A knowledge graph construction method for knowledge systematization of research institutes

Abstract

The application discloses a knowledge graph construction method for knowledge systematization of research institutes, and belongs to the technical field of knowledge graph construction. Firstly, a business process ontology skeleton and a synonymous mapping dictionary are constructed, and a semantic alignment sequence is obtained. On this basis, a document is divided into process fragments, and a context enhancement vector matrix with fusion process position perception is generated based on a process distance attenuation factor and a type compatibility coefficient. In the model training stage, a prior dictionary embedding is introduced to enhance the representation ability of structured naming, a boundary sensitive span identification method is adopted to accurately extract long entities containing version numbers by fusing left and right adjacent coding and boundary offset penalty, and a process gating vector is generated according to the process fragment distance and stage combination involved in the entity to dynamically modulate the relationship features, so that the cross-process misconnected edges are inhibited in the joint training framework. Finally, the accurate extraction and graph storage of business entities and their core relationships in the massive unstructured documents of research institutes are realized.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph construction technology, and in particular relates to a method for constructing a knowledge graph for the systematization of knowledge in research institutes. Background Technology

[0002] Over its long-term operation, the institute has accumulated a large number of business documents. These documents are diverse in type and format, including project proposals, weekly reports, test records, acceptance reports, equipment operation records, meeting minutes, and personnel resumes. These texts carry the institute's core knowledge assets, covering key business information such as "who was in which project, what equipment was used, and what deliverables were produced." However, because these documents are mostly stored in unstructured or semi-structured natural language formats across different business systems, a serious "knowledge silo" phenomenon has occurred, making it difficult for research management personnel to quickly trace project progress, equipment reuse, and personnel contributions from a holistic perspective.

[0003] However, existing conventional technologies have the following drawbacks:

[0004] Shortcoming 1: Existing technologies lack the ability to structurally perceive and model the temporal logic of business processes in vertical domains.

[0005] Existing information extraction methods based on pre-trained language models mostly treat text as a flat linear sequence or a fully connected graph structure. While their attention mechanisms can capture long-distance dependencies, they cannot distinguish whether such dependencies are "causal progression in business processes" or "meaningless cross-chapter co-occurrence." In research institute scenarios, documents are naturally written in an implicit order of "requirements-design-implementation-acceptance." Existing technologies cannot inject this stage prior into the lexical representation, resulting in a lack of guidance for the model in understanding "which relationships to focus on at which stage," easily leading to a large number of interfering edges that violate business logic during later relationship extraction.

[0006] Second shortcoming: Existing entity recognition methods do not have high accuracy in recognizing long, business-sensitive entities containing version numbers and codes, and there is a serious boundary drift problem.

[0007] General named entity recognition or span classification methods typically pool only the semantic information within an entity when constructing entity representations, neglecting information outside the entity's immediate boundaries. This might be feasible in general news corpora, but in research texts, entities are often closely linked to quantifiers, punctuation, or common nouns (e.g., "using a certain simulation platform V2.3"). Existing methods, lacking explicit encoding of the features of adjacent words, are prone to misidentifying subsequent common nouns (e.g., "using" or "system") as part of the entity, leading to erroneous entity nodes with redundant suffixes in the knowledge graph and disrupting the standardization and consistency of equipment models or deliverable names.

[0008] Shortcoming 3: Existing relation extraction methods rely on semantic co-occurrence and do not perform effective hard constraint pruning on cross-process entity pairs.

[0009] Existing relation extraction models typically evaluate any entity pair co-occurring within the same paragraph or sliding window as a candidate relation, primarily based on the semantic vector inner product or concatenation classification between entities. However, in a research director's document, a device might be mentioned in the "Requirements Analysis" section, while a person appears in the "Acceptance and Delivery" section; their semantic vectors may have high similarity due to belonging to the same project. Existing methods cannot effectively prune or suppress features of such candidate pairs using information about the document's process stages. This results in the extracted knowledge graph containing a large number of false business relationships that are "used beyond reach" or "not produced," severely reducing the graph's usability in project traceability and auditing scenarios. Summary of the Invention

[0010] To address the above technical problems, this invention provides a method for constructing a knowledge graph for the systematization of research institute knowledge, specifically including the following steps:

[0011] S1. Constructing Research Institute Business Text Samples and Establishing Business Process Ontology Skeleton: Collect research institute business texts to form a raw text set, preprocess the texts and retain the integrity of business-sensitive segments to obtain the raw text sequence; establish an ontology concept and relation constraint table with "project-personnel-equipment-deliverables" as the core; establish an entity synonym mapping dictionary, process stage trigger word list and abbreviation ambiguity adjudication rules;

[0012] S2. Text semantic alignment and process fragment context enhancement for business processes: Standard term replacement is performed on the original text sequence and ontology type tags are inserted to obtain a semantically aligned text sequence; the semantically aligned text sequence is divided into process fragments, and a context enhancement vector matrix is ​​generated according to the process fragment to which the word belongs and the word type;

[0013] S3. Training a model for entity recognition and process constraint extraction that integrates prior dictionary embedding and boundary-sensitive span recognition: Prior vectors are generated for lexical units based on a prior dictionary and fused with context-enhanced vectors. After sequence encoding, an entity recognition encoding matrix is ​​obtained. Entities are identified using candidate span enumeration. A span representation vector containing left and right adjacent position codes is constructed for each candidate span to classify entities. During training, a boundary offset penalty is applied to partially overlapping, hard-negative samples. Candidate entity pairs are generated based on the relationship constraint table. Process gating vectors are generated based on the distance between process segments involved in the entity pair and the stage combination. The basic representation of the entity pair is modulated, and the relationship classification result is output. Entity recognition and relationship extraction tasks are jointly trained, and the final model parameters are output.

[0014] S4. Online extraction of research business text and generation of knowledge results: Perform semantic alignment, type labeling and context enhancement processing on new documents in the same way as in the training phase, feed them into the trained model to obtain entity set and relation set, perform standard terminology primary key merging on entities and write them into the storage system to form a business knowledge network.

[0015] The present invention has the following beneficial effects:

[0016] This invention proposes an ontology skeleton and context enhancement mechanism with "process stage constraints" as the core.

[0017] This invention goes beyond general entity relationship definitions. Instead, it addresses the strong temporal nature of research institute operations by constructing a business process skeleton encompassing four ontologies: "projects, personnel, equipment, and deliverables." Building upon this, it innovatively proposes a context-enhancing method based on process segmentation. By calculating the "process distance decay factor" and "type compatibility coefficient" between lexical units and each process segment, it generates lexical representation vectors that incorporate process location awareness. This enables the model to determine, during the encoding phase, "which business stage the current text is in" and "which entity type should be considered at that stage," thus resolving the problem of general models lacking awareness of the temporal logic of business processes.

[0018] This invention proposes a boundary-sensitive span entity recognition method, specifically optimized for research director entities and structured naming.

[0019] To address the issues of equipment model names containing version numbers and lengthy project names that are easily conjured up with modifiers in research texts, this invention designs a unique span representation vector construction method. This representation vector not only includes conventional intra-span pooling features but also forcibly introduces left and right adjacent encodings for the entity's start and end positions. By enabling the model to perceive lexical features outside the entity boundary and combining this with a specially designed boundary offset penalty term, the tendency of the model to swallow preceding and following modifiers or punctuation marks into the entity is effectively suppressed. Compared to traditional CRF sequence labeling or classification methods that rely solely on intra-span semantics, this invention significantly improves boundary localization accuracy when identifying complex entities with version numbers, such as "a certain simulation platform V2.3".

[0020] This invention proposes a relation extraction method based on process gating vectors to suppress erroneous connections across stages.

[0021] This invention overcomes the limitation of existing relation extraction models that rely solely on the semantic similarity of entity pairs. It innovatively transforms the distance and stage combination of the process segments in which entity pairs reside into dynamic process gating vectors. Before relation classification, these gating vectors modulate the relation features of entity pairs element-by-element. For example, even if "personnel" and "equipment" are highly semantically related, if they belong to two distant process segments, "requirements analysis" and "acceptance and delivery," the gating vector will actively suppress the feature response of the "use" relation. This design, which transforms business process logic into differentiable computational signals, significantly reduces cross-process noise relations caused by text co-occurrence, ensuring the rigor of the final knowledge graph's business logic. Attached Figure Description

[0022] Figure 1 The method flowchart for the invention;

[0023] Figure 2 A comparative analysis chart of relation extraction performance at different process stages;

[0024] Figure 3 The ablation analysis diagram shows the contribution of the core components within the model. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0026] This invention proposes a knowledge graph construction method for the systematization of research institute knowledge, such as... Figure 1 As shown, it includes the following steps:

[0027] S1. Constructing research institute business text samples and establishing the business process ontology skeleton;

[0028] The research institute's business documents include various document types such as project proposals, weekly reports, test records, acceptance reports, equipment operation records, meeting minutes, and personnel resumes. Different documents show significant differences in the writing style for the same research object, and business relationships are often implicit in the chapter order and process stages. Without first establishing a unified business process ontology framework and training samples, subsequent models are prone to identifying the same object as multiple entities, or misjudging entities that only co-occur as business relationships.

