Project declaration content compliance review method and system combining named entity recognition and large model

Abstract

The application provides a project declaration content compliance review method and system combining named entity recognition and large models, and relates to the technical field of natural language processing. First, a project declaration full-cycle document set is obtained, including declaration documents, contracts, mid-term examination and acceptance documents, a fine-tuned named entity recognition model is used to extract structured entity triples, and a data set containing logical association relationships is constructed. Then, a compliance review prompt engineering instruction set is constructed, structured entity triples are input into a large language model for multi-document entity alignment and consistency verification, numerical evolution, semantic deviation and logical contradiction features are identified, and identification results are generated. Finally, a traceable review evidence chain set is constructed according to the identification results, the source of the problem is accurately located, and the association with the compliance rules is clarified. The application improves the review efficiency, accuracy and systematicness.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and more specifically, to a method and system for reviewing the compliance of project application content by combining named entity recognition and large model. Background Technology

[0002] In the project application and management process, ensuring the compliance of the application content is a crucial step in guaranteeing the smooth progress of the project, the rational allocation of resources, and compliance with relevant policies and regulations. Traditional methods for reviewing the compliance of project application content mainly rely on manual review. Reviewers need to carefully study all kinds of documents throughout the entire project application cycle, including the project application form, project contract, mid-term project review documents, and project completion and acceptance documents.

[0003] However, manual review has many limitations. On the one hand, the sheer volume and complexity of documents throughout the project application process make manual review not only inefficient but also prone to oversights, making it difficult to comprehensively and accurately identify compliance issues. On the other hand, for reviews involving key information such as technical specifications, deliverables, timelines, budget items, and regulatory references, manual review often lacks systematicity and consistency. Different reviewers may have differing understandings and judgments of the same information, leading to a high degree of subjectivity in the review results. Furthermore, when inconsistencies or violations are found in documents, manual review struggles to quickly and accurately pinpoint the source of the problem and establish a complete chain of evidence, hindering subsequent problem tracing and rectification. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, the present invention provides a method for reviewing the compliance of project application content by combining named entity recognition and large model, the method comprising:

[0005] Obtain the complete set of project application documents corresponding to the project to be reviewed. The complete set of project application documents includes project application documents, project contract documents, project mid-term inspection documents, and project completion and acceptance documents.

[0006] Based on the named entity recognition model that has been fine-tuned through text corpus in the project application field, a multi-stage structured entity extraction process is performed on the full-cycle document set of the project application to generate a set of structured entity triples containing logical relationships. Each structured entity triple in the set of structured entity triples contains at least two entity types from the following entities: technical indicator entity, deliverable entity, time node entity, budget item entity, and legal clause reference entity, as well as a descriptor of the logical relationship between the entities.

[0007] A set of compliance review prompts for engineering instructions is constructed and the set of structured entity triples is input into a large language model according to a standardized input format. The large language model is driven to perform multi-document entity alignment and consistency verification processing. The numerical evolution characteristics, semantic deviation characteristics and logical contradiction characteristics of the same entity appearing in different documents of the full life cycle document set of the project application are identified. Based on the preset compliance rule base, entity relationship alignment and logical contradiction identification results containing conflict type identifiers and citation identifiers are generated.

[0008] Based on the original text position mapping relationship corresponding to the inconsistencies in the entity relationship alignment and logical contradiction identification results, the traceable review evidence chain construction process is performed on the project application full-cycle document set to generate a traceable review evidence chain set containing the original text content of the inconsistencies, the coordinate positioning information of the inconsistencies in the documents, and the association relationship between the inconsistencies and compliance rules.

[0009] Furthermore, this invention also provides a project application content compliance review system that combines named entity recognition and large model, including:

[0010] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the aforementioned method for compliance review of project application content combining named entity recognition and large models by executing the machine-executable instructions.

[0011] In another aspect, the present invention also provides a computer program product, the computer program product including machine-executable instructions stored in a computer-readable storage medium, wherein a processor of the system for reviewing the compliance of project application content in conjunction with named entity recognition and large model reads the machine-executable instructions from the computer-readable storage medium, and the processor executes the machine-executable instructions, thereby causing the system for reviewing the compliance of project application content in conjunction with named entity recognition and large model to perform the aforementioned method for reviewing the compliance of project application content in conjunction with named entity recognition and large model.

[0012] Based on the above, by acquiring a complete set of project application documents and using a named entity recognition model fine-tuned from the text corpus in the project application domain for multi-stage structured entity extraction, key entities and their logical relationships can be extracted from various documents. This generates a set of structured entity triples containing logical relationships. Then, a compliance review prompt instruction set is constructed, and the set of structured entity triples is input into a large language model. This drives the large language model to perform multi-document entity alignment and consistency verification. It can efficiently identify features such as numerical evolution, semantic deviation, and logical contradictions of the same entity appearing in different documents. Based on a pre-set compliance rule base, detailed entity relationship alignment and logical contradiction identification results are generated, greatly improving the accuracy and comprehensiveness of the review and reducing the omissions and subjectivity of manual review. Finally, a traceable review evidence chain set is constructed based on the original text position mapping relationship of inconsistencies in the identification results. This allows for quick and accurate location of the problem source, determination of the relationship between the problem and compliance rules, and improvement of the efficiency, accuracy, and systematic nature of the compliance review of project application content. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the execution flow of the project application content compliance review method that combines named entity recognition and large model provided in the embodiments of the present invention.

[0014] Figure 2 This is a schematic diagram of exemplary hardware and software components of the project application content compliance review system that combines named entity recognition and large model provided in an embodiment of the present invention. Detailed Implementation

[0015] Figure 1 This is a flowchart illustrating a method for reviewing the compliance of project application content that combines named entity recognition and large model, provided by an embodiment of the present invention. A detailed description follows.

[0016] Step S110: Obtain the complete set of project application documents corresponding to the project to be reviewed. The complete set of project application documents includes project application documents, project contract documents, project mid-term inspection documents, and project completion acceptance documents.

[0017] In this embodiment, to address the full-cycle management needs of enterprise research projects, it is first necessary to systematically collect key documents generated at different stages of the project. The project application document, as the foundational document for project initiation, includes core content such as the project background, research objectives, technical route, and expected outcome indicators. The project contract document is a legally binding document signed with partners or supervisory units after project initiation, clarifying key information such as specific tasks, delivery standards, timelines, and budget allocation during project execution. The mid-term project review document records the project's progress during the execution phase, including completed interim results, indicators not met and their causes, and subsequent adjustment plans. The project completion and acceptance document is a summary document submitted upon project completion, including the overall project completion status, application of results, and fund usage. These documents are typically stored electronically in the enterprise's internal document management system or project management platform. They are then collected and consolidated through interface calls or batch import to form a complete set of project application lifecycle documents awaiting review. During the collection process, it is crucial to ensure the completeness and version accuracy of the documents to avoid impacting subsequent review processes due to missing documents or version errors.

[0018] Step S210: Perform a unified format conversion on each document in the project application full-cycle document set to obtain a plain text document with a unified format.

[0019] Because the documents in the project application lifecycle document collection may originate from different systems or be created by different personnel, they exist in various formats, such as DOCX, PDF, and TXT. To ensure the compatibility and processing efficiency of subsequent text processing algorithms, it is necessary to uniformly convert the formats of these documents. For DOCX format documents, the text content, paragraph structure, and table data are extracted by calling a document processing library; for PDF format documents, if it is a text-based PDF, the text content is directly extracted; if it is a scanned PDF, optical character recognition technology is used to convert the text in the image into editable text; for TXT format documents, their text content is directly retained. During the conversion process, the original chapter structure and text order of the document must be preserved, and formatting marks (such as fonts, colors, and layout styles) must be removed, retaining only the plain text information. For example, a PDF format project contract document is converted into a UTF-8 encoded plain text file, ensuring that the clause content, numbers, dates, and other information are complete and intact.

[0020] Step S220: Perform document structure parsing on each plain text document, identify chapter titles, paragraph boundaries and table areas, and generate a document structure tree containing chapter nodes, paragraph nodes and table nodes.

[0021] After format conversion, document structure parsing is performed on each plain text document. A rule-based and machine learning approach is used. First, chapter titles, such as "I. Project Overview" and "2.1 Research Content," with clear hierarchical characteristics are identified through keyword matching and regular expressions. Then, paragraph boundaries are identified using line breaks and punctuation marks (such as periods, semicolons, and exclamation marks) to divide continuous text into independent paragraph units. For table areas, table positions are located by recognizing table features in the text (such as row and column alignment of numbers, specific separators "|" and "-"), and the start and end line numbers of the table are recorded. Based on the recognition results, a tree-like data structure, i.e., the document structure tree, is constructed. The root node of the tree is the document itself, the first-level child nodes are chapter titles, and the second-level child nodes are paragraph and table nodes under the corresponding chapters. Node attributes include text content, the starting position index (character offset) in the document, length, and other information. For example, in the structure tree of the project application document, the "Research Objectives" section node contains 3 paragraph nodes and 1 table node, and each node records its specific location and content in the plain text document.

[0022] Step S230: Extract the table content from the table area corresponding to the table node, convert the table data into key-value pairs and insert them into the paragraph where the table node is located, generating enhanced document content containing the converted text of the table content.

[0023] For table nodes in the document structure tree, specialized content extraction and transformation processing is required. A table parsing algorithm is used to first identify the row and column boundaries of the table, determining the header row and data rows. Then, the cell content of the header row is used as the key, and the cell content of the corresponding column in the data row is used as the value, constructing a key-value pair structure. For example, a table containing columns for "Indicator Name," "Target Value," and "Assessment Method" will be transformed into text format: "Indicator Name: XX, Target Value: XX, Assessment Method: XX; Indicator Name: YY, Target Value: YY, Assessment Method: YY." The transformed key-value pair text is inserted at the end of the paragraph node containing the table node, forming enhanced document content. This processing allows the structured data in the table to participate in subsequent entity recognition and relation extraction processes in the form of natural language text, avoiding the omission or isolated processing of table information.

[0024] Step S240: Clean the enhanced document content to obtain a cleaned set of project application lifecycle documents.

