Knowledge graph enhanced intelligent question answering and writing assistance engine for innovation management

Abstract

The application discloses to the field of artificial intelligence, specifically to the knowledge graph enhanced innovation management intelligent question answering and writing assistance engine, first, collect internal and external multi-modal raw materials, and process into standardized units, then retain the backtracking meta-information, and synchronize the sensitive information, which lays the foundation for subsequent extraction and evidence tracing; then, the standardized units are converted into five-tuple structured facts containing time intervals and confidence, finally, the multi-modal information is extracted, the confidence is calculated, the modal embedding is generated, and the alignment is achieved through contrast learning, providing high-quality input for the construction of time-series knowledge graph; based on this, the application can significantly reduce the generation illusion, improve the fact consistency and evidence coverage, and ensure that the system has the ability of sustainable evolution and compliance use, and the overall technical effect is more superior.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a knowledge graph-enhanced intelligent question-answering and writing assistance engine for innovation management. Background Technology

[0002] In the innovation management of modern enterprises and research institutions, a large number of patents, papers, project reports, and experimental data are constantly generated in multimodal forms, containing rich innovative elements and technical evidence. How to manage this heterogeneous information in a structured manner, and achieve efficient retrieval, interpretable reasoning, and knowledge services while maintaining semantic integrity and temporal logic, has become an important direction for improving the quality of R&D decisions and technical documents. Meanwhile, the rise of intelligent question-answering and writing assistance systems has enabled knowledge graphs to not only play a role in information organization and representation, but also to demonstrate broad application prospects in supporting high-quality knowledge services, enhancing decision support capabilities, and promoting the standardization of innovation activities. Research surrounding multimodal knowledge fusion and intelligent generation has thus become a key background for promoting the development of innovation management and intelligent writing.

[0003] To address the aforementioned application needs, existing research has primarily progressed along several complementary directions. Firstly, embedding knowledge graphs (KGs) into retrieval-generation (RAG) frameworks has become the mainstream approach, with researchers using document graphs or external triples as retrieval units to organize the retrieval context and improve the factuality and citation quality of generated data. Secondly, graph structures and graph neural networks (GNNs) are directly introduced into the reader or inference stages to utilize explicit triples for evidence aggregation and interpretable path retrieval. Thirdly, temporal knowledge graphs (KGs) and time-sensitive inference methods have been proposed to handle scenarios where facts evolve over time. Fourthly, work on citation / review generation explores using KGs to represent relationships between papers to improve document organization and citation accuracy. Fifthly, a few studies have attempted to incorporate non-textual modalities such as tables and charts into the system, but these are mostly independent modules or simple preprocessing, and a unified cross-modal alignment and end-to-end fusion scheme has not yet been formed. Overall, these approaches have made significant progress in improving the utilization of question-and-answer evidence, enhancing the structure of generated text, and supporting time-sensitive queries. However, most academic and engineering implementations are still text-based and modular, and overall solutions for engineering, multimodal, and auditable applications are still under development.

[0004] While existing methods have made some progress in improving the factual accuracy of question-answering and writing assistance, several shortcomings remain. First, they lack the ability to uniformly process multimodal evidence. Existing technologies are mostly text-centric, making it difficult to effectively align tables, charts, experimental data, and text semantics, leading to loss and mismatch of cross-modal information during fusion and reasoning. Second, existing methods are weak in handling the temporal and uncertain modeling of facts, often relying on static filtering or simple threshold judgments, lacking systematic support for factual versioning, applicable time ranges, and evidence relevance, making it difficult to meet the auditability and robustness requirements of innovation management scenarios. Third, in the generation stage, existing solutions generally rely on prompts or retrieval context for guidance, lacking mechanisms to impose strict factual constraints and visualize the evidence chain during the decoding stage, making it difficult to fully guarantee the verifiability and traceability of the generated content. Summary of the Invention

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] The knowledge graph-enhanced intelligent question-answering and writing assistance engine for innovation management includes the following specific steps:

[0007] S1, Multi-source data acquisition and standardization: First, collect internal and external multimodal raw materials and process them into standardized units. Then, retain the retrospective metadata and isolate sensitive information simultaneously to lay the foundation for subsequent extraction and evidence tracing.

[0008] S2, Multimodal Information Extraction and Uncertainty Labeling: First, the standardized units are converted into five-tuple structured facts containing time intervals and confidence levels. Then, multimodal information is extracted, confidence levels are calculated, and modal embeddings are generated and aligned through contrastive learning to provide high-quality input for the construction of time-series knowledge graphs.

[0009] S3, Construction of Time-Sequential and Versionable Knowledge Graph: First, structured facts are inserted into a knowledge graph containing time, version, and uncertainty. Then, the weights of the facts are adjusted using a time decay function. Subsequently, the training entity relationship vectors are connected, and historical versions and audit trajectories are saved to achieve dynamic management and traceability of facts.

[0010] S4, Semantic parsing and subgraph retrieval: Maps the user's natural language query or writing instructions to retrieval constraints and pulls candidate subgraphs related to the question from the knowledge graph and vector index;

[0011] S5, Graph-enhanced Hybrid Reasoning: First, path scores are calculated on candidate subgraphs, and node representations are obtained through parallel GNN propagation. Then, vectorized scores and path scores are combined, and logical verification is performed using a symbolic rule base to generate candidate answers with interpretable evidence paths and quantified confidence, thereby improving reasoning accuracy.

