Scientific research literature intelligent question and answer system based on large language model

Abstract

The application discloses a scientific research literature intelligent question and answer system based on a large language model, comprising: a sentence position pressing-in module for generating a to-be-written sentence position; a power order control power module for rewriting five types of control power actions; a vote occupation hooking module for locking write-in vote occupation; a counter-evidence power seizing module for marking reverse literature evidence; a source locking module for completing sentence position opening and closing judgment; a re-entry forbidden writing module for blocking closed sentence positions; and a controlled generation module for outputting scientific research question and answer results and sentence-by-sentence reference sources. The application relates to the technical field of natural language processing, realizes controlled generation of scientific research question and answer, and improves reference accuracy and evidence credibility.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to an intelligent question-answering system for scientific research documents based on a large language model. Background Technology

[0002] Scientific literature question-answering systems are used to assist researchers in extracting research conclusions, experimental evidence, and limiting conditions from papers, reports, experimental records, and cited materials. Existing solutions often combine keyword retrieval, semantic vector retrieval, enhanced retrieval generation, and large language model question answering. This involves first retrieving relevant passages from a literature database, then inputting the retrieved passages into a large language model to generate answers. This approach improves literature reading efficiency and provides users with relatively complete natural language responses.

[0003] Existing solutions still have technical shortcomings. There is a lack of strict control over the writing rights between retrieved literature fragments and user questions; literature content with semantic similarity can easily enter the answer generation process. When literature conclusions involve experimental subjects, applicable conditions, evaluation indicators, and source locations, the existing system lacks control over content reversal, reverse literature evidence takeover, source locking, and failure re-entry. After the large language model generates the answer, there is a lack of sentence-by-sentence verification between the generated content and the literature evidence, which easily leads to problems such as answer sentences lacking sources, reverse evidence not being included, and evidence with mismatched conditions being cited.

[0004] Therefore, how to provide an intelligent question-answering system for scientific literature based on a large language model is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent question-answering system for scientific literature based on a large language model. This invention utilizes MixHop power-order propagation, a large language model, and literature source locking technology to construct a process of sentence position insertion, vote-taking and back-hooking, counter-evidence takeover, re-entry prohibition, and sentence-by-sentence citation. It has the advantages of strong evidence constraints, clear counter-evidence processing, and traceable answers.

[0006] According to an embodiment of the present invention, a scientific literature intelligent question-answering system based on a large language model includes: The sentence-position pushing module is used to perform rule validation on the limited content of the user's research question before the answer is generated, generate the sentence positions to be written, and push the sentence positions to be written into MixHop power-order propagation; The power order control module is used to rewrite MixHop's power order propagation into five types of control actions: vote taking, hooking back, taking over, locking, and re-entry. The ticket reservation and hook-back module is used to issue a write reservation ticket when the documentary evidence matches the sentence to be written, hook back the restricted content bound to the write reservation ticket to the sentence to be written, clear the write reservation ticket when the matching fails, and lock the write reservation ticket when the matching succeeds. The counter-evidence seizure module is used to read documentary evidence that is opposite to the direction of the locked writing vote, mark the documentary evidence with the opposite direction as reverse documentary evidence, and seize the right to write the sentence to be written when the reverse documentary evidence completes the content matching, and turn the sentence to be written into the disputed sentence. The source locking module is used to perform source locking on the document evidence that is locked and bound to the ticket holder. When the sentence to be written is converted into a disputed sentence, source locking is performed on the reverse document evidence. If the locking is successful, the sentence is opened; if the locking fails, the sentence is closed. The reentrant write-prohibited module is used to re-push closed clauses into MixHop power-order propagation, and set them to write-prohibited state if they are still not open after reentry; The controlled generation module is used to drive the large language model to fill only open sentence positions, and then deduct the filled content to the locked and written literature evidence bound to the vote. When the sentence position to be written becomes a disputed sentence position, it is simultaneously deducted to the reverse literature evidence. When the deduction is successful, the scientific research question and answer results and the source of each sentence citation are output. When the deduction fails, the filled content is cleared.

[0007] Optionally, the generation of sentence positions to be written in the sentence position pushing module specifically includes: The system performs question-and-answer segmentation for user research questions, dividing them into conclusion-based, basis-based, and limiting segments. The large language model is called to compress the conclusion entry point, the basis entry point, and the limitation entry point respectively, to obtain the conclusion query content, the basis query content, and the limitation query content. The rule validation submodule independently extracts the research objects, experimental conditions, and evaluation indicators from the limited entry point to obtain the limited validation content. The limited verification content and the limited query content are cross-compared. If the cross-comparison is consistent, the limited query content is used as the verified limited query content. If the cross-comparison is inconsistent, the limited query content is corrected with the limited verification content to obtain the verified limited query content. Push the conclusion query content into the conclusion sentence position, push the basis query content into the basis sentence position, and push the verified limitation query content into the limitation sentence position; Perform pre-generation sealing processing on the conclusion sentence, the basis sentence, and the limiting sentence, closing the text writing entry for the conclusion sentence, the basis sentence, and the limiting sentence, so that the conclusion sentence, the basis sentence, and the limiting sentence remain in an unopened state before obtaining the write reservation ticket; The conclusion, basis, and limitation sentences that have completed the pre-generation sealing process are combined into a sentence to be written. The sentence to be written is then pushed into MixHop power-order propagation and used as the judgment object for writing and occupancy.

[0008] Optionally, the rewriting process of the five types of control actions in the power-order control module specifically includes: Receive the sentence to be written as the object of the write ticket judgment, and replace the neighborhood read action of the sentence to be written in the MixHop power order propagation with the sentence write right judgment action. Rewrite MixHop's first-order propagation as a vote-grabbing action, MixHop's second-order propagation as a hook-back action, MixHop's third-order propagation as a power-grabbing action, MixHop's fourth-order propagation as a locking action, and MixHop's reentrant power order as a reentrant action. The write right status of the pending sentence position is changed in the order of the actions of occupying the ticket, hooking back, seizing the right, locking, and re-entering.

[0009] Optionally, the actions of ticket occupancy, hooking back, seizure, locking, and re-entry specifically include: In the ticket-grabbing action, when the first-order propagation result of MixHop hits the literature evidence, the write ticket-grabbing action is opened to the ticket-grabbing issuance state; when the first-order propagation result of MixHop does not hit the literature evidence, the pending sentence position remains in the closed state. In the hookback action, when the MixHop second-order propagation result reads the bound content to be written, the bound content to be written is hooked back to the pending sentence position. When the MixHop second-order propagation result does not read the bound content to be written, the pending sentence position is blocked from entering the takeover action, locking action and reentry action. In the power seizure action, when the MixHop third-order propagation result reads the documentary evidence that is opposite to the direction of the locked write vote, the documentary evidence with the opposite direction is used as the basis for the change of the writing right of the pending sentence position, triggering the pending sentence position to be converted into the disputed sentence position; In the locking action, when the MixHop fourth-order propagation result completes the source locking, the pending sentence position and the disputed sentence position are changed to open sentence positions; when the MixHop fourth-order propagation result does not complete the source locking, the pending sentence position and the disputed sentence position are changed to closed sentence positions. During the reentry action, the closed sentence is flowed back to the ticket-taking action, hook-back action, and lock action. The closed sentence that is not changed to an open sentence after reentry is output to the reentry write-prohibited module.

