position difference value to obtain the action object breakpoint degree;

[0073] Starting with the requirement of detecting the location of action words based on role consistency order, the process involves sequentially reading the role and action chain structure from the role consistency order, extracting the position index of the action word in the text along with its line number, and then performing a continuous scan of the words following the action word. Each word after the action word is compared sequentially to identify whether it is an object word. When several consecutive words fail to match a recognizable object noun, the line number of the action word is recorded as a breakpoint. To avoid omissions, the remaining part of the sentence containing the action word is scanned to confirm whether passive or ellipsis expressions imply an object. If no character segment that can be identified as an object word is detected, the line is determined to be an action object breakpoint. When processing cases where the same action word appears in different segments, the object words in those segments need to be compared. The algorithm extracts the character composition, position, and role binding relationship of object words to perform relative position judgment. For example, if the action "register" appears in line 12 of segment A as "application form" and in line 18 of segment B as "revised draft," the relative difference value of 6 is obtained by comparing the two line numbers. This value is recorded as part of the action object difference degree. The algorithm then judges whether the role and action within the segment can maintain a continuous relationship. When the same action appears repeatedly in a segment but the role word changes, the role change point is recorded as a new breakpoint item in the breakpoint degree list. If multiple breakpoints are generated in a single detection, the breakpoint items are sorted according to the order of action word appearance, arranged in ascending order of line. When a segment spans multiple line intervals, the correspondence between action words and object words is also confirmed line by line. The action object breakpoint degree is generated through continuous scanning, comparison, and recording.

[0074] S303: Call the action object breakpoint degree to perform calibration actions on the separated fragments and relative position differences according to the text paragraphs in sequence, establish a sequence structure with the missing object words or missing role usage parts and the original paragraph order, and record the fragment continuity change degree to create a record set, thus obtaining the expression chain break record set.

[0075] The initial requirement is to invoke the breakpoint degree of the action object to calibrate the separated segments and relative position difference items sequentially according to the text paragraphs. During execution, the line sequence of the separated points is read item by item from the breakpoint degree of the action object. The line number of each separated point is compared with the numerical range of the start and end line numbers of the original paragraph to confirm the paragraph number to which the line belongs. Then, the paragraph number is written into the record chain of the corresponding segment, establishing a fixed association between the separated segments and the paragraph order. The relative position difference items are read from the breakpoint degree and their order is judged. The positional relationship of the action nodes in different segments is determined by directly comparing the difference values. Action nodes with a difference value greater than zero are marked as moved backward, action nodes with a difference value less than zero are marked as moved forward, and action nodes with a difference value of zero are marked as moved forward. The items are recorded as position-preserving items. These moving items are then integrated with the roles and action words in the separated segments. A continuous record chain is constructed by corresponding the paragraph numbers in ascending order. During the formation of the continuous chain, the positions of missing object words need to be marked for completion. By recording the gaps in the positions where the action words should have been connected, the calibrated chain can show the connection gaps in the action chain. If there are missing roles within a segment, the missing point and the separation point need to be written into the calibration structure at the same time to reflect the break position of the role and action in the text. Then, the calibration structure is sorted segment by segment. The segment continuity variation is obtained by arranging the paragraph numbers in ascending order. The variation and the segment sequence are written into the record set to form the expression chain break record set.

[0076] Please see Figure 5 The steps to obtain the list of semantic conflict paths are as follows:

[0077] S401: Based on the record set of broken expression chains, extract the expression statements that involve repeated actions of the same character, call the statements in the text to perform the search action on the character behavior list, match the action roots in the expression statements with the behavior terms in the behavior list and record the matching differences. Establish a correspondence between the differences and the statement position number to generate action root matching groups.