[0012] S6, Controlled writing generation based on knowledge graph constraints: Taking the evidence set obtained through reasoning and writing instructions as input, a controlled decoding strategy is used to generate fluent and factually consistent text;

[0013] S7, Fact Verification, Human-Machine Confirmation and Closed-Loop Knowledge Graph Update: First, the assertions in the text are automatically verified and their support is calculated. Then, assertions with low support are prompted for human review. At the same time, the human confirmation information is written back to the KG with high confidence. Then, an immutable audit log is recorded.

[0014] As a preferred embodiment of the knowledge graph-enhanced innovation management intelligent question-answering and writing assistance engine described in this invention, the specific steps of S1 are as follows:

[0015] First, raw materials from internal and external sources are collected over a long period of time or in real time, and various unstructured and semi-structured inputs are converted into unified standardized document units, while a traceable evidence location index is retained on each standardized unit.

[0016] Next, sensitive information is detected and isolated according to pre-configured policies to meet compliance requirements.

[0017] As a preferred embodiment of the knowledge graph-enhanced innovation management intelligent question-answering and writing assistance engine described in this invention, the specific steps of S2 are as follows:

[0018] First, standardized document units are transcribed into structured fact triples, and each fact is assigned a confidence level and a time applicability range to facilitate the subsequent construction of a temporal knowledge graph. The extraction results are expressed in the form of quintuples:

[0019]

[0020] in, The main entity representing the fact triple; Indicates the relation type; Represents an object entity or attribute value; Indicate the time interval to which the facts apply, and give the start time. With end time ; Indicates the degree of confidence in the fact;

[0021] Next, the text extraction adopts a domain-fine-tuned named entity recognition and relation extraction model, the table extraction generates attribute declarations based on column semantic mapping, and the chart is converted into numerical facts and normalized units through OCR and coordinate conversion.

[0022] Then, increase the confidence level. The initial calculation can be obtained from the model output probability. With rules / source weight The fusion is given, using simple weighted fusion:

[0023]

[0024] in, This is a range mapping function; For weight hyperparameters, Indicates the initial credibility score of the source;

[0025] Simultaneously generate vector representations for text segments, table cells, and image regions. , , Furthermore, through contrastive learning, the representations of the same entity or fact are brought closer together across different modalities. The contrastive loss used can be InfoNCE.

[0026]

[0027] in, and These are encoders for different modes. For similarity function, , As a positive sample, Negative sample set.

[0028] As a preferred embodiment of the knowledge graph-enhanced innovation management intelligent question-answering and writing assistance engine described in this invention, the specific steps of S3 are as follows:

[0029] First, the extracted structured facts are inserted into a knowledge graph that supports time intervals, version control, and uncertainty annotation; the knowledge graph is a directed weighted graph. This means that each side of the diagram corresponds to a fact quintuple and carries confidence level and time information;

[0030] Next, a time decay function is introduced. Adjust the weights of the facts:

[0031]

[0032] in, Indicates the effective weight of the edge. The original confidence level of the facts. This represents the distance between the current query time and the actual time. The attenuation coefficient;

[0033] Subsequently, the vector representation of the entities in the diagram With relation vector A connection training strategy is employed, where one of the structural training objectives is link prediction based on an energy function, using the TransE scoring function:

[0034]

[0035] in, , These are the vector representations of the head entity and the tail entity, respectively. The relationship vector is used; during training, the margin-loss is used to minimize the difference between negative and positive samples.

[0036]

[0037] in, For the positive sample set, For the constructed negative sample set, This is the margin hyperparameter.

[0038] As a preferred embodiment of the knowledge graph-enhanced innovation management intelligent question-answering and writing assistance engine described in this invention, the specific steps of S4 are as follows:

[0039] First, the user inputs Encoded as semantic vectors ;

[0040] The retrieved score then combines semantic similarity with path-based structural score to form a hybrid scoring function:

[0041]

[0042] in, Represents vector similarity, For entities in the diagram The vector representation of , Indicates from candidate entities The score of the optimal path that can be obtained from the starting point to the relevant facts. This is the balance coefficient;

[0043] Next, subgraph retrieval uses vector retrieval to prioritize the location of highly semantically relevant entities, and combines graph indexing to perform width-constrained path expansion to generate candidate subgraphs. The candidate subgraphs are further filtered under time constraints and confidence thresholds and then passed to the hybrid inference module.

[0044] As a preferred embodiment of the knowledge graph-enhanced innovation management intelligent question-answering and writing assistance engine described in this invention, the specific steps of S5 are as follows:

[0045] First, in the candidate subgraph Run a hybrid reasoning process to produce answers with interpretable evidence paths and quantified confidence levels; for any path The path score is defined as the product of edge weights with a length penalty:

[0046]

[0047] in, For the first Effective weight of the edge, This is a path length penalty function. This is the path length attenuation factor;

[0048] In parallel, in the subgraph The upper execution graph neural network propagates to obtain the upper and lower representations of each node. The propagation of one layer of a GNN can be described as follows:

[0049]

[0050] in, For adjacency matrices with self-loops, for The degree matrix, For the first The node feature matrix of the layer, For learnable weight matrix, For activation functions;

[0051] The final candidate answer's overall score is a weighted combination of vectorized score and path score, further processed by a symbolic rule base. Perform a domain logic and time consistency check; entries that do not meet the symbol rules will be flagged or removed. The overall scoring format is as follows:

[0052]

[0053] in, As a weighting balancing factor, For vector similarity scoring based on GNN representation, For The set of candidate paths for endpoints.

