Dual-mode intelligent duplicate checking and similarity quantitative evaluation system for science and technology project management
By constructing a multimodal preprocessing and project semantic graph, combined with anchoring mechanism and local alignment, efficient similarity determination of text, image and tabular data is achieved, which solves the shortcomings of multimodal fusion in existing technologies and improves the accuracy and interpretability of science and technology project management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FOSHAN POWER SUPPLY BUREAU GUANGDONG POWER GRID
- Filing Date
- 2025-12-30
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies are difficult to effectively combine multimodal data such as text, images, and tables for similarity determination in science and technology project management, leading to missed or incorrect judgments. Furthermore, they lack interpretability and traceability, failing to meet the high requirements of science and technology projects.
A multimodal preprocessing and feature extraction module is adopted to establish a project semantic map and anchor point mechanism. Through local cross-modal alignment and uncertainty perception adaptive fusion, traceable evidence is generated and scenario-based risk assessment is carried out. Online learning is combined with human-machine closed loop.
It improves the accuracy and robustness of multimodal similarity determination, reduces determination variance, enhances interpretability and computational efficiency, and supports traceable auditing and online adaptive optimization.
Smart Images
Figure CN122045836A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of document similarity detection, knowledge management, and project management information technology, specifically a dual-modal intelligent plagiarism detection and similarity quantification evaluation system for science and technology project management. Background Technology
[0002] Science and technology project management is a systematic project, encompassing the entire process from project initiation, planning, resource allocation, schedule and quality control, deliverables acceptance, and intellectual property management. Unlike general engineering projects, science and technology projects are highly uncertain and knowledge-intensive, with deliverables exhibiting multimodal characteristics, including technical solution texts, experimental data, design diagrams, and software code. Furthermore, they place higher demands on originality, compliance, and traceability. Therefore, establishing a rigorous material review and similarity assessment mechanism in science and technology project management not only helps ensure research integrity and clarify the ownership of deliverables but also optimizes resource allocation, reduces redundant investment, and provides reliable evidence for decision-making. In the process of science and technology project management, to ensure research integrity, avoid duplicate project initiation, and clarify the ownership of deliverables, it is typically necessary to conduct similarity checks and comparisons on project application materials, technical solutions, design diagrams, and experimental data tables.
[0003] In the field of science and technology project management, document similarity detection has long been text-centric. Early plagiarism detection relied on string matching and fingerprinting techniques, such as n-gram or shingling-based comparisons, which could detect word-for-word plagiarism and partial repetitions. With the development of natural language processing and deep learning, research has gradually shifted to the semantic level, using sentence vectors and paragraph vectors for semantic comparison, thus achieving significant progress in paraphrasing and paraphrasing detection. However, science and technology project materials heavily rely on flowcharts, system architecture diagrams, tabular data, and hierarchical work breakdown structures to express technical routes and task arrangements. Purely textual methods struggle to capture the structured semantics carried by these non-textual modalities, thus easily leading to missed or false positives when diagrams are redrawn, processes are rearranged, or tabular structures are reused.
[0004] Incorporating images and diagrams into similarity analysis has been practiced in patent retrieval and engineering drawing recognition, with common strategies including primitive detection, visual feature extraction, and image-text fusion. Regarding structural similarity, tree edit distance and graph edit distance provide a theoretical basis for measuring process or WBS transformation; however, accurate calculation is costly in large-scale or online scenarios, and their semantic robustness to node naming differences and terminology substitutions is insufficient. In recent years, multimodal pre-trained models have provided new tools for cross-modal alignment, but the domain transfer problem and the specific topological semantics of primitives in engineering drawings make it difficult for general models to be directly adapted. Furthermore, multimodal fusion often relies on splicing or fixed weighting, lacking a mechanism for dynamically adjusting weights based on the credibility of evidence.