[0025] After enhancing the document content generation, text cleaning is required to remove noise and interfering information. Cleaning operations include: removing irrelevant characters, such as special symbols (#, @, ¥, etc.), control characters (except newline characters and tabs), and consecutive whitespace characters; correcting typos and spelling errors in the text, which can be done by using Chinese word segmentation tools and dictionaries for matching and verification; standardizing text formatting, such as standardizing date format to "YYYY-MM-DD" and standardizing numerical units to standard units (such as "ten thousand yuan", "item", "item", etc.); and filtering duplicate content, such as the same clauses or explanatory text appearing multiple times in the document. For example, "R&D funding 500,000 yuan" is corrected to "R&D funding 500,000 yuan", and "December 2023" is standardized to "2023-12". After cleaning, the text content in the resulting project application full-cycle document collection is more standardized and cleaner.

[0026] Step S250: Assign a globally unique document identifier to each cleaned document, and store them in the corresponding document library according to document type, and establish a document index table.

[0027] To facilitate the management and retrieval of the cleaned documents, each document is assigned a globally unique document ID. Document IDs can be generated using UUIDs (Universally Unique Identifiers) to ensure they are not duplicated throughout the system. Simultaneously, documents are stored in corresponding document repositories based on document type (project proposals, project contracts, mid-term reviews, project completion acceptance), such as creating directories like "Proposal Repository" and "Contract Repository" in the file system, or establishing different document tables in the database. Furthermore, a document index table is created to record metadata such as document ID, document name, document type, creation time, storage path, and summary information. The index table can be implemented using a relational database, supporting queries and sorting by document type, creation time, and other criteria for quick location and access to specific documents.

[0028] Step S120: Based on the named entity recognition model that has been fine-tuned through the text corpus of the project application domain, perform multi-stage structured entity extraction processing on the full-cycle document set of the project application to generate a set of structured entity triples containing logical relationships. Each structured entity triple in the set of structured entity triples contains at least two entity types from the following: technical indicator entity, deliverable entity, time node entity, budget item entity, and legal clause reference entity, as well as a descriptor of the logical relationship between the entities.

[0029] After document preprocessing, the next stage is structured entity extraction. This step aims to extract entities of specific types (such as technical indicators, deliverables, and timelines) from the entire project application lifecycle document collection, identify the logical relationships between these entities, and ultimately organize them into structured entity triples. This process relies on a named entity recognition model fine-tuned with text corpora from the project application domain to improve the accuracy and domain adaptability of entity recognition.

[0030] Step S121: Obtain a pre-trained general named entity recognition basic model, and fine-tune the pre-trained general named entity recognition basic model using a historical project application full-cycle document sample set containing entity type labels and entity relationship labels to obtain a named entity recognition model that has been fine-tuned by the text corpus in the project application field.

[0031] First, a pre-trained general-purpose named entity recognition (NAME) model is selected; in this embodiment, the BERT-Base model is chosen as the base model. This base model is pre-trained on large-scale general-purpose text corpora (such as Wikipedia and BookCorpus) and possesses strong contextual semantic understanding capabilities. Next, a fine-tuned dataset for the project application domain is constructed: historical project application documents throughout their entire lifecycle (including application forms, contracts, mid-term inspections, and final acceptance documents) are collected. Domain experts annotate the entities in these documents, including technical indicator entities, deliverable entities, timeline entities, budget item entities, and legal clause reference entities. The types of relationships between entities are also annotated (such as numerical constraints and temporal dependencies). Brat or LabelStudio is used as the annotation tool to generate annotation files containing entity boundaries, entity types, and relationship types.

[0032] The fine-tuning process consists of two stages: The first stage fine-tunes the entity recognition task by converting labeled data to BIO format (B-entity type, I-entity type, O-non-entity) and inputting it into the model for training, optimizing the entity recognition head. The second stage fine-tunes the relation extraction task by using entity pairs and their relation types as training samples, optimizing the relation classification head. During training, the cross-entropy loss function is used, the optimizer is AdamW, the initial learning rate is set to 2e-5, the batch size is 16, the training epochs are 50, and an early stopping strategy is adopted (training stops when the F1 score on the validation set fails to improve for 5 consecutive epochs). Through fine-tuning, the model can learn entity features and relation patterns specific to the project application domain, thereby improving the performance of entity recognition and relation extraction in this domain.

[0033] Step S122: Divide the project application full-cycle document set into a project application document subset, a project contract document subset, a project mid-term inspection document subset, and a project completion acceptance document subset according to document type. Perform document content segmentation processing on each document subset to obtain multiple document paragraph units corresponding to each document subset.

[0034] The pre-processed project application lifecycle document set is divided into four subsets according to document type: the project application document subset contains all application materials submitted during the project initiation stage; the project contract document subset contains various contract texts signed after project initiation; the project mid-term inspection document subset contains mid-term progress reports submitted during project execution; and the project completion and acceptance document subset contains the completion report and acceptance materials submitted after project completion.

[0035] For each document subset, the document content is segmented. Segmentation is based on paragraph nodes in the document structure tree; each paragraph node's corresponding text content is treated as an independent document paragraph unit. For example, the three paragraph nodes under the "Research Objectives" section of a project application document correspond to three document paragraph units; the converted key-value pairs of the "Budget Table" are treated as an independent paragraph unit. Each document paragraph unit includes attributes such as text content, its document identifier, and its starting position index within the document. This segmentation process breaks down long documents into shorter text units, reducing the complexity of subsequent model processing and improving the efficiency and accuracy of entity recognition.

[0036] Step S123: Input multiple document paragraph units corresponding to each document subset into the named entity recognition model, convert each document paragraph unit into a paragraph unit embedding vector representation through the embedding layer of the named entity recognition model, and perform context feature extraction processing on the paragraph unit embedding vector representation through the bidirectional long short-term memory network layer of the named entity recognition model to generate a paragraph unit feature vector sequence containing context semantic information.

[0037] Each document paragraph unit is taken as input and fed into a finely tuned named entity recognition model. The model's embedding layer first converts each Chinese character in the paragraph unit into a character-level embedding vector, and simultaneously fuses pre-trained word embedding vectors (such as Word2Vec or GloVe) to form an initial word embedding representation. For out-of-vocabulary words, embedding vectors are generated using random initialization. The word embedding sequence output by the embedding layer has dimensions of [sequence length, embedding dimension], where the sequence length is the number of Chinese characters in the paragraph unit, and the embedding dimension is set to 768.

[0038] Next, the embedded vector sequence is input into a Bidirectional Long Short-Term Memory (BiLSTM) layer. The BiLSTM layer contains two directions: a forward LSTM and a backward LSTM, with each direction having a hidden layer dimension of 256. The forward LSTM processes the word embedding sequence from left to right, capturing the positive semantic information of the context; the backward LSTM processes from right to left, capturing the negative semantic information of the context. The hidden states of the forward and backward LSTMs are concatenated along the feature dimension to obtain the context feature vector at each time step, forming a paragraph unit feature vector sequence with dimensions [sequence length, 512]. This paragraph unit feature vector sequence contains the semantic information of each Chinese character within its context.

[0039] Step S124: Perform sequence labeling and decoding processing on the feature vector sequence of the paragraph unit through the conditional random field layer of the named entity recognition model, and output the entity type label corresponding to each word segmentation unit in each document paragraph unit. The entity type label includes technical indicator entity label, deliverable entity label, time node entity label, budget item entity label, legal clause reference entity label and non-entity label.

[0040] The paragraph unit feature vector sequence is input into the Conditional Random Field (CRF) layer for sequence labeling and decoding. The CRF layer takes the feature vector sequence output by BiLSTM as input, learns the transition probabilities between labels, and thus predicts the optimal entity label sequence. The model outputs the entity type label corresponding to each word segmentation unit (in this embodiment, Chinese characters are used as the smallest word segmentation unit). The label set includes: B-TechInd (start of technical indicator entity), I-TechInd (inside technical indicator entity), B-Deliverable (start of deliverable entity), I-Deliverable (inside deliverable entity), B-TimeNode (start of time node entity), I-TimeNode (inside time node entity), B-BudgetItem (start of budget item entity), I-BudgetItem (inside budget item entity), B-Regulation (start of regulatory clause reference entity), I-Regulation (inside regulatory clause reference entity), and O (non-entity).

[0041] For example, for the paragraph unit "The R&D cycle of this project is 24 months, and 3 core technical indicators need to be completed", the tag sequence output by the CRF layer might be: OOOOB-TimeNodeI-TimeNodeOOOB-TechIndI-TechIndI-TechIndO. Through this tag sequence, it can be identified that "24 months" is a time node entity and "3 core technical indicators" is a technical indicator entity.

[0042] Step S125: Based on the entity type tag corresponding to each word segmentation unit in each document paragraph unit, extract all consecutive word segmentation unit combinations with entity type tags from each document paragraph unit to form a set of original entity units to be associated. Each original entity unit to be associated in the set of original entity units to be associated contains entity text content, the starting position index of the entity in the document, and entity type tag.

[0043] Based on the entity type labels output by the CRF layer, entity extraction is performed on each document paragraph unit. Each word segmentation unit within the paragraph unit is traversed. When a tag starting with "B-" is encountered, it is marked as the start of an entity. Traversal continues until an "O" tag or another "B-" tag is encountered. The word segmentation units corresponding to consecutive "I-" tags and the initial "B-" tag are combined to form an entity text content. For example, the word segmentation units corresponding to the tag sequence "B-TechIndI-TechIndI-TechInd" combined form "core technical indicators".

[0044] Each extracted raw entity unit to be associated contains the following attributes: entity text content (e.g., "24 months", "3 core technical indicators"), the entity's starting position index in the document (i.e., the character offset of the first word segmentation unit of the entity in the document), and entity type label (e.g., TimeNode, TechInd). The raw entity units to be associated extracted from all document paragraph units are then aggregated to form a set of raw entity units to be associated.

[0045] Step S126: Perform entity core word extraction processing on each original entity unit to be associated in the set of original entity units to be associated, remove the modifying words in the entity text content, retain the core keywords that can uniquely identify the meaning of the entity, and generate the entity core word representation corresponding to each original entity unit to be associated.