[0010] Optionally, the function of hooking back the restricted content bound to the ticket reservation in the ticket reservation hook-back module specifically includes: Bind the relevant documentary evidence to the sentence to be written, and read the claim direction and limiting content of the relevant documentary evidence; A write staking is generated based on the claim direction of the hit document evidence, and the write staking is bound to the hit document evidence; Place the write ticket in the pending state, and block the write ticket in the pending state from entering the takeover and locking actions; The specified content written to the ticket-holding binding will be returned to the specified sentence position for execution and deduction; If the write operation fails, clear the write reservation and restore the pending sentence to an unopened state; When the binding is successful, the writing of the ballot is changed to the locking of the ballot. The claim direction of the document evidence is written to the locking of the ballot direction. The document evidence bound to the locking of the ballot is attached to the power seizure action and the binding action.

[0011] Optionally, the reverse document evidence in the counter-evidence seizure module seizes the right to write the sentence to be written when the limited content is matched, specifically including: Using the locked writing direction as the benchmark for reverse puncturing, reverse puncturing is performed on documentary evidence. When the puncture hits literature evidence in the opposite direction, the literature evidence in the opposite direction is marked as reverse literature evidence; when the puncture does not hit literature evidence in the opposite direction, the writing right status of the sentence to be written is maintained. The reverse literature evidence is attached to the sentence to be written, and the limiting content of the reverse literature evidence is returned to the limiting sentence to perform the matching. When the connection fails, the link between the reverse literature evidence and the sentence to be written is released, and the right to write the locked sentence to be written is maintained. When the connection is successful, the one-way writing right of the pending sentence position is sealed and locked, and the writing right of the pending sentence position is taken over by the reverse literature evidence; The sentence to be written that has been deprived of its writing rights is converted into a disputed sentence, and the reverse literature evidence is attached to the source locking module.

[0012] Optionally, the process of performing source locking in the source locking module specifically includes: Receive the document evidence that is locked and written to the ticket-bound document evidence, extract the document source from the document evidence that is locked and written to the ticket-bound document evidence, and lock the document source to the sentence to be written; When the sentence to be written is transformed into the disputed sentence, reverse literature evidence is received, the source of the reverse literature is extracted from the reverse literature evidence, and the source of the reverse literature is locked to the disputed sentence. If the source of the literature is successfully locked and the sentence to be written has not been converted into a disputed sentence, the sentence to be written will be changed into an open sentence. If the source citation fails to be locked and the sentence to be written has not been converted into a disputed sentence, the sentence to be written will be changed into a closed sentence. When a sentence to be written is converted into a disputed sentence, and both the source and the reverse source are successfully locked, the disputed sentence is converted into an open sentence. If a sentence to be written becomes a disputed sentence and either the source or the reverse source fails to lock, the disputed sentence is changed to a closed sentence.

[0013] Optionally, the reentrancy write-prohibited module specifically includes: Receive the closed sentence and re-push it into the MixHop power order propagation; Set a reentry counter with a maximum count of 3 times. Each time the closed position is re-push in MixHop power propagation, the reentry counter is incremented by 1. The closed position is driven to re-execute the ticket-grabbing action, hook-back action, and locking action; If the re-entry counter has been entered no more than 3 times, and the process of re-acquiring the write ticket, re-completing the binding of the write ticket, and re-completing the source lock has been completed, the closed sentence position will be changed to an open sentence position. If the reentry counter reaches 3 times and the closed sentence position has not been changed to an open sentence position, the closed sentence position will be put into a write-prohibited state. Block the write-prohibited closed sentences from being generated outside the controlled generation module.

[0014] Optionally, the controlled generation module specifically includes: Receive open sentences and keep write-prohibited closed sentences outside the controlled generation module; The large language model is driven to fill in open sentence positions to obtain the filled content; Extract the answer claims from the filled content, and then apply the answer claims back to the document evidence that is locked and written into the voting binding; When an open sentence is a controversial sentence, the answer claim should be simultaneously linked to the opposing literature evidence; If an open sentence is not converted into a disputed sentence and the answer is reversed and written into the locked document evidence bound to the vote, the fill content is retained. If an open sentence is a disputed sentence and the answer claims to be reversed to the locked sentence, and both the documentary evidence and the reverse documentary evidence for voting are successful, the fill content will be retained. The answer suggests that if the reverse deduction fails, the filled content should be cleared, and the open sentences corresponding to the filled content should not be output.

[0015] Optionally, the source of the sentence-by-sentence citation specifically includes: After receiving the successfully deducted content, retain the filled content and solidify it into the answer sentence; Link the answer sentence to the source of the literature to form a sentence-by-sentence citation; When the open sentence is a disputed sentence, the answer sentence will be simultaneously linked to the source of the reverse literature; Clear the answer sentence if the source is missing; Clear the answer sentence when the open sentence is a controversial sentence and the reverse literature source is missing; Output the retained answer sentence, reference source, and sentence-by-sentence citation source. When the open sentence is a disputed sentence, output the reverse reference source simultaneously.

[0016] The beneficial effects of this invention are: (1) This invention first converts the user's research question into a sentence to be written, and then pushes the sentence into the MixHop power order propagation. The right to write the answer is controlled by five types of control actions: voting, hooking back, taking over, locking, and re-entry. When the literature evidence has not been obtained and written into the voting process, it cannot enter the large language model generation process. Unlike the simple search-enhanced generation method, this can reduce the number of semantically similar but unsupported literature fragments being written into the answer.

[0017] (2) This invention introduces a rebuttal-based right-of-way process after the writing vote is locked. The locked writing vote direction is used as the reverse piercing benchmark. After the literature evidence with the opposite direction completes the content matching, it seizes the right to write the sentence position to be written and turns the sentence position to be written into the disputed sentence position. This process makes literature with opposite conclusions no longer just ordinary supplementary material, but directly changes the state of the answer sentence position, which can avoid the one-way conclusion covering the reverse evidence in scientific research Q&A.

[0018] (3) This invention performs controlled generation and answer claim reversal for open sentence positions. The filled content must be reversed to the locked and written literature evidence bound to the vote; under disputed sentence positions, it must also be reversed to the reverse literature evidence at the same time. When the reversal fails, the filled content is cleared. When the literature source is missing, the answer sentence is cleared. This can form a closed loop of answer sentence, literature source and sentence-by-sentence citation source, which improves the verifiability of scientific research question and answer results. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the modules of an intelligent question-answering system for scientific literature based on a large language model proposed in this invention. Figure 2 This is a schematic diagram of sentence position pushing and power-order control in a scientific literature intelligent question-answering system based on a large language model proposed in this invention. Figure 3 This invention provides a closed-loop diagram for the proposed ticket grabbing, counter-evidence seizure, and controlled generation. Detailed Implementation

[0020] 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.