[0054] As a preferred embodiment of the knowledge graph-enhanced innovation management intelligent question-answering and writing assistance engine described in this invention, the specific steps of S6 are as follows:

[0055] First, the basic language model provides the input conditions. With historical generation At that time, for the next token The original probability is ;

[0056] Next, fact scores are introduced. Consistency penalty item Construct a controlled decoding distribution:

[0057]

[0058] in, Hyperparameters are used to balance smoothness and factual constraints; For token A score indicating how well the information matches the facts or entity of KG. Penalty scores for conflicting evidence with KG or violating symbol rules;

[0059] Subsequently, the writing training uses joint loss to optimize language fluency and factual consistency:

[0060]

[0061] in, The negative log-likelihood of the language model. For fact consistency loss, based on KG coverage or cross-entropy of fact-based questions and answers, These are the weighting coefficients;

[0062] Finally, the generated text is output along with an evidence citation index for each assertion to ensure traceability.

[0063] As a preferred embodiment of the knowledge graph-enhanced innovation management intelligent question-answering and writing assistance engine described in this invention, the specific steps of S7 are as follows:

[0064] First, the assertions in the generated text undergo secondary automatic validation, followed by manual verification, to complete the incremental update of the KG; the automatic validation module verifies each assertion in the text. Perform retrieval validation and calculate its support. The support level is the coverage of high-confidence evidence for this assertion in KG:

[0065]

[0066] in, To support the assertion The set of candidate evidence, As evidence confidence level For the preset threshold, For indicator functions;

[0067] Next, assertions with low support are labeled and highlighted in the user interface to prompt users for manual review; simultaneously, the user's adoption or modification is recorded as a manual confirmation signal and written back to the KG according to the configured policy. When writing back, the manual confirmation evidence can be assigned a higher initial confidence level. ;

[0068] Subsequently, all write-back operations are saved as audit logs in an immutable log.

[0069] Compared with existing technologies:

[0070] This invention offers the following advantages: Highly reliable knowledge management with traceability and time-awareness: Through a unified meta-information model and a time-seriesd, versionable knowledge graph, each assertion can be traced back to its specific source, location, and version. It can also automatically adjust fact weights based on time sensitivity, significantly improving the verifiability of conclusions and the ability to judge timeliness in different application scenarios. Robust evidence resulting from multimodal fusion and uncertainty quantification: The system aligns multimodal evidence such as text, tables, and images and assigns confidence levels to each fact. It can quantify uncertainty and trigger corresponding strategies (such as manual review) when evidence is sparse or conflicting, thus providing more comprehensive and robust evidence support in complex heterogeneous information environments. Explainable and efficient inference combining vectorized reasoning and symbolic rule verification: By using vector-based semantic reasoning and path-based symbolic evidence chains in parallel, it retains the flexibility of semantic retrieval while providing clear evidence chains and logical verification, balancing retrieval / reasoning effectiveness and interpretability, and reducing the risk of factual errors caused by black-box generation. Controlled writing and closed-loop updates ensure factual consistency and long-term reliability: Controlled generation and constrained search based on knowledge graph constraints significantly reduce the illusion of generated content. At the same time, through automatic verification and human-machine closed-loop confirmation, highly credible information is written back by humans and sensitive fields are protected by differential strategies. This not only improves the factual consistency of generated text, but also ensures compliance and privacy protection in continuous iterative updates.

[0071] In summary, this invention introduces mechanisms such as traceability, explainability, multimodal fusion, and human-machine collaboration throughout the entire knowledge acquisition, storage, reasoning, generation, and updating chain, forming a highly reliable generation method that balances accuracy, timeliness, and compliance. Compared with existing technologies that rely solely on large models or static knowledge bases, this invention significantly reduces generation illusions, improves factual consistency and evidence coverage, and ensures the system's ability to evolve sustainably and be used compliantly, resulting in superior overall technical performance. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0073] This invention provides a knowledge graph-enhanced intelligent question-answering and writing assistance engine for innovation management, including the following specific steps:

[0074] S1, Multi-source data acquisition and standardization: First, collect internal and external multimodal raw materials and process them into standardized units. Then, retain the retrospective metadata and isolate sensitive information simultaneously to lay the foundation for subsequent extraction and evidence tracing.

[0075] The specific steps of S1 are as follows:

[0076] First, raw materials from internal and external sources are collected over a long period or in real-time. Various unstructured and semi-structured inputs are converted into unified standardized document units, with a traceable evidence location index retained on each standardized unit. To achieve this, a format recognizer is used to parse the input files to obtain document text, table structures, and image annotations. Character encoding, noise character filtering, and sentence segmentation are performed on the text. Semantic header recognition and cell type determination are performed on the tables. OCR and semantic image processing are performed on the images. The standardized units record the following metadata during storage: Document ID, Source system, original document location index (Page / Offset / CharRange), collection timestamp, and original document checksum, serving as unique identifiers for subsequent extraction and evidence tracing. The standardized data stream generated in this step is written to the indexing service and extraction pipeline for use in the next step. Then, sensitive information is detected and isolated according to a pre-configured strategy to meet compliance requirements.