[0029] This step first constructs the original business text set and manually annotated results, then establishes four categories of ontology concepts and relational constraints: "project-personnel-equipment-deliverables," providing a unified input standard for subsequent semantic alignment, entity recognition, and relation extraction. The specific steps are as follows:

[0030] S101, Construction of Research Institute Business Text Samples;

[0031] The research institute's business documents suffer from diverse formats and implicit business relationships that are difficult to obtain directly. Key information such as project codes, equipment models, and personnel ranks are often mixed with ordinary text in the original documents, and business relationships (such as responsibility and participation) are implicit in chapter order or paragraph descriptions. Directly inputting these into the model would lead to blurred entity boundaries and misjudged relationships. Therefore, a technical approach combining text standardization preprocessing and structured manual annotation is adopted: First, all types of documents are converted into a unified plain text stream while preserving the integrity of business-sensitive segments. Then, four types of core entities and four types of core business relationships are manually annotated, forming a training sample set with accompanying process location clues. This provides consistent, clearly defined, and relationship-standardized foundational data for subsequent semantic alignment and model training. The specific steps are as follows:

[0032] 1) Collect research institute business documents to form a set of original business documents. , Represents the set of original documents to be processed, containing This document Indicates the number of documents.

[0033] Original business text collection Each document retains its title, chapter number, paragraph breaks, table cell content, and time information, but does not delete project number, equipment model, version number, date, job title suffix, or internal code.

[0034] As an example, in a "Weekly Report of XX Pre-research Project," the original text includes fields such as "Project Number: JG2025-07," "Test Equipment: A Certain Type of Simulation Platform V2.3," "Reporter: Senior Engineer Zhang," and "Document Version: V1.2." According to this procedure, the project number, equipment model, version number, job title suffix, and internal code in the above-mentioned fields must be completely retained without deletion or standardized rewriting. This is to prevent situations where "JG2025-07" is mistakenly split into "JG2025" and "-07," or "A Certain Type of Simulation Platform V2.3" is mistakenly split into "A Certain Type of Simulation Platform" and "V2.3," thus disrupting the integrity of the business process.

[0035] 2) Collect the original business text set Each document in the process is converted to plain text and then segmented according to periods, semicolons, colons, line breaks, numbered item boundaries, and table cell boundaries to obtain the original text sequence of the current document. . This represents the word sequence of the current document after basic preprocessing, with a length denoted as . , This indicates the number of terms in the current document.

[0036] During basic preprocessing, sensitive business segments such as "project-year-serial number", "equipment model-version number", and "name + title" are not separated. For example, "JG2025-07", "a certain type of simulation platform V2.3", and "Zhang Moumou Senior Engineer" are all retained as continuous segments.

[0037] As an example, suppose a paragraph in a document reads: "Project JG2025-07 has completed its solution design; test cases are detailed in Attachment A: Test Report V1.2." The resulting original text sequence after segmentation... It may contain the following consecutive segments (with " / " indicating the boundary):

[0038] [JG2025-07] / [Project solution design completed] / [Test cases detailed in Attachment A] / [Test Report V1.2] / [。 ]

[0039] Among them, “JG2025-07” and “Test Report V1.2” are business-sensitive segments that remain intact and are not broken up by periods or line breaks; the content within the table cells is independently segmented according to the cell boundaries to ensure that the project code or equipment model within the cell is not truncated.

[0040] 3) For the original business text set Manual annotation is required, and the annotation content must include at least four types of entities and four types of business relationships. The four types of entities are "Project," "Personnel," "Equipment," and "Deliverables"; the four types of business relationships are "Responsible," "Participating," "Using," and "Output." Among them, "Responsible" indicates that personnel bear primary responsibility for the project; "Participating" indicates that personnel have a collaborative relationship with the project; "Using" indicates that personnel use the equipment during implementation, testing, or verification; and "Output" indicates the deliverables formed by the project.

[0041] 4) Label each entity with its start and end positions, entity category, and the document it belongs to; label each business relationship with its head entity, tail entity, relationship category, and the range of the statement or adjacent process segment in which the relationship is located.

[0042] For example, regarding the text fragment: "Senior Engineer Zhang was responsible for writing the overall technical solution for the JG2025-07 project," during manual annotation:

[0043] a) Entity labeling: The personnel entity “Zhang Moumou Senior Engineer” is labeled with starting position 0 and ending position 5 (in characters); the project entity “JG2025-07” is labeled with starting position 8 and ending position 16; the deliverable entity “Overall Technical Solution” is labeled with starting position 21 and ending position 26.

[0044] b) Relationship labeling: Head entity "Zhang Moumou Senior Engineer", tail entity "JG2025-07", relationship category "responsible", the scope of the statement in which the relationship is located is this sentence; at the same time, record the chapter title to which the statement belongs as "personnel division of labor", so as to provide clues about the process location.

[0045] It should be noted that, in order to ensure the feasibility of extracting constraints in the subsequent process, manual annotation should prioritize recording the first occurrence of an entity and the process segment in which the entity is located. If the document contains chapter titles, the chapter names should also be recorded.

[0046] 5) Bind the text content, entity annotations, and relation annotations of each document to obtain the research institute's business text training samples. The research institute's business text training samples not only contain text, but also entity boundaries, entity categories, relation categories, and process location clues. All subsequent training processes use the research institute's business text training samples as input.

[0047] S102. Establishment of the business process entity skeleton;

[0048] To address the pain points of diverse descriptions of the same object in research institute business documents (coexistence of full name / abbreviation / code) and strict constraints on relationship types due to business stages, a business process ontology framework centered on "project-personnel-equipment-deliverables" is constructed. The specific steps are as follows:

[0049] 1) Establish the core framework of the business process;

[0050] The business process ontology framework is based on four core ontology concepts: "project, personnel, equipment, and deliverables." "Project" refers to a research topic, special topic, task, or pre-research project; "personnel" refers to the person in charge, chief engineer, engineer, tester, or other participants; "equipment" refers to a platform, system, device, terminal, module, or instrument; and "deliverables" refers to technical solutions, interface documents, simulation models, test reports, acceptance materials, or operation manuals.

[0051] As examples, the four types of ontology concepts are shown below:

[0052] Projects: For example, "JG2025-07" (project number), "New Radar Pre-research" (topic abbreviation), "XX Topic-03 Sub-project" (task decomposition number).

[0053] Personnel: For example, "Zhang Moumou Senior Engineer" (name + title), "Li Moumou Chief Engineer" (name + role), "Wang Moumou Tester".

[0054] Equipment: For example, “a certain type of simulation platform V2.3” (platform + version number), “high frequency signal source AV1442” (instrument model), “data acquisition terminal DT-100” (module model).

[0055] Deliverables: For example, “Overall Technical Solution V2.3” (document + version number), “Test Report (2025-03)” (document + date), “Interface Design Specification IDD-1.2” (document + number).

[0056] 2) Create a relation constraint table;

[0057] The relationship constraint table is used to limit which combinations of entity types are allowed to generate business relationships. In a simplified implementation approach, only "personnel-project" is retained, corresponding to "responsible" or "participating," "personnel-equipment" to "using," and "project-deliverable" to "output." When an entity type combination is not in the relationship constraint table, it will not be considered a high-priority candidate entity pair in subsequent relationship extraction stages to reduce invalid combinations.

[0058] As an example, the relational constraint table can be designed as follows:

[0059]

[0060] When a combination of "project" and "equipment" appears (such as "JG2025-07 uses a certain type of simulation platform"), it will not be considered a high-priority candidate in the subsequent relationship extraction stage because it is not in the constraint table. This avoids the model misjudging non-standard expressions such as "project uses equipment" as valid business relationships.

[0061] 3) Establish an entity synonym mapping dictionary;

[0062] The entity thesaurus is used to store standard terms and their variant spellings. Each entry includes at least the standard term, entity category, abbreviation, internal code, historical name, English abbreviation, numbering pattern, and optional contextual qualifiers. For example, the same project can simultaneously register the full project name, project initiation number, and internal code; the same equipment can simultaneously register the full model name, equipment abbreviation, and equipment serial number prefix; and the same deliverable can simultaneously register the document template name and common version spellings.

[0063] As an example, here is a sample entry for an entity synonym mapping dictionary:

[0064]

[0065] For the abbreviation "project report", the dictionary can add contextual qualifiers such as "if 'project' or 'topic' appears before or after, it should refer to the deliverable first" to assist in subsequent decisions.