[0021] refer to Figures 1-3 A scientific research literature intelligent question-answering system based on a large language model includes: The sentence-position pushing module is used to perform rule validation on the limited content of the user's research question before the answer is generated, generate the sentence positions to be written, and push the sentence positions to be written into MixHop power-order propagation; MixHop power-order propagation is a graph propagation process that reads content according to different power orders. In this paper, MixHop first-order propagation reads the literature evidence, MixHop second-order propagation reads the limited content bound by the ticket, MixHop third-order propagation reads the literature evidence in the opposite direction, MixHop fourth-order propagation reads the source locking result, and MixHop reentrant power-order receives the closed sentence position. In MixHop power-order propagation, the graph nodes include the sentence to be written node, the document evidence node, the content-limiting node, the document evidence node with the opposite direction, the source-locking result node, and the closed sentence node. A first-order adjacency edge is established between the sentence to be written node and the document evidence node. The sentence to be written node forms a second-order path from the document evidence node to the content-limiting node. The sentence to be written node forms a third-order path from the document evidence node corresponding to the locked writing direction to the document evidence node with the opposite direction. The sentence to be written node forms a fourth-order path from the document evidence node to the source-locking result node. A re-entrant adjacency edge is established between the closed sentence node and the sentence to be written node. The first-order propagation result of MixHop is the document evidence node read by the sentence node to be written through the first-order adjacent edge; the second-order propagation result of MixHop is the limited content node read by the sentence node to be written through the second-order path; the third-order propagation result of MixHop is the document evidence node in the opposite direction read by the sentence node to be written through the third-order path; the fourth-order propagation result of MixHop is the source locking result node read by the sentence node to be written through the fourth-order path; the re-entrant power-order propagation result of MixHop is the re-judged object formed by the backflow of the closed sentence node through the re-entrant adjacent edge. When a first-order propagation result contains a document evidence node, a vote-taking action is triggered; when a second-order propagation result contains a content-limited node, a hook-back action is triggered; when a third-order propagation result contains a document evidence node in the opposite direction, a power-seizure action is triggered, and the document evidence in the opposite direction is marked as reverse document evidence during the power-seizure action; when a fourth-order propagation result contains a source-locking result node, a locking action is triggered; when a re-entry power-order propagation result contains a closed sentence node, a re-entry action is triggered; each power-order propagation result only changes the writing right status of the sentence to be written, and does not directly generate the answer text; In other implementations, MixHop power-order propagation can be replaced by a graph propagation structure with multi-hop node reading capability; the graph propagation structure with multi-hop node reading capability is a propagation structure that can read graph node content according to a one-hop path, two-hop path, three-hop path, four-hop path, or re-entry path; the replaced graph propagation structure still uses the pending sentence node, document evidence node, limited content node, document evidence node with opposite direction, source locking result node, and closed sentence node as reading objects, and triggers voting action, hooking action, takeover action, locking action, and re-entry action according to the propagation result; the write right status of the pending sentence node still changes according to the processing result of the voting action, hooking action, takeover action, locking action, and re-entry action; The power order control module is used to rewrite MixHop's power order propagation into five types of control actions: vote taking, hooking back, taking over, locking, and re-entry. The ticket reservation and hook-back module is used to issue a write reservation ticket when the documentary evidence matches the sentence to be written, hook back the restricted content bound to the write reservation ticket to the sentence to be written, clear the write reservation ticket when the matching fails, and lock the write reservation ticket when the matching succeeds. The counter-evidence seizure module is used to read documentary evidence that is opposite to the direction of the locked writing vote, mark the documentary evidence with the opposite direction as reverse documentary evidence, and seize the right to write the sentence to be written when the reverse documentary evidence completes the content matching, and turn the sentence to be written into the disputed sentence. The source locking module is used to perform source locking on the document evidence that is locked and bound to the ticket holder. When the sentence to be written is converted into a disputed sentence, source locking is performed on the reverse document evidence. If the locking is successful, the sentence is opened; if the locking fails, the sentence is closed. The reentrant write-prohibited module is used to re-push closed clauses into MixHop power-order propagation, and set them to write-prohibited state if they are still not open after reentry; The reentrant write-prohibited module includes a failure log recording unit, which records the failure types when the closing position is in a write-prohibited state. Failure types include failure to reacquire a write pass, failure to re-complete the restricted content locking, and failure to re-complete the source locking. The failure log recording unit is the recording unit in the reentrant write-prohibited module. When recording the failure type, the closing position identifier, the current count value of the reentrant counter, and the failure type are written to the write-prohibited record. The closing position identifier is the number corresponding to the closing position, and the write-prohibited record is the record content formed after the closing position enters the write-prohibited state. The controlled generation module is used to drive the large language model to fill only open sentence positions, and then deduct the filled content to the locked and written literature evidence bound to the vote. When the sentence position to be written becomes a disputed sentence position, it is simultaneously deducted to the reverse literature evidence. When the deduction is successful, the scientific research question and answer results and the source of each sentence citation are output. When the deduction fails, the filled content is cleared.

[0022] In this embodiment, the generation of sentence positions to be written in the sentence position pushing module specifically includes: The system performs question-and-answer segmentation for user research questions, dividing them into conclusion-based, basis-based, and limiting segments. The question-and-answer segmentation is an operation that segments the user's research question. During execution, the segmentation is based on the conclusion question, the basis question, and the limiting question in the user's research question. The conclusion segment is the sentence segment that requires obtaining research judgment; the basis segment is the sentence segment that requires obtaining literature support; and the limiting segment is the sentence segment that includes the research object, experimental conditions, and evaluation indicators. The large language model is called to compress the conclusion entry point, the basis entry point, and the limitation entry point respectively, to obtain the conclusion query content, the basis query content, and the limitation query content. When using a large language model for compression, the text of a single entry point is input into the large language model; the output content of the conclusion entry point retains the research judgment and forms the conclusion inquiry content; the output content of the basis entry point retains the literature support requirements and forms the basis inquiry content; the output content of the limitation entry point retains the research object, experimental conditions and evaluation indicators and forms the limitation inquiry content; the three types of inquiry content are only written as sentence positions and are not output as answer text. The rule validation submodule independently extracts the research objects, experimental conditions, and evaluation indicators from the limited entry point to obtain the limited validation content. The rule validation submodule is the validation unit in the sentence insertion module. It is used to perform independent extraction of the research objects, experimental conditions and evaluation indicators in the limited entry point, and cross-compare the independent extraction results with the limited query content. The limited validation content is the content independently extracted from the limited entry point by the rule validation submodule, including the research object, experimental conditions, and evaluation indicators; the limited validation content is used to correct the limited query content. During independent extraction, the rule validation submodule performs field identification on the defined cuts. The field identification objects include the research object field, experimental condition field, and evaluation index field. The research object field is identified through entity name recognition, the experimental condition field is identified through condition words, unit values, interval values, and percentage values, and the evaluation index field is identified through index name recognition. Entity name recognition reads the research object name, experimental material name, and method name from the defined cut. Unit value, interval value, and percentage value recognition reads the numerical content with units of measurement, range symbols, upper and lower limits, and percentage expressions from the defined cut. Field identification is performed through rule matching and entity name recognition. Rule matching includes matching condition words, units of measurement, range symbols, upper and lower limits, and percentage expressions, used to read the unit values, interval values, and percentage values ​​from the defined cut. The limited verification content and the limited query content are cross-compared. If the cross-comparison is consistent, the limited query content is used as the verified limited query content. If the cross-comparison is inconsistent, the limited query content is corrected with the limited verification content to obtain the verified limited query content. Cross-comparison involves comparing the research subjects, experimental conditions, and evaluation indicators in the limited verification content with those in the limited inquiry content to determine their consistency. Cross-comparison is considered consistent when the research subjects, experimental conditions, and evaluation indicators are all consistent. Cross-comparison is considered inconsistent when any of the following situations exist: inconsistent research subjects, inconsistent experimental conditions, inconsistent evaluation indicators, or missing corresponding content in the limited inquiry content. Correcting the limited query content with limited verification content means replacing the inconsistent or missing research subjects, experimental conditions, and evaluation indicators in the limited query content with the corresponding research subjects, experimental conditions, and evaluation indicators in the limited verification content. Push the conclusion query content into the conclusion sentence position, push the basis query content into the basis sentence position, and push the verified limitation query content into the limitation sentence position; The conclusion sentence, the basis sentence, and the limiting sentence are controlled writing locations established before the answer is generated; Perform pre-generation sealing processing on the conclusion sentence, the basis sentence, and the limiting sentence, closing the text writing entry for the conclusion sentence, the basis sentence, and the limiting sentence, so that the conclusion sentence, the basis sentence, and the limiting sentence remain in an unopened state before obtaining the write reservation ticket; The pre-generation sealing process closes the text writing entry points for the conclusion sentence, the basis sentence, and the limiting sentence; the text writing entry points are the writing channels for the large language model to fill the answer text into the sentence positions; during the pre-generation sealing process, a "not open" status flag is written to the conclusion sentence, the basis sentence, and the limiting sentence; the "not open" status flag indicates that the large language model prohibits the filling of answer text into the corresponding sentence position; The conclusion, basis, and limitation sentences that have completed the pre-generation sealing process are combined into a sentence to be written. The sentence to be written is then pushed into MixHop power-order propagation and used as the judgment object for writing and occupancy.