[0078] Based on the premise of extracting expressions involving repeated actions of the same character from a broken record set, the process involves reading each broken item in the record set one by one. The character words in each record entry are bound to the corresponding action words for that character, and the text positions are scanned line by line from the original broken chain of the record set to confirm whether the character's actions are repeated in multiple positions. The content of the sentences containing the repeated action words is then extracted according to their line numbers to form a candidate sentence set. The verb roots within these candidate sentences are then split, obtaining root items by breaking the complete verb into character sequences of roots and adjuncts. Roots are repeatedly extracted for different sentences of the same character, and these roots are matched with the character behavior terms recorded in the behavior list. This matching action determines whether the character sequences of the two are consistent by comparing them character by character. If there is a difference between the root character sequence and the behavior words in the behavior list, the difference is recorded as a difference item and matched with the sentence. The location number establishes the association. When performing a matching action, if similar but different verbs appear, such as "register", "login", "login", or "review" and "verify", then each word needs to be compared character by character to confirm the level of difference. The degree of difference is judged by the number of inconsistent characters. For example, if "register" and "login" differ in the third character, it is recorded as a single difference, while if "login" and "register" differ in both the second and third characters, it is recorded as a double difference. The position of the differing character is then bound and saved to the compared root word. In some sentences, although the roles are the same, the action words are slightly different, such as "organize data" and "integrate data". In this case, character matching needs to be performed again. When the root word "integrate" is the same and the second character changes, it is recorded as a difference point and placed under the sentence number corresponding to that role. This series of difference records is arranged in the order of the segments, so that each segment corresponds to a set of "role-action root word difference items". After scanning all segments, the difference sets are arranged in row order to form action root word matching groups.

[0079] S402: Based on the action root matching group, perform detection actions on the position of the negative adverb in the sentence, split the negative adverb and the following action words and record the split items, and at the same time, perform one-way judgment on the relationship between the role and the object in the sentence connection structure and record the relationship direction item to obtain the role relationship structure group;

[0080] Based on the premise that the position of the negation adverb in the sentence is detected by the action root matching group, the execution process reads the line order and root difference of each sentence in the matching group, splits the sentence into continuous word segments and scans them word by word to confirm the position of the negation adverb. When the negation adverb is scanned, its character position is recorded as n1, and the position of the immediately following action word is recorded as v1. The closeness of the two is judged by recording the character distance between them. Then, the negation adverb and the action word are split. The split item is formed by marking the negation adverb as an independent semantic unit and marking the subsequent action word as the negated semantic unit. The split item is bound and saved with the sentence number. At the same time, a one-way judgment action is performed on the role and object relationship in the internal connection structure of the sentence. This judgment action obtains the relationship direction by comparing the position of the role and the position of the object word in the connection structure. For example, in the structure "Zhang did not organize the record content while Li continued to register the attachments", the relationship direction is obtained by scanning. The phrase "not organized" is used to confirm that the combination relationship between "Zhang" and "recorded content" is a top-down action chain. At the same time, scanning "Li registered attachments" after the connecting word obtains the action chain of the character "Li" towards the object "attachments". The directions of the two chains are marked as "forward chain" and "backward chain" respectively, and the order of the character and object in the sentence is recorded to form direction items. When judging the direction, if the object directions of the two character chains are opposite, for example, one action chain appears in the order of character-action-object, while the other action chain appears in the order of character-negation-action, then the difference in direction between the two is recorded as a reverse item and included in the segment structure. When the negative adverb appears at the paragraph transition in the sentence, such as "it was not submitted and was handled by others", then the split "not submitted" needs to be determined as a forward action chain, and "handled by others" needs to be determined as a backward chain. The relationship between the two is recorded as two direction points. The split items and direction items are combined and organized in the order of sentence number to obtain the character relationship structure group.

[0081] S403: Call the role relationship structure group to perform a combination action on the matching difference items and relationship direction items in the sentence fragment, arrange the combined structure in text order and record the conflict association items, and build the arranged content into a path list to obtain the semantic conflict path list.

[0082] Under the premise of calling the role relation structure group to perform a combination action on the matching difference items and relation direction items in the statement fragment, during the execution process, the statement number, direction item, and split item are extracted from each record in the relation structure group. The position of the negation adverb in the split item and the position of the action word are combined into a binary structure. Then, the role direction and object direction in the direction item are combined into another structure. The combination action is performed on the two. By comparing the binary structure and the direction structure according to the position correspondence of the statement number from small to large, they are merged into a semantic chain. If a matching difference item (i.e., a word root inconsistency item) appears in the statement during the combination process, a difference marker needs to be inserted into the chain, and the position of the difference marker is bound to the corresponding direction item. For example, in statement number 12 If the statement "Zhang Mou did not organize the recorded content" appears in the text and his behavior list records "Zhang Mou organized", then during the chain combination process, "organize" and "not organize" are identified as opposite items and the conflict point is marked in the chain. Then, the directional items appearing in the same statement are continuously judged. By confirming whether the same role performs an action or negates an action in a continuous position, it is determined whether two directional items need to be merged into a continuous conflict group. If the roles involved in two adjacent directional items are identified to be different, the role switching point needs to be recorded as a disconnection item. Then, all conflict points, directional difference points, and role switching points are sorted according to the order of text appearance to form a list of conflict-related items. Then, these conflict-related items are constructed into a path list according to the fragment order. A semantic conflict path list is generated by arranging each conflict path in the list according to the text from front to back.

