A procurement document checking method based on a large language model

Abstract

The application discloses a procurement document checking method based on a large language model, which comprises the following steps: receiving a procurement document to be checked, using a large language model to analyze the content of the procurement document, and generating a procurement content set; obtaining a checking rule set and a historical audit case set, and generating a rule item set and a case knowledge set; based on the procurement content set, the rule item set and the case knowledge set, a unified semantic representation set is constructed; the unified semantic representation set is input into an improved end-to-end memory network to form a multi-slot memory representation; a semantic query representation vector is constructed and input into the improved end-to-end memory network to generate an audit reasoning result; based on the audit reasoning result, a checking conclusion of the procurement document is output; and according to artificial audit feedback, newly added checking rules or newly added audit cases, the improved end-to-end memory network is updated. The application uses a large language model and a memory network to realize intelligent semantic checking of procurement documents.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent auditing and natural language processing, and particularly relates to a procurement document checking method based on a large language model. BACKGROUND

[0002] In the field of procurement document auditing, the existing technology mainly relies on manual auditing and rule-based automated tools. Manual auditing relies on the detailed interpretation and experience judgment of the auditing personnel on the content of the procurement document, and manually compares the contract terms, bid information, technical parameters, chart structure and picture attachments. Although this method has a certain flexibility, due to the high work intensity, long auditing period, and differences in knowledge and experience of the auditing personnel, it is difficult to unify the checking standards, and the accuracy and consistency of the auditing results are difficult to guarantee. In order to improve the auditing efficiency, some systems introduce rule or template-based automated auditing methods, which screen and preliminarily judge the procurement content by setting standard rules, but such methods can only be effective for formatted or structured content, and have limited processing capacity for unstructured text, complex semantic expression and logical association between multiple contents.

[0003] At the same time, the existing automated auditing system usually lacks effective use of historical auditing cases, and cannot fully mine and reuse previous auditing knowledge, resulting in insufficient identification ability of the system for new or boundary compliance issues. In addition, the rule maintenance and updating process is complicated, and it is difficult to adapt to the dynamic changes of procurement policies and business processes. When the complexity of the procurement document content increases or the auditing scene changes, the traditional method shows obvious shortcomings in semantic understanding, content modeling, knowledge reasoning and self-adaptive evolution.

[0004] Therefore, how to provide a procurement document checking method based on a large language model is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a procurement document checking method based on a large language model. The present application combines structured content modeling, unified semantic representation construction and improved end-to-end memory network reasoning mechanism, and describes in detail the semantic association modeling method between procurement content, rule items and historical cases, and the reasoning and updating process based on variable structure memory slot. The present application has the advantages of high degree of automation in the checking process, strong semantic association expression ability, good explainability of the auditing conclusion and strong continuous adaptation ability of the model.

[0006] According to the procurement document checking method based on a large language model of the present application, the method comprises the following steps:

[0007] receiving a procurement document to be checked, using a large language model to analyze the content of the procurement document, and generating a set of procurement content;

[0008] obtain a set of verification rules and a set of historical audit cases related to procurement business, generate a set of rule items and a set of case knowledge;

[0009] based on the set of procurement content, the set of rule items and the set of case knowledge, construct a set of unified semantic representations;

[0010] input the set of unified semantic representations into an improved end-to-end memory network, the improved end-to-end memory network dynamically generates a plurality of memory slots to form a multi-slot memory representation;

[0011] receive a verification task request, construct a semantic query representation vector, input the semantic query representation vector into the improved end-to-end memory network, and generate an audit reasoning result;

[0012] based on the audit reasoning result, output a verification conclusion of the procurement file;

[0013] According to the feedback of artificial audit, the newly added verification rules or the newly added audit cases, the improved end-to-end memory network is updated.

[0014] Optionally, the set of procurement content includes type tags, field attribute sets, potential mapping relationships with rule items, and context meta information of all procurement content.

[0015] Optionally, the generation of the set of rule items and the set of case knowledge specifically includes:

[0016] obtain a set of verification rules corresponding to the procurement file, perform structured analysis on each rule item in the set of verification rules, and split it into rule number, rule description, applicable scope, judgment condition, compliance requirement, and associated field tag to form a set of rule items;

[0017] for each rule item in the set of verification rules, establish a mapping relationship with each procurement content in the set of procurement content;

[0018] obtain a set of historical audit cases, perform abstract extraction on each case in the set of historical audit cases to obtain case abstract text, associated rule number, involved field attribute, case processing conclusion and time information, and form a set of case knowledge;

[0019] for each case in the set of case knowledge, establish a mapping relationship between the set of rule items and the set of procurement content.