[0066] 4) Establish a list of trigger terms for each stage of the process;

[0067] The process phase trigger term list is used to determine which business process phase a text fragment is closer to. In one implementation, five process phases can be set: "Requirements Analysis," "Solution Design," "Implementation and Integration," "Acceptance and Delivery," and "Other." Specifically, when words such as "requirements," "scenarios," "constraints," and "goals" appear, it is prioritized for "Requirements Analysis"; when words such as "solution," "architecture," "interface," "selection," and "design" appear, it is prioritized for "Solution Design"; when words such as "deployment," "setup," "integration," "experimentation," "verification," and "testing" appear, it is prioritized for "Implementation and Integration"; and when words such as "acceptance," "delivery," "review," "summary," and "archiving" appear, it is prioritized for "Acceptance and Delivery."

[0068] For example, a paragraph might begin with "According to the requirements specification, this phase requires the completion of interface selection and preliminary design...", which includes the trigger words "requirements" (categorized under requirements analysis) and "selection / design" (categorized under solution design). Since the trigger words "solution design" appear more frequently, this phase of the process will be tagged as "solution design." Similarly, if a text mentions "deploying the test environment and conducting integration testing," the trigger words "deployment / integration testing" primarily point to the "implementation integration testing" phase, and it will be directly categorized under that type.

[0069] 5) Establish rules for resolving ambiguities in abbreviations;

[0070] When the same abbreviation may correspond to different entity categories, the decision is made based on the contextual trigger words. For example, if an abbreviation is preceded or followed by "project establishment," "topic," or "special topic," it is prioritized as a project; if preceded or followed by "platform," "module," "device," or "system," it is prioritized as equipment; and if preceded or followed by "report," "plan," "document," or "record," it is prioritized as a deliverable. The decision window can take 5 to 8 tokens before and after the current segment.

[0071] As an example, suppose the text fragment is: "...the functional verification of the XX platform has been completed, and the platform has been deployed in the laboratory...". For the abbreviation "XX platform", if words such as "deployment" and "experiment" appear in the five terminology windows before and after it, it is prioritized as equipment. If another fragment is: "...the funding for the XX platform project has been allocated, and the project leader is...", and the window contains "project", "funding", and "project", then the entity synonym mapping dictionary is used to check whether "XX platform" is registered as a standard term or alias for the project. If so, it is prioritized as a "project".

[0072] It should be noted that the research institute's business texts are characterized by closed terminology, numerous abbreviations, multiple internal numbering systems, and strong constraints on business processes. A general semantic model alone is insufficient to reliably understand "different ways of writing the same object" and "entity combinations that are allowed to establish relationships." This step, by clarifying the extension of ontology concepts, limiting the legal combinations of relationships, unifying multi-source heterogeneous naming, and injecting prior knowledge from the process stages, provides unified terminology, unified type boundaries, and unified relationship constraints for subsequent steps, reducing the learning difficulty of subsequent models.

[0073] S2, Business process-oriented text semantic alignment and process fragment context enhancement;

[0074] In research institute business documents, the same object often appears in multiple ways, such as full name, abbreviation, code name, and version name, and entity relationships often span across single sentences and depend on the context of the process. If the original text is directly fed into the model, problems such as entity boundary drift, synonym entity splitting, and lack of process support for cross-sentence relationships can easily occur.

[0075] This step first performs semantic alignment and type tagging, then divides the document into process segments, and generates a context-enhanced vector matrix based on "process distance + type compatibility". The specific steps are as follows:

[0076] S201, Semantic Alignment and Type Markup;

[0077] In research institute business documents, the same project, equipment, or deliverable often appears in different forms. If they are not expressed in a unified way first, subsequent entity recognition will misidentify the same object in the business as multiple entities.

[0078] This sub-step performs standard term replacement on the original text sequence and inserts ontology type tags to obtain a semantically aligned text sequence. The specific steps are as follows:

[0079] 1) Read the original text sequence of the current document. The scanning process proceeds from left to right along the word order. During scanning, it prioritizes matching longer consecutive segments, then attempts numbered pattern matching, and finally fuzzy matching with context constraints. Matching longer consecutive segments is used to prioritize the identification of full project names, complete device names, and complete deliverable names, avoiding situations where short segments are matched first and then the identification of longer segments is compromised.

[0080] As an example, for the fragment "test record of a certain type of simulation platform V2.3" in the original text sequence, the scan first tries the fragment "a certain type of simulation platform V2.3" with a length of 5 words. If the complete device name exists in the entity synonym mapping dictionary, it is directly matched and identified as a device entity. If the fragment "a certain type" with a length of 2 words is tried first, it may mistakenly trigger the matching of ordinary nouns, which will destroy the complete recognition of long entities. Therefore, the longest matching priority strategy is adopted.

[0081] 2) Use an entity thesaurus to map the original text sequence The system performs standard term replacement on entity fragments. In its implementation, the length of consecutive fragments can be set from 2 to 12 lexical units. For each starting position, fragments of different lengths are tried sequentially and matched with standard terms and aliases in the entity synonym mapping dictionary. If multiple candidates are matched, candidates with longer coverage, more consistent contextual constraints, and more specific entity categories are given priority.

[0082] Furthermore, for fragments that do not directly match the entity synonym mapping dictionary, rule matching is introduced to supplement the identification. Rule matching includes: project number rules, equipment model rules, deliverable template rules, and personnel expression rules. Project number rules can match the format "letter + year + serial number" or "topic-sub-item-version"; equipment model rules can match the format "type + platform / module / terminal / system"; deliverable template rules can match the format "technical solution / test report / acceptance report / operating procedure" followed by a version number; personnel expression rules can match the format "name + title" or "name + role". Once a rule match is found, the fragment is replaced with standard terminology or a standardized name generated by the rule.

[0083] In one embodiment, dictionary matching is exemplified as follows: For the text "Project JG2025-07 adopts the HIL platform", if the dictionary has registered "HIL platform" as an alias for the standard term "Hardware-on-a-Loop Simulation Platform V2.3", then when "HIL platform" is scanned, it will be replaced with "Hardware-on-a-Loop Simulation Platform V2.3", while retaining the original fragment location information. Rule matching is exemplified as follows: For the unregistered fragment "a certain type of integrated verification platform", if the equipment model rule "a certain type + platform / system" is matched, then the standardized name "a certain type of integrated verification platform" will be generated and marked as equipment; for "Test Report V1.2", the deliverable template rule will be matched, standardized as "Test Report V1.2", and marked as a deliverable.

[0084] 3) Perform context-based adjudication for ambiguous abbreviations. In practice, the current abbreviation is the center, and a context window consisting of 5 to 8 words before and after it is read. Then, the entity category is determined according to the abbreviation ambiguity adjudication rules. For example, if "XX platform" is preceded by "deployment, joint debugging, terminal access", it is preferentially classified as equipment; if "XX platform" is preceded by "project, task book, funding", the entity synonym mapping dictionary can be further checked. If there is a project with the same name in the dictionary, it is preferentially classified as a project.

[0085] As an example, the text is: "...48 hours of stability testing was conducted on the XX platform in the joint debugging environment...". Taking "XX platform" as the center, five units before and after it are taken to form a window: "...48 hours of stability testing was conducted on the XX platform in the joint debugging environment...". If words such as "joint debugging" and "testing" appear in the window, it is preferentially judged as a device according to the adjudication rules. If the dictionary also contains an entity named "XX platform", it is still judged as a device category because the context strongly points to the device semantics.

[0086] 4) After the standard terminology replacement is completed, type tags are inserted into the identified entities to obtain text fragments with boundary hints. The four type tags are [PROJ], [PER], [EQUIP], and [DELIV]. [PROJ] represents the project type tag, [PER] represents the personnel type tag, [EQUIP] represents the equipment type tag, and [DELIV] represents the deliverable type tag. The insertion method involves inserting the corresponding type tag once before and once after the entity fragment to clearly define the entity's start and end boundaries.

[0087] For example, the original text is: "Senior Engineer Zhang Moumou uses a certain type of simulation platform." After inserting the type tag, it becomes:

[0088] [PER] Senior Engineer Zhang Moumou used [EQUIP] a certain type of simulation platform;

[0089] Among them, [PER] and [EQUIP] serve as boundary cues, allowing subsequent models to directly perceive the start and end positions of entities without relying solely on semantic inference.

[0090] 5) Conflict resolution is performed on overlapping or nested fragments. In the specific implementation, if a long fragment has already been identified by the entity synonym mapping dictionary or rule matching, the long fragment is retained first; if a fragment satisfies multiple entity categories, the decision is made in the order of "projects take precedence over deliverables, equipment takes precedence over common nouns, and personnel take precedence over role descriptions"; if long fragments and short fragments belong to different categories and nesting is allowed in the business, the long fragment is retained for semantic alignment, and the short fragment is reserved for fine-grained judgment in the subsequent cross-entity recognition stage.