[0005] At the industry practice level, although several general technical frameworks for multimodal processing exist, they are mostly geared towards general tasks and lack refined designs for science and technology project management, particularly in terms of interpretability and auditability. Most systems only output a single similarity score or partial annotation, failing to provide a complete cross-modal evidence chain and node-level alignment results. They also struggle to map the judgment results to scenario-based handling strategies in management stages such as project initiation, review, and acceptance. Based on these shortcomings, there is an urgent need for a comprehensive solution that combines textual semantics, image vision, and structured parsing; supports alignment of project semantic graphs with semantically meaningful tree graphs; introduces learnable adaptive fusion; and can generate traceable evidence and scenario-based risk indicators to improve the accuracy, interpretability, and engineering applicability of similarity judgments.
[0006] In summary, existing technologies still have significant shortcomings in multimodal similarity determination for science and technology project management. Text-based methods struggle to cover the semantics of diagrams and tables; image-text fusion technology lacks task-level structural alignment capabilities; structural distance methods still have limitations in computational scale and semantic robustness; multimodal fusion mechanisms lack adaptability and interpretability; and the review system lacks a traceable chain of evidence and scenario-based risk indicators. Therefore, this invention proposes a bimodal intelligent deduplication and similarity quantification evaluation system for science and technology project management. Summary of the Invention
[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A dual-modal intelligent plagiarism detection and similarity quantification evaluation system for science and technology project management, comprising: The multimodal preprocessing and feature extraction module is used to preprocess and extract features from data in three modalities: text, graphics, and tables; then, it simultaneously associates source metadata. The project semantic graph and anchor mechanism module is used to first map entities in text, diagrams, and tables to a unified project semantic graph to form a set of nodes; then, key nodes are selected as anchors by node centrality and embedding cluster density to build an inverted index; at the same time, deep local matching is carried out in the alignment stage. The local cross-modal alignment module is used to perform local alignment of text paragraphs, illustration subgraphs, and table fragments centered on anchor points: text similarity is calculated using the cosine similarity of segment vectors; illustration similarity is calculated by combining visual embedding cosine similarity with primitive label matching ratio, while structural channels use a learned model to calculate weighted edit distance; tables are calculated using corresponding methods according to cell type. The uncertainty-aware adaptive fusion module is used to first input modal similarity and confidence into two layers of MLP to generate dynamic weight vectors, and then obtain the overall similarity by weighted summation and Sigmoid normalization; The traceable evidence generation and scenario-based risk index module is used to first generate structured evidence records in local alignment, then calculate the project similarity risk index, and map it to management strategies. The human-machine closed-loop and online learning module is used to apply the results of manual review to fine-tune the learning-based structural editing cost model and fusion weights, and to generate audit reports and training samples regularly.
[0008] As a preferred embodiment of the dual-modal intelligent plagiarism detection and similarity quantification evaluation system for science and technology project management described in this invention, the specific steps of the multimodal preprocessing and feature extraction module are as follows: First, preprocessing and feature extraction are performed on the data in three modalities: text, graphs, and tables. After character encoding standardization, redundant symbol removal, and format normalization, the text data is segmented. Semantic information at the paragraph level is extracted from each segment, and named entities are identified and subject-action-object structures are extracted. Each text segment is then represented by a segment vector embedding using a domain-adjusted BERT dataset. (1) in, This is the semantic vector of the text paragraph. It is the vector dimension; The data is divided into visual and structural channels. The visual channel extracts text information using OCR, and then uses a ResNet50 convolutional network to generate image embeddings. (2) in, Represents the image embedding vector. It is the visual feature dimension; the structural channel extracts nodes, connections, and labels through primitive detection to construct local subgraphs. The node features include type, text label, and topological information, and node embeddings are generated through two layers of GCN: (3) in, Represents the node feature matrix, This represents adding a self-loop adjacency matrix. Represents the node degree matrix, Represents the weight matrix. Represents graph nodes. Indicates node embedding; The table data identifies header and column types, standardizes units and numerical ranges, maps cells to vector embeddings, and generates table embeddings through row and column aggregation. ; Next, all modal features are simultaneously accompanied by document ID, author, timestamp, and node ID for evidence tracing.