[0046] To accurately resolve entity coreference, it is necessary to extract core words from the text content of each original entity unit to be associated and remove redundant modifiers. For example, in the entity text content "3 core technical indicators that this project plans to complete", "3" are modifiers, and the core word should be "core technical indicators".

[0047] Step S1261: Perform word segmentation and part-of-speech tagging on the entity text content to obtain each word segmentation unit and its part-of-speech tag. The part-of-speech tags include noun tags, verb tags, adjective tags, numeral tags, quantifier tags, and auxiliary word tags.

[0048] Use a Chinese word segmentation tool (such as Jieba) to segment the entity text content, splitting the continuous text into independent word segmentation units. For example, after segmenting the text "The 3 core technical indicators to be completed in this project", we get "this / project / planned / complete / of / 3 / items / core / technology / indicators". Then, use a词性标注工具 (such as HanLP) to perform词性标注 on each word segmentation unit, obtaining the词性标签 of each word segmentation unit, such as "this (pronoun)", "project (noun)", "planned (verb)", "complete (verb)", "of (particle)", "3 (numeral)", "items (measure word)", "core (adjective)", "technology (noun)", "indicators (noun)".

[0049] Step S1262: According to the词性标签, screen out the word segmentation units with词性标签 of noun label, verb label, and numeral label from all word segmentation units as candidate core word units, and filter out the word segmentation units of other词性.

[0050] Screen candidate core word units according to the词性标签, retaining nouns, verbs, and numerals, and filtering out modifying or auxiliary words such as pronouns, particles, adjectives, and measure words. Taking the above example, the candidate core word units obtained after screening are "project (noun)", "planned (verb)", "complete (verb)", "3 (numeral)", "technology (noun)", "indicators (noun)".

[0051] Step S1263: Calculate the term frequency-inverse document frequency weight value of each candidate core word unit in the entity text content, sort them from high to low according to the term frequency-inverse document frequency weight value, and select the top pre-set number of candidate core word units with the highest term frequency-inverse document frequency weight value as the core keyword units.

[0052] Calculate the term frequency-inverse document frequency (TF-IDF) weight value of the candidate core word units. Term frequency (TF) refers to the number of times a candidate core word unit appears in the current entity text content divided by the total number of words in the entity text content; inverse document frequency (IDF) refers to the total number of documents in the project declaration full-cycle document set divided by the number of documents containing the candidate core word unit, and then taking the logarithm. The TF-IDF weight value is the product of TF and IDF. For example, "indicators" appears frequently in multiple documents, and its IDF value is relatively low; while "core technology" appears with a relatively high frequency in specific domain documents, and its TF value is relatively high. Sorting according to the TF-IDF weight value, the pre-set number is set to 3, and the top 3 candidate core word units with the highest weight value are selected, such as "technology", "indicators", "project".

[0053] It should be noted that the Chinese terms "词性标注工具" and "词性标签" in the original text seem to be specific terms in a certain context. Here, I have directly translated them as "词性标注工具" and "词性标签" according to the rules, but it may be more appropriate to use more accurate and specific translations if there are more detailed definitions or common translations in the relevant field. If you can provide more context or clarify these terms, the translation can be further optimized.Step S1264: Concatenate the core keyword units according to their original order in the entity text content to generate the entity core word representation corresponding to each original entity unit to be associated, and perform stop word filtering and word form restoration on the entity core word representation to obtain the final version of the entity core word representation.

[0054] The selected core keyword units are concatenated according to their original order in the entity text content to form a preliminary entity core word representation, such as "project technical indicators". Then, stop word filtering is performed to remove common words such as "project" that may affect the uniqueness of the entity, resulting in "technical indicators". Finally, lexical restoration is performed (e.g., restoring plural forms to singular forms and verbs to their base forms). Since Chinese does not have strict lexical changes, this step mainly ensures the consistency of the core word representation, ultimately yielding the entity core word representation "technical indicators".

[0055] Step S127: Based on the entity core word representation and entity type label corresponding to each original entity unit to be associated, perform cross-document and cross-paragraph entity core reference resolution on the set of original entity units to be associated, calculate the semantic similarity of any two original entity units to be associated at the entity core word representation level and the consistency at the entity type label level, determine the original entity units to be associated whose semantic similarity exceeds the preset similarity threshold and whose entity type labels are consistent as core reference entity units pointing to the same entity object, and assign a unified global entity identifier to each group of core reference entity units.

[0056] Entity coreference resolution aims to associate different original entity units that point to the same entity object. For example, "R&D cycle" and "project execution cycle" may point to the same entity at the same point in time.

[0057] Step S1271: For any two original entity units to be associated, convert their respective entity core word representations into core word embedding vectors, and calculate the cosine similarity between the two core word embedding vectors as the core word semantic similarity score.

[0058] Pre-trained word vector models (such as Word2Vec or BERT word vectors) are used to convert entity core word representations into core word embedding vectors. For example, the core word embedding vector for "R&D cycle" is V1, and the core word embedding vector for "project execution cycle" is V2. The cosine similarity between V1 and V2 is calculated using the formula: Cosine Similarity = (V1·V2) / (||V1||||V2||), where "·" represents the vector dot product, and "||||" represents the magnitude of the vectors. The cosine similarity ranges from -1 to 1, with values ​​closer to 1 indicating greater semantic similarity.

[0059] Step S1272: Compare whether the entity type labels of the two original entity units to be associated are completely identical, and generate an entity type consistency identifier.

[0060] If the entity type labels of the two original entity units to be associated are both TimeNode, then the entity type consistency is marked as "consistent"; if one is TechInd and the other is Deliverable, then it is marked as "inconsistent".

[0061] Step S1273: If the semantic similarity score of the core words exceeds the preset semantic similarity threshold and the entity type consistency is marked as consistent, then the two original entity units to be associated are marked as candidate core-referenced entity pairs, and the distance penalty factor is calculated based on the positional distance of the two entities in the document, and then the comprehensive core-reference confidence score is calculated.

[0062] The preset semantic similarity threshold is set to 0.8. When the semantic similarity score of the core words exceeds 0.8 and the entity types are consistent, they are marked as candidate core-referenced entity pairs. Locational distance refers to the physical distance between two entities in a document set (e.g., the number of documents or paragraphs spanned); the closer the distance, the higher the likelihood of core-reference. The distance penalty factor is calculated as: Distance Penalty Factor = 1 / (1 + Distance), where distance is the sum of the number of documents and paragraphs between the two entities. The overall core-reference confidence score = Core Word Semantic Similarity Score × Distance Penalty Factor.

[0063] Step S1274: If the overall coreference confidence score exceeds the preset overall coreference confidence threshold, then the two original entity units to be associated are determined to be coreference entity units pointing to the same entity object, and a coreference relationship record is established for them.

[0064] The preset overall common reference confidence threshold is set to 0.7. When the overall common reference confidence score exceeds 0.7, it is determined to be a common reference entity, and information such as the entity identifier and overall common reference confidence score of both entities is recorded.

[0065] Step S1275: Repeat the above process for each pair of original entity units to be associated in the set of original entity units to be associated, construct a global coreference graph containing all coreference records, perform connected component analysis on the global coreference graph, and assign a unified global entity identifier to all original entity units to be associated within the same connected component.

[0066] In the global coreference graph, nodes are the original entity units to be associated, and edges are coreference relationship records. Through connected component analysis (such as depth-first search or breadth-first search), entity units that have coreference relationships with each other are divided into connected components. Each connected component corresponds to a unique entity object, which is assigned a global entity identifier (such as Ent_001, Ent_002, etc.).

[0067] Step S128: Obtain the co-occurrence distance, co-occurrence frequency, and syntactic dependency relationship between any two original entity units to be associated in each document paragraph unit. Construct a set of candidate logical association relationships between any two original entity units to be associated based on the co-occurrence distance, co-occurrence frequency, and syntactic dependency relationship. Each candidate logical association relationship in the set of candidate logical association relationships includes the entity identifier of the relationship initiator, the entity identifier of the relationship receiver, and the initial probability distribution of the relationship type.

[0068] Co-occurrence refers to the phenomenon where two original entity units to be associated co-occur within the same document paragraph unit. Co-occurrence distance refers to the difference in position (e.g., number of characters or sentences) between the two entities within the paragraph; the smaller the distance, the higher the likelihood of association. Co-occurrence frequency refers to the number of times the two entities co-occur in all document paragraph units; the higher the frequency, the higher the likelihood of association. Syntactic dependency relations are obtained through dependency parsing tools (such as Stanford Parser), such as subject-verb, verb-object, and modifier-head relations. Different syntactic relations may correspond to different types of logical associations.

[0069] Based on co-occurrence distance, co-occurrence frequency, and syntactic dependency, an initial probability distribution of relation types is calculated for each pair of co-occurring entity units. For example, if two entities have a verb-object relation (such as "completion-indicator"), the initial probability of a "numerical constraint relation" is higher; if a symmetric relation (such as "24 months-R&D cycle") exists, the initial probability of a "temporal dependency relation" is higher. The initial probability distribution of relation types is a probability vector containing all possible relation types (numerical constraint, temporal dependency, delivery correspondence, budget allocation, regulatory reference), with the sum of all probabilities being 1.

[0070] Step S129: Input each document paragraph unit into the pre-trained language model that has been fine-tuned by the relation extraction task, perform relation classification processing on any two original entity units to be associated in each document paragraph unit, and output the target logical association type and relation confidence score between any two original entity units to be associated. The target logical association type includes numerical constraint relationship type, temporal dependency relationship type, delivery correspondence relationship type, budget allocation relationship type and regulatory reference relationship type.

[0071] A BERT model fine-tuned for relation extraction was selected for relation classification. Two original entity units to be associated from a document paragraph and their context text were used as model input. The model outputs the probability that the entity pair belongs to each target logical relation type. The type with the highest probability was taken as the target logical relation type, and its probability value was used as the relation confidence score. For example, given the input "R&D cycle is 24 months," where the entity pair is "R&D cycle" (TimeNode) and "24 months" (numerical value), the model outputs "Time-series dependency," with a relation confidence score of 0.92.