[0023] When composing the sentences to be written, the order of the three types of sentences and the "not open" status marker are retained.

[0024] In this embodiment, the rewriting process of the five types of control actions in the power-order control module specifically includes: Receive the sentence to be written as the object of the write ticket judgment, and replace the neighborhood read action of the sentence to be written in the MixHop power order propagation with the sentence write right judgment action. The neighborhood read action represents the operation of reading the aforementioned content in a power order by the MixHop power order propagation; the sentence position write right judgment action represents the judgment of the write right status of the sentence position to be written, without generating the answer text; during replacement, the sentence position to be written is used as the judgment object, and the first-order propagation, second-order propagation, third-order propagation, fourth-order propagation and re-entry power order are respectively connected to the corresponding control actions. Rewrite MixHop's first-order propagation as a vote-grabbing action, MixHop's second-order propagation as a hook-back action, MixHop's third-order propagation as a power-grabbing action, MixHop's fourth-order propagation as a locking action, and MixHop's reentrant power order as a reentrant action. The write right status of the pending sentence position is changed in the order of the actions of occupying the ticket, hooking back, seizing the right, locking, and re-entering.

[0025] The write right status represents the processing result of the pending write position in the five types of control actions. Write right status includes: not open, write pass issued, open position, closed position, write-prohibited, and disputed position. The not open status indicates that the pending write position has not obtained write permission; the write pass issued status indicates that the pending write position is allowed to generate a write pass; the open position indicates that the pending write position is allowed to enter the controlled generation module; the closed position indicates that the pending write position needs to enter the reentry action; the write-prohibited status indicates that the pending write position is blocked from entering the controlled generation module; and the disputed position indicates that the pending write position has had its write right seized by reverse document evidence.

[0026] In this embodiment, the actions of ticket taking, hooking back, seizing power, locking, and re-entering specifically include: In the ticket-grabbing action, when the first-order propagation result of MixHop hits the literature evidence, the write ticket-grabbing action is opened to the ticket-grabbing issuance state; when the first-order propagation result of MixHop does not hit the literature evidence, the pending sentence position remains in the closed state. Hit literature evidence refers to literature evidence read by MixHop first-order propagation, and the claims in the literature evidence are consistent with the research judgment in the conclusion sentence, and the literature sources in the literature evidence are consistent with the literature support requirements in the basis sentence; Miss literature evidence refers to literature evidence that MixHop first-order propagation did not read, which meets the above consistency conditions; Write ticket issuance status indicates that the ticket issuance action allows the generation of write ticket based on hit literature evidence. In the hookback action, when the MixHop second-order propagation result reads the bound content to be written, the bound content to be written is hooked back to the pending sentence position. When the MixHop second-order propagation result does not read the bound content to be written, the pending sentence position is blocked from entering the takeover action, locking action and reentry action. The limiting content of the written vote binding comes from the literature evidence of the written vote binding, and the limiting content includes the research subjects, experimental conditions and evaluation indicators; the hook in the hook action is to return the limiting content of the written vote binding to the limiting sentence position; In the power seizure action, when the MixHop third-order propagation result reads the documentary evidence that is opposite to the direction of the locked write vote, the documentary evidence with the opposite direction is used as the basis for the change of the writing right of the pending sentence position, triggering the pending sentence position to be converted into the disputed sentence position; Locked write to the ballot is the ballot state formed after the limited content of the ballot is completed; locked write to the ballot direction is the direction value of the locked write to the ballot record, and the direction value includes the support direction and the negation direction; document evidence with the opposite direction is document evidence that forms a judgment opposite to the locked write to the ballot direction. In the locking action, when the MixHop fourth-order propagation result completes the source locking, the pending sentence position and the disputed sentence position are changed to open sentence positions; when the MixHop fourth-order propagation result does not complete the source locking, the pending sentence position and the disputed sentence position are changed to closed sentence positions. Source locking is the operation of locking the source of a document with the sentence to be written; the source of a document includes the document identifier, paragraph position, and original text excerpt index; source locking is successful when all three items are bound to the sentence to be written, and source locking fails when any one item is missing; During the reentry action, the closed sentence is flowed back to the ticket-taking action, hook-back action, and lock action. The closed sentence that is not changed to an open sentence after reentry is output to the reentry write-prohibited module.

[0027] Closed sentence reflow is the operation of re-entering the closed sentence into the ticket-taking, hook-back, and locking actions; if it is not changed to an open sentence after re-entry, it means that the closed sentence still does not meet the opening conditions after re-executing the ticket-taking, hook-back, and locking actions.

[0028] In this embodiment, the process of hooking back the restricted content bound to the ticket reservation in the ticket reservation hook-back module specifically includes: Bind the relevant documentary evidence to the sentence to be written, and read the claim direction and limiting content of the relevant documentary evidence; The proposed direction is to support or refute the research judgment formed by the literature evidence in the conclusion sentence; the limited content follows the aforementioned research subjects, experimental conditions and evaluation indicators. A write staking is generated based on the claim direction of the hit document evidence, and the write staking is bound to the hit document evidence; The placeholder record is a record created between the hit document evidence and the sentence to be written; after the placeholder record is bound to the hit document evidence, it records the claim direction and limiting content of the hit document evidence; Place the write ticket in the pending state, and block the write ticket in the pending state from entering the takeover and locking actions; The pending deduction status indicates that the written ticket has not yet completed the deduction of the specified content; the written ticket in the pending deduction status will only enter the hook-back action; The specified content written to the ticket-holding binding will be returned to the specified sentence position for execution and deduction; The process involves determining the consistency between the research subjects, experimental conditions, and evaluation indicators and the corresponding content in the specified sentence. If the write operation fails, clear the write reservation and restore the pending sentence to an unopened state; Matching failures include inconsistencies in research subjects, experimental conditions, evaluation indicators, and missing content for any corresponding item; clearing the write ticket involves deleting the binding record between the write ticket and the matching literature evidence; When the binding is successful, the writing of the ballot is changed to the locking of the ballot. The claim direction of the document evidence is written to the locking of the ballot direction. The document evidence bound to the locking of the ballot is attached to the power seizure action and the binding action.