[0009] As a preferred embodiment of the dual-modal intelligent plagiarism detection and similarity quantification evaluation system for science and technology project management described in this invention, the specific steps of the project semantic graph and anchor point mechanism module are as follows: First, based on the multimodal feature extraction results, entities in text, diagrams, and tables are mapped to the project semantic graph, forming a set of nodes containing tasks, milestones, modules, parameters, and deliverables, with edges representing task dependencies or parameter associations. Next, an anchor point mechanism is introduced, with the anchor point being the core node; the anchor point is selected based on node centrality and embedding cluster density. (4) in, Indicates node centrality. This indicates the embedding cluster density; after selecting the anchor point, an inverted index is built to achieve fast location of local subgraphs and document fragments; during the alignment stage, deep local matching is performed with the anchor point as the center.
[0010] As a preferred embodiment of the dual-modal intelligent plagiarism detection and similarity quantification evaluation system for science and technology project management described in this invention, the specific steps of the local cross-modal alignment module are as follows: At each anchor point, perform local alignment of text paragraphs, subplots, and table fragments: Text similarity is calculated using the cosine of segment vectors: (5) in, and Indicates a text paragraph; The visual channel diagram illustrates image similarity calculation based on embedding cosine similarity, combined with primitive label matching ratio: (6) in, Represents an image. Represents a set of primitive labels. Indicates visual channel weights; The structural channel uses a learned structural edit cost model to calculate the weighted edit distance, where the node replacement cost, node insertion cost, and node deletion cost are respectively: (7) (8) in, Indicates the editing cost weight. Indicates the similarity of knowledge graphs. Represents node confidence; normalized structural similarity is calculated as follows: (9) in, Represents a subgraph. Indicates the edit distance of the subgraph; Table similarity is calculated using normalized distance for numeric cells, vector cosine for text cells, and tree-structured alignment for formula cells, ultimately generated by weighted average. .
[0011] As a preferred embodiment of the dual-modal intelligent plagiarism detection and similarity quantification evaluation system for science and technology project management described in this invention, the specific steps of the uncertainty-aware adaptive fusion module are as follows: Calculate the similarity of each modality. With confidence level The confidence level is calculated based on anchor coverage, matching segment length, and time difference; and a modal weight vector is generated through a two-layer MLP. (10) (11) in, Represents the weight matrix. This represents the bias vector. Represents the similarity vector and confidence vector of modalities. Indicates the output of the hidden layer. Represents the modal weight vector; The final overall similarity is calculated by weighted summation and normalized using Sigmoid: (12) in, These represent the local similarity and total similarity of each modality, respectively.
[0012] As a preferred embodiment of the dual-modal intelligent plagiarism detection and similarity quantification evaluation system for science and technology project management described in this invention, the specific steps of the traceable evidence generation and scenario-based risk index module are as follows: First, evidence records are generated during the local alignment process: (13) in, Indicates the document number. Align text fragments, Indicates the anchor point number. Represents modal similarity, Indicates modal confidence. Indicates time difference, Indicates the document author; chained storage facilitates auditing. Next, the Project Similarity Risk Index (PSRI) is defined as follows: (14) in, Mapping time window effect Indicates the score for similar coverage area. Assign weights to structure reuse, visual redrawing, and table reuse. Configurable weights.
[0013] As a preferred embodiment of the dual-modal intelligent plagiarism detection and similarity quantification evaluation system for science and technology project management described in this invention, the specific steps of the human-machine closed-loop and online learning module are as follows: The results of manual review are fed back as a supervisory signal to fine-tune the learning-based structural editing cost model and fusion weights; The loss function for training the edit cost model is: (15) The fusion weight network is fine-tuned using regression loss; audit reports and training samples are generated periodically.