[0020] Optionally, the set of unified semantic representations based on the set of procurement content, the set of rule items and the set of case knowledge specifically includes:

[0021] vectorize the set of procurement content, the set of rule items and the set of case knowledge respectively to obtain procurement semantic vector, rule semantic vector and case semantic vector;

[0022] performing a multi-layer semantic fusion operation on the procurement semantic vector, the rule semantic vector and the case semantic vector;

[0023] According to the type label, the field attribute set in the procurement content set, and the rule number, the field attribute, and the case summary in the rule item set and the case knowledge set, a semantic association mapping relationship is established to form a triple set;

[0024] Each triple in the triple set includes a procurement content object, a rule item object and a case knowledge object, and the procurement content object and the rule item object and the case knowledge object are mapped through semantic association information;

[0025] Based on the semantic vector representation of the procurement content object, the rule item object and the case knowledge object in each triple, dimension alignment is performed respectively, then a three-order semantic interaction tensor is generated through tensor product operation, and all three-order semantic interaction tensors are aggregated according to the semantic association weight to form a unified semantic representation set.

[0026] Optionally, the improved end-to-end memory network forms a multi-slot memory representation, which specifically includes:

[0027] The unified semantic representation set is input into the improved end-to-end memory network, which includes an input encoding layer, a variable structure memory slot layer, a memory retrieval and update layer, and an output reasoning layer;

[0028] The input encoding layer performs vectorization processing on the unified semantic representation set to obtain semantic feature representations of each content type grouping;

[0029] The variable structure memory slot layer dynamically generates a plurality of memory slots according to the type label, the field attribute set and the semantic association information of each grouping to construct a memory slot set;

[0030] The semantic feature representation of each content type grouping in the unified semantic representation set is written into the corresponding memory slot, and all memory slot contents jointly form a multi-slot memory representation.

[0031] Optionally, the variable structure memory slot layer dynamically generates a plurality of memory slots, which specifically includes:

[0032] According to the semantic feature representation of each content type grouping transmitted by the input encoding layer, the corresponding type label, field attribute set and semantic association information are extracted, a structured semantic feature representation is constructed, and a semantic complexity index is evaluated based on the structured semantic feature representation;

[0033] Based on the total sum of the semantic complexity indexes of all content groupings, the number of memory slots to be allocated in this round is dynamically calculated according to a variable structure memory slot allocation function;

[0034] According to the number of memory slots allocated in the current round, a dynamic allocation operation of slot resources is performed to generate a plurality of memory slots matching the semantic complexity, constituting a memory slot set;

[0035] Based on the content feature distribution and the slot usage state, an adaptive adjustment operation of the memory slot is performed.

[0036] Optionally, the writing of the semantic feature representation of each content type grouping into the corresponding memory slot includes, for each content type grouping in the unified semantic representation set, writing the semantic feature representation into the designated memory slot according to the slot type label correspondence; after the semantic content in all memory slots is sequentially written by type, a multi-slot memory representation is jointly constituted.

[0037] Optionally, the input of the semantic query representation vector into the improved end-to-end memory network to generate an audit reasoning result specifically includes:

[0038] Receiving a verification task request for a specific content object in the procurement document, parsing the verification task request, determining the type label, field attribute and verification target of the target content unit;

[0039] Based on the verification task request, a semantic query representation is constructed, which is matched with the slot type label, field attribute set and content index information in the multi-slot memory representation, and the memory slots and the content in the slots highly related to the verification task are screened;

[0040] In the screened memory slots, based on the semantic feature representation and the verification task request, a multi-hop attention reasoning operation is performed, and the procurement content nodes, audit rule nodes and case knowledge nodes are gradually reasoned and information is aggregated according to the context dependency and association path of the task semantics, to generate an audit reasoning intermediate result;

[0041] According to the audit reasoning intermediate result, the corresponding relationship between the verification rule item, the case knowledge and the target content unit and the compliance requirement are further compared, and finally an audit reasoning result is generated.

[0042] Optionally, the output of the verification conclusion of the procurement document based on the audit reasoning result specifically includes:

[0043] Receiving the audit reasoning result output by the improved end-to-end memory network, the audit reasoning result includes the reasoning path, the compliance comparison state, the rule matching score and the related case link for the target content unit;

[0044] Based on the rule matching score and the reasoning path in the audit reasoning result, a compliance judgment is performed on the target content unit, and a compliance judgment result of compliance, partial compliance or non-compliance is output, with the number identification of the non-compliant field and the violated clause, constituting the compliance judgment result.

[0045] According to the sensitivity of the association content of the verification task, the rule severity level and the case influence range, the verification risk level is quantified, and the level labels of low risk, medium risk, high risk or serious warning are output, and all level labels constitute risk level information;

[0046] The memory slot access sequence experienced in the audit reasoning process is extracted, and audit link information is constructed, which includes procurement content node identification, slot number, jump order and reasoning weight;

[0047] The compliance judgment result, risk level information and audit link information are packaged to form a verification conclusion.