[0072] Step S1210: Based on the target logical relationship type and relationship confidence score, the candidate logical relationships in the candidate logical relationship set are screened, and the candidate logical relationships with relationship confidence scores exceeding the preset relationship confidence threshold are retained as confirmed logical relationships, and a relationship identifier is assigned to each confirmed logical relationship.

[0073] The preset relationship confidence threshold is set to 0.85. When the relationship confidence score exceeds 0.85, the corresponding candidate logical relationship is confirmed as a confirmed logical relationship and a unique relationship identifier (such as Rel_001, Rel_002, etc.) is assigned.

[0074] Step S1211: Take each entity object with a global entity identifier as the subject entity of the structured entity triple, take the entity object with a confirmed logical relationship with the subject entity as the object entity of the structured entity triple, and take the confirmed logical relationship type between the subject entity and the object entity as the predicate relation of the structured entity triple. Combine them to generate a structured entity triple containing logical relationships. Each structured entity triple contains a subject entity global identifier, a predicate relation type identifier, an object entity global identifier, and a relation confidence score. At the same time, the subject entity and object entity in the structured entity triple carry at least one entity type label from the following categories: technical indicator entity label, deliverable entity label, time node entity label, budget item entity label, and regulatory clause reference entity label.

[0075] For example, if the global identifier for the subject entity is Ent_001 (entity type label is TimeNode, entity core term is "R&D cycle"), the predicate relation type identifier is Rel_001 (temporal dependency), the global identifier for the object entity is Ent_002 (entity type label is numeric, entity core term is "24 months"), and the relation confidence score is 0.92, then the generated structured entity triples are (Ent_001, Rel_001, Ent_002, 0.92). All structured entity triples are combined to form a set of structured entity triples.

[0076] Step S130: Construct a set of compliance review prompt engineering instructions and input the set of structured entity triples into the large language model according to a standardized input format. Drive the large language model to perform multi-document entity alignment and consistency verification processing, identify the numerical evolution characteristics, semantic deviation characteristics and logical contradiction characteristics of the same entity in different documents of the full life cycle document set of the project application, and generate entity relationship alignment and logical contradiction identification results containing conflict type identifiers and citation identifiers according to the preset compliance rule base.

[0077] After obtaining the set of structured entity triples, it is necessary to use a large language model to verify the consistency of entities across documents. This step guides the large language model to compare and analyze the behavior of the same entity in different documents by constructing a set of prompting engineering instructions, thereby identifying potential compliance issues.

[0078] Step S131: Parse each structured entity triple in the set of structured entity triples, extract the subject entity global identifier, predicate relation type identifier, and object entity global identifier contained in each structured entity triple, and obtain the entity text content, entity type label, and the starting position index of the entity in the document corresponding to each entity object from the set of original entity units to be associated based on the subject entity global identifier and the object entity global identifier, and construct a structured entity information mapping table containing complete entity information.

[0079] Iterate through each triple in the set of structured entity triples, extracting the subject entity global identifier (e.g., Ent_001), the predicate relation type identifier (e.g., Rel_001), and the object entity global identifier (e.g., Ent_002). Based on these identifiers, locate the corresponding entity object in the set of original entity units to be associated, and obtain its entity text content (e.g., "R&D cycle", "24 months"), entity type label (e.g., TimeNode, numerical value), and its starting position index in the document (e.g., characters 100-105 of document Doc_001). Integrate the above information into a structured entity information mapping table, which uses the entity global identifier as the key and the entity details as the value, facilitating quick querying of the entity's complete attributes.

[0080] Step S132: Based on the entity type label corresponding to each entity object in the structured entity information mapping table, arrange the entity text content of entity objects with the same global entity identifier in different documents according to the document time order to generate a multi-document entity text content sequence for each global entity object. The multi-document entity text content sequence includes the entity text content in the project application document, the entity text content in the project contract document, the entity text content in the project mid-term inspection document, and the entity text content in the project completion acceptance document.

[0081] Based on the chronological order of document types (project proposal → project contract → mid-term review → final acceptance), the entity text content of the same global entity identifier in different documents is sorted. For example, the text content of the global entity identifier Ent_001 (R&D cycle) is "24 months" in the project proposal, "2 years" in the project contract, "executed for 12 months" in the mid-term review, and "actual completion cycle 26 months" in the final acceptance. Then the multi-document entity text content sequence is ["24 months", "2 years", "executed for 12 months", "actual completion cycle 26 months"].

[0082] Step S133: Perform numerical entity content detection processing on the multi-document entity text content sequence of each global entity object, identify whether each entity text content in the multi-document entity text content sequence contains numerical information, and if the entity text content contains numerical information, extract the numerical part and the numerical unit part from the entity text content to generate a numerical entity representation corresponding to each entity text content containing numerical information.

[0083] Regular expressions are used to match numerical information in entity text content. For example, in "24 months", "24" is the numerical part and "months" is the unit; in "2 years", "2" is the numerical part and "years" is the unit. For entity text content containing numerical information, it is converted into a numerical entity representation, such as (24, months) or (2, years). For entity text content that does not contain numerical information (such as "core technology"), it is marked as a non-numerical entity.

[0084] Step S134: Combine and encode the multi-document entity text content sequence of each global entity object and the numerical entity representation corresponding to each entity text content containing numerical information according to a preset prompt engineering template format to generate entity alignment and consistency verification task prompt word unit corresponding to each global entity object. The entity alignment and consistency verification task prompt word unit includes a task instruction prefix, a global identifier of the entity to be verified, a list of entity text content of the entity to be verified in different documents, a list of numerical entity representations of the entity to be verified in different documents, and an output format requirement descriptor.

[0085] The default prompt template format is as follows: "Task: Verify the consistency of entity {Entity ID} across different documents. Entity type: {Type}. Document text sequence: {Text list}. Numerical representation sequence: {Number list}. Please analyze whether there are numerical evolutions, semantic deviations, or logical contradictions, and output them in the following format: Conflict type | Affected documents | Description | Confidence level."

[0086] For example, for Ent_001, the generated prompt word unit is: "Task: Verify the consistency of entity Ent_001 across different documents. Entity type: TimeNode. Document text sequence: ["24 months", "2 years", "Executed for 12 months", "Actual completion period 26 months"]. Numerical representation sequence: [(24, months), (2, years), (12, months), (26, months)]. Please analyze whether there are numerical evolutions, semantic deviations, or logical contradictions, and output them in the following format: Conflict type | Affected documents | Description | Confidence."

[0087] Step S135: Aggregate the entity alignment and consistency verification task prompt word units corresponding to each global entity object to generate a batch entity alignment and consistency verification task prompt word set containing all global entity objects to be verified, and add global task context description information and multi-task collaborative processing instructions to the batch entity alignment and consistency verification task prompt word set to form a compliance review prompt engineering instruction set.

[0088] All entity alignment and consistency verification task prompts are grouped into a batch task set, and a global context description is added: "The following are entity consistency verification tasks in the full lifecycle documents of an enterprise scientific research project, involving entity types such as technical indicators, deliverables, and time nodes. Please analyze the performance of each entity in the application form, contract, mid-term inspection, and final acceptance documents to identify potential compliance issues." A multi-task collaborative processing instruction is also added: "If there are relationships between multiple entities (such as the correspondence between technical indicators and deliverables), please comprehensively consider their consistency."

[0089] Step S136: Input the compliance review prompt engineering instruction set as input prompt text into the large language model. Convert the input prompt text into a prompt text embedding vector sequence through the input embedding layer of the large language model. Process the prompt text embedding vector sequence through the self-attention mechanism and feedforward neural network through the Transformer decoder layer of the large language model to generate a prompt text decoding feature representation containing instruction understanding semantics.

[0090] GPT-4 was chosen as the large language model. The input prompt text first passes through an input embedding layer, converting each token in the text into an embedding vector with a dimension of 1024. The sequence of embedding vectors is then fed into a Transformer decoder layer (containing 24 decoder blocks), each block comprising a multi-head self-attention sublayer and a feedforward neural network sublayer. The self-attention mechanism allows the model to focus on different parts of the prompt text, capturing the relationships between entities and contextual information; the feedforward neural network performs a non-linear transformation on the attention output, enhancing feature representation capabilities. After multi-layer decoding, a decoded feature representation of the prompt text containing the semantic information required for instruction understanding and task execution is generated.

[0091] Step S137: Based on the decoding feature representation of the prompt text, the large language model performs a multi-document entity alignment and consistency verification task. For each global entity object, the large language model first compares the entity text content of the global entity object in different documents, calculates the semantic similarity of the entity text content between different documents, and if the semantic similarity is lower than the preset semantic similarity threshold, it is determined that there is a semantic deviation feature, and generates a semantic deviation type identifier and semantic deviation description text.

[0092] Large language models utilize their internal semantic understanding capabilities to calculate the semantic similarity of the text content of the same entity in different documents. For example, comparing the semantic similarity between "R&D cycle" (application form) and "project execution cycle" (contract), if the similarity is lower than a preset threshold (e.g., 0.7), a semantic deviation is determined, generating a semantic deviation type identifier (e.g., SD_001) and the descriptive text "The application form uses 'R&D cycle,' while the contract uses 'project execution cycle,' indicating a name inconsistency."

[0093] Step S138: For each global entity object containing a numerical entity representation, the large language model compares the numerical entity representation of the global entity object in different documents, calculates the amount and trend of change of the numerical part between different documents, and if the amount of change exceeds the preset allowable range of numerical change or the trend of change does not conform to the project progress logic, it is determined that there is a numerical evolution feature, and a numerical evolution type identifier and a numerical evolution description text are generated.

[0094] Step S1381: Obtain the time-sorted sequence of numerical entity representations corresponding to each global entity object containing numerical entity representations.

[0095] For example, the numerical entity representation sequence of Ent_001 is [(24, months), (2, years), (12, months), (26, months)]. First, unify the units (e.g., convert "year" to "month") to get [(24, months), (24, months), (12, months), (26, months)].

[0096] Step S1382: Extract the numerical part of the two adjacent numerical entities in the sequence, calculate the difference between the next value and the previous value as the change in value between adjacent documents, and determine whether the trend of the change in value is an increasing trend or a decreasing trend based on the sign of the difference.