[0014] Compared with existing technologies: This invention offers the following advantages: Reduced decision variance and improved robustness: By treating each modality as an independent estimate of "true similarity," weighted fusion based on uncertainty awareness reduces the impact of high-variance (unreliable) estimates and increases the weight of low-variance (reliable) estimates, statistically reducing overall decision variance and thus improving the system's robustness under different noise scenarios. Increased discriminative dimensions and improved discriminative boundaries: Textual semantics, visual appearance, structural topology, and table values provide different discriminative cues, which are orthogonal or complementary dimensions in the feature space. This allows for better differentiation between samples that are "superficially similar but semantically different" and samples that are "superficially different but functionally equivalent," thereby reducing the false positive rate and improving the recall rate. Localized processing reduces error propagation and computational burden: Local alignment centered on anchor points theoretically reduces error propagation paths (local differences do not spread into global errors), while decomposing the complex global matching problem into multiple sub-problems that can be processed in parallel, facilitating distributed computing and horizontal scaling. The learnable cost function is closer to human judgment: by adjusting the cost term of structural comparison through supervised learning, the similarity measure of the model can be made to continuously approach the judgment of experts. Theoretically, it can converge to a similarity measure function that is closer to human judgment, thus making up for the limitations of heuristic rules.
[0015] In summary, this invention achieves complementary fusion of text, illustration, structure, and table information through the construction of a multimodal semantic graph and an anchor-driven local matching mechanism, significantly improving the accuracy and robustness of similarity determination. Through a trainable structural comparison and confidence weighting mechanism, the determination results are closer to human experience and have interpretability. At the same time, it outperforms existing technologies in terms of computational efficiency, scalability, and traceability. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the overall framework of the present invention. Detailed Implementation
[0017] 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 with reference to the accompanying drawings.
[0018] This invention provides a dual-modal intelligent plagiarism detection and similarity quantification evaluation system for science and technology project management. Please refer to [link / reference]. Figure 1 ,include: The multimodal preprocessing and feature extraction module is used to preprocess and extract features from three modalities: text, image, and table. For text, semantic information is extracted by segmentation and named entities are identified. Then, BERT with domain fine-tuning is used to generate segment vector embeddings. For images, visual and structural channels are processed separately. For tables, embeddings are generated by row and column aggregation. Then, source metadata is associated. The specific steps of the multimodal preprocessing and feature extraction module are as follows: After character encoding standardization, redundant symbol removal, and format normalization, the text data is segmented. Semantic information at the paragraph level is extracted from each segment, and named entities are identified, along with the subject-action-object structure, covering tasks, milestones, parameters, and deliverables. Each text segment is then represented by a domain-fine-tuned BERT vector embedding. (1) in, This is the semantic vector of the text paragraph. It is the vector dimension; The data is divided into visual and structural channels. The visual channel extracts text information using OCR, and then uses a ResNet50 convolutional network to generate image embeddings. (2) in, Represents the image embedding vector. It is the visual feature dimension; the structural channel extracts nodes, connections, and labels through primitive detection to construct local subgraphs. The node features include type, text label, and topological information, and node embeddings are generated through two layers of GCN: (3) in, Represents the node feature matrix, This represents adding a self-loop adjacency matrix. Represents the node degree matrix, Represents the weight matrix. Represents graph nodes. Indicates node embedding; The table data identifies header and column types, standardizes units and numerical ranges, maps cells to vector embeddings, and generates table embeddings through row and column aggregation. ; All modal features are accompanied by document ID, author, timestamp, and node ID for evidence tracing.
[0019] The above processing provides a foundation for building a multimodal project semantic graph, ensuring that the semantic and structural information between different modalities can be aligned.