[0048] Optionally, the adaptive updating of the slot structure and the in-slot storage content in the improved end-to-end memory network specifically includes:

[0049] Receiving artificial audit feedback information corresponding to the current verification task, the feedback information includes verification conclusion correction record, misjudgment correction mark, supplementary trial suggestion and corresponding timestamp;

[0050] Obtain new rules or new cases, perform format processing, generate structured rule items or structured case summaries, and align and compare with the original rule item set or case knowledge set, and mark the update difference segment;

[0051] According to the type label, field attribute set and semantic representation feature in the artificial audit feedback information and the new content, the corresponding memory slot in the current improved end-to-end memory network is located, and the slot number, in-slot item index and context association relationship to be updated are determined;

[0052] According to the positioning result, the in-slot content updating operation is performed, the old item is replaced, the new item is added, the index order is adjusted or the semantically similar content is merged, and the slot control parameters are adjusted according to the content update density and access frequency;

[0053] When the updating operation causes the memory slot configuration and content structure to deviate from the preset rule or the access performance to decrease, a slot structure reconstruction mechanism is triggered, and an updated improved end-to-end memory network is formed.

[0054] The beneficial effects of the present application are:

[0055] The procurement document checking method based on the large language model can automatically analyze contract terms, bid information, technical parameters, chart structures and picture attachments in the procurement document, convert them into a structured content set, and combine with checking rules and historical cases to perform unified semantic modeling.

[0056] Based on the structured content set, the rule item set and the case knowledge set, the method uses an end-to-end memory network to establish a multi-slot memory representation, realizes efficient organization and management of heterogeneous audit information. In actual checking tasks, the application can automatically match related content, rules and cases according to specific task requests, realize intelligent screening and compliance judgment of content through a multi-hop attention reasoning mechanism, and effectively improve the automation degree of complex procurement document checking and the traceability of the audit process. For new rules or new cases encountered in the checking process, and artificial audit feedback, the application can also dynamically adjust and optimize the memory network structure, continuously improve the accuracy and adaptive ability of the checking conclusion.

[0057] Through the above method, the application effectively overcomes the shortcomings of the prior art in content structured modeling, complex semantic understanding, rule adaptation and knowledge continuous evolution, realizes the automation and intelligentization of the whole process of intelligent checking of procurement documents. The output checking conclusion has compliance judgment, risk level quantization and reasoning link traceability, can provide strong decision support for procurement management and risk control, and significantly improves the work efficiency and result quality of procurement document audit. BRIEF DESCRIPTION OF DRAWINGS

[0058] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application together with the embodiments of the application, and are used to explain the application, and do not constitute a limitation on the application. In the drawings:

[0059] Fig. 1 A flowchart of a procurement document checking method based on a large language model according to the application;

[0060] Fig. 2 An improved end-to-end memory network structure diagram for introducing a variable structure memory slot in a procurement document checking method based on a large language model according to the application;

[0061] Fig. 3 A multi-hop attention reasoning process schematic diagram for a procurement checking task in a procurement document checking method based on a large language model according to the application. DETAILED DESCRIPTION

[0062] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic illustrations of the basic structure of the application and therefore only show the components relevant to the application.

[0063] Reference Figs. 1-3 A procurement document verification method based on a large language model, comprising the following steps:

[0064] Receiving a procurement document to be verified, using a large language model to analyze the content of the procurement document, and generating a procurement content set;

[0065] Obtaining a set of verification rules related to procurement business and a set of historical audit cases, generating a set of rule items and a set of case knowledge;

[0066] Based on the procurement content set, the rule item set and the case knowledge set, a unified semantic representation set is constructed;

[0067] The unified semantic representation set is input into an improved end-to-end memory network, which dynamically generates multiple memory slots according to the structural features and type labels of the input content, forming a multi-slot memory representation;

[0068] Receiving a verification task request for a target content unit of the procurement document, constructing a semantic query representation vector based on the verification task request, inputting the semantic query representation vector into the improved end-to-end memory network, and using the multi-slot memory representation to perform a multi-hop attention reasoning operation to generate an audit reasoning result;

[0069] Based on the audit reasoning result, output the verification conclusion of the procurement document;

[0070] According to the feedback of artificial audit, new rule content or new case content, the slot structure and slot storage content in the improved end-to-end memory network are adaptively updated.

[0071] In this embodiment, the procurement content set includes the type label of all procurement content, the field attribute set, the potential mapping relationship with the rule item, and the content context meta information;

[0072] The type label includes contract terms, bid information, technical parameters, chart structure, and picture attachments;

[0073] The field attribute set includes the field name, field value, field position, field length, and field hierarchical relationship of each content unit;

[0074] The potential mapping relationship with the rule item includes a mapping identifier, a rule reference number, and an applicable condition description for describing the correspondence or association between the content unit and the verification rule item;

[0075] The context meta-information includes a page number, a paragraph position, a chapter to which the procurement content unit belongs, an adjacent procurement content unit identifier, and time information of generation or change of the procurement content unit in the procurement document.