[0083] Please see Figure 6 The steps for obtaining the intelligent review task path set are as follows:

[0084] S501: Based on the semantic conflict path list and the expression chain break record set, relevant content is extracted and the action repetition and role change segments are monitored. For these segments containing result words, the sentence is sorted according to the role position and verb intersection point. The sorted sentence structure is then matched with the original text line order to generate the role verb intersection sequence.

[0085] Based on the premise of extracting repeated action segments involving the same role from the semantic conflict path list and expression chain break record set, the process compares the two types of records one by one. A searchable index is established by the line order, role words, and action word positions in the records. Then, the positions pointed to by these indexes are scanned in the text line by line, grouping action statements of the same role appearing in different lines into a comparable unit. By judging the composition and position of role words character by character, it is confirmed whether the role maintains a consistent identity in multiple statements. Then, root word splitting is performed on the action words, removing additional components from the action words according to their character structure to extract the root words. These root words are compared one by one with the action words in the corresponding action list of the role. A point-by-point character sequence matching method is used to determine whether the two are consistent. When inconsistencies are found, the difference is recorded as a difference item, and a correspondence is established between the difference item and the line order of the statement in which it appears. If a change in the role is detected between two statements, for example, "Wang registered the information" in the first sentence and "Li registered the attachments" in the second sentence, then a character-by-character comparison of the role words in the two sentences is required. For actions, the difference in the position of the character is recorded as the character change item. Sentences containing result words are further segmented. By locating the character of the result word, the sentence is divided into a first-segment character area and a second-segment action area. Then, the character position extraction action is performed on the character area, and the start and end positions of the character word in the sentence are recorded as the character index. At the same time, the verb index extraction action is performed on the action area, and its character position is recorded in the same way. The cross-judgment action is performed on the character index and the verb index. By judging the order of the two in the sentence and their mutual pointing relationship, the character verb cross point is obtained. The cross point is regarded as the core representation of the sentence structure. A mapping relationship is established between it and the line order of the original text, so that each sentence structure can be located to the original text position. If a cross-sentence jump relationship is found between the character and the action during the generation of cross structures, such as the character being in the previous sentence and the action being in the next sentence, the line order of the two sentences needs to be judged by the spacing action, and they are classified into the same cross unit. After all cross structures are sorted out, they are arranged in line order to form a character verb cross sequence.

[0086] S502: Call the character verb cross sequence to perform judgment actions on expression segments containing multiple actions, arrange the order of verbs according to line position and record the order items, and at the same time perform screening processing on paragraphs where the verb actions are not connected by subjects or adverbs and record the screening items to obtain action connection sorting groups;

[0087] Under the premise of calling the cross-sequence of role verbs to perform judgment on expression fragments containing multiple actions, the process reads the role index, verb index, and line order information of each cross-record during execution. Then, it performs a sorting action on multiple action words appearing in the same fragment according to their line order position. By comparing the line numbers, the order of action occurrence is obtained. Actions with earlier line numbers are recorded as earlier actions, and actions with later line numbers are recorded as later actions. These sorting results are recorded as action sequence items. At the same time, a subject judgment action is performed on the words preceding the action words. By scanning the consecutive words preceding the action word, it is determined whether there is a character sequence that can match the role index. When the match fails, the action is recorded as a missing subject item. Then, an adverb concatenation judgment action is performed on the area before and after the action word. By retrieving the adverb position and matching it with the action word position, it performs a concatenation judgment action. Character distance comparison is performed. When the distance exceeds two words or the adverb character sequence cannot form a semantic connection with the action, it is recorded as an adverb not connected item. The continuity of the relationship between multiple action words is judged by calculating the difference in the line order of adjacent actions and whether the corresponding roles are the same to determine whether a continuous action chain is formed. For example, if "Zhang Mou checks" appears in line 15 and "Zhang Mou registers" appears in line 16, the two actions are recorded as a continuous chain. However, if "Li Mou supplements" appears in line 18, the role change item needs to be recorded. At the same time, if there is a line order jump between actions, such as between line 10 and line 15, it is recorded as an action discontinuity item. Then, these action order items, subject missing items, adverb not connected items, and action discontinuity items are uniformly organized into a sorting structure, and the action connection sorting group is formed by arranging the line order from small to large.