[0091] As an example, the fragment "A certain type of simulation platform V2.3" is identified as a long entity called "equipment"; its internal sub-fragment "simulation platform" may also satisfy the common noun matching. In this case, the long fragment "A certain type of simulation platform V2.3" is retained first, and the independent entity candidate of the short fragment "simulation platform" is discarded. If a nested situation such as "project technical solution" occurs ("project" is the project entity, and "technical solution" is the deliverable entity), then the long fragment "project technical solution" is retained for semantic alignment, and the short fragment is reserved for subsequent span recognition stage to determine whether it should be output as an independent deliverable entity.

[0092] 6) Output semantically aligned text sequence , This represents a text sequence after standard terminology replacement and type tag insertion, which serves as direct input for subsequent process segmentation and context enhancement.

[0093] As an example, the original text fragment "A certain pre-research project was led by Senior Engineer Zhang, and the verification of technical solution V2.3 was completed using a certain type of simulation platform" can be transformed into "[PROJ] A certain pre-research project [PROJ] was led by Senior Engineer Zhang [PER], and the verification of technical solution V2.3 [DELIV] was completed using a certain type of simulation platform [EQUIP]". Here, "a certain pre-research project" is standardized as standard project terminology, "Senior Engineer Zhang" is identified as a person according to the "name + title" rule, "a certain type of simulation platform" is identified as equipment according to the equipment model rule, and "technical solution V2.3" is identified as a deliverable according to the deliverable template rule.

[0094] It should be noted that this step is not simply a dictionary replacement, but rather injects the skeleton of the research institute's business process ontology into the input text in advance. This enables the subsequent model to have the ability to unify terminology, clarify boundaries, and perceive types before entering training, thereby significantly reducing the entity splitting problem caused by synonyms and improving the stability of subsequent entity boundary localization.

[0095] S202, Process Segmentation and Context Enhancement;

[0096] Research institute business documents are often written in the order of "requirements-design-implementation-acceptance," and even without headings, the process progression is implied in the paragraph order. The relationships between entities depend not only on local neighboring terms, but also on the context of the current process stage and adjacent process stages.

[0097] This sub-step triggers a vocabulary based on the document structure and process stages, divides the semantically aligned text sequence into process segments, and then generates process-sensitive and type-sensitive context-enhanced vectors for each term. The specific steps are as follows:

[0098] 1) Semantically aligned text sequences The process is divided into segments, resulting in a set of process segments. The total number of process segments is denoted as . , No. Each process segment is denoted as , Indicates the first Each process segment contains a set of terms. When dividing process segments, priority is given to using chapter titles, number levels, paragraph boundaries, and table area boundaries; if the current document lacks titles, it is divided by paragraph, and a process stage label is assigned to each paragraph in conjunction with the process stage trigger terminology.

[0099] Furthermore, process stage labels can be selected from five categories: "Requirements Analysis," "Solution Design," "Implementation and Integration," "Acceptance and Delivery," and "Other." When a paragraph hits multiple trigger words simultaneously, the process stage label with the higher hit frequency is prioritized; if the hit frequencies are similar, the labels of the two paragraphs before and after the paragraph are combined for a smooth decision to avoid irregular changes in labels between adjacent paragraphs.

[0100] In one embodiment, the division is exemplified as follows: If a document contains chapters "1. Requirements Analysis", "2. Solution Design", and "3. Testing and Verification", then the process segments are directly divided according to the chapter boundaries, and the tags "Requirements Analysis", "Solution Design", and "Implementation and Integration" are assigned sequentially (because the trigger word "Testing and Verification" points to Implementation and Integration). A smooth decision is exemplified as follows: If a paragraph contains both "Requirements" (2 times) and "Design" (3 times), with similar frequency, and if the preceding paragraph is tagged "Requirements Analysis" and the following paragraph is tagged "Solution Design", then the current paragraph is smoothly tagged as "Solution Design" to avoid tag jumps.

[0101] 2) Semantically align text sequences Each lexical element in the record belongs to the process segment number, denoted as . .in, Indicates the semantic alignment of lexical units in a text sequence. The position index in the middle, Indicates the first The process segment number to which each word belongs. This number is used for subsequent calculation of process distance.

[0102] As an example, suppose a document is divided into three process segments: segment 1 contains words 1-50, segment 2 contains words 51-120, and segment 3 contains words 121-200. Then:

[0103] For lexical index ,That =1;

[0104] For lexical index ,That =2;

[0105] For lexical index ,That =3;

[0106] This number will be used for distance calculations in subsequent processes.

[0107] 3) Semantically align text sequences Each lexical in the vector is mapped to an initial lexical vector. . Indicates the first The initial vector representation of each word, denoted by dimension . , The feature dimension representing a single word can be 128 or 256. Initial word vector. It can be generated from a frozen pre-trained Chinese word vector model, character vector model, or character-level encoder.

[0108] At the same time, for the first Each lexical assignment type gain vector . Indicates the first The type enhancement coefficient vector of each lexical unit, with dimension and The same applies. For tokens within the [PROJ] boundary, the first quarter of the vector can be assigned a higher coefficient, for example, 1.3, and the other dimensions can be 1.0; for tokens within the [PER] boundary, the second quarter can be assigned a higher coefficient; for tokens within the [EQUIP] boundary, the third quarter can be assigned a higher coefficient; for tokens within the [DELIV] boundary, the last quarter can be assigned a higher coefficient; ordinary tokens use a vector of all 1s. Then... and Element-wise multiplication yields the type-sensitive meta-representation.

[0109] As an example, let the word vector dimension be... For the term “JG2025-07” located within the [PROJ] boundary, its type gain vector forward Maintenance settings The remaining 96 dimensions are Initial word vectors and After element-wise multiplication, the type-sensitive representation of this term is obtained on the "item-related feature dimension". The increased amplitude makes it easier for the model to distinguish entity types.

[0110] 4) Generate a process segment representation for each process segment. . Indicates the first Each process segment is represented by a representative vector, used to summarize the overall semantics and main ontology components of that process segment. In specific implementations, the process segments... All sensitive word types within the entity are represented by a weighted average. Ordinary word types have a weight of 1, while word types located within entity boundaries have a weight of 1.3 to 1.5, thereby improving the representation of process segments by entity words. . contributions.

[0111] As an example, for the first A process segment Assuming the dataset contains 20 tokens, with 5 tokens within the [EQUIP] boundary, 3 tokens within the [PER] boundary, and the remainder being ordinary tokens. In the weighted average, the weights of tokens within the device entity are taken as follows: The weight of the internal word units of the personnel entity is taken Ordinary word weights are taken The process segment obtained from this calculation represents This will more significantly reflect the semantic composition of equipment and personnel within the segment.

[0112] 5) Calculate the attention weight of each word element for each process segment. For the first... First, calculate its relationship with the word element. The process distance between each process segment is then adjusted according to type compatibility to obtain an unnormalized score, expressed as:

[0113] ;

[0114] in, Indicates the first The word pair Unnormalized attention scores for each process segment; Represents the natural exponential function; This represents the process distance attenuation coefficient, used to control the rate attenuation of the influence of long-distance process segments, and can be taken from 0.6 to 1.2; Indicates the first The process segment number to which each word belongs; Indicates the target process segment number; The type compatibility coefficient describes the relationship between the entity type of the current lexical and the type of the lexical unit. The degree of compatibility between the dominant semantics of each process segment.

[0115] In one implementation, if the first... The terminology is located within the project entity, and the first terminology is located within the project entity. Each process segment focuses on deliverables and project-related semantics. 1.3 can be taken; if the first The word element is located within the personnel entity, and the first word element is located within the personnel entity. Each process segment focuses on device and implementation semantics. A value between 1.2 and 1.4 is acceptable; if the first... If a word is a common word or its type cannot be clearly determined, then Take 1; if the types are obviously incompatible, then A value of 0.7 to 0.9 is acceptable.

[0116] Furthermore, all Normalization yields the first... The attention weight of each word element to each process segment The sum of the weights of each focus is 1.

[0117] 6) Generate the number based on the attention weight. Context enhancement vectors of 1 word This allows the current lexical unit to retain its own semantics while also incorporating information about the preceding and following segments related to its process stage, as shown below:

[0118] ;

[0119] in, Indicates the first Context enhancement vectors for each word; This indicates element-wise multiplication; Indicates the first The word pair The attention weight of each process segment; Indicates the first The representative vector of each process segment.

[0120] 7) Context-enhanced vectors for all lexical units Stack them in order to obtain the context-enhanced vector matrix. . This represents a matrix representing the current document after context enhancement of the process fragment, with a size of [size missing]. Each row corresponds to a context enhancement vector for a word.