[0029] Successful locking means that the research object, experimental conditions, and evaluation indicators are all consistent; binding means that the locked document evidence bound to the ticket is registered as the reading object of the takeover action and locking action.

[0030] In this embodiment, when the reverse documentary evidence in the counter-evidence seizure module completes the content matching, it seizes the right to write the sentence to be written, specifically including: Using the locked writing direction as the benchmark for reverse puncturing, reverse puncturing is performed on documentary evidence. Reverse puncture is an operation that compares the direction of each piece of documentary evidence according to the locked writing direction; during the comparison, the supporting direction and the negative direction are treated as opposite judgments, and the negative direction and the supporting direction are treated as opposite judgments. When the puncture hits literature evidence in the opposite direction, the literature evidence in the opposite direction is marked as reverse literature evidence; when the puncture does not hit literature evidence in the opposite direction, the writing right status of the sentence to be written is maintained. If the puncture hits document evidence in the opposite direction, it means that the direction of the document evidence being compared is opposite to the direction of the locked write ticket; if the puncture does not hit document evidence in the opposite direction, it means that no document evidence satisfying the opposite judgment was read. The reverse literature evidence is attached to the sentence to be written, and the limiting content of the reverse literature evidence is returned to the limiting sentence to perform the matching. When reverse literature evidence is attached to the sentence to be written, it means that the reverse literature evidence is registered as the reading object of the sentence to be written. When the limited content of the reverse literature evidence is returned to the limited sentence for matching, the matching rule follows the consistency judgment of the aforementioned research object, experimental conditions and evaluation indicators. When the connection fails, the link between the reverse literature evidence and the sentence to be written is released, and the right to write the locked sentence to be written is maintained. Unlinking the reverse literature evidence from the sentence to be written means deleting the registration record of the reverse literature evidence as the object to be read by the sentence to be written; if the matching fails, the right to write the sentence to be written remains unchanged. When the connection is successful, the one-way writing right of the pending sentence position is sealed and locked, and the writing right of the pending sentence position is taken over by the reverse literature evidence; One-way write right means that the lock write ticket controls the state of the write right of the pending sentence; sealing one-way write right means suspending the control of the write right of the pending sentence by the lock write ticket, and retaining the document evidence bound to the lock write ticket; reverse document evidence takes over the write right of the pending sentence, meaning that the reverse document evidence has acquired the write right of the pending sentence. The sentence to be written that has been deprived of its writing rights is converted into a disputed sentence, and the reverse literature evidence is attached to the source locking module.

[0031] The disputed sentence position is the sentence position state formed after the right to write a sentence position is seized by reverse literature evidence. The reverse literature evidence is attached to the source locking module, indicating that the reverse literature evidence is registered as a read object of the source locking module.

[0032] In this embodiment, the process of performing source locking in the source locking module specifically includes: Receive the document evidence that is locked and written to the ticket-bound document evidence, extract the document source from the document evidence that is locked and written to the ticket-bound document evidence, and lock the document source to the sentence to be written; The source of the document uses the aforementioned document identifier, paragraph position, and original text fragment index; extracting the source of the document involves reading the document identifier, paragraph position, and original text fragment index from the document evidence that is locked and bound to the ticket; locking the source of the document to the sentence to be written means that the source of the document is registered to the sentence to be written. When the sentence to be written is transformed into the disputed sentence, reverse literature evidence is received, the source of the reverse literature is extracted from the reverse literature evidence, and the source of the reverse literature is locked to the disputed sentence. The source of the reverse reference is the document identifier, paragraph position, and original text fragment index read from the reverse reference evidence; the reverse reference source is locked to the disputed sentence position, indicating that the reverse reference source is registered at the disputed sentence position; If the source of the literature is successfully locked and the sentence to be written has not been converted into a disputed sentence, the sentence to be written will be changed into an open sentence. The source of the document has been successfully locked, meaning that the document identifier, paragraph position, and original text excerpt index have all been registered to the sentence to be written. If the source citation fails to be locked and the sentence to be written has not been converted into a disputed sentence, the sentence to be written will be changed into a closed sentence. The failure to match the source of a document is due to the absence of any one of the following: document identifier, paragraph position, or original text excerpt index; or any one of these items is not registered in the sentence to be written. When a sentence to be written is converted into a disputed sentence, and both the source and the reverse source are successfully locked, the disputed sentence is converted into an open sentence. The reverse source locking was successful, meaning that the document identifier, paragraph position, and original text fragment index have all been registered to the disputed sentence position; If a sentence to be written becomes a disputed sentence and either the source or the reverse source fails to lock, the disputed sentence is changed to a closed sentence.

[0033] The failure to lock the source of the reverse reference is due to the absence of any one of the following: the reference identifier, the paragraph position, or the original text fragment index, or any one of these is not registered to the disputed sentence position.

[0034] In this embodiment, the reentrancy write-prohibited module specifically includes: Receive the closed sentence and re-push it into the MixHop power order propagation; The closed position is re-push into MixHop power-order propagation, which means that the closed position is used as a new write to the ticket-taking judgment object, and the ticket-taking action, hook-back action and locking action are re-entered. Set a reentry counter with a maximum count of 3 times. Each time the closed position is re-push in MixHop power propagation, the reentry counter is incremented by 1. The closed position is driven to re-execute the ticket-grabbing action, hook-back action, and locking action; If the re-entry counter has been entered no more than 3 times, and the process of re-acquiring the write ticket, re-completing the binding of the write ticket, and re-completing the source lock has been completed, the closed sentence position will be changed to an open sentence position. After re-acquiring the write ticket as a closed sentence position and re-executing the ticket acquisition action, a write ticket is generated; the bound content of the re-acquisition of the write ticket is matched with the corresponding content in the bound sentence position, including the research object, experimental conditions and evaluation indicators; the source is locked again, and the document identifier, paragraph position and original text fragment index are all registered to the closed sentence position. If the reentry counter reaches 3 times and the closed sentence position has not been changed to an open sentence position, the closed sentence position will be put into a write-prohibited state. The write-prohibited state prevents closed sentence positions from entering the controlled generation module; after a closed sentence position is in the write-prohibited state, the large language model will not fill the closed sentence position with answer text. Block the write-prohibited closed sentences from being generated outside the controlled generation module.

[0035] Blocked outside the controlled generation module, the closed position indicating a write-prohibited state is prohibited from entering the controlled generation module.