[0077] S2, Multimodal Information Extraction and Uncertainty Labeling: First, the standardized units are converted into five-tuple structured facts containing time intervals and confidence levels. Then, multimodal information is extracted, confidence levels are calculated, and modal embeddings are generated and aligned through contrastive learning to provide high-quality input for the construction of time-series knowledge graphs.

[0078] The specific steps of S2 are as follows:

[0079] First, standardized document units are transcribed into structured fact triples, and each fact is assigned a confidence level and a time applicability range to facilitate the subsequent construction of a temporal knowledge graph. The extraction results are expressed in the form of quintuples:

[0080]

[0081] in, The head entity representing the fact triple; Indicates the relation type (relation); Represents an object entity or attribute value (tail entity / value); Indicate the time interval to which the facts apply, and give the start time. With end time ; Indicates the degree of confidence in the fact;

[0082] Next, text extraction employs a domain-fine-tuned Named Entity Recognition (NER) and Relation Extraction (RE) model; table extraction generates attribute declarations based on column semantic mapping; and charts are converted into numerical facts and normalized units through OCR and coordinate transformation.

[0083] Then, increase the confidence level. The initial calculation can be obtained from the model output probability. With rules / source weight The fusion is given, using simple weighted fusion:

[0084]

[0085] in, This is a range mapping function; For weight hyperparameters, Indicates the initial credibility score of the source;

[0086] To enhance cross-modal retrieval and alignment capabilities, this step generates vector representations for text segments, table cells, and image regions simultaneously. , , Furthermore, through contrastive learning, the representations of the same entity or fact are brought closer together across different modalities. The contrastive loss used can be InfoNCE.

[0087]

[0088] in, and These are encoders for different modes. For similarity function, , As a positive sample, Negative sample set.

[0089] S3, Time-sequential and versionable knowledge graph construction (T-KG): First, structured facts are inserted into a knowledge graph containing time, version, and uncertainty. Then, the weights of the facts are adjusted using a time decay function. Subsequently, the training entity relationship vectors are connected, and historical versions and audit trajectories are saved to achieve dynamic management and traceability of facts.

[0090] The specific steps of S3 are as follows:

[0091] First, the extracted structured facts are inserted into a knowledge graph that supports time intervals, version control, and uncertainty annotation; the knowledge graph is a directed weighted graph. This means that each side of the diagram corresponds to a fact quintuple and carries confidence level and time information;

[0092] Next, to reflect the importance of facts changing over time in retrieval and reasoning, a time decay function is introduced. Adjust the weights of the facts:

[0093]

[0094] in, Indicates the effective weight of the edge. The original confidence level of the facts. This represents the distance between the current query time and the actual time. The attenuation coefficient;

[0095] Subsequently, the vector representation of the entities in the diagram With relation vector A connection training strategy is employed, where one of the structural training objectives is link prediction based on an energy function, using the TransE scoring function:

[0096]

[0097] in, , These are the vector representations of the head entity and the tail entity, respectively. The relationship vector is used; during training, the margin-loss is used to minimize the difference between negative and positive samples.

[0098]

[0099] in, For the positive sample set, For the constructed negative sample set, This is the margin hyperparameter. To ensure the temporal sequence and versionability of the graph, the graph database stores the version number and source list for each edge, and all historical versions and audit trails are persistently saved for backtracking.

[0100] S4, Semantic parsing and subgraph retrieval: Maps the user's natural language query or writing instructions to retrieval constraints and pulls candidate subgraphs related to the question from the knowledge graph and vector index;

[0101] The specific steps of S4 are as follows:

[0102] First, the user inputs Encoded as semantic vectors ;

[0103] The retrieved score then combines semantic similarity with path-based structural score to form a hybrid scoring function:

[0104]

[0105] in, Represents vector similarity, For entities in the diagram The vector representation of , Indicates from candidate entities The score of the optimal path that can be obtained from the starting point to the relevant facts. This is the balance coefficient;

[0106] Next, subgraph retrieval uses vector retrieval (vector database) to prioritize the location of highly semantically relevant entities, and combines graph indexing to perform width-limited path expansion to generate candidate subgraphs. The candidate subgraphs are further filtered under time constraints and confidence thresholds and then passed to the hybrid inference module.

[0107] S5, Graph-enhanced Hybrid Reasoning (Vector Reasoning and Symbol Verification): First, path scores are calculated on candidate subgraphs, and node representations are obtained through parallel GNN propagation. Then, vectorized scores and path scores are combined, and logical verification is performed using a symbol rule base to generate candidate answers with interpretable evidence paths and quantified confidence, thereby improving reasoning accuracy.

[0108] The specific steps of S5 are as follows:

[0109] First, in the candidate subgraph Run a hybrid reasoning process to produce answers with interpretable evidence paths and quantified confidence levels; for any path The path score is defined as the product of edge weights with a length penalty:

[0110]

[0111] in, For the first Effective weight of the edge, This is a path length penalty function. This is the path length decay factor; the path confidence can be approximately obtained by multiplying the edge confidences, or a more conservative evidence fusion model (such as Dempster–Shafer or Bayesian network) can be used to handle the correlation between edges.