[0088] S503: Based on the action connection sorting group, perform combined actions on the sorted segments and the filtered items, build the structure of the combined content segment by segment according to the text paragraph order, record the action and role arrangement changes, organize these structures into a task path set, and obtain the intelligent review task path set.

[0089] Based on the premise that the action connection sorting group combines the sorted segments and the filtered items, the execution process reads records such as action order, role index, and connection status from the sorting group. Then, the role segments and action segments in the sorted segments are mapped to the action nodes of the sorting group according to their row order. By inserting each action node into the sorted segment structure in the order of segments, a segment structure chain is formed. The missing subject items and adverb unconnected items recorded in the filtered items are inserted below the corresponding action nodes in the structure chain according to their row order, so that each action node has complete contextual information. Record actions are executed for changes in role arrangement. The role index of adjacent nodes is compared item by item to determine whether a role has been cut. When the character sequence is different or the character position index changes, it is recorded as a character change item. At the same time, position tracking is performed on the action sequence. By judging whether the action sequence is continuously ascending, if there is a sequence back jump or sequence jump, it is recorded as an action arrangement anomaly item. Then, paragraph-level sorting is performed on all structural items. By arranging all segments in ascending order according to the original paragraph number of the segment, all segments are linked into a continuous task path. The character change item, action arrangement change item, and filter item are all filled into the path structure, so that each action node has dual arrangement information of character and action. The above content is organized into a task path set to obtain the intelligent review task path set.

[0090] An intelligent document review system based on large-scale model fusion causal reasoning, comprising:

[0091] The collection and extraction module collects all paragraphs in the document, extracts the first line format and blank line spacing of each paragraph, extracts sentences with role words and their corresponding actions one by one, retrieves the first object noun for phrases containing action verbs, uses a role correspondence table to handle semantic repetition, merges the results according to the original paragraph position, and outputs a clause behavior positioning table.

[0092] The causal association module sequentially retrieves text fields involving causal conjunctions from the clause behavior positioning table, compares the order of occurrence of action words, performs screening operations to determine whether the roles and actions in the text are continuous, groups segments with no more than three lines between verb phrases, connects continuous text actions with their corresponding positions, and outputs a causal association segment sequence group.

[0093] The consistency check module extracts action words and role combinations from the causal relationship sequence group, checks the consistency of role usage between the first and last sentences, separates the expression of the action without the object word, performs relative position calibration for the same behavior that appears multiple times in the text but with different objects, confirms the missing parts in the order of the text paragraphs in which they appear, and outputs a record set of expression chain breaks.

[0094] The semantic conflict analysis module extracts expressions involving repeated actions of the same character from the record set of broken expression chains. It calls the list of character behaviors according to the distance between them in the text, performs root matching between the expression actions and the words used in the preceding and following behaviors, performs splitting of the position of the negative adverb in the sentence and the collocation of the behavior words, performs one-way sorting of the subject-object relationship in the sentence connection structure, and outputs a list of semantic conflict paths.

[0095] The intelligent review module extracts relevant content from the list of semantic conflict paths and the record set of expression chain breaks, separates the segments involving repeated actions and role changes, organizes the sentences containing result words according to their intersection with the role and verb, arranges the expression segments containing multiple actions according to the order of the verbs, and filters the segments where the verb actions are not connected by subjects or adverbs, and outputs the intelligent review task path set.