[0076] In the embodiment, the generation rule item set and the case knowledge set specifically include:

[0077] A verification rule set corresponding to the procurement document is obtained, and structured analysis is performed on each rule item in the verification rule set to split the rule item into a rule number, a rule description, an applicable scope, a judgment condition, a compliance requirement, and an associated field label to form a rule item set;

[0078] For each rule item in the verification rule set, a mapping relationship between the rule item and the procurement content in the procurement content set is established, and the mapping relationship includes an associated identifier and constraint logic between the rule item and the type label and the field attribute set;

[0079] A historical audit case set is obtained, and abstract extraction is performed on each case in the historical audit case set to obtain a case abstract text, an associated rule number, an involved field attribute, a case processing conclusion, and time information to form a case knowledge set;

[0080] For each case in the case knowledge set, a potential mapping relationship between the case and the rule item set and the procurement content set is established, and the mapping relationship includes a multi-dimensional associated identifier of the case abstract and the rule item number, the type label, and the field attribute set.

[0081] In the embodiment, the construction of the unified semantic representation set based on the procurement content set, the rule item set, and the case knowledge set specifically includes:

[0082] The procurement content set, the rule item set, and the case knowledge set are respectively represented by vectors to obtain a procurement semantic vector, a rule semantic vector, and a case semantic vector;

[0083] A multi-layer semantic fusion operation is performed on the procurement semantic vector, the rule semantic vector, and the case semantic vector, and the semantic fusion operation includes cross-set information alignment based on an attention mechanism, content association enhancement, and context dependency relationship modeling;

[0084] According to the type label and the field attribute set in the procurement content set, a semantic association mapping relationship is established between the rule number, the field attribute, and the case abstract in the rule item set and the case knowledge set to form a triple set;

[0085] Each triplet in the triplet set includes a procurement content object, a rule entry object, and a case knowledge object. The procurement content object, the rule entry object, and the case knowledge object are mapped through semantic association information such as type tags, field attributes, rule numbers, and case summaries. This is used to describe the semantic association and correspondence between the procurement content, the audit rules, and the historical knowledge.

[0086] Based on the set of triples, the semantic vector representations of the procurement content object, rule entry object and case knowledge object in each triple are aligned in dimension. Then, a third-order semantic interaction tensor is generated through tensor product operation. All tensor results are aggregated according to semantic association weight to form a unified semantic representation set.

[0087] In this embodiment, the improved end-to-end memory network specifically includes:

[0088] A unified semantic representation set is input into an improved end-to-end memory network, which includes an input encoding layer, a variable structure memory slot layer, a memory retrieval and update layer, and an output inference layer.

[0089] The input encoding layer vectorizes the unified semantic representation set to obtain semantic feature representations for each content type group, and then passes the semantic feature representations to the variable structure memory slot layer.

[0090] The variable structure memory slot layer dynamically generates multiple memory slots based on the type label, field attribute set, and semantic association information of each group, and constructs a memory slot set;

[0091] The semantic feature representation of each type group in the unified semantic representation set is written into the corresponding memory slot, and the contents of all memory slots are used to form a multi-slot memory representation.

[0092] The memory retrieval and update layer constructs a semantic query representation based on the verification task request, performs relevance matching with the slot type labels and index structure in the multi-slot memory representation, selects relevant slots and content items, performs multi-hop attention reasoning operation, and outputs the intermediate results of the verification reasoning.

[0093] In each hop, the multi-hop attention reasoning operation associates the procurement content node, the review rule node, and the case knowledge node through an attention mechanism. It aggregates information according to the semantic path and reasoning context to support the updating, replacement, or re-indexing of the content in the slot under feedback-driven conditions, thereby achieving progressive optimization of the memorized content.

[0094] The output reasoning layer performs semantic path tracking, rule matching and scoring, and compliance status judgment based on the intermediate results of the review reasoning, and finally outputs the structured review reasoning results.

[0095] The end-to-end memory network adopts a variable structure memory slot mechanism, so that the network can automatically adapt to different procurement document structures and content types, realize efficient storage and organization of heterogeneous information, deep expression and reasoning of complex semantic relationships, and significantly improve the generalization ability, scalability and interpretability of the intelligent auditing method, which is different from traditional fixed structure or static allocation end-to-end memory networks.

[0096] In the embodiment, the variable structure memory slot layer dynamically generates a plurality of memory slots, and specifically includes:

[0097] The semantic feature representation of each content type group is received from the input encoding layer, the type label, field attribute set and semantic association information corresponding to each content type group are extracted, the structured semantic feature representation is constructed, and the semantic complexity index is evaluated based on the structured semantic feature representation;

[0098] Based on the complexity index sum of all groups, the number of memory slots to be allocated in this round is dynamically calculated according to the variable structure memory slot allocation function:

[0099] ;

[0100] Wherein, is the number of memory slots to be allocated in this round, is the slot allocation threshold, is the number of content type groups in the unified semantic representation set, is the index number of the content type group, is the proportion coefficient of the adjustment structure dimension influence weight, is the structure dimension function value of the th content type group, is the proportion coefficient of the semantic association strength weight, is the semantic association strength function value of the th content type group, is the upward rounding symbol;

[0101] In order to realize the slot dynamic allocation mechanism of the structured semantic feature group in the memory network, a variable structure memory slot allocation function is designed, which is used to adaptively determine the number of memory slots to be generated according to the semantic complexity of the input group;