[0112] In parallel, in the subgraph The upper-level graph neural network (GNN) propagates to obtain the upper and lower representations of each node. The propagation of one layer of a GNN can be described as follows:

[0113]

[0114] in, For adjacency matrices with self-loops, for The degree matrix, For the first The node feature matrix of the layer, For learnable weight matrix, For activation functions;

[0115] The final candidate answer's overall score is a weighted combination of vectorized score and path score, further processed by a symbolic rule base. Perform a domain logic and time consistency check; those that do not meet the symbol rules will be marked or removed. An example of the comprehensive scoring format is as follows:

[0116]

[0117] in, As a weighting balancing factor, For vector similarity scoring based on GNN representation, For The set of candidate paths for the endpoints; the output of this step is the sorted candidate answers, the evidence paths corresponding to each answer, and the combined confidence score.

[0118] S6, Controlled writing generation based on knowledge graph constraints: Taking the evidence set obtained through reasoning and writing instructions (including writing templates, style constraints and time range) as input, a controlled decoding strategy is used to generate fluent and factually consistent text;

[0119] The specific steps of S6 are as follows:

[0120] First, the basic language model provides the input conditions. With historical generation At that time, for the next token The original probability is ;

[0121] Next, to guide the generation of decoded distributions in a direction consistent with KG evidence, fact scores are introduced. Consistency penalty item Construct a controlled decoding distribution:

[0122]

[0123] in, Hyperparameters are used to balance smoothness and factual constraints; For token A score indicating how well the information matches the facts or entity of KG. Penalty scores for conflicting evidence with KG or violating symbol rules;

[0124] Subsequently, the writing training uses joint loss to optimize language fluency and factual consistency:

[0125]

[0126] in, The negative log-likelihood of the language model. For fact consistency loss, based on KG coverage or cross-entropy of fact-based questions and answers, These are the weighting coefficients;

[0127] Finally, the generated text is output along with an evidence citation index for each assertion to ensure traceability.

[0128] S7, Fact Verification, Human-Machine Confirmation and Closed-Loop Knowledge Graph Update: First, the assertions in the text are automatically verified and their support is calculated. Then, assertions with low support are prompted for human review. At the same time, the human confirmation information is written back to the KG with high confidence. Then, an immutable audit log is recorded.

[0129] The specific steps of S7 are as follows:

[0130] First, the assertions in the generated text undergo secondary automatic validation, followed by manual verification, to complete the incremental update of the KG; the automatic validation module verifies each assertion in the text. Perform retrieval validation and calculate its support. For example, support can be defined as the coverage of high-confidence evidence for the assertion in the KG:

[0131]

[0132] in, To support the assertion The set of candidate evidence, As evidence confidence level For the preset threshold, For indicator functions;

[0133] Next, assertions with low support are labeled and highlighted in the user interface to prompt users for manual review; simultaneously, the user's adoption or modification is recorded as a manual confirmation signal and written back to the KG according to the configured policy. When writing back, the manual confirmation evidence can be assigned a higher initial confidence level. ;

[0134] Subsequently, all write-back operations are stored as audit logs in an immutable log, with log entries including change time, operator ID, old fact version, modified content, and corresponding source index. This closed-loop mechanism ensures that human-machine collaborative judgment can improve the accuracy and reliability of the KG in the long term.

[0135] To achieve end-to-end consistency, each step relies on a unified metadata model and a shared index. Let any fact quintuple... Meta-information vectors are always present in the system. During the data flow process, the retrieval, reasoning, and generation modules all reference and update the data. The Version and Timestamp fields in the data; confidence level. The updates are performed sequentially during multi-source fusion, inference propagation, and manual verification. These updates can be formalized as follows:

[0136]

[0137] in, The evidence fusion function can be defined as a weighted average, a Dempster-Shafer combination, or a Bayesian update rule. For the newly added source confidence set, This is a manual adoption marker. The above design ensures that the system can reference a consistent factual representation at any stage and that its credibility can be updated traceably.

[0138] Based on the above, the present invention includes, but is not limited to, the following embodiments:

[0139] Example 1:

[0140] The USPTO public text dataset (Patent Grant / Application text collection, publicly available full-text text) is used as the primary knowledge source and evidence base. DocVQA (Document Image Question Answering dataset) is used for training / validation of information extraction from images and charts (such as text / annotation recognition in attached images). WikiTableQuestions / TableQA datasets are used to train table structure extraction and table question answering capabilities (adapting to table information in instruction manuals). SQuAD / Natural Questions serves as fine-tuning corpus for question answering, improving natural language understanding and short text retrieval capabilities.

[0141] The specific steps for implementing this invention are as follows:

[0142] (1) Multi-source data acquisition and standardization

[0143] Retrieve USPTO XML / text files and import them into the database; import PDF attachments (drawings, tables) into document storage; annotate source metadata (document ID, page number, file time, original checksum). Perform format recognition on input files: parse XML fields, use OCR on PDFs (adapted to DocVQA models and rules), and run a header recognition process on tables to standardize cell semantics. Record the source index of each standardized cell for later backtracking.

[0144] (2) Multimodal information extraction and uncertainty labeling

[0145] Text: Extract main entities (inventory elements, components, process steps), relationships (e.g., "contains," "used for," "generated"), and attributes using a corpus-fine-tuned NER and relation extraction model. Tables: Map table cells to attribute declarations (e.g., performance parameters, test results) and standardize units and semantic column names. Images / Attachments: Extract annotations and notes from images using a DocVQA-style model, identifying component numbers and associated text. Assign an initial confidence level to each fact (higher confidence levels for official USPTO text with higher source weights, and relatively lower confidence levels for text extracted by OCR), and record timestamps and document version information.