[0121] As an example, suppose a document is divided into four process segments: "Requirements Analysis," "Solution Design," "Implementation and Integration," and "Acceptance and Delivery." If the current term is located in the "Implementation and Integration" process segment and belongs to the equipment type, then the current term receives the highest attention weight for the "Implementation and Integration" process segment; it retains a relatively high attention weight for the "Solution Design" process segment because the equipment selection is usually already specified in the solution design; it retains a medium attention weight for the "Acceptance and Delivery" process segment because the equipment may be related to the test results; and it receives a significantly lower attention weight for the "Requirements Analysis" process segment because the process is far away and the type compatibility is weak.

[0122] It should be noted that this step does not rely on complex trainable attention structures, but directly utilizes the natural process sequence of the research documents. It generates context enhancement vectors through two-level rules of "process distance + type compatibility", which can still allow the model to obtain context support that closely resembles real business processes when the training samples are limited.

[0123] S3. Train an entity recognition and process constraint relationship extraction model that integrates prior dictionary embedding and boundary-sensitive span recognition;

[0124] After completing S2, the text already possesses unified terminology, type boundaries, and process-sensitive context, but model training is still required to complete entity recognition and relation extraction. The research institute's business texts exhibit issues such as numerous long entities, complex equipment model and version number structures, frequent appearances of unregistered entities, and relationships strongly constrained by process stages.

[0125] This step constructs a joint training model consisting of "prior dictionary embedding enhanced encoding - boundary-sensitive span entity recognition - process constraint relation extraction". The specific steps are as follows:

[0126] S301, Prior dictionary embedding enhanced context encoding;

[0127] Context-dependent augmentation vector matrix When performing sequence encoding, the model may still be insufficient in recognizing low-frequency equipment models, project codes, and versioned deliverable names.

[0128] This sub-step utilizes a priori dictionary to supplement domain knowledge, and then generates an entity recognition coding matrix through a bidirectional long short-term memory network. The specific steps are as follows:

[0129] 1) Establish a priori dictionary. The priori dictionary stores high-value domain terms, including project standard terminology, project codes, equipment models, equipment aliases, deliverable template names, common version spellings, personnel names, and job title expressions. Each term must be associated with at least a standard term and an entity category. The priori dictionary can be expanded from an entity thesaurus or supplemented with high-frequency terms from historical documents.

[0130] As an example, the following is a sample of prior dictionary entries (expanded from an entity thesaurus to include high-frequency historical terms):

[0131]

[0132] 2) Generate term vectors offline for each term in the prior dictionary. In one implementation, the term is first split into a character sequence, and then the character sequence is mapped to a character embedding vector; then, a 1D convolution is used to extract local structural patterns, with kernel lengths of 2, 3, and 4; then, max pooling is performed on the convolution output; finally, a fixed-length term vector is output through a fully connected layer. This method is more stable for structured naming such as "model + version number" and "code + number".

[0133] As an example, the term "a certain type of simulation platform V2.3" is split into a character sequence ['certain', 'type', 'simulation', 'real', 'platform', 'platform', '', 'V', '2', '.', '3'], which is then embedded into a matrix. The convolution kernel length is used... , , One-dimensional convolutions extract local patterns (such as "simulation", "platform", "V2.3"), which are then processed by max pooling and fully connected layers to output a fixed-dimensional (e.g., 128-dimensional) term vector. This approach has better generalization ability for structured naming such as "model + version number".

[0134] 3) Context-enhanced vector matrix Generate a query vector for each term in the dictionary. For the first term... Take the nth word as the center and extract two words before and after it to form a local window of length 5; average all context enhancement vectors within the window and then perform linear mapping to obtain the nth word. The query vector consists of individual words. This is done because research entities are often multi-word phrases, and using only a single word for dictionary retrieval can easily lose phrase-level information.

[0135] As an example, for the first Take the word "simulation" and extract the two words before and after it to form a window: "using a certain type of simulation platform V2.3". Average the context-enhanced vectors of the five words within this window, and then perform a linear transformation to obtain the query vector. At this point, the query vector already contains phrase-level information of "a certain type of simulation platform V2.3", and has a higher similarity to the vectors of equipment-related entries in the prior dictionary, effectively recalling relevant domain entries.

[0136] 4) Retrieve the prior dictionary based on the query vector and generate the prior vector. In the specific implementation, first filter out the preceding words from the prior dictionary based on word overlap, prefix matching, type tag consistency, or character edit distance. 10 candidate terms The similarity can be set to 5 or 10; then calculate the similarity between the candidate term vector and the query vector; then normalize the similarity, and sum the candidate term vectors according to the normalized weights to obtain the result. Prior vectors of each word . Indicates that the prior dictionary is for the first Each lexical element provides a domain-specific supplementary knowledge vector.

[0137] For example, if the word "simulation platform" appears in the text, and the prior dictionary already contains entries such as "a certain type of simulation platform", "semi-physical simulation platform", and "integrated verification platform", then the equipment category entries most similar to "simulation platform" will participate in the weighted summation, so that the current word unit obtains supplementary features that are clearly biased towards the equipment domain. Even if the specific expression does not appear sufficiently in the training set, it will still help to identify it as an equipment entity in the future.

[0138] 5) The first Context enhancement vectors of 1 word With prior vector By concatenating the vectors, a fused vector is obtained. Then merge all vectors. Input the data sequentially into a single-layer bidirectional long short-term memory network, and output an entity recognition encoding matrix. . This represents the matrix after dictionary enhancement and sequence encoding of the current document, the nth... The row corresponds to the first Encoding vectors of each word The hidden dimension of a bidirectional long short-term memory network can be 128 or 256, used to simultaneously encode the left and right contexts.

[0139] It should be noted that this sub-step introduces domain prior knowledge into the model during the encoding stage rather than the post-processing stage, so that structured naming such as device model, project code and deliverable template has prior support before entering entity recognition, thereby improving the recognizability of unregistered entities and low-frequency entities.

[0140] S302, Boundary-sensitive span entity recognition;

[0141] Entities in research institute business documents can be short entities, such as names, or long entities, such as "overall technical solution V2.3 for a certain pre-research project". If word-by-word annotation is used, problems such as incomplete boundaries, missing version numbers, or excessive modifiers being included in the entity are likely to occur.

[0142] This sub-step uses candidate span enumeration to identify entities and incorporates boundary-sensitive features and a boundary penalty mechanism to improve boundary localization accuracy. The specific steps are as follows:

[0143] 1) Generate candidate spans within each sentence. Each candidate span is denoted as ,in, Indicates the starting position of the candidate span. Indicates the ending position of the candidate span, and To control computational load, a maximum span length is set, which can be 8, 12, or 16 tokens. Candidate spans are generated without crossing sentences, and [PROJ], [PER], [EQUIP], and [DELIV] themselves are not considered part of the entity content.

[0144] As an example, consider the sentence: "[PROJ] New type radar pre-research project [PROJ] is led by [PER] Senior Engineer Zhang [PER]". When generating candidate spans, the type tags such as [PROJ] are not considered part of the entity content. For the lexical sequence ['new', 'radar', 'pre-research', 'project'] (located between two [PROJ] tags), the candidate spans are enumerated as follows:

[0145] Span 1: Start a=0, end b=3 (covering all 4 word elements);

[0146] Span 2: Start a=0, End b=2 (covering the first 3 words);

[0147] ...

[0148] The maximum span length is set to 12 lexical units, which is sufficient to cover common long entities in the research.

[0149] 2) Construct a span representation vector for each candidate span. In practical implementation, the span represents a vector. It is obtained by concatenating at least 6 parts: the starting position encoding vector Termination position encoding vector Max pooling result of the encoding vector within the span, and encoding vector of the left adjacent position. Right adjacent position encoding vector And the span length embedding vector. If the candidate span is at the beginning or end of the sentence, the out-of-bounds position is replaced with a zero vector. By introducing left adjacent position encoding vectors and right adjacent position encoding vectors, the model can more sensitively perceive whether role words, version numbers, or ordinary modifiers appear outside the entity boundary.

[0150] As an example, for candidate span Its span represents a vector Here is an example of concatenation (assuming the encoding vector dimension is 256):

[0151] Start position encoding (256 dimensions);

[0152] Termination position code (256 dimensions);

[0153] Max pooling results within the span (256 dimensions);

[0154] Left Adjacency Position Encoding (256 dimensions; if it exceeds the limit, it will be set to zero);

[0155] Right Adjacency Position Encoding (256 dimensions; if it exceeds the limit, it will be set to zero);

[0156] Span length embedding vector (e.g., 16-dimensional);

[0157] The total dimension after splicing is The dimensions are further compressed and classified by a multilayer perceptron.

[0158] 3) Represent the span vector The input is a multilayer perceptron, and the output is the predicted probability that the candidate span belongs to one of five categories: "Project", "Personnel", "Equipment", "Deliverables", and "Non-entity". The multilayer perceptron can use one hidden layer with a dimension of 256 and an activation function of ReLU.