[0036] In this embodiment, the controlled generation module specifically includes: Receive open sentences and keep write-prohibited closed sentences outside the controlled generation module; Open sentences are those that are allowed to be filled with answer text after source locking is completed; closed sentences in a forbidden state are kept outside the controlled generation module, indicating that closed sentences in a forbidden state do not enter the large language model filling process; The large language model is driven to fill in open sentence positions to obtain the filled content; The populated content is the answer text written into open sentence positions by the large language model; Extract the answer claims from the filled content, and then apply the answer claims back to the document evidence that is locked and written into the voting binding; The answer statement is the text portion expressing the research judgment in the fill-in content; the direction marker is the supporting or opposing direction; the reverse check is to verify the correspondence between the answer statement and the content of the statement written into the literature evidence bound to the vote and the direction marker; the correspondence check is to determine whether the answer statement retains the content of the statement in the corresponding literature evidence and whether the direction marker is consistent; When determining the reverse deduction, the direction mark in the document evidence that is locked and bound to the ticket is used as the benchmark; when the open sentence is a disputed sentence, the direction mark in the reverse document evidence is used as the benchmark for synchronous reverse deduction; when the direction mark in the answer claim is inconsistent with the direction mark in the corresponding document evidence, the reverse deduction is deemed to have failed. The fill-in content generated by the large language model is only used as the answer text to be verified, and is not used as the basis for judging the writing right status, the locked writing direction, the reverse literature evidence direction, the source locking result, and the reverse deduction result; the writing right status, the locked writing direction, the reverse literature evidence direction, the source locking result, and the reverse deduction result are all determined by the literature evidence, the reverse literature evidence, the limited content locking result, and the source locking result. When an open sentence is a controversial sentence, the answer claim should be simultaneously linked to the opposing literature evidence; Synchronous reverse deduction involves verifying the corresponding claims and directional markers in the document evidence and reverse document evidence that are locked and bound to the vote-taking process. If an open sentence is not converted into a disputed sentence and the answer is reversed and written into the locked document evidence bound to the vote, the fill content is retained. Successful counter-deduction retains the claim content locked in the documentary evidence bound to the vote, and the direction markers are consistent; If an open sentence is a disputed sentence and the answer claims to be reversed to the locked sentence, and both the documentary evidence and the reverse documentary evidence for voting are successful, the fill content will be retained. The successful counter-offering under the disputed sentence position retains the claim content and direction marker written into the documentary evidence bound to the vote, and also retains the claim content and direction marker in the reverse documentary evidence; The answer suggests that if the reverse deduction fails, the filled content should be cleared, and the open sentences corresponding to the filled content should not be output.

[0037] A failed reverse deduction occurs when either the answer claim fails to retain the claim content from the corresponding literature evidence, or the direction marker of the answer claim is inconsistent with the direction marker of the corresponding literature evidence; clearing the fill content means deleting the answer text in the open sentence; open sentences are not output, indicating that the open sentence does not form the answer sentence in the research question and answer result.

[0038] In this implementation method, the source is cited sentence by sentence, specifically including: After receiving the successfully deducted content, retain the filled content and solidify it into the answer sentence; After a successful reverse deduction, the retained fill content is the answer text that was successfully reversed and not cleared in the controlled generation module; the answer sentence is a single-sentence answer text formed by the fill content; "fixed as answer sentence" means that the fill content is determined as a single-sentence answer text; Link the answer sentence to the source of the literature to form a sentence-by-sentence citation; The answer sentence is bound to the source of the literature, which means that a corresponding record is established between the answer sentence and the literature identifier, paragraph position and original text excerpt index in the source of the literature; the source of each sentence is the corresponding record formed between the answer sentence and the source of the literature; When the open sentence is a disputed sentence, the answer sentence will be simultaneously linked to the source of the reverse literature; The answer sentence is synchronously bound to the source of the reverse reference, which means that a corresponding record is established between the answer sentence and the reference identifier, paragraph position and original text fragment index in the source of the reverse reference; Clear the answer sentence if the source is missing; Missing citation means any one of the following is missing: citation identifier, paragraph position, or original text excerpt index; Clearing the answer sentence means deleting the answer sentence. Clear the answer sentence when the open sentence is a controversial sentence and the reverse literature source is missing; Missing reverse source refers to the absence of any one of the following in the reverse source: the reference identifier, paragraph position, or original text fragment index. Output the retained answer sentence, reference source, and sentence-by-sentence citation source. When the open sentence is a disputed sentence, output the reverse reference source simultaneously.

[0039] Example 1: To verify the feasibility of this invention in practice, it was applied to a scientific research literature question-and-answer knowledge base. This knowledge base includes abstracts, method sections, experimental result sections, conclusion sections, and reference information from papers in different disciplines. The usage scenario involves researchers inputting a question with conclusions, follow-up questions, and limiting conditions. The system needs to find evidence from candidate literature to support the answer and provide an answer with sentence-by-sentence citations. Common problems in this scenario are: the large language model mixes conclusions from different literature when generating answers, resulting in seemingly cited answers without a sentence-by-sentence correspondence between sentences and sources; when encountering contradictory conclusions, the model tends to retain only one side of the evidence, omitting controversial information; and when the question includes experimental conditions, evaluation indicators, or research object limitations, the model tends to cite similar but different limiting conditions literature, resulting in an imprecise answer.

[0040] In this scenario, after receiving a user's research question, the system first breaks it down into three segments: conclusion, basis, and limitation. The conclusion segment retains the research conclusions that need to be judged; the basis segment retains the content that needs to be supported by cited literature; and the limitation segment retains the research object, experimental conditions, and evaluation indicators. After compression, these three segments are entered into the conclusion, basis, and limitation sentence positions, respectively. All three sentence positions are in a closed state before the answer is generated, and the large language model cannot directly write them into the answer text. After the sentence positions enter the MixHop power-order propagation, the system no longer directly uses the propagation results to generate the answer. Instead, it rewrites the different power-order propagations into five control actions: vote taking, hooking back, seizing power, locking, and re-entry.

[0041] When the first-order propagation reads literature evidence that matches the conclusion and basis sentences, the system issues a write ticket. Issuing a write ticket does not immediately generate an answer; the system will also backtrack the specified content bound to the write ticket to the specified sentence for matching. Only when the research subjects, experimental conditions, and evaluation indicators all match the content corresponding to the specified sentence is the write ticket locked. If the matching fails, the write ticket is cleared, and the pending sentence returns to an inactive state. This prevents the misuse of literature with similar conclusions but different experimental conditions.

[0042] When opposing documentary evidence appears, the system uses the locked writing direction as the reverse puncture benchmark and compares the direction of each candidate documentary evidence. If there is documentary evidence with a supporting direction opposite to the negative direction, and the limited content of the opposing documentary evidence can also be matched with the limited sentence position, the opposing documentary evidence obtains the right to write the sentence position, and the sentence position becomes the disputed sentence position. The disputed sentence position will not be directly output; it must be locked through source matching to ensure that both the positive and opposing documentary sources can be registered in the corresponding sentence position. If any of the documentary identifier, paragraph position, or original text fragment index is missing, the sentence position becomes a closed sentence position, and the closed sentence position re-enters the MixHop power-order propagation; if the voting, matching, and source matching cannot be completed after re-entry, the closed sentence position is in a forbidden state and does not enter the controlled generation module.