[0146] (3) Construction of time-series and versionable knowledge graphs (T-KG)

[0147] For each candidate document pair, deep alignment is performed only at the anchor point and its neighborhood: vector similarity is compared between paragraphs and similar segments are recorded; visual embedding matching is performed on figures and the coverage rate of figure labels is combined; cell-level alignment of column names and key value columns is performed on tables. The alignment output is an "evidence unit", such as "similarity between paragraph 3 of document A and paragraph 5 of document B is 0.87; figure A and figure B figure matching rate is 72%; table column x matching rate is 83%".

[0148] (4) Semantic parsing and subgraph retrieval

[0149] Users submit natural language queries (e.g., "What are the key steps to implement function B using technology A?"). The system encodes the query into a vector and retrieves relevant entities / facts from the vector library. It then performs width-constrained path expansion using a graph index to generate candidate subgraphs. Simultaneously, it considers time constraints (prioritizing recently published evidence) and confidence thresholds to filter evidence.

[0150] (5) Graph-enhanced hybrid reasoning (vector reasoning and symbol verification)

[0151] A graph neural network is run in parallel on candidate subgraphs to obtain contextualized vector representations and to compute path scores based on evidence chains to generate interpretable evidence paths. Conflict detection is performed on multiple candidate conclusions, and a hybrid uncertainty fusion strategy is employed to handle dependencies between sources.

[0152] (6) Controlled writing generation based on KG constraints

[0153] Given writing instructions, the language model's decoding distribution is constrained based on high-confidence facts provided by the KG (Knowledge-Gathering) algorithm to ensure consistency between key terms and evidence. The generated results also include an evidence index (page number, paragraph) for each assertion. A constraint search mechanism is used to avoid expressions that conflict with the KG algorithm while maintaining linguistic fluency.

[0154] (7) Fact verification, human-machine confirmation and closed-loop update

[0155] The system automatically performs a support retrieval for each assertion in the generated draft, highlighting assertions with low support in the interface and requiring reviewers to confirm or modify them. Manual confirmation after review is written back to the KG with a high-confidence status and recorded in an immutable audit log (time, operator, previous version, and modification details). Sensitive business information is differentiated or subject to access control according to policy.

[0156] (8) Output and Delivery

[0157] Provides drafters with: a draft with evidence, a clickable evidence path for each assertion, review suggestions and uncertainty tips, and supports exporting to Word / PDF and audit reports.

[0158] The following baseline was designed for experimental testing. The results of this approach are compared with those of several model baselines as follows:

[0159] Comparison Methods Answer F1 Fact Precision Evidence Coverage Hallucination Rate Method of the present invention 0.93 0.95 0.90 0.05 Baseline A (LM-only) 0.78 0.70 0.20 0.25 Baseline B (LM + Static Knowledge Base) 0.85 0.85 0.65 0.15 Baseline C (LM + Dynamic Knowledge Base) 0.88 0.88 0.75 0.12

[0160] Example 2:

[0161] The full-text literature (including research papers of various genres) from the publicly available PubMed Central Open Access (PMC OA) dataset is used as the main knowledge base and evidence source. DocVQA and ChartQA (or chart data from PlotQA / ICDAR) are used for training on text / numerical extraction from charts and scientific images. WikiTableQuestions / TableQA are used to improve table question answering and extraction capabilities. The specific steps to implement this invention are as follows:

[0162] (1) Multi-source data acquisition and standardization

[0163] The system retrieves XML and PDF documents from PMC OA, uniformly parsing them to extract titles, methods, results, tables, figures, and attachments, recording source information (DOI, publication date, journal). It performs structured parsing of tables and figures within the papers: semanticizing table headers, standardizing numerical units, extracting curves / annotations from figures using OCR, and aligning them with legends. All standardized units are written to the indexing service while retaining their positional information within the document.

[0164] (2) Multimodal information extraction and uncertainty labeling

[0165] Text extraction: Use domain-fine-tuned NER / RE to extract experimental subjects, parameters, results, and conclusive assertions (e.g., "Method X improves upon control by Z% under condition Y"). Chart extraction: Use ChartQA / PlotQA-style models to extract curve peaks, trends, and annotations from charts, normalize them to numerical facts, and assign confidence levels (image OCR confidence levels are typically lower than text confidence levels). Table extraction: Map key columns to attributes and extract statistical indicators (mean, standard deviation, significance markers), recording their source and confidence level.

[0166] (3) Construction of time-series and versionable knowledge graphs (T-KG)

[0167] Facts extracted from the papers are inserted into the T-KG: facts include a time interval (publication date or data collection date) and confidence level (e.g., results from peer-reviewed journals may be assigned higher initial weights). Entity alignment is performed on similar research results (i.e., naming variants of the same method in different papers are mapped to the same entity).

[0168] (4) Semantic parsing and subgraph retrieval

[0169] When a user submits a task (e.g., "Please write a technical review of method X in field Y based on existing literature" or "Give the key advantages and disadvantages of method X and supporting evidence"), the system will query the code and retrieve relevant entities, and generate candidate subgraphs (containing factual fragments and graphical evidence from multiple papers) by combining time / confidence constraints. It supports filtering by time window (retrieving only studies from the last 5 years) or by evidence type (prioritizing numerical values ​​in graphs or experimental tables).