[0020] The project semantic graph and anchor mechanism module is used to first map entities in text, illustrations, and tables to a unified project semantic graph based on multimodal features, forming a set of nodes; then, key nodes are selected as anchors through node centrality and embedding cluster density to build an inverted index; at the same time, deep local matching is carried out during the alignment stage. The specific steps of the project semantic graph and anchor point mechanism module are as follows: Based on the multimodal feature extraction results, entities in text, diagrams and tables are mapped to the project semantic graph, forming a set of nodes containing tasks, milestones, modules, parameters and deliverables, with edges representing task dependencies or parameter associations. To efficiently handle large-scale project repositories, an anchor mechanism is introduced. Anchors are core nodes, such as key parameters, milestones, or module names. Anchors are selected based on node centrality and embedding cluster density. (4) in, Indicates node centrality. It represents the embedding cluster density; after selecting the anchor point, an inverted index is built to achieve fast location of local subgraphs and document fragments; in the alignment stage, deep local matching is performed with the anchor point as the center, which reduces the computational scale, allows for local naming differences and structural adjustments, ensures accurate capture of key entities, and provides a reliable foundation for cross-modal similarity calculation.
[0021] The local cross-modal alignment module is used to perform local alignment of text paragraphs, illustration subgraphs, and table fragments centered on anchor points: text similarity is calculated using the cosine similarity of segment vectors; illustration similarity is calculated by combining visual embedding cosine similarity with primitive label matching ratio, while structural channels use a learned model to calculate weighted edit distance; tables are calculated using corresponding methods according to cell type. The specific steps of the local cross-modal alignment module are as follows: At each anchor point, perform local alignment of text paragraphs, subplots, and table fragments: Text similarity is calculated using the cosine of segment vectors: (5) in, and Indicates a text paragraph; The visual channel diagram illustrates image similarity calculation based on embedding cosine similarity, combined with primitive label matching ratio: (6) in, Represents an image. Represents a set of primitive labels. Indicates visual channel weights; The structural channel uses a learned structural edit cost model to calculate the weighted edit distance, where the node replacement cost, node insertion cost, and node deletion cost are respectively: (7) (8) in, Indicates the editing cost weight. Indicates the similarity of knowledge graphs. Represents node confidence; normalized structural similarity is calculated as follows: (9) in, Represents a subgraph. Indicates the edit distance of the subgraph; Table similarity is calculated using normalized distance for numeric cells, vector cosine for text cells, and tree-structured alignment for formula cells, ultimately generated by weighted average. ; Local cross-modal alignment ensures that even with naming differences or local structural adjustments, it accurately reflects the correspondence between multimodal semantics and structures, providing reliable input for fusion computing.
[0022] The uncertainty-aware adaptive fusion module is used to first input modal similarity and confidence into two layers of MLP to generate dynamic weight vectors, and then obtain the overall similarity by weighted summation and Sigmoid normalization; The specific steps of the uncertainty-aware adaptive fusion module are as follows: Calculate the similarity of each modality. With confidence level The confidence level is calculated based on anchor coverage, matching segment length, and time difference; a modal weight vector is generated through a two-layer MLP. (10) (11) in, Represents the weight matrix. This represents the bias vector. Represents the similarity vector and confidence vector of modalities. Indicates the output of the hidden layer. This represents the modal weight vector; the final overall similarity is calculated by weighted summation and normalized using the Sigmoid function. (12) in, These represent the local similarity and total similarity of each modality, respectively. The fusion unit dynamically adjusts modality weights based on the quality of evidence, with high-confidence modalities contributing more to the overall similarity score, thus ensuring the accuracy and reliability of similarity determination.
[0023] The traceable evidence generation and scenario-based risk index module is used to first generate structured evidence records in local alignment, then calculate the project similarity risk index, and map it to management strategies. The specific steps of the traceable evidence generation and scenario-based risk index module are as follows: Evidence records are generated during local alignment: (13) in, Indicates the document number. Align text fragments, Indicates the anchor point number. Represents modal similarity, Indicates modal confidence. Indicates time difference, Indicates the document author; chained storage facilitates auditing. The Project Similarity Risk Index (PSRI) is defined as follows: (14) in, Mapping time window effect Indicates the score for similar coverage area. Assign weights to structure reuse, visual redrawing, and table reuse. Configurable weights; PSRI is used to trigger automatic approval, manual review, or rejection, enabling scenario-based and differentiated management decision-making.