[0102] Firstly, the semantic complexity index of a single group is defined as :

[0103] ;

[0104] Wherein, has a dimension, is dimensionless, , are dimensionless, have dimensional units;

[0105] based on the sum of the semantic complexity of all groups, a slot allocation threshold is introduced , with dimensional units, for controlling the maximum holding complexity of a single slot, thereby determining the number of memory slots to be allocated in this round :

[0106] ;

[0107] The formula structure is referenced from the task binning model and task grouping assembly strategy in a computer system, combined with the complexity measurement mechanism of semantic grouping, to form the variable structure memory slot allocation function used in the present application;

[0108] According to the number of memory slots to be allocated in this round, a dynamic allocation operation of slot resources is performed to generate a plurality of memory slots matching the semantic complexity, constituting a memory slot set;

[0109] Each memory slot includes a slot type label, a semantic content tensor in the slot, content index information and slot control parameters; the slot type label is used to identify the content type label carried by the slot; the semantic content tensor in the slot is used to store a set of triple semantic vectors of the corresponding type, supporting high-dimensional semantic interaction and fast retrieval; the content index information is used to identify the position, field relationship and context dependence path of the semantic item in the slot; the slot control parameters are used to manage the write state, content capacity, priority and update strategy of the slot;

[0110] After the memory slots are generated, an adaptive adjustment operation of the memory slots is performed based on the current content feature distribution and the slot usage state. The adaptive adjustment operation includes: dynamically adding slot positions according to the slot usage frequency to improve the representation ability of high-frequency content; performing a merging operation on memory slots with similar semantics or high redundancy to improve overall expression efficiency; releasing low-frequency slots that have not been accessed for a long time to release resources; and performing cross-slot migration on redundant or unevenly distributed semantic items to achieve optimal management of memory slot resources and structured organization of semantic content.

[0111] In the present embodiment, the writing of the semantic feature representation of each content type grouping into the corresponding memory slot to form a multi-slot memory representation specifically includes:

[0112] Each content type grouping in the unified semantic representation set is traversed, and according to the slot type label information carried by each grouping, its corresponding memory slot is determined, and the semantic feature representation is written into the corresponding slot to ensure that the semantic features are collected in order by category in the slot layer, avoiding semantic interference or information redundancy;

[0113] In the writing process, according to the content field attribute, semantic correlation strength and index identification, the writing order and the in-slot physical storage position of each semantic entry are determined to realize accurate positioning and fast retrieval of in-slot entries. The writing strategy gives priority to semantic density distribution and context consistency to optimize the space utilization and access efficiency in the slot;

[0114] After the semantic content in all memory slots is written in order by type, it jointly constitutes a multi-slot memory representation, which serves as the basic data structure for multi-hop reasoning and intelligent decision-making in subsequent procurement document verification tasks, and provides efficient support for relevance retrieval, content aggregation and task-driven reasoning operations;

[0115] During the subsequent verification task execution process, the content arrangement order in the slot can be dynamically adjusted, the index information can be updated, and part of the content can be migrated from one memory slot to another memory slot in real time according to the content changes or feedback results, so as to ensure the structural rationality and content accuracy of the multi-slot memory representation.

[0116] In the embodiment, the input of the semantic query representation vector into the improved end-to-end memory network generates an audit reasoning result, which specifically includes:

[0117] The improved end-to-end memory network receives a verification task request for a specific content object in the procurement document, analyzes the verification task request, and determines the type label, field attribute and verification target of the target content unit;

[0118] Based on the verification task request, a semantic query representation is constructed, which is matched with the slot type label, field attribute set and content index information in the multi-slot memory representation to filter the memory slots and in-slot content highly related to the verification task;

[0119] The memory retrieval and update layer of the improved end-to-end memory network constructs a semantic query representation according to the target content unit type, field attribute and verification target specified in the verification task request, and the semantic query representation includes a task semantic vector, a field label embedding vector and a context prompt vector;

[0120] The memory retrieval and update layer matches the semantic query representation with the slot type label, content field attribute set and content index structure in the multi-slot memory representation. This matching process is based on semantic similarity calculation and field attribute comparison to filter out the target memory slot and its in-slot candidate content entry set with high semantic relevance and field correspondence to the current verification task;

[0121] In the target memory slot screened out, a multi-hop attention reasoning operation is performed according to the interaction relationship between the semantic query representation and the semantic content in the slot. The reasoning operation takes the current query as the starting point, and according to the context dependency relationship in the task semantics, the procurement content node and the associated audit rule node and case knowledge node are activated hop by hop to form a layer-by-layer reasoning path. In each hop, the attention distribution is calculated to determine which content or rule is most critical to the current task, thereby realizing information aggregation and path tracking. The process dynamically adjusts the reasoning path and aggregation mechanism by accumulating the attention score and the semantic consistency score to avoid semantic loss or misdirection caused by single-hop reasoning.