[0170] (5) Graph-enhanced hybrid reasoning

[0171] Running GNNs on candidate subgraphs captures inter-document correlations (e.g., method citation networks, experimental condition similarities) and provides interpretable support chains with path evidence (e.g., "Paper A provides preliminary evidence → Paper B replicates the verification → Paper C proposes improvements"). Hybrid uncertainty fusion methods are employed to handle cross-paper dependencies (e.g., reducing overall confidence gain when multiple papers come from the same research group).

[0172] (6) Controlled writing generation based on KG constraints

[0173] Given writing instructions (technical overview, project report paragraph, experimental comparison), the system decodes them using high-confidence fact constraints in the Knowledge Base (KG); the generated text includes evidence citations (DOI, page number, figure / table number) for each conclusion paragraph. If the user requires counterfactual analysis, the system constructs a counterfactual query in the KG and returns the corresponding evidence path and visualization suggestions.

[0174] (7) Fact verification, human-machine confirmation and closed-loop update

[0175] Automatically validate the support of generated text assertions; highlight assertions with low support and provide candidate evidence for quick expert review; modifications adopted by experts will be recorded and written back to KG, and the identity of the adopter and the time will be written to the audit log; differentiate or restrict the export of data containing sensitive data (such as unpublished experimental data) according to the strategy.

[0176] (8) Output and Delivery

[0177] Provide researchers with: review paragraphs with evidence, original text excerpts and figure screenshots for each assertion, counterfactual analysis reports, and exportable audit reports. Design the following baseline for experimental testing:

[0178] The following baseline was designed for experimental testing. The results of this approach are compared with those of several model baselines as follows:

[0179] Comparison Methods Answer F1 Fact Precision Evidence Coverage Hallucination Rate Method of the present invention 0.94 0.96 0.92 0.04 Baseline A (LM-only) 0.76 0.72 0.25 0.28 Baseline B (LM + Static Knowledge Base) 0.86 0.86 0.68 0.14 Baseline C (LM + Dynamic Knowledge Base) 0.89 0.89 0.78 0.10

[0180] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A knowledge graph-enhanced intelligent question-answering and writing assistance engine for innovation management, characterized by: The specific steps are as follows: S1, Multi-source data acquisition and standardization: First, collect internal and external multimodal raw materials and process them into standardized units. Then, retain the retrospective metadata and isolate sensitive information simultaneously to lay the foundation for subsequent extraction and evidence tracing. S2, Multimodal Information Extraction and Uncertainty Labeling: First, the standardized units are converted into five-tuple structured facts containing time intervals and confidence levels. Then, multimodal information is extracted, confidence levels are calculated, and modal embeddings are generated and aligned through contrastive learning to provide high-quality input for the construction of time-series knowledge graphs. S3, Construction of Time-Sequential and Versionable Knowledge Graph: First, structured facts are inserted into a knowledge graph containing time, version, and uncertainty. Then, the weights of the facts are adjusted using a time decay function. Subsequently, the training entity relationship vectors are connected, and historical versions and audit trajectories are saved to achieve dynamic management and traceability of facts. S4, Semantic parsing and subgraph retrieval: Maps the user's natural language query or writing instructions to retrieval constraints and pulls candidate subgraphs related to the question from the knowledge graph and vector index; S5, Graph-enhanced Hybrid Reasoning: First, path scores are calculated on candidate subgraphs, and node representations are obtained through parallel GNN propagation. Then, vectorized scores and path scores are combined, and logical verification is performed using a symbolic rule base to generate candidate answers with interpretable evidence paths and quantified confidence, thereby improving reasoning accuracy. S6, Controlled writing generation based on knowledge graph constraints: Taking the evidence set obtained through reasoning and writing instructions as input, a controlled decoding strategy is used to generate fluent and factually consistent text; S7, Fact Verification, Human-Machine Confirmation and Closed-Loop Knowledge Graph Update: First, the assertions in the text are automatically verified and their support is calculated. Then, assertions with low support are prompted for human review. At the same time, the human confirmation information is written back to the KG with high confidence. Then, an immutable audit log is recorded.

2. The knowledge graph-enhanced intelligent question-answering and writing assistance engine for innovation management according to claim 1, characterized in that, The specific steps of S1 are as follows: First, raw materials from internal and external sources are collected over a long period of time or in real time, and various unstructured and semi-structured inputs are converted into unified standardized document units, while a traceable evidence location index is retained on each standardized unit. Next, sensitive information is detected and isolated according to pre-configured policies to meet compliance requirements.