[0159] As an example, the structure of a multilayer perceptron is as follows: 1296-dimensional input, 256-dimensional hidden layer (ReLU activation), and 5-dimensional output layer (corresponding to projects, personnel, equipment, deliverables, and non-entities). For example, the probability distribution output by the multilayer perceptron for the span "a certain type of simulation platform V2.3" is... Among them, the "device" category has the highest probability and is predicted to be this category.

[0160] 4) During training, training batches are constructed using a combination of "all positive samples + some hard-negative samples". Positive samples represent candidate spans that completely match the start and end positions of manually labeled entities; hard-negative samples represent candidate spans that partially overlap with real entities but whose boundaries are not completely consistent. For example, if the real entity span is [5,8], candidate spans could be [5,9], [4,8], or [6,8]. These types of candidate spans are most likely to cause boundary confusion and should be prioritized for training.

[0161] As an example, suppose the real entity "a certain type of simulation platform" corresponds to a span of... ,but:

[0162] Positive samples: candidate span It is completely consistent with the real entity;

[0163] Hard-to-bear samples: candidate span (Multiple swallowing of the word "system") (Multiple swallowing of the word "adopt") (Missing word "type");

[0164] In the training batch, all positive samples are retained, and hard negative samples are sampled and added at a certain proportion, so that the model focuses on learning to distinguish boundary deviations.

[0165] 5) Add boundary penalties to partially overlapping hard-to-bear samples to obtain entity recognition loss. , represented as:

[0166] ;

[0167] in, This represents the cross-entropy loss for candidate span 5 categories; This represents the boundary penalty weight, which can range from 0.2 to 0.5; This represents the boundary offset penalty term. The boundary offset penalty term is calculated as follows: for each hard-to-bear sample that is incorrectly predicted as an entity by the model, calculate the absolute value of the deviation between its starting position and the starting position of the nearest real entity, and the absolute value of the deviation between its ending position and the ending position of the nearest real entity. Then sum the two values ​​and average them over the hard-to-bear samples.

[0168] Based on this, the model not only learns "whether it is an entity", but also "where the boundary should fall".

[0169] In one embodiment, if the real entity is "a certain type of simulation platform", the corresponding span is [10,12], and the candidate span [10,13] also includes the subsequent common modifier "system" in the entity range, even if the model classifies it as a device, it will still be penalized for the boundary offset because there is one extra word on the right boundary. This can suppress the model's habitual expansion of the boundary.

[0170] 6) During the inference phase, candidate spans with predicted probabilities exceeding a threshold are retained as the entity candidate set. For completely overlapping candidate spans of the same category, the one with the higher score is retained. For candidate spans with an inclusion relationship but different categories and whose nesting is allowed in business terms, both can be retained. For candidate spans with slight overlap and incompatible coexistence, the one with the higher score and more compact boundaries is retained. The final entity set is obtained. . This represents the set of entities identified in the current document. Each entity includes at least the standard terminology text, entity category, start position, end position, and the process segment number to which it belongs.

[0171] As an example, during the inference phase, suppose three candidate spans are generated for the same text region, as follows:

[0172] Span A: Predicted category "device", score ;

[0173] Span B: Predicted category "device", score (If the score completely overlaps with A and belongs to the same category, keep A, which has the higher score).

[0174] Span C: Predicted category "device", score (Compared with A), since they cannot coexist and A has a higher score and a tighter boundary, A is kept and C is discarded;

[0175] Span D: Predicted category "Deliverables", score If the business allows for nested projects / deliverables, then retain it along with A.

[0176] Final entity set Each entity records standard terminology, category, start and end positions, and the process segment number to which it belongs.

[0177] It should be noted that this sub-step combines "in-boundary semantics + out-of-boundary adjacency semantics + length prior + boundary penalty" for the recognition of business text entities in the research institute, which can simultaneously improve the boundary accuracy of long entities, equipment model entities, and deliverable entities with version numbers.

[0178] S303, Extraction of process constraint relationships;

[0179] After entity identification, it is also necessary to determine whether there are business relationships such as "responsibility," "participation," "use," and "output" between entities. Business relationships within the research institute are not established simply by arbitrarily co-occurring; rather, they are clearly constrained by the combination of process stages and entity types. For example, projects and deliverables typically establish an "output" relationship in the "solution design" or "acceptance and delivery" process segments, while personnel and equipment typically establish a "use" relationship in the "implementation and commissioning" process segment.

[0180] This sub-step uses a process gating mechanism to suppress erroneous cross-process connections. The specific steps are as follows:

[0181] 1) Based on entity set Candidate entity pairs are generated from the relation constraint table. Only entity type combinations that satisfy the relation constraint table are retained, with priority given to entity pairs within the same sentence, the same process segment, or adjacent process segments. Entity pairs with excessively large process segment intervals are directly excluded from high-priority relation classification. This reduces a large number of obviously invalid candidate combinations.

[0182] As an example, the document identifies the following entities: project entity "JG2025-07", personnel entity "Senior Engineer Zhang", equipment entity "a certain type of simulation platform", and deliverable entity "technical solution V2.3".

[0183] When generating candidate entity pairs, the following are first retained based on the relation constraint table:

[0184] (Zhang, Senior Engineer, JG2025-07) → Possibly related to being "responsible for" or "involved in";

[0185] (Zhang, Senior Engineer, a certain type of simulation platform) → Possibly related to "use";

[0186] (JG2025-07, Technical Solution V2.3) → Possible relationship "output".

[0187] (JG2025-07, a certain type of simulation platform) is not in the constraint table, so no high-priority candidate pair is generated, and it is directly excluded.

[0188] For candidate pairs (Zhang Moumou, Senior Engineer, JG2025-07), if the two are located in two process segments with a large interval between "implementation and commissioning" and "acceptance and delivery", they will not participate in the high-priority relationship classification to avoid cross-stage misjudgment.

[0189] 2) Construct a basic representation for each candidate entity pair. In the specific implementation, for the current candidate entity pair, extract the span representation vectors of the head entity and the tail entity, respectively. Then, extract the pooling result of the encoded vectors of the text region between the two entities, and append the intra-sentence distance embedding and the process segment distance embedding. Concatenate these to obtain the basic representation of the entity pair. If the two entities are not in the same sentence but are located in adjacent process segments, the pooling result of the encoded vectors of the text region between the two entities can be replaced by the connection region between the two sentences or the starting context of the next sentence.

[0190] As an example, for the candidate entity pair ("Senior Engineer Zhang", "A certain type of simulation platform"), then:

[0191] Head entity span representation vector (Originated from the S302 span representation layer);

[0192] Tail entity span representation vector ;

[0193] The text region (“use”) between the two entities is pooled using the encoding matrix to obtain a context vector;

[0194] Intra-sentence distance embedding (both are in the same sentence, the distance is 2 words);

[0195] Process segment distance embedding (both belong to the "Implementation and Integration" segment, the distance is 0);

[0196] The above concatenation forms the basic representation of entity pairs, which are then input into the relationship classification module.

[0197] 3) Generate process gating vectors based on process segment distance, process stage combination, and entity type combination. The process gating vector for the current candidate entity pair is denoted as... , This represents the gating vector used to modulate the relational features of the current candidate entities, with a value range of 0 to 1. In its implementation, the distance between the process segments of two entities is first discretized into three categories: "same process segment," "adjacent process segments," and "separated by two or more process segments." Then, the process stage labels of the two entities are combined to form a stage combination feature. Next, the entity type combination is encoded into a type combination feature. Finally, these features are input into a small feedforward network, activated by a Sigmoid function, and the output is a process gating vector. .

[0198] In one implementation, if the current candidate entity pair is "personnel-equipment", and the two entities are located in the "implementation and commissioning" process segment or an adjacent "scheme design-implementation and commissioning" process segment, then the process gating vector... A higher pass value is given to the "use" related feature dimensions; if the current candidate entity pair is "project-deliverable" and is located in the "acceptance and delivery" process segment, then the process gating vector is... A higher pass value is given to the "output" related feature dimensions; if the entity type combination itself does not match, or the process segments are too far apart, the process gating vector is adjusted. Overall suppression relationship characteristic response.

[0199] 4) Perform relation classification on the entity pair basic representation. Specifically, first, input the entity pair basic representation into the relation feature transformation layer to obtain the relation hidden vector; then, combine the relation hidden vector with the flow gating vector... Element-wise multiplication yields the relational features constrained by the process; finally, these features are input into a classification layer, which outputs predicted probabilities for five categories: "responsible," "participating," "using," "producing," and "no relation." The "no relation" category is used to absorb invalid entity pairs, preventing the model from being forced to misclassify them as a business relationship.