[0097] Calculate the numerical changes between adjacent documents: the change between the contract and the application is 24-24=0 (no change); the change between the mid-term inspection and the contract is 12-24=-12 (reducing trend); the change between the final acceptance and the mid-term inspection is 26-12=14 (increasing trend).

[0098] Step S1383: Based on the entity type tag of the global entity object, query the corresponding preset value change allowable range from the preset value change allowable range library, divide the value change between adjacent documents by the previous value to obtain the change percentage, and if the change percentage exceeds the preset value change allowable range, generate a value out of range flag.

[0099] In the preset numerical change allowable range library, the allowable change percentage for time node entities is ±10%. The change percentage for mid-term inspections and contracts is (12-24) / 24=-50%, which exceeds the allowable range, generating a numerical out-of-range flag.

[0100] Step S1384: Based on the project stage information of the global entity object, determine whether the trend of numerical change conforms to the project progress logic. If not, generate a trend anomaly flag.

[0101] During the mid-term review phase of the project, the time node value should be less than the contractual value (24 months). Currently, it is 12 months, which is consistent with the progress logic. During the final acceptance phase, the time node value of 26 months is greater than the contractual value of 24 months, and there is no reasonable explanation. This does not conform to the progress logic, so an abnormal trend indicator is generated.

[0102] Step S1385: If there is at least one of the numerical out-of-range identifier or the trend abnormal identifier, it is determined that there is a numerical evolution feature, and a corresponding numerical evolution type identifier and numerical evolution description text are generated for the global entity object.

[0103] Due to the presence of out-of-range and abnormal trend indicators, Ent_001 is determined to have numerical evolution characteristics. A numerical evolution type identifier (such as NE_001) and the descriptive text "The progress was only 50% complete (12 / 24) during the mid-term check, which is below the allowable deviation range; the project completion period of 26 months exceeds the contractually agreed 24 months without a reasonable explanation" are generated.

[0104] Step S139: For each global entity object, the large language model analyzes whether the logical association between the global entity object and other entity objects remains consistent in different documents based on the structured entity triples formed by the global entity object. If the type of logical association between the global entity object and the same other entity object changes in different documents, or if the logical association between the global entity object and different entity objects in different documents conflicts, then it is determined that there is a logical contradiction feature, and a logical contradiction type identifier and logical contradiction description text are generated.

[0105] For example, if global entity Ent_003 (technical indicator A) has a "delivery correspondence" with Ent_004 (deliverable X) in the application form, but has a "delivery correspondence" with Ent_005 (deliverable Y) in the contract, and X and Y are different deliverables, then a logical contradiction is determined, and a logical contradiction type identifier (such as LC_001) and the descriptive text "Technical indicator A corresponds to deliverable X in the application form and deliverable Y in the contract, the deliverables are inconsistent" are generated.

[0106] Step S1310: The large language model matches the identified semantic deviation features, numerical evolution features, and logical contradiction features with the rule conditions in the compliance rule library according to the preset compliance rule library. The compliance rule library contains multiple compliance rules for project application content. Each compliance rule includes the applicable entity type, applicable feature type, rule judgment conditions, and a description of the consequences of rule violation. The large language model calculates the matching degree between each identified feature and each compliance rule. If the matching degree exceeds the preset rule matching degree threshold, it is determined that the feature violates the corresponding compliance rule.

[0107] Example rule from the compliance rule library: Rule R001, applicable entity type TimeNode, applicable feature type Numerical Evolution, rule judgment condition "the completion period exceeds the contract period by more than 10%", rule violation consequence "an explanation for the extension must be provided". The large language model calculates the matching degree between the numerical evolution feature of Ent_001 and R001 (e.g., 0.95). If it exceeds the preset rule matching degree threshold (0.8), it is determined that R001 is violated.

[0108] Step S1311: For each feature determined to violate compliance rules, the large language model generates a conflict record unit containing a conflict type identifier, a violated rule identifier, a feature description text involved, and a global entity object identifier corresponding to the feature. Based on the global entity object identifier, the model obtains the starting position index of the entity text content corresponding to the feature in the document from the structured entity information mapping table, and generates a reference citation identifier containing the document name, paragraph number, and sentence offset.

[0109] Example of a conflict record unit: Conflict type identifier NE_001, rule violation identifier R001, feature description text "completion period 26 months, exceeding the contractual agreement of 24 months", global entity object identifier Ent_001. Based on the citation identifier, it includes the document name "completion acceptance report.docx", paragraph number 3, and sentence offset 15-20 (corresponding to "actual completion period 26 months").

[0110] Step S1312: The large language model outputs all conflict record units and the corresponding reference citation identifiers for each conflict record unit in a structured manner according to the preset output format requirements, generating entity relationship alignment and logical contradiction identification results. The entity relationship alignment and logical contradiction identification results include multiple conflict record entries. Each conflict record entry includes a conflict type identifier, a violated rule identifier, the feature description text involved, a global entity object identifier, and a reference citation identifier.

[0111] The results of entity relationship alignment and logical contradiction identification are output in JSON format. Each conflict record corresponds to a compliance issue, which facilitates subsequent processing.

[0112] Step S140: Based on the original text position mapping relationship corresponding to the inconsistency item in the entity relationship alignment and logical contradiction identification results, perform traceable review evidence chain construction processing on the project application full-cycle document set to generate a traceable review evidence chain set containing the original text content of the inconsistency item, the coordinate positioning information of the inconsistency item in the document, and the association relationship between the inconsistency item and the compliance rules.

[0113] To ensure the traceability of the review results, it is necessary to build a chain of evidence that links conflict records to the content of the original documents.

[0114] Step S141: Parse the entity relationship alignment and logical contradiction identification results, and extract the conflict type identifier, violated rule identifier, involved feature description text, global entity object identifier and basis citation identifier contained in each conflict record entry. The basis citation identifier includes document name field, chapter number field, paragraph number field and statement number field.

[0115] Parse the JSON data in the entity relationship alignment and logical contradiction identification results to extract detailed information for each conflict record entry. For example, extract the conflict type identifier NE_001, the rule violation identifier R001, the feature description text "completion period 26 months exceeds the contractual agreement of 24 months" from a certain conflict record entry, the global entity object identifier Ent_001, and the citation identifier {document name: "completion acceptance report.docx", chapter number: "2.1", paragraph number: 3, statement number: 2}.

[0116] Step S142: Based on the document name field contained in the citation identifier in each conflict record entry, locate the corresponding target document from the project application full-cycle document set. Based on the chapter number field, paragraph number field, and statement number field contained in the citation identifier, locate the corresponding target statement unit in the target document and extract the complete text content of the target statement unit as the original content of the inconsistency item.

[0117] The target document, "Completion and Acceptance Report.docx", was located in the document library. The "Project Execution Status" section was located based on the chapter number "2.1". The third paragraph of this section was found based on the paragraph number 3. The second sentence of this paragraph, "The actual completion period of this project was 26 months, which was 2 months longer than the contractual agreement," was extracted as the original text of the inconsistency item.

[0118] Step S143: Based on the document name field, chapter number field, paragraph number field and statement number field included in the citation identifier, generate the coordinate location information of the inconsistency item in the document corresponding to each conflict record entry. The coordinate location information includes the document identifier, chapter path string, paragraph index value and statement sequence number in the paragraph. The coordinate location information is used to uniquely determine the precise location of the inconsistency item in the project application full-cycle document set.

[0119] The document identifier is the globally unique ID of the target document (e.g., Doc_004), the chapter path string is "Project Execution Status / 2.1 Execution Cycle", the paragraph index value is 3 (counting from 0), and the statement's sequence number within the paragraph is 2 (counting from 1). The coordinate positioning information is combined as "Doc_004|Project Execution Status / 2.1 Execution Cycle|3|2".

[0120] Step S144: Obtain the global entity object identifier in each conflict record entry. Based on the global entity object identifier, query all entity text content corresponding to the global entity object and the document name and position index corresponding to each entity text content from the structured entity information mapping table. Generate a multi-document entity occurrence position list of the global entity object. The multi-document entity occurrence position list contains the position information of the global entity object in the project application document, project contract document, project mid-term inspection document, and project completion acceptance document.

[0121] For Ent_001, retrieve its location information from the structured entity information mapping table, including its location information in the application (Doc_001, Chapter 1.2, Paragraph 2, Statement 3), contract (Doc_002, Chapter 3.1, Paragraph 1, Statement 1), mid-term inspection (Doc_003, Chapter 2.3, Paragraph 4, Statement 2), and final acceptance (Doc_004, Chapter 2.1, Paragraph 3, Statement 2), and generate a list of the locations where the multi-document entity appears.

[0122] Step S145: Based on the conflict type identifier and the rule identifier violated in each conflict record entry, extract the corresponding detailed rule description text and rule violation consequence text from the preset compliance rule base, and use the detailed rule description text and rule violation consequence text as the description of the association between the conflict record entry and the compliance rule.

[0123] Extract the detailed description of rule R001 from the compliance rule library: "The project completion period shall not exceed 10% of the contractually agreed period. For any excess, a written extension application and approval document must be provided." and the text of the consequences of violating the rule: "Failure to provide an extension explanation does not comply with project management regulations." Use these as related descriptions.

[0124] Step S146: Construct a traceable review evidence chain node unit for each conflict record entry. The traceable review evidence chain node unit includes a node identifier, a conflict type identifier field, a rule violation identifier field, a relationship description field, an inconsistency original text field, a coordinate positioning information field, and a multi-document entity occurrence location list field. The node identifier is generated based on the combination of the global entity object identifier and the conflict type identifier of the conflict record entry to ensure that each evidence chain node unit has a unique identifier.

[0125] The node identifier is generated using the rule "Ent_001_NE_001" to ensure uniqueness. Each field of the evidence chain node unit is filled with the corresponding content to form a complete node information.

[0126] Step S147: Perform node association construction processing on the traceable review evidence chain node units. Based on the global entity object identifier in each traceable review evidence chain node unit, associate the traceable review evidence chain node units corresponding to different conflict types with the same global entity object identifier to generate a set of evidence chain node association relationships based on the same global entity object.