3. The knowledge graph-enhanced intelligent question-answering and writing assistance engine for innovation management according to claim 1, characterized in that, The specific steps of S2 are as follows: First, standardized document units are transcribed into structured fact triples, and each fact is assigned a confidence level and a time applicability range to facilitate the subsequent construction of a temporal knowledge graph. The extraction results are expressed in the form of quintuples: (1) in, The main entity representing the fact triple; Indicates the relation type; Represents an object entity or attribute value; Indicate the time interval to which the facts apply, and give the start time. With end time ; Indicates the degree of confidence in the fact; Next, the text extraction adopts a domain-fine-tuned named entity recognition and relation extraction model, the table extraction generates attribute declarations based on column semantic mapping, and the chart is converted into numerical facts and normalized units through OCR and coordinate conversion. Then, increase the confidence level. The initial calculation can be obtained from the model output probability. With rules / source weight The fusion is given, using simple weighted fusion: (2) in, This is a range mapping function; For weight hyperparameters, Indicates the initial credibility score of the source; Simultaneously generate vector representations for text segments, table cells, and image regions. , , Furthermore, through contrastive learning, the representations of the same entity or fact are brought closer together across different modalities. The contrastive loss used can be InfoNCE. (3) in, and These are encoders for different modes. For similarity function, , As a positive sample, Negative sample set.

4. The knowledge graph-enhanced intelligent question-answering and writing assistance engine for innovation management according to claim 1, characterized in that, The specific steps of S3 are as follows: First, the extracted structured facts are inserted into a knowledge graph that supports time intervals, version control, and uncertainty annotation; the knowledge graph is a directed weighted graph. This means that each side of the diagram corresponds to a fact quintuple and carries confidence level and time information; Next, a time decay function is introduced. Adjust the weights of the facts: (4) in, Indicates the effective weight of the edge. The original confidence level of the facts. This represents the distance between the current query time and the actual time. The attenuation coefficient; Subsequently, the vector representation of the entities in the diagram With relation vector A connection training strategy is employed, where one of the structural training objectives is link prediction based on an energy function, using the TransE scoring function: (5) in, , These are the vector representations of the head entity and the tail entity, respectively. The relationship vector is used; during training, the margin-loss is used to minimize the difference between negative and positive samples. (6) in, For the positive sample set, For the constructed negative sample set, This is the margin hyperparameter.

5. The knowledge graph-enhanced intelligent question-answering and writing assistance engine for innovation management according to claim 1, characterized in that, The specific steps of S4 are as follows: First, the user inputs Encoded as semantic vectors ; The retrieved score then combines semantic similarity with path-based structural score to form a hybrid scoring function: (7) in, Represents vector similarity, For entities in the diagram The vector representation of , Indicates from candidate entities The score of the optimal path that can be obtained from the starting point to the relevant facts. This is the balance coefficient; Next, subgraph retrieval uses vector retrieval to prioritize the location of highly semantically relevant entities, and combines graph indexing to perform width-constrained path expansion to generate candidate subgraphs. The candidate subgraphs are further filtered under time constraints and confidence thresholds and then passed to the hybrid inference module.

6. The knowledge graph-enhanced intelligent question-answering and writing assistance engine for innovation management according to claim 1, characterized in that, The specific steps of S5 are as follows: First, in the candidate subgraph Run a hybrid reasoning process to produce answers with interpretable evidence paths and quantified confidence levels; for any path The path score is defined as the product of edge weights with a length penalty: (8) in, For the first Effective weight of the edge, This is a path length penalty function. This is the path length attenuation factor; In parallel, in the subgraph The upper execution graph neural network propagates to obtain the upper and lower representations of each node. The propagation of one layer of a GNN can be described as follows: (9) in, For adjacency matrices with self-loops, for The degree matrix, For the first The node feature matrix of the layer, For learnable weight matrix, For activation functions; The final candidate answer's overall score is a weighted combination of vectorized score and path score, further processed by a symbolic rule base. Perform a domain logic and time consistency check; entries that do not meet the symbol rules will be marked or removed. The overall scoring format is as follows: (10) in, As a weighting balancing factor, For vector similarity scoring based on GNN representation, For The set of candidate paths for endpoints.

7. The knowledge graph-enhanced intelligent question-answering and writing assistance engine for innovation management according to claim 1, characterized in that, The specific steps of S6 are as follows: First, the basic language model provides the input conditions. With historical generation At that time, for the next token The original probability is ; Next, fact scores are introduced. Consistency penalty item Construct a controlled decoding distribution: (11) in, Hyperparameters are used to balance fluency and factual constraints; For token A score indicating how well the information matches the facts or entity of KG. Penalty scores for conflicting evidence with KG or violating symbol rules; Subsequently, the writing training uses joint loss to optimize language fluency and factual consistency: (12) in, The negative log-likelihood of the language model. For fact consistency loss, based on KG coverage or cross-entropy of fact-based questions and answers, These are the weighting coefficients; Finally, the generated text is output along with an evidence citation index for each assertion to ensure traceability.

8. The knowledge graph-enhanced intelligent question-answering and writing assistance engine for innovation management according to claim 1, characterized in that, The specific steps of S7 are as follows: First, the assertions in the generated text undergo secondary automatic validation, followed by manual verification, to complete the incremental update of the KG; the automatic validation module verifies each assertion in the text. Perform retrieval validation and calculate its support. The support level is the coverage of high-confidence evidence for this assertion in KG: (13) in, To support the assertion The set of candidate evidence, As evidence confidence level For the preset threshold, For indicator functions; Next, assertions with low support are labeled and highlighted in the user interface to prompt users for manual review; simultaneously, the user's adoption or modification is recorded as a manual confirmation signal and written back to the KG according to the configured policy. When writing back, the manual confirmation evidence can be assigned a higher initial confidence level. ; Subsequently, all write-back operations are saved as audit logs in an immutable log.

Images