[0122] The memory retrieval and update layer outputs an audit reasoning intermediate result, which includes the semantic matching relationship between the verification task request and the target content unit, the content node, the rule node and the case node set involved in the activated path, the attention score of each node and its reasoning contribution value.

[0123] In the present application, the "procurement content node" refers to the content unit in the structured content set after semantic mapping processing; the "audit rule node" refers to the rule item in the verification rule set with semantic label and semantic vector representation; and the "case knowledge node" refers to the representation unit formed after semantic abstract conversion in the case knowledge set.

[0124] According to the audit reasoning intermediate result, the corresponding relationship and compliance requirement between the verification rule item, the case knowledge and the target content unit are further compared, and finally the audit reasoning result is generated, which includes the reasoning path for the target procurement content unit, the compliance comparison state, the rule matching score and the related case link, for supporting the output and explanation of the procurement document verification conclusion.

[0125] In the present embodiment, the output of the procurement document verification conclusion based on the audit reasoning result specifically includes:

[0126] The audit reasoning result output by the improved end-to-end memory network is received, which includes the reasoning path for the target procurement content unit, the compliance comparison state, the rule matching score and the related case link;

[0127] Based on the rule matching score and the reasoning path in the audit reasoning result, the compliance judgment for the target procurement content unit is performed, and the compliance judgment result of compliance, partial compliance or non-compliance is output, with the non-compliance field and the number identification of the violated clause, to constitute the compliance judgment result;

[0128] According to the sensitivity of the association content of the verification task, the rule severity level and the case influence range, the verification risk level is quantified, and the low risk, medium risk, high risk or serious warning level labels are output. All level labels constitute risk level information, which is used to evaluate the potential problems of procurement behavior;

[0129] The memory slot access sequence experienced in the extraction review reasoning process is extracted to construct review link information, which includes procurement content node identification, slot number, jump order and reasoning weight, and is used to trace the review process and support explainable output;

[0130] The compliance judgment result, risk level information and review link information are packaged to form a verification conclusion, which is output to the upper business system or user terminal, and is used to assist procurement management decision and risk warning feedback.

[0131] In the embodiment, the adaptive updating of the slot structure and the in-slot storage content in the improved end-to-end memory network specifically includes:

[0132] Receiving artificial review feedback information corresponding to the current verification task, the feedback information includes verification conclusion correction record, misjudgment correction mark, supplementary review suggestion and corresponding timestamp;

[0133] Obtain the newly added rule content or newly added case content of the system, perform format processing on the new information, generate structured rule items or structured case summaries, and align and compare with the original rule item set or case knowledge set, and mark the updated difference segment;

[0134] According to the content type label, field attribute set and semantic representation feature in the artificial review feedback information and the new content, the corresponding memory slot in the current improved end-to-end memory network is located, and the slot number, in-slot item index and context association relationship to be updated are determined;

[0135] According to the positioning result, perform in-slot content updating operation, replace old items, add new items, adjust index order or merge semantically similar content, and adjust slot control parameters according to content update density and access frequency to maintain network operation stability;

[0136] When the update operation causes the memory slot configuration and content structure to deviate from the preset rules or the access performance to decrease, trigger the slot structure reconstruction mechanism, automatically evaluate the current slot distribution state, and perform slot merging, splitting or reconstruction to form an updated improved end-to-end memory network for subsequent tasks.

[0137] Embodiment 1:

[0138] To verify the feasibility of the application in implementation, the application is applied to the annual bidding document review process of a provincial large-scale enterprise centralized procurement platform. The platform needs to process more than 8000 procurement documents submitted by branch agencies every year, covering IT equipment procurement, engineering construction, office supplies and professional services of various types, with non-uniform file formats, complex content structure, heavy manual review pressure, low efficiency, and problems such as missing review rules, insufficient case reuse, and lack of supporting evidence for review conclusions found in multiple rounds of manual sampling. The application realizes the semantic analysis of procurement document content, rule mapping and automatic reasoning of historical case linkage by introducing large language models and multi-slot structure memory networks, effectively improving the review efficiency and consistency of conclusions.

[0139] In actual deployment, the system automatically identifies contract terms, bid structure, key technical parameters and attachment chart content after receiving each procurement document, converts them into uniform semantic vectors through pre-trained language models, and generates a procurement content set. Combined with the check rule items and real audit cases maintained in the knowledge base, a rule item set and a case knowledge set are constructed. Taking an office automation equipment procurement document as an example, the system identifies that there are multiple specification fields in the content that are not filled in accordance with the platform's unified technical standards, and some clauses use expressions that have been judged as "potential supplier discrimination" in history. When performing the check task, the system matches the relevant rules and cases through semantic query representation in the memory network, and outputs the "non-compliance" judgment conclusion combined with the attention reasoning mechanism, while labeling the corresponding rule number and historical case path. In the same batch of test data, the manual review team confirms that the conclusion is accurate and complete, has good explainability and guiding significance.