[0024] The human-machine closed-loop and online learning module is used to apply the results of manual review to fine-tune the learning-based structural editing cost model and fusion weights, and to generate audit reports and training samples regularly. The specific steps of the human-machine closed-loop and online learning module are as follows: The results of manual review are fed back as a supervisory signal to fine-tune the learning-based structural editing cost model and fusion weights; The loss function for training the edit cost model is: (15) The fusion weight network uses regression loss for fine-tuning; at the same time, it regularly generates audit reports and training samples to achieve online optimization as the project library size and business changes, ensuring that similarity judgment is accurate, interpretable, and can be engineered.
[0025] Based on the implementation of the above technical solutions, high-precision similarity determination and risk assessment of multimodal project data can be achieved, ensuring accurate identification of key entities, effective alignment of cross-modal information, and supporting traceable auditing and online adaptive optimization, thereby improving the intelligence level and decision-making efficiency of project management.
[0026] Based on the above, the present invention includes, but is not limited to, the following embodiments: Example 1: We used the public dataset PMCOA (PubMed Central Open Access subset) in conjunction with PubLayNet. We randomly selected 12,000 papers (including text, figures, and tables) from PMCOA to construct sample pairs, and then constructed 10,000 pairs from these 12,000 papers for training / validation / testing. Specifically, the dataset was divided into a training set of 7,000 pairs (1,400 positive / 5,600 negative), a validation set of 1,500 pairs (300 positive / 1,200 negative), and a test set of 1,500 pairs (300 positive / 1,200 negative). The positive sample construction strategy was as follows: directly selecting figure + paragraph pairs reused by the same author group in different papers; or paraphrasing the original paragraphs and moving the figures; or rewriting the table values while preserving the semantics. Manual double annotation ensured label quality (two people were consistent; if inconsistent, a third party arbitrated). The distribution of chart types is as follows: flowcharts / schematic diagrams 32%, experimental setup diagrams 28%, and curves / results diagrams 40%.
[0027] The specific steps for implementing this invention are as follows: (1) Multimodal preprocessing and feature extraction Text: Unified encoding, sentence segmentation, and paragraph segmentation; for each paragraph, SciBERT is used for domain fine-tuning before extracting segment vectors and performing named entity recognition (extracting meta-information such as tasks, parameters, and deliverables). Visual representation: PubLayNet page segmentation results are used to locate figure regions; OCR is performed on each figure (extracting figure titles and annotations), and ResNet50 is used to extract visual vectors (and image type labels are saved). Structure representation: A primitive detector (detecting bounding boxes, arrows, and node labels) is run on flowcharts / schematic diagrams; the detection results are converted into local subgraphs, and text and topological attributes are extracted for each node. Tables: Table headers / column names are parsed, units of measurement are standardized, and key cells are vectorized (for subsequent alignment). All modal vectors are accompanied by document ID, figure ID, author, and timestamp for traceability.
[0028] (2) Project semantic graph and anchor point mechanism
[0029] Map text entities (tasks / parameters / milestones) to the same set of semantic graph nodes. Automatically evaluate the importance of each node (based on node frequency, paragraph / chart location, and semantic cluster density), and select Top-K anchors (K=8) to build an inverted index. Anchors are used for candidate retrieval and local alignment acceleration.
[0030] (3) Local cross-modal alignment
[0031] 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%".
[0032] (4) Uncertainty-aware adaptive fusion
[0033] For each piece of evidence, a confidence score is calculated (based on anchor coverage, matching fragment length, OCR confidence, time difference, etc.). The similarity score and confidence score are input into a two-layer MLP to obtain modal weights, and then weighted aggregation is performed to obtain the overall similarity. The output also includes an evidence list and a scenario-based risk index (PSRI). The PSRI comprehensively maps factors such as overall similarity, similarity coverage, time difference, and evidence type to "low / medium / high" risk labels.
[0034] (5) Traceable evidence generation and human-machine closed loop
[0035] The system stores the chain of evidence as structured records, facilitating auditing and manual review. The results of manual review are fed back to fine-tune the learning-based structural cost model and fusion parameters (incremental fine-tuning weekly).