[0043] The controlled generation module only accepts open sentence positions. After the large language model fills the open sentence position with answer text, the system extracts the answer claim from the filled content and then applies the answer claim back to the locked, vote-bound literature evidence. If the open sentence position is a disputed sentence position, the answer claim must also be back-bound to the reverse literature evidence simultaneously. Only after successful back-bound is the filled content retained and solidified as the answer sentence; if back-bound fails, the filled content is cleared and no answer is output. The answer sentence must also be bound to a literature source, and if the open sentence position is a disputed sentence position, a reverse literature source must also be bound simultaneously. If the literature source or the reverse literature source is missing, the answer sentence is cleared. Through this process, answer generation no longer depends solely on the model's language fluency, but is jointly controlled by sentence position openness, vote-bound locking, reverse takeover, source locking, and back-bound verification.

[0044] In the validation, samples were selected that included biomedical, material properties, experimental conditions, model effects, contrasting conclusions, interpretation of evaluation indicators, cross-document evidence integration, conditional retrieval, source tracing, and comprehensive scientific research questions and answers. Conventional question-and-answer methods only require the model to provide the answer and citation source; this invention requires each open sentence to complete the following: entry reservation, content matching, source locking, reverse verification, and sentence-by-sentence citation binding. In the statistical definitions, the proportion of answers without sources indicates the proportion where the generated answer cannot be traced to a clear document identifier, paragraph position, and original text fragment index; the erroneous claim output rate indicates the proportion where the answer claim is inconsistent with the corresponding document evidence claim; the sentence-by-sentence citation completeness rate indicates the proportion where every sentence in the answer corresponds to a document source; and the disputed evidence identification rate indicates the proportion where the system successfully identifies and retains the reverse document source when contradictory evidence exists.

[0045] Table 1: Comparison of the effects of controlled sentence-based question-and-answer and sentence-by-sentence quotation

[0046] The table data shows that this invention significantly improves the control of answers without citations. Conventional question-and-answer methods have a high proportion of answers without citations in tasks such as cross-document evidence integration, contrasting conclusions, and conditional retrieval, reaching as high as 27.5%. This invention controls answer output through unopened states, write-to-reserve states, citation locking, and write-prohibited states. Sentences without citation locking cannot enter the controlled generation module, reducing the proportion of answers without citations in all ten test categories to below 6.4%. This result indicates that the system does not simply require the model to include the source after the answer, but rather uses document citation as a prerequisite for open sentences.

[0047] The rate of incorrect claims also decreased significantly. The rate of incorrect claims in the conventional question-and-answer method was 18.2% in contrasting conclusion analysis, 16.9% in cross-document evidence integration, and 15.6% in experimental condition follow-up questions. This invention introduces limited content matching and answer claim reversal; only when the answer claim corresponds to the claim content and direction markers in the documentary evidence is the filled content retained. The corresponding indicators for this invention in the table are all below 4.0%, indicating that reversal verification has a direct effect on reducing incorrect claim output.

[0048] The sentence-by-sentence citation integrity rate remained between 91.6% and 96.1%, indicating a relatively stable correspondence between the answer sentence and the bibliographic identifier, paragraph position, and original text fragment index. The disputed evidence identification rate reached 89.8% and 88.4% in scenarios involving opposing conclusion analysis and cross-document evidence integration, respectively, demonstrating that reverse puncturing, linking reverse bibliographic evidence, and locking disputed sentence positions can preserve opposing evidence and prevent the answer from presenting only one-sided conclusions. In summary, this invention provides targeted improvements to the problems of "incomplete citations, mismatched evidence, and omitted disputed evidence" in scientific research question answering. The data results correspond to sentence position control, vote-taking and reversal, counter-evidence overriding, source locking, and controlled generation processes.

[0049] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A scientific research literature intelligent question-answering system based on a large language model, characterized in that, include: The sentence-position pushing module is used to perform rule validation on the limited content of the user's research question before the answer is generated, generate the sentence positions to be written, and push the sentence positions to be written into MixHop power-order propagation; The power order control module is used to rewrite MixHop's power order propagation into five types of control actions: vote taking, hooking back, taking over, locking, and re-entry. The ticket reservation and hook-back module is used to issue a write reservation ticket when the documentary evidence matches the sentence to be written, hook back the restricted content bound to the write reservation ticket to the sentence to be written, clear the write reservation ticket when the matching fails, and lock the write reservation ticket when the matching succeeds. The counter-evidence seizure module is used to read documentary evidence that is opposite to the direction of the locked writing vote, mark the documentary evidence with the opposite direction as reverse documentary evidence, and seize the right to write the sentence to be written when the reverse documentary evidence completes the content matching, and turn the sentence to be written into the disputed sentence. The source locking module is used to perform source locking on the document evidence that is locked and bound to the ticket holder. When the sentence to be written is converted into a disputed sentence, source locking is performed on the reverse document evidence. If the locking is successful, the sentence is opened; if the locking fails, the sentence is closed. The reentrant write-prohibited module is used to re-push closed clauses into MixHop power-order propagation, and set them to write-prohibited state if they are still not open after reentry; The controlled generation module is used to drive the large language model to fill only open sentence positions, and then deduct the filled content to the locked and written literature evidence bound to the vote. When the sentence position to be written becomes a disputed sentence position, it is simultaneously deducted to the reverse literature evidence. When the deduction is successful, the scientific research question and answer results and the source of each sentence citation are output. When the deduction fails, the filled content is cleared.

2. The intelligent question-answering system for scientific research literature based on a large language model according to claim 1, characterized in that, The generation of sentence positions to be written in the sentence position pushing module specifically includes: The system performs question-and-answer segmentation for user research questions, dividing them into conclusion-based, basis-based, and limiting segments. The large language model is called to compress the conclusion entry point, the basis entry point, and the limitation entry point respectively, to obtain the conclusion query content, the basis query content, and the limitation query content. The rule validation submodule independently extracts the research objects, experimental conditions, and evaluation indicators from the limited entry point to obtain the limited validation content. The limited verification content and the limited query content are cross-compared. If the cross-comparison is consistent, the limited query content is used as the verified limited query content. If the cross-comparison is inconsistent, the limited query content is corrected with the limited verification content to obtain the verified limited query content. Push the conclusion query content into the conclusion sentence position, push the basis query content into the basis sentence position, and push the verified limitation query content into the limitation sentence position; Perform pre-generation sealing processing on the conclusion sentence, the basis sentence, and the limiting sentence, closing the text writing entry for the conclusion sentence, the basis sentence, and the limiting sentence, so that the conclusion sentence, the basis sentence, and the limiting sentence remain in an unopened state before obtaining the write reservation ticket; The conclusion, basis, and limitation sentences that have completed the pre-generation sealing process are combined into a sentence to be written. The sentence to be written is then pushed into MixHop power-order propagation and used as the judgment object for writing and occupancy.