[0127] For example, Ent_001 may simultaneously contain node units with two conflict types: numerical evolution (NE_001) and semantic deviation (SD_001). These two node units are linked together through a global entity object identifier to form a set of evidence chain node association relationships.

[0128] Step S148: Based on the list of multiple document entity occurrence locations in each traceable review evidence chain node unit, extract the document stage information corresponding to each occurrence location, sort the evidence chain node units of the same global entity object according to the time order of the document stages, and generate a time-sequential evidence chain structure containing evidence chain sub-nodes of the project application stage, evidence chain sub-nodes of the project contract stage, evidence chain sub-nodes of the project mid-term inspection stage, and evidence chain sub-nodes of the project completion and acceptance stage.

[0129] Step S1481: Parse the global entity object identifier corresponding to each conflict record entry, obtain the list of multiple document entity occurrence positions of the global entity object, and each occurrence position record includes a document name field and the starting position index in the document.

[0130] For example, the multi-document entity occurrence location list of Ent_001 contains four occurrence location records, which correspond to the application form, contract, mid-term inspection, and final acceptance document, respectively.

[0131] Step S1482: Match the document name field with the preset document stage mapping table, assign a corresponding document stage label to each occurrence location record, and obtain a set of occurrence location records carrying document stage labels.

[0132] A pre-defined document stage mapping table is used: Project Application Form → "Application Stage", Project Contract → "Contract Stage", Mid-term Inspection Document → "Mid-term Stage", Project Completion and Acceptance Document → "Completion Stage". A corresponding label is assigned to each occurrence location.

[0133] Step S1483: Copy the traceable review evidence chain node unit corresponding to the conflict record entry multiple times, so that the number of copies is the same as the number of occurrence location records, and associate each copied evidence chain node unit with an occurrence location record, fill in the coordinate positioning information, and generate a staged evidence chain sub-node corresponding to each occurrence location.

[0134] Copy the evidence chain node unit of Ent_001 four times, associate it with the four occurrence location records respectively, fill in their respective coordinate positioning information, and generate staged evidence chain sub-nodes for the application stage, contract stage, mid-term stage, and final stage.

[0135] Step S1484: Based on the document stage tags, classify the staged evidence chain sub-nodes into the project application stage, project contract stage, project mid-term inspection stage, and project completion acceptance stage sub-node sets respectively, and select the sub-node with the smallest starting position index from each stage sub-node set as the evidence chain sub-node of the corresponding stage.

[0136] The earliest appearing child node from the set of child nodes in each stage is selected as the representative evidence chain child node for that stage.

[0137] Step S1485: Arrange the selected evidence chain sub-nodes from the project application stage, project contract stage, project mid-term inspection stage, and project completion acceptance stage in chronological order according to the project stage to generate a chronological evidence chain structure.

[0138] The sub-nodes are arranged in the order of "application stage → contract stage → mid-term stage → project completion stage" to form a time-sequential evidence chain structure, which intuitively shows the changes of the entity throughout the entire cycle.

[0139] Step S149: Perform hyperlink mapping processing on the original text content of the inconsistency item corresponding to each evidence chain sub-node in the time-series evidence chain structure and the coordinate positioning information to generate a clickable positioning identifier corresponding to each evidence chain sub-node. The clickable positioning identifier is used to directly jump to the corresponding original text position in the full life cycle document set of the project application after clicking in the visualization interface.

[0140] Clickable location identifiers are URL links containing coordinate location information, such as "doc: / / Doc_004#Project Execution Status / 2.1 Execution Cycle|3|2". Clicking this identifier in the visualization interface will directly open the corresponding document and locate the original text of the inconsistency.

[0141] Step S1410: Describe the evolution relationship between adjacent evidence chain sub-nodes in the time-series evidence chain structure. Based on the conflict type identifiers and feature description texts corresponding to the adjacent evidence chain sub-nodes, generate entity evolution description texts between adjacent stages. The entity evolution description texts between adjacent stages are used to explain the specific manifestations of numerical changes, semantic shifts, or logical contradictions of the same entity object in different project stages.

[0142] For example, the sub-nodes for both the application stage and the contract stage are "24 months", and the evolution description text is "The R&D cycle for the application stage and the contract stage is consistent, both being 24 months"; the sub-nodes for the contract stage and the mid-term stage change from "24 months" to "12 months", and the evolution description text is "The R&D cycle progress at the mid-term inspection is 12 months, which is no deviation from the contractually agreed progress (12 months)"; the sub-nodes for the mid-term stage and the completion stage change from "12 months" to "26 months", and the evolution description text is "The actual cycle at the completion and acceptance stage is 26 months, which is 2 months longer than the contractually agreed period, exceeding the allowable deviation range".

[0143] Step S1411: Aggregate and encapsulate all the completed traceable review evidence chain node units, node relationship sets, temporal evidence chain structures, and entity evolution description texts between adjacent stages to generate a traceable review evidence chain set. The traceable review evidence chain set contains multiple independent evidence chain entries. Each evidence chain entry corresponds to a global entity object with compliance issues. Each evidence chain entry contains the original text of inconsistencies of the entity object at each stage of the project lifecycle, coordinate positioning information, relationship with compliance rules, stage evolution descriptions, and clickable positioning identifiers.

[0144] By integrating all the above components, a traceable audit evidence chain is formed, with each evidence chain entry fully recording the compliance issues of an entity and the evidence thereto.

[0145] For example, the method may also include:

[0146] Step S150: Perform an integrity check on each evidence chain entry in the traceable review evidence chain set to check whether it contains evidence chain sub-nodes from at least two different project phases.

[0147] Integrity checks ensure that the chain of evidence reflects changes to an entity at least in two phases, avoiding isolated points of failure. For example, an issue that only arises in the closing phase may not constitute a compliance conflict if there is no comparative data from other phases.

[0148] Step S151: Assess the severity level of the evidence chain entries that pass the integrity check, and calculate the comprehensive severity score based on the number of conflict type identifiers and the rule weights corresponding to the rule violation identifiers.

[0149] The more types of conflict type identifiers there are (e.g., simultaneous numerical evolution and logical contradictions), the higher the rule weight (e.g., a violation of core rule R001 has a weight of 5, while a violation of a general rule has a weight of 2), and the higher the overall severity score. The calculation formula is: Overall Severity Score = Σ (Number of Conflict Types × Rule Weight).

[0150] Step S152: Sort the evidence chain items in descending order according to the comprehensive severity score, extract the top preset number of evidence chain items with the highest comprehensive severity scores as key evidence chain items, and add key attention marks.

[0151] The preset quantity is set to 10, and the top 10 evidence chains with the highest comprehensive severity scores are marked as key concerns and given priority processing.

[0152] Step S153: Integrate the key evidence chain items and other evidence chain items according to the report template format to generate a project application compliance review report, and store the project application compliance review report in association with the traceable review evidence chain set.

[0153] The report template includes basic project information, a list of key concerns, a summary of all evidence items, and compliance conclusions. The review report is linked to the traceable evidence chain set via evidence chain node identifiers, facilitating access to detailed evidence.

[0154] Step S154: Receive the review feedback information returned by the review terminal, associate the review feedback information with the corresponding evidence chain item and store it, and update the traceable review evidence chain set.

[0155] Reviewers manually verify the evidence chain items through the review terminal and provide feedback (such as "verified, requires supplementary explanation" or "misjudgment, conflict revoked"). The system stores the feedback information in the corresponding evidence chain item and updates the evidence chain status.

[0156] In one exemplary embodiment, a project application content compliance review system combining named entity recognition and large modeling is provided. This system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2 As shown, this project application content compliance review system, which combines named entity recognition and large modeling, includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a project application content compliance review method that combines named entity recognition and large modeling. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or it can be a button, trackball, or touchpad set on the shell of the project application content compliance review system that combines named entity recognition and large model, or it can be an external keyboard, touchpad, or mouse, etc.

[0157] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for reviewing the compliance of project application content by combining named entity recognition and large-scale modeling, characterized in that, The method includes: Obtain the complete set of project application documents corresponding to the project to be reviewed. The complete set of project application documents includes project application documents, project contract documents, project mid-term inspection documents, and project completion and acceptance documents. Based on the named entity recognition model that has been fine-tuned through text corpus in the project application field, a multi-stage structured entity extraction process is performed on the full-cycle document set of the project application to generate a set of structured entity triples containing logical relationships. Each structured entity triple in the set of structured entity triples contains at least two entity types from the following entities: technical indicator entity, deliverable entity, time node entity, budget item entity, and legal clause reference entity, as well as a descriptor of the logical relationship between the entities. A set of compliance review prompts for engineering instructions is constructed and the set of structured entity triples is input into a large language model according to a standardized input format. The large language model is driven to perform multi-document entity alignment and consistency verification processing. The numerical evolution characteristics, semantic deviation characteristics and logical contradiction characteristics of the same entity appearing in different documents of the full life cycle document set of the project application are identified. Based on the preset compliance rule base, entity relationship alignment and logical contradiction identification results containing conflict type identifiers and citation identifiers are generated. Based on the original text position mapping relationship corresponding to the inconsistencies in the entity relationship alignment and logical contradiction identification results, a traceable review evidence chain construction process is performed on the project application full-cycle document set to generate a traceable review evidence chain set containing the original text content of the inconsistencies, the coordinate positioning information of the inconsistencies in the documents, and the association relationship between the inconsistencies and compliance rules. This includes: constructing a traceable review evidence chain node unit for each conflict record entry. The traceable review evidence chain node unit includes a node identifier, a conflict type identifier field, a violation rule identifier field, an association relationship description field, an inconsistency original text content field, a coordinate positioning information field, and a multi-document entity occurrence position list field. The node identifier is generated based on the combination of the global entity object identifier and the conflict type identifier of the conflict record entry to ensure that each evidence chain node unit has a unique identifier. The process of constructing inter-node associations for the traceable review evidence chain node units is carried out. Based on the global entity object identifier in each traceable review evidence chain node unit, the traceable review evidence chain node units with different conflict types having the same global entity object identifier are associated to generate a set of evidence chain node associations based on the same global entity object. Based on the list of multiple document entity occurrence locations in each traceable review evidence chain node unit, extract the document stage information corresponding to each occurrence location, sort the evidence chain node units of the same global entity object according to the time order of the document stages, and generate a time-sequential evidence chain structure containing evidence chain sub-nodes of the project application stage, evidence chain sub-nodes of the project contract stage, evidence chain sub-nodes of the project mid-term inspection stage, and evidence chain sub-nodes of the project completion and acceptance stage. The original text of the inconsistency item corresponding to each sub-node of the time-sequential evidence chain structure is hyperlinked and mapped to the coordinate positioning information to generate a clickable positioning identifier corresponding to each sub-node of the evidence chain. The clickable positioning identifier is used to directly jump to the corresponding original text position in the full-cycle document set of the project application after clicking in the visualization interface. The evolution relationship between adjacent evidence chain sub-nodes in the time-series evidence chain structure is described. Based on the conflict type identifiers and feature description texts corresponding to the adjacent evidence chain sub-nodes, entity evolution description texts between adjacent stages are generated. The entity evolution description texts between adjacent stages are used to explain the specific manifestations of numerical changes, semantic shifts or logical contradictions of the same entity object in different project stages. All completed traceable review evidence chain node units, node relationship sets, temporal evidence chain structures, and entity evolution description texts between adjacent stages are aggregated and encapsulated to generate a traceable review evidence chain set. The traceable review evidence chain set contains multiple independent evidence chain entries. Each evidence chain entry corresponds to a global entity object with compliance issues. Each evidence chain entry contains the original text of inconsistencies of the entity object at each stage of the project lifecycle, coordinate location information, relationship with compliance rules, stage evolution descriptions, and clickable location identifiers.