[0036] Design the following baseline for experimental testing: Baseline 1: Text-only (SciBERT segment vector similarity) — Similarity is determined using only the full-text text vectors.
[0037] Baseline 2: Simple text + visual concatenation (concatenating text vectors and image vectors and then using MLP for judgment) — no local anchor strategy, no confidence adaptive.
[0038] Baseline 3: Traditional layout / keyword search (rule system based on key terms / title matching).
[0039] The results of this approach are compared with those of several model baselines as follows:
[0040] Example 2: We used the publicly available AI2D (Allen Institute Diagram Dataset) combined with DocBank / DocVQA (for accompanying text and tables). The original corpus consisted of approximately 40,000 diagrams from AI2D, with 12,000 diagrams containing process / step illustrations selected as the main samples. Sample pair construction: 10,000 training / validation / test pairs were constructed: 7,000 pairs for training (1,750 positive / 5,250 negative), 1,500 pairs for validation (375 positive / 1,125 negative), and 1,500 pairs for testing (375 positive / 1,125 negative). Positive sample construction method: The original images underwent "style redrawing + synonymous label replacement + local node order adjustment" (preserving the core topology), resulting in significant visual differences while maintaining structural semantic reuse; manual double annotation confirmed whether "structural reuse" was achieved. Primitive annotation quality: AI2D annotations (nodes / connections / text boxes) were used, and manual correction was performed on 1,200 samples to ensure training quality.
[0041] The specific steps for implementing this invention are as follows: (1) Multimodal preprocessing and feature extraction Graph parsing: Detects nodes, connections, and text labels (OCR) for each node; converts the graph into a directed / undirected graph, with nodes carrying text vectors and position / shape information. Node text is extracted and embedded using BERT, and node topological features (in-degree / out-degree, neighbor type distribution) are encoded together. Visual features (overall style) are still extracted using ResNet to determine if it is a "visual redraw." If there is a table / paragraph next to the graph (located using DocBank), the key columns of the table are mapped to the graph nodes for cross-modal validation.
[0042] (2) Project semantic graph and anchor point mechanism
[0043] Within the graph structure, "key milestone nodes / resource nodes" are selected as anchor points (based on node centrality and text importance scores), using a Top-K (K=6) approach. The search space for calculating edit distances is limited by using these anchor points as centers, reducing computational complexity and improving positioning accuracy.
[0044] (3) Local cross-modal alignment
[0045] A learnable node replacement / insertion / deletion cost (input includes node text similarity, neighborhood similarity, and node confidence) is used to calculate the "minimum edit path" between graphs. Local alignment outputs a node-level alignment table (which node matches which node, replacement / insertion / deletion cost, and confidence).
[0046] (4) Uncertainty-aware adaptive fusion
[0047] Node-level / graph-level similarity and corresponding confidence scores are input into a two-layer MLP fusion processor to output an overall similarity score and a structural reuse risk label. Special prompts are given for "visual redrawing" cases (low visual similarity but high structural similarity) to encourage manual review.
[0048] (5) Traceable evidence generation and human-machine closed loop
[0049] The system stores the chain of evidence as structured records, facilitating auditing and manual review. The results of manual review are fed back to fine-tune the learning-based structural cost model and fusion parameters (incremental fine-tuning weekly).
[0050] Design the following baseline for experimental testing: Baseline A: Traditional fixed-cost tree / graph edit distance (such as Zhang-Shasha or APTED).
[0051] Baseline B: Graph hash / fingerprint (fast retrieval but coarse) + manual confirmation.
[0052] Baseline C: Text-only (compares only node text, without considering topology).