[0140] The following table is a "procurement document review efficiency and accuracy comparison data table", reflecting the core performance indicators in the actual operation of the system:

[0141] Table 1 Procurement document review efficiency and accuracy comparison data table

[0142] ;

[0143] As can be seen from the above table, the application has significant advantages over traditional manual review methods in several key performance indicators of procurement document verification. First, in terms of review efficiency, the application method takes only 32 hours to process 1200 procurement documents, which is much lower than the 156 hours required for manual review. The average processing time per document is shortened to 1 minute and 36 seconds, which is significantly faster than the 7 minutes and 48 seconds of manual review, with a processing speed improvement of more than 5 times, significantly relieving the human pressure under centralized review tasks.

[0144] Secondly, in terms of the accuracy of the review and the reliability of the compliance judgment, the accuracy rate of the system review reaches 96.2%, which is significantly better than the 88.5% of the artificial method, effectively reducing the missed and misjudged phenomena. At the same time, in terms of the explainability of the review conclusion, the system generated review link and compliance comparison information have good traceability ability, and the explanation rate of inconsistent conclusions in actual test is 91.4%, which is much higher than the artificial backtracking explanation rate of 58.3%, so that each judgment conclusion can be traced to specific rules and historical cases, and the transparency and persuasiveness of the results are enhanced.

[0145] In addition, in terms of rule application ability, the system realizes full coverage of 97 audit rules and can automatically quote 516 historical cases to support the audit process. The automatic case reference ability reaches 83.7%, while the effective reference rate in the artificial audit process is only 41.9%, indicating that the method of the application can more fully activate the knowledge assets for the current verification task, and improve the knowledge utilization efficiency. At the same time, in terms of system adaptive ability, the response cycle based on artificial audit feedback or new rules is controlled within 2 hours, which is significantly better than the maintenance cycle of 3 to 5 days in the artificial method, showing strong dynamic updating and self-evolution ability. In summary, the intelligent verification system based on large language model and multi-slot memory network constructed by the application has high efficiency, high accuracy, strong explainability and good rule adaptability, which can effectively replace the traditional artificial method and significantly improve the intelligent level of procurement document review.

[0146] The above describes only the preferred specific embodiments of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes to the technical solutions and inventive concepts of the application within the technical scope disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A procurement document verification method based on a large language model, characterized in that, The method comprises the following steps: receiving a procurement document to be checked, using a large language model to analyze the content of the procurement document, and generating a procurement content set; obtaining a set of verification rules and a set of historical audit cases related to procurement business, generating a set of rule items and a set of case knowledge; based on the procurement content set, the rule item set and the case knowledge set, a unified semantic representation set is constructed; input the unified semantic representation set into the improved end-to-end memory network, the improved end-to-end memory network dynamically generates a plurality of memory slots to form a multi-slot memory representation; receiving a verification task request, constructing a semantic query representation vector, inputting the semantic query representation vector into the improved end-to-end memory network, and generating an audit reasoning result; based on the audit reasoning result, output the verification conclusion of the procurement document; according to the feedback of artificial audit, new verification rules or new audit cases, update the improved end-to-end memory network; the improved end-to-end memory network dynamically generates a plurality of memory slots to form a multi-slot memory representation, which specifically comprises: input the unified semantic representation set into the improved end-to-end memory network, the improved end-to-end memory network comprises an input encoding layer, a variable structure memory slot layer, a memory retrieval and update layer and an output reasoning layer; the input encoding layer performs vectorization processing on the unified semantic representation set to obtain semantic feature representation of each content type grouping; the variable structure memory slot layer dynamically generates a plurality of memory slots according to the type label, field attribute set and semantic association information of each grouping to construct a memory slot set; write the semantic feature representation of each content type grouping in the unified semantic representation set into the corresponding memory slot, and all memory slot contents jointly form a multi-slot memory representation; the variable structure memory slot layer dynamically generates a plurality of memory slots according to the type label, field attribute set and semantic association information of each grouping to construct a memory slot set, which specifically comprises: according to the semantic feature representation of each content type grouping transmitted by the input encoding layer, extract the corresponding type label, field attribute set and semantic association information, construct a structured semantic feature representation, and evaluate the semantic complexity index based on the structured semantic feature representation; based on the total sum of semantic complexity indexes of all content groupings, dynamically calculate the number of memory slots to be allocated in this round according to the variable structure memory slot allocation function; according to the number of memory slots to be allocated in this round, perform dynamic allocation operation of slot resource, generate a plurality of memory slots matched with semantic complexity, and constitute a memory slot set; based on the content feature distribution and the slot usage state, perform adaptive adjustment operation of memory slot.

2. The procurement document verification method based on a large language model according to claim 1, characterized in that, The procurement content set comprises type label, field attribute set, potential mapping relationship with rule item, and context meta information of all procurement contents.