2. The method for reviewing the compliance of project application content combining named entity recognition and large model as described in claim 1, characterized in that, The named entity recognition model, based on fine-tuned text corpus in the project application domain, performs multi-stage structured entity extraction processing on the full-cycle document set of project applications, generating a set of structured entity triples containing logical relationships, including: A pre-trained general named entity recognition basic model is obtained, and the pre-trained general named entity recognition basic model is fine-tuned using a historical project application full-cycle document sample set containing entity type labels and entity relationship labels to obtain a named entity recognition model that has been fine-tuned by the text corpus in the project application field. The project application lifecycle document set is divided into a project application document subset, a project contract document subset, a project mid-term inspection document subset, and a project completion and acceptance document subset according to document type. The document content of each document subset is segmented to obtain multiple document paragraph units corresponding to each document subset. The multiple document paragraph units corresponding to each document subset are input into the named entity recognition model. Each document paragraph unit is converted into a paragraph unit embedding vector representation through the embedding layer of the named entity recognition model. The paragraph unit embedding vector representation is then processed by the bidirectional long short-term memory network layer of the named entity recognition model to extract context features, thereby generating a paragraph unit feature vector sequence containing context semantic information. The named entity recognition model performs sequence labeling and decoding on the feature vector sequence of the paragraph unit through the conditional random field layer, and outputs the entity type label corresponding to each word segmentation unit in each document paragraph unit. The entity type label includes technical indicator entity label, deliverable entity label, time node entity label, budget item entity label, legal clause reference entity label and non-entity label. Based on the entity type tag corresponding to each word segmentation unit in each document paragraph unit, extract all consecutive word segmentation unit combinations with entity type tags from each document paragraph unit to form a set of original entity units to be associated. Each original entity unit to be associated in the set of original entity units to be associated contains entity text content, the starting position index of the entity in the document, and entity type tag. For each original entity unit to be associated in the set of original entity units to be associated, entity core word extraction processing is performed to remove the modifying words in the entity text content, retain the core keywords that can uniquely identify the meaning of the entity, and generate the entity core word representation corresponding to each original entity unit to be associated. Based on the entity core word representation and entity type label corresponding to each original entity unit to be associated, cross-document and cross-paragraph entity core reference resolution processing is performed on the set of original entity units to be associated. The semantic similarity of any two original entity units to be associated at the entity core word representation level and the consistency at the entity type label level are calculated. Original entity units to be associated with semantic similarity exceeding a preset similarity threshold and with the same entity type label are determined as core reference entity units pointing to the same entity object, and a unified global entity identifier is assigned to each group of core reference entity units. Obtain the co-occurrence distance, co-occurrence frequency, and syntactic dependency relationship between any two original entity units to be associated in each document paragraph unit. Construct a set of candidate logical association relationships between any two original entity units to be associated based on the co-occurrence distance, co-occurrence frequency, and syntactic dependency relationship. Each candidate logical association relationship in the set of candidate logical association relationships includes the entity identifier of the relationship initiator, the entity identifier of the relationship receiver, and the initial probability distribution of the relationship type. Each document paragraph unit is input into a pre-trained language model that has been fine-tuned by the relation extraction task. Relationship classification is performed on any two original entity units to be associated in each document paragraph unit. The target logical relationship type and relationship confidence score between any two original entity units to be associated are output. The target logical relationship type includes numerical constraint relationship type, temporal dependency relationship type, delivery correspondence relationship type, budget allocation relationship type and regulatory reference relationship type. Based on the target logical relationship type and relationship confidence score, the candidate logical relationships in the candidate logical relationship set are filtered, and the candidate logical relationships with relationship confidence scores exceeding the preset relationship confidence threshold are retained as confirmed logical relationships, and a relationship identifier is assigned to each confirmed logical relationship. Each entity object with a global entity identifier is used as the subject entity of a structured entity triple. Entities with a confirmed logical relationship with the subject entity are used as the object entities of the structured entity triple. The type of the confirmed logical relationship between the subject entity and the object entity is used as the predicate relation of the structured entity triple. These are combined to generate structured entity triples containing logical relationships. Each structured entity triple contains a global identifier for the subject entity, a predicate relation type identifier, a global identifier for the object entity, and a relation confidence score. Simultaneously, the subject entity and object entity in the structured entity triple each carry at least one entity type label from the following categories: technical indicator entity label, deliverable entity label, time node entity label, budget item entity label, and regulatory clause reference entity label.

3. The method for reviewing the compliance of project application content combining named entity recognition and large model as described in claim 1, characterized in that, The process involves constructing a set of compliance review prompts and inputting the structured entity triple set into a large language model according to a standardized input format. This drives the large language model to perform multi-document entity alignment and consistency verification processing. The process identifies the numerical evolution characteristics, semantic deviation characteristics, and logical contradiction characteristics of the same entity appearing in different documents across the entire project application lifecycle document set. Based on a preset compliance rule base, it generates entity relationship alignment and logical contradiction identification results, including conflict type identifiers and citation identifiers. Each structured entity triple in the set of structured entity triples is parsed, and the subject entity global identifier, predicate relation type identifier, and object entity global identifier contained in each structured entity triple are extracted. Based on the subject entity global identifier and the object entity global identifier, the entity text content, entity type label, and the starting position index of the entity in the document corresponding to each entity object are obtained from the set of original entity units to be associated, and a structured entity information mapping table containing complete entity information is constructed. Based on the entity type label corresponding to each entity object in the structured entity information mapping table, the entity text content of entity objects with the same global entity identifier in different documents is arranged in document time order to generate a multi-document entity text content sequence for each global entity object. The multi-document entity text content sequence includes entity text content in the project application document, entity text content in the project contract document, entity text content in the project mid-term inspection document, and entity text content in the project completion acceptance document. Numerical entity content detection processing is performed on the multi-document entity text content sequence of each global entity object to identify whether each entity text content in the multi-document entity text content sequence contains numerical information. If the entity text content contains numerical information, the numerical part and the numerical unit part are extracted from the entity text content to generate a numerical entity representation corresponding to each entity text content containing numerical information. The multi-document entity text content sequence of each global entity object and the numerical entity representation corresponding to each entity text content containing numerical information are combined and encoded according to a preset prompt engineering template format to generate entity alignment and consistency verification task prompt word unit corresponding to each global entity object. The entity alignment and consistency verification task prompt word unit includes a task instruction prefix, a global identifier of the entity to be verified, a list of entity text content of the entity to be verified in different documents, a list of numerical entity representations of the entity to be verified in different documents, and an output format requirement descriptor. The entity alignment and consistency verification task prompt word units corresponding to each global entity object are aggregated to generate a batch entity alignment and consistency verification task prompt word set containing all global entity objects to be verified. Global task context description information and multi-task collaborative processing instructions are added to the batch entity alignment and consistency verification task prompt word set to form a compliance review prompt engineering instruction set. The set of compliance review prompts is input as prompt text into a large language model. The input prompt text is converted into a prompt text embedding vector sequence through the input embedding layer of the large language model. The prompt text embedding vector sequence is then processed by a self-attention mechanism and a feedforward neural network through the Transformer decoder layer of the large language model to generate a prompt text decoding feature representation containing the semantics of instruction understanding. Based on the decoding feature representation of the prompt text, the large language model performs a multi-document entity alignment and consistency verification task. For each global entity object, the large language model first compares the entity text content of the global entity object in different documents, calculates the semantic similarity of the entity text content between different documents, and if the semantic similarity is lower than the preset semantic similarity threshold, it determines that there is a semantic deviation feature and generates a semantic deviation type identifier and semantic deviation description text. For each global entity object containing a numerical entity representation, the large language model compares the numerical entity representation of the global entity object in different documents, calculates the amount and trend of change of the numerical part between different documents, and if the amount of change exceeds the preset allowable range of numerical change or the trend of change does not conform to the project progress logic, it is determined that there is a numerical evolution feature, and a numerical evolution type identifier and a numerical evolution description text are generated. For each global entity object, the large language model analyzes whether the logical association between the global entity object and other entity objects remains consistent in different documents based on the structured entity triples formed by the global entity object. If the type of logical association between the global entity object and the same other entity object changes in different documents, or if the logical association between the global entity object and different entity objects in different documents conflicts, then it is determined that there is a logical contradiction feature, and a logical contradiction type identifier and logical contradiction description text are generated. The large language model matches the identified semantic deviation features, numerical evolution features, and logical contradiction features with the rule conditions in the compliance rule library based on a preset compliance rule library. The compliance rule library contains multiple compliance rules for project application content. Each compliance rule includes the applicable entity type, applicable feature type, rule judgment conditions, and a description of the consequences of rule violation. The large language model calculates the matching degree between each identified feature and each compliance rule. If the matching degree exceeds a preset rule matching degree threshold, the feature is determined to violate the corresponding compliance rule. For each feature that is determined to violate compliance rules, the large language model generates a conflict record unit containing a conflict type identifier, a violated rule identifier, the feature description text involved, and a global entity object identifier corresponding to the feature. Based on the global entity object identifier, the model obtains the starting position index of the entity text content corresponding to the feature in the document from the structured entity information mapping table, and generates a reference citation identifier containing the document name, paragraph number, and sentence offset. The large language model outputs all conflict record units and the corresponding citation identifiers for each conflict record unit in a structured manner according to the preset output format requirements, generating entity relationship alignment and logical contradiction identification results. The entity relationship alignment and logical contradiction identification results contain multiple conflict record entries. Each conflict record entry contains a conflict type identifier, a violated rule identifier, the feature description text involved, a global entity object identifier, and a citation identifier.