[0096] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An intelligent document review method based on large model fusion causal reasoning, characterized in that, It comprises the following steps: S1: based on collecting all paragraphs of the document, extracting the first line format and line spacing of the paragraph in turn, extracting the sentence containing the character word and its corresponding action one by one, searching the first object noun after the verb group containing the action, using the character corresponding table to process the semantic repetition expression, merging the results according to the paragraph position, and outputting the clause behavior positioning table; S2: based on the clause behavior positioning table, the paragraphs containing causal conjunction words are searched in turn, the appearance order of action words is compared, the continuity of characters and actions is screened, the segments with a verb group interval of not more than three lines are listed, the continuous text action and the corresponding position are connected, and the causal correlation segment sequence group is output; S3: based on the causal correlation segment sequence group, the action word and character combination of each paragraph are extracted, the consistency of the first sentence and the last sentence character usage is checked, the action expression of the unconnected object word is separated, the relative position of the same behavior but different objects is calibrated, the missing part is confirmed according to the paragraph order, and the expression chain fracture record set is output; S4: extract the expression chain fracture record set, extract the repeated action expression sentence of the same character, call the character behavior list according to the distance between the front and back of the text, match the action and the aforementioned word root, split the negative adverb and behavior word combination, and comb the connection relationship between subject and object structure. The result is a semantic conflict path list.

2. The method of claim 1, wherein the method is performed by a computer system. The clause behavior positioning table comprises the first line format of the paragraph, line spacing, character word, action word, object noun, semantic repetition expression, and paragraph position identification; the causal correlation segment sequence group comprises causal conjunction words, action word order, character and action connection, verb group interval, and continuous action text; the expression chain fracture record set comprises character and action combination, consistency of first sentence and last sentence character, unconnected object, action object inconsistency, position calibration, and missing part identification; The semantic conflict path list comprises repeated action, character behavior list, word root matching, negative adverb use, and main and guest relationship combing.

3. The method of claim 1, wherein the method is performed by a computer system comprising one or more processors. The acquisition steps of the clause behavior positioning table are: S101: acquire the paragraph content of the document and synchronously acquire the first line format item and line spacing item of each paragraph, call the paragraph content and format item to perform checking action on the paragraph position sequence, perform corresponding record on the character word and its action verb phrase appearing in the paragraph content, and establish sequence relationship with the paragraph serial number, generate paragraph action structure; S102: based on the paragraph action structure, call the position of the action verb phrase to perform retrieval action on the subsequent word group in order, confirm the first noun item in the retrieval path and the object of the phrase, record the sequence structure between the character word, action phrase and noun item, and generate the character action object chain; S103: according to the character action object chain, call the character corresponding table to perform comparison action on the repeated expression item, merge the comparison established item with the original paragraph position serial number and keep the original order of the paragraph, and get the clause behavior positioning table.

4. The method of claim 1, wherein the method further comprises: The acquisition steps of the causal correlation segment sequence group are: S201: Detecting text fields involving causal conjunctions based on the clause behavior positioning table, performing comparison actions on the order of action words in these text fields, performing screening operations on the connection relationship between role words and action words in adjacent sentences and recording order differences, and generating action connection degrees; S202: Calling the action connection degree to perform detection actions on the line sequence where each segment action word is located, performing grouping and recording the first and last line sequences of the segment whose adjacent action word line sequence interval is not more than three lines, performing judgment operations on the order of role words and action words in the grouped segment, and establishing a continuous structure to obtain a continuous action interval degree; S203: According to the continuous action interval degree, combine the grouped segment with the original paragraph position sequence number, arrange the combined structure according to the text appearance order and record the position of the causal conjunction in the segment, establish the arrangement result as a sequence group, and obtain a causal correlation segment sequence group.

5. The method of claim 1, wherein: The expression chain rupture record set acquisition step is: S301: Extracting the action word and role combination in each segment of the causal correlation segment sequence group, calling the role words in the first and last sentences of the segment to perform comparison actions on their usage, recording the role usage and corresponding action words that appear different after comparison as difference items and establishing a corresponding relationship with the segment sequence number, and generating a role consistency degree; S302: According to the role consistency degree, perform detection actions on the position of the action word, perform separation processing on the expression segment after the action word without connecting object words and record the separation point line sequence, and simultaneously perform relative position judgment on the expressions that appear in different segments and have different objects of the same behavior and record the relative position difference value, to obtain an action object breakpoint degree; S303: Calling the action object breakpoint degree to perform calibration actions on the separated segment and the relative position difference item according to the text paragraph sequence, establishing a sequence structure for the missing object words or missing role usage part and the original paragraph sequence and recording the segment continuity change degree, and establishing it as a record set to obtain an expression chain rupture record set.