3. The procurement document verification method based on a large language model according to claim 1, characterized in that, The generation of rule item set and case knowledge set specifically comprises: obtaining a set of verification rules corresponding to the procurement document, performing structured analysis on each rule item in the set of verification rules, and splitting into rule number, rule description, applicable scope, judgment condition, compliance requirement and associated field label to form a rule item set; for each rule item in the set of verification rules, a mapping relationship with each procurement content in the procurement content set is established; Obtaining a historical audit case set, performing abstract extraction on each case in the historical audit case set to obtain a case abstract text, an associated rule number, an involved field attribute, a case processing conclusion and time information, and forming a case knowledge set; For each case in the case knowledge set, a mapping relationship between the rule item set and the procurement content set is established.

4. The procurement document verification method based on a large language model according to claim 1, characterized in that, The construction of the unified semantic representation set based on the procurement content set, the rule item set and the case knowledge set specifically includes: The procurement content set, the rule item set and the case knowledge set are respectively represented by vectors to obtain procurement semantic vectors, rule semantic vectors and case semantic vectors; Performing multi-layer semantic fusion operation on the procurement semantic vectors, the rule semantic vectors and the case semantic vectors; According to the type label in the procurement content set, the field attribute set, the rule number, the field attribute and the case abstract in the rule item set and the case knowledge set, a semantic association mapping relationship is established to form a triple set; Each triple in the triple set includes a procurement content object, a rule item object and a case knowledge object, and the mapping relationship between the procurement content object and the rule item object and the case knowledge object is established through semantic association information; Based on the semantic vector representation of the procurement content object, the rule item object and the case knowledge object in each triple, dimension alignment is performed respectively, then a three-order semantic interaction tensor is generated through tensor product operation, and all three-order semantic interaction tensors are aggregated according to the semantic association weight to form a unified semantic representation set.

5. The procurement document verification method based on a large language model according to claim 1, characterized in that, The semantic feature representation of each content type grouping in the unified semantic representation set is respectively written into the corresponding memory slot, which includes writing the semantic feature representation into the specified memory slot according to the slot type label correspondence for each content type grouping in the unified semantic representation set; After the semantic content in all memory slots is written in order by type, a multi-slot memory representation is formed.

6. The procurement document verification method based on a large language model according to claim 1, characterized in that, The semantic query representation vector is input into the improved end-to-end memory network to generate an audit reasoning result, which specifically includes: Receiving a verification task request for a specific content object in the procurement document, analyzing the verification task request, determining the type label, field attribute and verification target of the target content unit; Based on the verification task request, a semantic query representation is constructed, which is matched with the slot type label, field attribute set and content index information in the multi-slot memory representation to filter the memory slots and the content in the slots that are highly related to the verification task; In the filtered memory slots, based on the semantic feature representation and the verification task request, a multi-hop attention reasoning operation is performed, and the procurement content node, the audit rule node and the case knowledge node are gradually reasoned and information is aggregated according to the context dependence relationship and the associated path of the task semantics, to generate an audit reasoning intermediate result; According to the audit reasoning intermediate result, the corresponding relationship and compliance requirement between the verification rule item, the case knowledge and the target content unit are further compared, and finally an audit reasoning result is generated.

7. The procurement document verification method based on a large language model according to claim 1, characterized in that, The verification conclusion of the procurement document is output based on the audit reasoning result, which specifically includes: Receiving an audit inference result output by the improved end-to-end memory network, the audit inference result including an inference path for a target content unit, a compliance comparison state, a rule matching score, and a related case link; Based on the rule matching score and the inference path in the audit inference result, performing a compliance judgment on the target content unit, outputting a compliance judgment result of compliance, partial compliance, or non-compliance, and attaching a non-compliance field and a number identification of violated clauses to constitute the compliance judgment result; According to the sensitivity of the associated content of the verification task, the rule severity level, and the case influence range, quantifying the verification risk level, and outputting a level label of low risk, medium risk, high risk, or serious warning, all level labels constituting risk level information; Extracting a memory slot access sequence experienced in the audit inference process, and constructing audit link information including procurement content node identification, slot number, jump order, and inference weight; Packing the compliance judgment result, risk level information, and audit link information into a verification conclusion.

8. The procurement document verification method based on a large language model according to claim 1, characterized in that, The updating of the improved end-to-end memory network specifically includes: Receiving artificial audit feedback information corresponding to the current verification task, the feedback information including verification conclusion correction records, misjudgment correction marks, supplementary review suggestions, and corresponding time stamps; Obtaining new rules or new cases, performing format processing, generating structured rule items or structured case summaries, and aligning and comparing with the original rule item set or case knowledge set to mark the updated difference segments; According to the type label, field attribute set, and semantic representation features in the artificial audit feedback information and the new content, positioning the corresponding memory slot in the current improved end-to-end memory network, determining the slot number, slot item index, and context association relationship to be updated; According to the positioning result, performing in-slot content update operations, replacing old items, adding new items, adjusting index order, or merging semantically similar content, while adjusting slot control parameters according to content update density and access frequency; When the update operation causes the memory slot configuration and content structure to deviate seriously from the preset rules or the access performance to decrease, triggering a slot structure reconstruction mechanism to form an updated improved end-to-end memory network.

Images