4. The method for reviewing the compliance of project application content combining named entity recognition and large model as described in claim 1, characterized in that, The process involves constructing a traceable review evidence chain for the entire project application document set based on the original text position mapping relationship corresponding to the inconsistencies in the entity relationship alignment and logical contradiction identification results. This generates a traceable review evidence chain set containing the original text content of the inconsistencies, the coordinate location information of the inconsistencies in the documents, and the association between the inconsistencies and compliance rules. The entity relationship alignment and logical contradiction identification results are analyzed, and the conflict type identifier, the violated rule identifier, the feature description text involved, the global entity object identifier, and the basis citation identifier contained in each conflict record entry are extracted. The basis citation identifier includes the document name field, the chapter number field, the paragraph number field, and the statement number field. Based on the document name field contained in the citation identifier in each conflict record entry, the corresponding target document is located from the project application full-cycle document set. Based on the chapter number field, paragraph number field and sentence number field contained in the citation identifier, the corresponding target sentence unit is located in the target document, and the complete text content of the target sentence unit is extracted as the original content of the inconsistency item. Based on the document name field, chapter number field, paragraph number field and statement number field included in the citation identifier, coordinate location information of the inconsistency item in the document is generated for each conflict record entry. The coordinate location information includes the document identifier, chapter path string, paragraph index value and statement sequence number in the paragraph. The coordinate location information is used to uniquely determine the precise location of the inconsistency item in the entire project application document set. Obtain the global entity object identifier in each conflict record entry, and query all entity text content corresponding to the global entity object and the document name and position index corresponding to each entity text content from the structured entity information mapping table based on the global entity object identifier. Generate a multi-document entity occurrence position list of the global entity object. The multi-document entity occurrence position list contains the position information of the global entity object in the project application document, project contract document, project mid-term inspection document and project completion acceptance document. Based on the conflict type identifier and the rule identifier violated in each conflict record entry, the corresponding detailed rule description text and rule violation consequence text are extracted from the preset compliance rule base. The detailed rule description text and rule violation consequence text are used as the content describing the association between the conflict record entry and the compliance rule.

5. The method for reviewing the compliance of project application content combining named entity recognition and large model as described in claim 2, characterized in that, The process of extracting entity core words for each original entity unit in the set of original entity units to be associated, removing modifiers from the entity text content, retaining core keywords that uniquely identify the entity's meaning, and generating an entity core word representation for each original entity unit to be associated includes: The entity text content is segmented and tagged with parts of speech to obtain each segmentation unit and its part of speech tag. The part of speech tags include noun tags, verb tags, adjective tags, numeral tags, quantifier tags, and auxiliary word tags. Based on the part-of-speech tags, word segmentation units with noun tags, verb tags, and numeral tags are selected from all word segmentation units as candidate core word units, while word segmentation units with other parts of speech are filtered out. Calculate the inverse document frequency weight value of each candidate core word unit in the entity text content, sort the candidates according to the inverse document frequency weight values ​​from high to low, and select the top preset number of candidate core word units with the highest inverse document frequency weight values ​​as core keyword units; The core keyword units are concatenated in their original order within the entity text content to generate the entity core word representation corresponding to each original entity unit to be associated. Stop word filtering and word form restoration are then performed on the entity core word representation to obtain the final version of the entity core word representation.

6. The method for reviewing the compliance of project application content combining named entity recognition and large model as described in claim 2, characterized in that, The process involves performing cross-document and cross-paragraph core reference resolution on the set of original entity units to be associated based on the entity core word representation and entity type label corresponding to each original entity unit to be associated. This includes calculating the semantic similarity of any two original entity units at the entity core word representation level and their consistency at the entity type label level. Original entity units to be associated with a semantic similarity exceeding a preset similarity threshold and identical entity type labels are identified as core-referenced entity units pointing to the same entity object. A unified global entity identifier is assigned to each group of core-referenced entity units, including: For any two original entity units to be associated, the corresponding entity core word representations are converted into core word embedding vectors, and the cosine similarity between the two core word embedding vectors is calculated as the core word semantic similarity score. Compare whether the entity type labels of the two original entity units to be associated are exactly the same, and generate an entity type consistency identifier; If the semantic similarity score of the core words exceeds the preset semantic similarity threshold and the entity type consistency is marked as consistent, then the two original entity units to be associated are marked as candidate core-referenced entity pairs, and the distance penalty factor is calculated based on the positional distance of the two entities in the document, and then the comprehensive core-reference confidence score is calculated. If the overall coreference confidence score exceeds the preset overall coreference confidence threshold, then the two original entity units to be associated are determined to be coreference entity units pointing to the same entity object, and a coreference relationship record is established for them. For each pair of original entity units to be associated in the set of original entity units to be associated, a global coreference graph containing all coreference records is constructed. Connectivity analysis is performed on the global coreference graph, and all original entity units to be associated within the same connected component are assigned a unified global entity identifier.

7. The method for reviewing the compliance of project application content combining named entity recognition and large model as described in claim 3, is characterized in that, For each global entity object containing a numerical entity representation, the large language model compares the numerical entity representations of the global entity object in different documents, calculates the amount and trend of change of the numerical part between different documents, and if the amount of change exceeds the preset allowable range of numerical change or the trend of change does not conform to the project progress logic, it is determined that there are numerical evolution characteristics, and a numerical evolution type identifier and numerical evolution description text are generated, including: Get the time-sorted sequence of numerical entity representations for each global entity object that contains numerical entity representations; Extract the numerical portion of two adjacent numerical entities in the sequence, calculate the difference between the latter and the former values ​​as the change in numerical value between adjacent documents, and determine whether the trend of numerical change is an increasing or decreasing trend based on the sign of the difference. Based on the entity type label of the global entity object, the corresponding preset value change allowable range is queried from the preset value change allowable range library. The value change between adjacent documents is divided by the previous value to obtain the change percentage. If the change percentage exceeds the preset value change allowable range, a value out of range flag is generated. Based on the project stage information of the global entity object, determine whether the trend of numerical change is consistent with the project progress logic. If not, generate a trend anomaly flag. If at least one of the following exists: a value out of range flag or a trend anomaly flag, then a value evolution feature is determined to exist, and a corresponding value evolution type identifier and value evolution description text are generated for the global entity object.

8. The method for reviewing the compliance of project application content combining named entity recognition and large model as described in claim 1, characterized in that, The process involves extracting document stage information corresponding to each occurrence position from the list of multiple document entity occurrence positions in each traceable evidence chain node unit, sorting the evidence chain node units of the same global entity object according to the chronological order of the document stages, and generating a temporal evidence chain structure containing evidence chain sub-nodes for the project application stage, the project contract stage, the project mid-term inspection stage, and the project completion and acceptance stage. Parse the global entity object identifier corresponding to each conflict record entry, obtain the list of multiple document entity occurrence positions of the global entity object, and each occurrence position record includes a document name field and the starting position index in the document; Match the document name field with the preset document stage mapping table, assign a corresponding document stage label to each occurrence location record, and obtain a set of occurrence location records carrying document stage labels; Make multiple copies of the traceable review evidence chain node unit corresponding to the conflict record entry, so that the number of copies is the same as the number of occurrence location records, and associate each copy of the evidence chain node unit with an occurrence location record, fill in the coordinate positioning information, and generate a staged evidence chain sub-node corresponding to each occurrence location. Based on the document stage tags, the staged evidence chain sub-nodes are classified into the project application stage, project contract stage, project mid-term inspection stage, and project completion acceptance stage sub-node sets, and the sub-node with the smallest starting position index is selected from each stage sub-node set as the evidence chain sub-node of the corresponding stage. The selected evidence chain sub-nodes from the project application stage, project contract stage, project mid-term inspection stage, and project completion and acceptance stage are arranged in chronological order according to the project stage to generate a chronological evidence chain structure.

9. The method for reviewing the compliance of project application content combining named entity recognition and large model as described in claim 1, characterized in that, After obtaining the full-cycle document set of project applications corresponding to the projects to be reviewed, before performing multi-stage structured entity extraction processing on the full-cycle document set of project applications based on the named entity recognition model that has been fine-tuned by the text corpus in the project application field, the process further includes: Each document in the project application lifecycle document set is uniformly converted to obtain a plain text document with a unified format. For each plain text document, perform document structure parsing, identify chapter titles, paragraph boundaries and table areas, and generate a document structure tree containing chapter nodes, paragraph nodes and table nodes; Extract the table content from the table area corresponding to the table node, convert the table data into key-value pairs and insert them into the paragraph where the table node is located, generating enhanced document content containing the converted text of the table content; The enhanced document content is cleaned to obtain a cleaned set of project application lifecycle documents; Each cleaned document is assigned a globally unique document identifier and stored in the corresponding document library according to its document type, thus creating a document index table.

10. A project application content compliance review system combining named entity recognition and large model, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the project application content compliance review method combining named entity recognition and large model as described in any one of claims 1 to 9 by executing the machine-executable instructions.

Images