[0200] 5) Train the relation classifier using cross-entropy loss to obtain the relation classification loss. The training samples include both positive entity pairs with real business relationships and negative entity pairs with no business relationship, extracted at a certain proportion. Through joint training with both related and unrelated samples, the model can learn to suppress entity pairs that co-occur in the same segment but are not related in business.

[0201] 6) During the inference phase, for each candidate entity pair, the relationship category with the highest predicted probability exceeding a threshold is selected as the output; if the probability of "no relationship" is the highest, then the business relationship of that entity pair is not output. This ultimately yields the relationship set. . This represents the set of business relationships extracted from the current document. Each relationship includes at least a head entity, a tail entity, a relationship category, a confidence level, and information about the process segment to which it belongs.

[0202] In one embodiment, if "Senior Engineer Zhang" and "a certain type of simulation platform" are located in the "Implementation and Debugging" process segment, the process gating vector will enhance the "use" relationship feature, and the model will more easily output "Senior Engineer Zhang - Use - a certain type of simulation platform"; if "a certain pre-research project" and "technical solution V2.3" are located in the "Acceptance and Delivery" process segment, the process gating vector will enhance the "output" relationship feature, and the model will more easily output "a certain pre-research project - Output - technical solution V2.3"; if a certain piece of equipment only appears in the "Requirements Analysis" process segment, and a certain deliverable only appears in the distant "Acceptance and Delivery" process segment, the process gating vector will significantly suppress the entity pair and avoid erroneous edge connections.

[0203] It should be noted that this sub-step transforms the combination of process location and entity type into a gating signal for relation feature modulation, rather than relying solely on entity co-occurrence. This can significantly reduce erroneous cross-process connections and improve the consistency between relation extraction results and the actual business process.

[0204] S304, Joint Training and Model Output;

[0205] To achieve coordinated optimization of entity recognition and relation extraction, this sub-step jointly trains the two tasks and outputs the final model parameters. The specific steps are as follows:

[0206] 1) Process the research business text training samples sequentially through steps S2 and S3 to obtain entity recognition prediction results and relation extraction prediction results, and calculate the entity recognition loss for each. And relational classification loss .

[0207] 2) Calculate the total loss , represented as:

[0208] ;

[0209] in, This indicates the total losses from the joint training exercise; This represents the weighting coefficient of the relation classification loss, used to balance the optimization intensity of entity recognition and relation extraction tasks. It can be taken from 0.3 to 0.8, for example, 0.5.

[0210] 3) Update the model parameters using an adaptive moment estimator. The initial learning rate can be set to... The batch size can be 8, 16, or 32, and the total number of training cycles can be 50 to 150. Gradient clipping, learning rate decay, and early stopping strategies can be used during training to improve training stability.

[0211] 4. After each training cycle, the F1 score for entity recognition and the F1 score for relation extraction are calculated using the validation set. Then, a weighted composite score is calculated, and the model parameters corresponding to the highest composite score are saved as the final output model. The final output model includes the parameters of the prior dictionary lookup layer, bidirectional long short-term memory network, span entity recognition layer, flow gating network, and relation classification layer.

[0212] It should be noted that the accuracy of entity boundaries directly affects the quality of relation extraction, and the relation extraction results, in turn, constrain entity recognition to be closer to real business objects. By conducting joint training, the two tasks can promote each other and reduce the error propagation caused by individual training.

[0213] In one embodiment, a performance comparison analysis of different methods on entity recognition tasks is conducted to verify the effectiveness of the entity recognition module in the knowledge graph construction method proposed in this invention. Horizontal comparisons with other typical methods and ablation configurations clarify the contribution of each technical component to the final recognition performance. The experiment compares five different processing strategies: first, the traditional bidirectional long short-term memory network combined with a conditional random field model, which serves as a classic baseline for early sequence labeling tasks; second, a deep learning method based on a bidirectional encoder representation transformation model combined with bidirectional long short-term memory networks and conditional random fields, representing the current mainstream pre-trained language model enhancement scheme; third, the complete method proposed in this invention, including semantic alignment, process enhancement, prior dictionary embedding, and boundary-sensitive span recognition; fourth, an ablation version that removes process fragment context enhancement; and fifth, an ablation version that removes prior dictionary embedding. All methods are trained and evaluated on the same set of research business text annotation samples. The evaluation metric is the harmonic mean score of entity recognition, and the mean and standard deviation of five randomized repeated experiments are used as the basis for presentation. In the experimental figures, the horizontal axis labels the specific names of the five methods, and the vertical axis represents the harmonic mean score of entity recognition, in percentage form. The average performance of each method is presented in a bar chart, with error bars added to the top of the bars to reflect the fluctuation range of multiple experimental results. The bar height clearly shows that the bar corresponding to the complete method of this invention is significantly taller than the other four configurations, and the error bars are relatively shorter, indicating that this method not only has the best entity recognition performance on research business texts but also exhibits stable performance. The version without the prior dictionary shows a performance decline compared to the complete method, indicating that structured naming rules and prior knowledge of domain terminology play an important supporting role in recognizing long-tail entities such as equipment models and project codes. The version without process enhancement also shows a performance drop, reflecting that process stage information helps the model accurately locate entity boundaries within chapter boundaries and implicit processes.

[0214] In one embodiment, a comparative analysis of relation extraction performance at different process stages is conducted to verify the effectiveness of the process gating mechanism for relation extraction tasks. The gain of the proposed method relative to the baseline method is observed by evaluating the harmonic mean score of relation extraction at different business process stages. The experiment selected four research business process stages with clear temporal characteristics: requirements analysis, scheme design, implementation and commissioning, and acceptance and delivery. The baseline method used for comparison was a relation classifier based on a bidirectional encoder representation transformation model. This baseline only utilizes the contextual semantics of entity pairs for relation discrimination, without introducing process position and stage constraint information. The proposed method adds a process gating vector to the same encoding, dynamically modulating relation features based on the distance of the entity pair to the process segment and the stage combination. Both methods use the same training data and evaluation criteria, and independently calculate performance indicators and standard deviations at each stage. The horizontal axis of the experimental graph labels the names of the four business process stages, and the vertical axis represents the harmonic mean score of relation extraction, in percentage form. Line graphs with error bars are used to plot the performance changes of the baseline method and the proposed method at each stage. As can be seen from the line trend, the line corresponding to the method of this invention is generally above the baseline and relatively stable, while the baseline method fluctuates more significantly between different stages. Especially in the scheme design and implementation integration stages, the leading margin of the method of this invention is more prominent, because the relationships between personnel and equipment usage, and between project and deliverable outputs are most intensive in these two stages, and the process gating mechanism can effectively suppress the tendency of erroneous connections between cross-stage co-existing entities. Furthermore, the error bars added at the line nodes reflect the stability in repeated experiments, such as... Figure 2 As shown, the error bars of the method of the present invention are generally shorter, which further confirms the positive effect of process constraints on reducing the uncertainty of relation extraction.

[0215] In one embodiment, ablation analysis of the contribution of core components within the model is performed. Using the paradigm of ablation studies, the contribution weight of each key design component to the final performance is quantified. The experimental results are presented as a horizontal bar chart. The vertical axis lists five different model configurations: "Full Model," "Removal of Prior Lexicon Augmentation," "Removal of Flow Gating," "Removal of Boundary Sensitivity Mechanism," and "Removal of Type Marking." The horizontal axis represents the "Harmonic Mean Value," with red bars representing entity recognition and green bars representing relation extraction plotted side-by-side. The length of the bars directly corresponds to the performance value. In the experimental chart, the "Full Model" bar at the top is the longest, representing the performance benchmark. Figure 3 As shown, when a component is removed, the length of the bars is shortened to varying degrees.

[0216] S4. Online extraction of research institute business texts and generation of knowledge results;

[0217] After completing model training, this invention is deployed in the research institute's business text processing system to perform automatic knowledge extraction on newly received project initiation materials, weekly reports, test records, acceptance reports, and equipment operation records. The specific steps are as follows:

[0218] 1) For a new document, first perform semantic alignment, type marking, process segmentation, and context enhancement according to S201 and S202 to obtain the context enhancement vector matrix of the document. Since the online processing phase still uses the same entity synonym mapping dictionary, process phase trigger vocabulary, and business process ontology skeleton as the training phase, the new document maintains consistency with the training document in terms of input specifications.

[0219] 2) Context-enhanced vector matrix The trained model is fed into the model to generate an entity recognition encoding matrix. Then perform boundary-sensitive span entity recognition to obtain the entity set. Subsequently, candidate entity pairs are constructed according to the relation constraint table, and the relation set is obtained through process gating relation classification. .

[0220] 3) For entity sets The system performs standard terminology primary key merging on entities. In practice, it uses "standard terminology + entity category" as the entity primary key to merge entities with the same name within the same document and across documents. If the same object has both a historical name and a current name, it is merged under the standard terminology name, and the historical alias field is retained for easy retrieval and tracing later.