6. The method of claim 1, wherein: The semantic conflict path list acquisition step is: S401: Extracting the expression sentences involving repeated actions of the same role based on the expression chain rupture record set, calling the role behavior list to perform retrieval actions on the sentences separated by line sequence in the text, performing matching between the action word root in the expression sentence and the behavior usage in the behavior list and recording the matching difference item, establishing a corresponding relationship between the difference item and the sentence position sequence number, and generating an action word root matching group; S402: According to the action word root matching group, perform detection actions on the position of the negative adverb in the sentence, perform splitting between the negative adverb and the action word after it and record the splitting item, and simultaneously perform one-way judgment on the role and object relationship in the sentence connection structure and record the relationship direction item to obtain a role relationship structure group; S403: Calling the role relationship structure group to perform grouping actions on the matching difference item and the relationship direction item in the sentence segment, arranging the grouped structure according to the text order and recording the conflict association item, establishing the arrangement content as a path list, and obtaining a semantic conflict path list.

7. The method of claim 1, wherein the method further comprises: Also includes: S5: based on the semantic conflict path list and the expression chain fracture record set, separate the action repetition and the role change segment, arrange the result word sentence at the role and verb intersection point, arrange the verb sequence of the multi-action expression segment, screen the unconnected subject or adverb paragraph, and the result is an intelligent review task path set; The intelligent review task path set includes action repetition, role change, result word extraction, verb intersection point identification, action arrangement order, and unconnected subject or adverb paragraph.

8. The method of claim 7, wherein the method further comprises: The acquisition step of the intelligent review task path set is: S501: based on the semantic conflict path list and the expression chain fracture record set, monitor the action repetition and the role change segment, arrange the action of the sentence containing the result word at the role position and the verb intersection point, establish the corresponding relationship between the arranged sentence structure and the original text line sequence, and generate the role verb intersection sequence; S502: call the role verb intersection sequence to judge the action of the expression segment containing multiple actions, arrange and record the arrangement order item according to the line sequence position, and screen and record the unconnected subject or adverb paragraph between the verb behaviors, and get the action connection sorting group; S503: according to the action connection sorting group, perform the combination action on the arranged segment and the screened item, establish the structure of the combined content according to the text paragraph sequence, record the action and role arrangement change item, arrange these structures into a task path set, and get the intelligent review task path set.

9. An intelligent document review system based on large model fusion causal reasoning, characterized in that, The system is used for the intelligent document review method of large model fusion causal reasoning according to any one of claims 1-8, and the system comprises: A collection and extraction module collects all paragraphs in a document, sequentially extracts the first line format and line spacing of each text, extracts the sentence with role words and its corresponding action one by one, searches the first item object noun after the action verb group, processes the semantic repeated expression using the role correspondence table, merges the results according to the original text paragraph position, and outputs a clause behavior positioning table; A causal correlation module sequentially searches the text field related to the causal connection word in the clause behavior positioning table, compares the text appearance order of the action word, performs screening operation on whether the roles and actions in the text are continuous, classifies the segments with a verb group interval of no more than three lines, and processes the continuous text action and corresponding position, and outputs a causal correlation segment sequence group; An identity verification module extracts each action word and role combination in the correlation segment sequence group, verifies the consistency of the role usage between the first sentence and the last sentence, separates the expression of the object word after the action, aligns the relative position of the same behavior appearing multiple times but with different objects, confirms the missing part according to the appearance of the text paragraph, and outputs an expression chain fracture record set. The semantic conflict analysis module extracts the expression statements in the expression chain break record set involving the repeated actions of the same role, calls the role behavior list according to the distance between the front and back in the text, performs root matching on the expression action and the upper behavior language, performs splitting on the use position and the behavior word collocation of the negative adverb in the sentence, performs one-way combing on the subject-object relationship appearing in the sentence connection structure, and outputs a semantic conflict path list. The intelligent review module separates the relevant content from the semantic conflict path list and the expression chain break record set, separates the sentences involving action repetition and role change, arranges the sentences containing result words according to the intersection of the role and the verb, arranges the expression segments containing multiple actions according to the sequence relationship of the verbs before and after, and performs screening processing on the paragraphs in which the subjects are not connected between the verb behaviors or the adverbs, and outputs an intelligent review task path set.

Images