3. The intelligent question-answering system for scientific research literature based on a large language model according to claim 2, characterized in that, The rewriting process for the five types of control actions in the power-order control module specifically includes: Receive the sentence to be written as the object of the write ticket judgment, and replace the neighborhood read action of the sentence to be written in the MixHop power order propagation with the sentence write right judgment action. Rewrite MixHop's first-order propagation as a vote-grabbing action, MixHop's second-order propagation as a hook-back action, MixHop's third-order propagation as a power-grabbing action, MixHop's fourth-order propagation as a locking action, and MixHop's reentrant power order as a reentrant action. The write right status of the pending sentence position is changed in the order of the actions of occupying the ticket, hooking back, seizing the right, locking, and re-entering.

4. The intelligent question-answering system for scientific research literature based on a large language model according to claim 3, characterized in that, The actions of ticket grabbing, hooking back, seizing power, locking, and re-entering specifically include: In the ticket-grabbing action, when the first-order propagation result of MixHop hits the literature evidence, the write ticket-grabbing action is opened to the ticket-grabbing issuance state; when the first-order propagation result of MixHop does not hit the literature evidence, the pending sentence position remains in the closed state. In the hookback action, when the MixHop second-order propagation result reads the bound content to be written, the bound content to be written is hooked back to the pending sentence position. When the MixHop second-order propagation result does not read the bound content to be written, the pending sentence position is blocked from entering the takeover action, locking action and reentry action. In the power seizure action, when the MixHop third-order propagation result reads the documentary evidence that is opposite to the direction of the locked write vote, the documentary evidence with the opposite direction is used as the basis for the change of the writing right of the pending sentence position, triggering the pending sentence position to be converted into the disputed sentence position; In the locking action, when the MixHop fourth-order propagation result completes the source locking, the pending sentence position and the disputed sentence position are changed to open sentence positions; when the MixHop fourth-order propagation result does not complete the source locking, the pending sentence position and the disputed sentence position are changed to closed sentence positions. During the reentry action, the closed sentence is flowed back to the ticket-taking action, hook-back action, and lock action. The closed sentence that is not changed to an open sentence after reentry is output to the reentry write-prohibited module.

5. The intelligent question-answering system for scientific research literature based on a large language model according to claim 4, characterized in that, The "ticket reservation hook-back" module hooks back the restricted content bound to the ticket reservation to the pending sentence position, specifically including: Bind the relevant documentary evidence to the sentence to be written, and read the claim direction and limiting content of the relevant documentary evidence; A write staking is generated based on the claim direction of the hit document evidence, and the write staking is bound to the hit document evidence; Place the write ticket in the pending state, and block the write ticket in the pending state from entering the takeover and locking actions; The specified content written to the ticket-holding binding will be returned to the specified sentence position for execution and deduction; If the write operation fails, clear the write reservation and restore the pending sentence to an unopened state; When the binding is successful, the writing of the ballot is changed to the locking of the ballot. The claim direction of the document evidence is written to the locking of the ballot direction. The document evidence bound to the locking of the ballot is attached to the power seizure action and the binding action.

6. The intelligent question-answering system for scientific research literature based on a large language model according to claim 5, characterized in that, The counter-evidence seizure module, when the reverse documentary evidence completes the content matching, seizes the right to write the sentence to be written, specifically including: Using the locked writing direction as the benchmark for reverse puncturing, reverse puncturing is performed on documentary evidence. When the puncture hits literature evidence in the opposite direction, the literature evidence in the opposite direction is marked as reverse literature evidence; when the puncture does not hit literature evidence in the opposite direction, the writing right status of the sentence to be written is maintained. The reverse literature evidence is attached to the sentence to be written, and the limiting content of the reverse literature evidence is returned to the limiting sentence to perform the matching. When the connection fails, the link between the reverse literature evidence and the sentence to be written is released, and the right to write the locked sentence to be written is maintained. When the connection is successful, the one-way writing right of the pending sentence position is sealed and locked, and the writing right of the pending sentence position is taken over by the reverse literature evidence; The sentence to be written that has been deprived of its writing rights is converted into a disputed sentence, and the reverse literature evidence is attached to the source locking module.

7. The intelligent question-answering system for scientific research literature based on a large language model according to claim 6, characterized in that, The process of performing source locking in the source locking module specifically includes: Receive the document evidence that is locked and written to the ticket-bound document evidence, extract the document source from the document evidence that is locked and written to the ticket-bound document evidence, and lock the document source to the sentence to be written; When the sentence to be written is transformed into the disputed sentence, reverse literature evidence is received, the source of the reverse literature is extracted from the reverse literature evidence, and the source of the reverse literature is locked to the disputed sentence. If the source of the literature is successfully locked and the sentence to be written has not been converted into a disputed sentence, the sentence to be written will be changed into an open sentence. If the source citation fails to be locked and the sentence to be written has not been converted into a disputed sentence, the sentence to be written will be changed into a closed sentence. When a sentence to be written is converted into a disputed sentence, and both the source and the reverse source are successfully locked, the disputed sentence is converted into an open sentence. If a sentence to be written becomes a disputed sentence and either the source or the reverse source fails to lock, the disputed sentence is changed to a closed sentence.

8. The intelligent question-answering system for scientific research literature based on a large language model according to claim 7, characterized in that, The reentrant write-prohibition module specifically includes: Receive the closed sentence and re-push it into the MixHop power order propagation; Set a reentry counter with a maximum count of 3 times. Each time the closed position is re-push in MixHop power propagation, the reentry counter is incremented by 1. The closed position is driven to re-execute the ticket-grabbing action, hook-back action, and locking action; If the re-entry counter has been entered no more than 3 times, and the process of re-acquiring the write ticket, re-completing the binding of the write ticket, and re-completing the source lock has been completed, the closed sentence position will be changed to an open sentence position. If the reentry counter reaches 3 times and the closed sentence position has not been changed to an open sentence position, the closed sentence position will be put into a write-prohibited state. Block the write-prohibited closed sentences from being generated outside the controlled generation module.

9. A scientific research literature intelligent question-answering system based on a large language model according to claim 8, characterized in that, The controlled generation module specifically includes: Receive open sentences and keep write-prohibited closed sentences outside the controlled generation module; The large language model is driven to fill in open sentence positions to obtain the filled content; Extract the answer claims from the filled content, and then apply the answer claims back to the document evidence that is locked and written into the voting binding; When an open sentence is a controversial sentence, the answer claim should be simultaneously linked to the opposing literature evidence; If an open sentence is not converted into a disputed sentence and the answer is reversed and the document evidence bound to the locked sentence is successfully written, the fill content is retained. If an open sentence is a disputed sentence and the answer claims to be reversed to the locked sentence, and both the documentary evidence and the reverse documentary evidence for voting are successful, the fill content will be retained. The answer suggests that if the reverse deduction fails, the filled content should be cleared, and the open sentences corresponding to the filled content should not be output.

10. A scientific research literature intelligent question-answering system based on a large language model according to claim 9, characterized in that, The sources of the line-by-line citations specifically include: After receiving the successfully deducted content, retain the filled content and solidify it into the answer sentence; Link the answer sentence to the source of the literature to form a sentence-by-sentence citation; When the open sentence is a disputed sentence, the answer sentence will be simultaneously linked to the source of the reverse literature; Clear the answer sentence if the source is missing; Clear the answer sentence when the open sentence is a controversial sentence and the reverse literature source is missing; Output the retained answer sentence, reference source, and sentence-by-sentence citation source. When the open sentence is a disputed sentence, output the reverse reference source simultaneously.

Images