[0221] 4) Entity set and relation set Write the data into a structured storage system or graph database to form a business knowledge network centered around four ontology concepts: "project-personnel-equipment-deliverables". Each entity record must store at least the standard terminology, entity category, source document, first occurrence position, and process segment number; each relation record must store at least the head entity, tail entity, relation category, confidence level, source document, and process segment number.

[0222] For example, given the text "A certain pre-research project is managed by Senior Engineer Zhang, and the verification of technical solution V2.3 is completed using a certain simulation platform," the system can ultimately output the project entity "certain pre-research project," the personnel entity "Senior Engineer Zhang," the equipment entity "certain simulation platform," and the deliverable entity "technical solution V2.3." It will also output three business relationships: "Senior Engineer Zhang - Responsible for - certain pre-research project," "Senior Engineer Zhang - Uses - certain simulation platform," and "certain pre-research project - Output - technical solution V2.3." If this text is located within the "Implementation and Integration" process segment, the relationship record will simultaneously save the "Implementation and Integration" process segment information for subsequent process traceability and project overview retrieval.

[0223] Finally, it should be noted that this invention does not simply perform generic named entity recognition and relationship classification on research institute texts. Instead, it focuses on the core framework of the research institute's business process ontology, and links standardized terminology, type boundary hints, enhanced context of process segments, sensitive boundary span recognition, and extraction of process constraint relationships into a complete technical chain. Under the conditions of closed terminology, dense abbreviations, strong process constraints, and limited labeled samples unique to research institutes, it can stably extract four types of entities, namely "project-personnel-equipment-deliverables," and their core business relationships, and output structured knowledge results that can be directly stored in the database.

[0224] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a knowledge graph for the systematization of research institute knowledge, characterized in that, Includes the following steps: S1. Constructing Research Institute Business Text Samples and Establishing Business Process Ontology Skeleton: Collect research institute business texts to form a raw text set, preprocess the texts and retain the integrity of business-sensitive segments to obtain the raw text sequence; establish an ontology concept and relation constraint table with "project-personnel-equipment-deliverables" as the core; establish an entity synonym mapping dictionary, process stage trigger word list and abbreviation ambiguity adjudication rules; S2. Text semantic alignment and process fragment context enhancement for business processes: Standard term replacement is performed on the original text sequence and ontology type tags are inserted to obtain a semantically aligned text sequence; the semantically aligned text sequence is divided into process fragments, and a context enhancement vector matrix is ​​generated according to the process fragment to which the word belongs and the word type; S3. Training a model for entity recognition and process constraint relationship extraction that integrates prior dictionary embedding and boundary-sensitive span recognition: Prior vectors are generated for lexical units based on the prior dictionary and fused with context-enhanced vectors. After sequence encoding, an entity recognition encoding matrix is ​​obtained. Entities are identified using a candidate span enumeration method. A span representation vector containing left and right adjacent position encodings is constructed for candidate spans to classify entities. Boundary offset penalties are applied to partially overlapping hard negative samples during training. Candidate entity pairs are generated based on the relationship constraint table. Process gating vectors are generated based on the distance between process segments involved in the entity pair and the stage combination. The basic representation of the entity pair is modulated before outputting the relationship classification result. Jointly train entity recognition and relation extraction tasks, and output the final model parameters; S4. Online extraction of research business text and generation of knowledge results: Perform semantic alignment, type labeling and context enhancement processing on new documents in the same way as in the training phase, feed them into the trained model to obtain entity set and relation set, perform standard terminology primary key merging on entities and write them into the storage system to form a business knowledge network.

2. The method according to claim 1, characterized in that, The process segment context enhancement in S2 specifically includes: The semantically aligned text sequence is divided into multiple process segments according to document structure and process stage trigger words; the process segment number to which each word belongs is recorded; a type gain coefficient is applied to the initial vector of the word to obtain a type-sensitive word representation; for the j-th word, the product of the process distance attenuation factor and the type compatibility coefficient between it and the k-th process segment is calculated, and after normalization, the attention weight of the word to the k-th process segment is obtained; the representative vectors of each process segment are weighted and summed according to the attention weight, and then fused with the type-sensitive word representation to generate the context enhancement vector.

3. The method according to claim 1, characterized in that, The entity recognition with boundary-sensitive span in S3 specifically includes: For each sentence, candidate spans are enumerated to generate a span representation vector. The span representation vector is composed of at least the start position encoding vector, the end position encoding vector, the max pooling result of the span internal encoding vector, the left adjacent position encoding vector, the right adjacent position encoding vector, and the span length embedding vector. The span representation vector is input into a multilayer perceptron, which outputs the predicted probability that the candidate span belongs to various entities or non-entities. During training, the boundary offset penalty is calculated based on the absolute value of the deviation between the hard negative sample predicted by the model as an entity and the start and end positions of the nearest real entity, in order to suppress entity boundary drift.

4. The method according to claim 1, characterized in that, The extraction of process constraint relationships in S3 specifically includes: Candidate entity pairs are generated based on entity types and the relation constraint table. For each candidate entity pair, a basic entity pair representation is constructed, including a head entity span representation, a tail entity span representation, inter-entity text pooling results, intra-sentence distance embedding, and process segment distance embedding. Based on the discrete distance values ​​of the process segments to which the two entities belong, the combination of process stage labels, and the combination of entity types, the process gating vector is generated through a feedforward network and an activation function. The basic entity pair representation is then multiplied element-wise with the process gating vector after feature transformation to obtain the relation features after process constraints, which are then input into the classification layer to output the relation category.

5. The method according to claim 1, characterized in that, The entity synonym mapping dictionary entries in S1 include standard terms, entity categories, abbreviations, internal codes, numbering patterns, and contextual constraints; the process stage trigger word list includes at least the trigger word sets corresponding to the four stages of "requirements analysis", "solution design", "implementation and commissioning", and "acceptance and delivery"; the abbreviation ambiguity adjudication rule is as follows: take a context window of a preset length centered on the current abbreviation, and adjudicate the entity category based on the trigger word categories appearing in the window and the pre-registered information in the entity synonym mapping dictionary.

6. The method according to claim 1, characterized in that, The preprocessing in S1 specifically includes: converting each document in the original business text set into a plain text stream, and segmenting it according to periods, semicolons, colons, line breaks, numbered item boundaries, and table cell boundaries to obtain the original text sequence; during segmentation, continuous business-sensitive segments composed of project numbers, equipment models, version numbers, dates, job title suffixes, and internal codes are not split, so as to keep them as complete word units.

7. The method according to claim 1, characterized in that, The prior dictionary embedding generation method in S3 is specifically as follows: For each word in the prior dictionary, it is split into a character sequence and mapped to a character embedding vector. Local structural patterns are extracted by one-dimensional convolution with multiple convolution kernels of different lengths. Max pooling is performed on the convolution output, and then a fixed-dimensional word vector is output through a fully connected layer. During the encoding stage, a local window of a preset length is taken centered on the current word, and the query vector is obtained by averaging all context enhancement vectors within the window and performing linear mapping. Based on the query vector, candidate words are retrieved from the prior dictionary, and the prior vector is generated by weighted summation according to similarity.

8. The method according to claim 1, characterized in that, The joint training in S3 adopts an alternating optimization strategy: in each training batch, the parameters of the relation extraction module are fixed first, and only the parameters of the entity recognition module are updated and the entity recognition loss is calculated; then the parameters of the entity recognition module are fixed, the parameters of the relation extraction module are updated and the relation classification loss is calculated; finally, the two losses are weighted and summed before the overall parameters are updated; the weight ratio of the entity recognition loss to the relation classification loss is 1:0.3 to 1:0.

8.

9. The method according to claim 1, characterized in that, The standard terminology primary key merging in S4 specifically includes: using standard terms and entity categories as entity primary keys to merge entities identified within the same document and across documents; when the same object has standard terms, abbreviations, internal codes, and historical names, they are uniformly merged under the standard term name, and the alias field and alias type label are retained during storage; if two different standard terms are confirmed to be the same business object through contextual adjudication, a manual review process is triggered or they are automatically merged according to preset rules, and a merge traceability log is recorded.

10. The method according to claim 1, characterized in that, After completing S4, the process also includes knowledge graph quality assessment and incremental update steps: For the business knowledge network written to the storage system, the entity density, relationship density and cross-stage relationship ratio within each stage are statistically analyzed according to the process stage dimension. If the proportion of cross-stage relationships exceeds a preset threshold, an alarm will be triggered and the corresponding relationship will be marked for manual review. When new business documents are added to the database, existing model parameters are reused for incremental reasoning, and the newly extracted entities and relationships are compared with the existing knowledge network in terms of semantic similarity. For candidate nodes whose similarity exceeds the threshold, automatic fusion or candidate conflict prompts are performed to achieve continuous evolution and consistency maintenance of the knowledge graph.

Images