[0053] The results of this approach are compared with those of several model baselines as follows:
[0054] 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 dual-modal intelligent plagiarism detection and similarity quantification evaluation system for science and technology project management, characterized in that: include: The multimodal preprocessing and feature extraction module is used to preprocess and extract features from data in three modalities: text, graphs, and tables. Then, the source traceability metadata is simultaneously associated; The project semantic graph and anchor mechanism module is used to first map entities in text, diagrams, and tables to a unified project semantic graph to form a set of nodes; then, key nodes are selected as anchors by node centrality and embedding cluster density to build an inverted index; at the same time, deep local matching is carried out in the alignment stage. The local cross-modal alignment module is used to perform local alignment of text paragraphs, illustration subgraphs, and table fragments centered on anchor points: text similarity is calculated using the cosine similarity of segment vectors; illustration similarity is calculated by combining visual embedding cosine similarity with primitive label matching ratio, while structural channels use a learned model to calculate weighted edit distance; tables are calculated using corresponding methods according to cell type. The uncertainty-aware adaptive fusion module is used to first input modal similarity and confidence into two layers of MLP to generate dynamic weight vectors, and then obtain the overall similarity by weighted summation and Sigmoid normalization; The traceable evidence generation and scenario-based risk index module is used to first generate structured evidence records in local alignment, then calculate the project similarity risk index, and map it to management strategies. The human-machine closed-loop and online learning module is used to apply the results of manual review to fine-tune the learning-based structural editing cost model and fusion weights, and to generate audit reports and training samples regularly.
2. The dual-modal intelligent deduplication and similarity quantification evaluation system for science and technology project management according to claim 1, characterized in that, The specific steps of the multimodal preprocessing and feature extraction module are as follows: First, preprocessing and feature extraction are performed on the data in three modalities: text, graphs, and tables. After character encoding standardization, redundant symbol removal, and format normalization, the text data is segmented. Semantic information at the paragraph level is extracted from each segment, and named entities are identified and subject-action-object structures are extracted. Each text segment is then represented by a segment vector embedding using a domain-adjusted BERT dataset. (1) in, This is the semantic vector of the text paragraph. It is the vector dimension; The data is divided into visual and structural channels. The visual channel extracts text information using OCR, and then uses a ResNet50 convolutional network to generate image embeddings. (2) in, Represents the image embedding vector. It is the visual feature dimension; the structural channel extracts nodes, connections, and labels through primitive detection to construct local subgraphs. The node features include type, text label, and topological information, and node embeddings are generated through two layers of GCN: (3) in, Represents the node feature matrix, This represents adding a self-loop adjacency matrix. Represents the node degree matrix, Represents the weight matrix. Represents graph nodes. Indicates node embedding; The table data identifies header and column types, standardizes units and numerical ranges, maps cells to vector embeddings, and generates table embeddings through row and column aggregation. ; Next, all modal features are simultaneously accompanied by document ID, author, timestamp, and node ID for evidence tracing.
3. The dual-modal intelligent deduplication and similarity quantification evaluation system for science and technology project management according to claim 1, characterized in that, The specific steps of the project semantic graph and anchor point mechanism module are as follows: First, based on the multimodal feature extraction results, entities in text, diagrams, and tables are mapped to the project semantic graph, forming a set of nodes containing tasks, milestones, modules, parameters, and deliverables, with edges representing task dependencies or parameter associations. Next, an anchor point mechanism is introduced, with the anchor point being the core node; the anchor point is selected based on node centrality and embedding cluster density. (4) in, Indicates node centrality. This indicates the embedding cluster density; after selecting the anchor point, an inverted index is built to achieve fast location of local subgraphs and document fragments; during the alignment stage, deep local matching is performed with the anchor point as the center.
4. The dual-modal intelligent deduplication and similarity quantification evaluation system for science and technology project management according to claim 1, characterized in that, The specific steps of the local cross-modal alignment module are as follows: At each anchor point, perform local alignment of text paragraphs, subplots, and table fragments: Text similarity is calculated using the cosine of segment vectors: (5) in, and Indicates a text paragraph; The visual channel diagram illustrates image similarity calculation based on embedding cosine similarity, combined with primitive label matching ratio: (6) in, Represents an image. Represents a set of primitive labels. Indicates visual channel weights; The structural channel uses a learned structural edit cost model to calculate the weighted edit distance, where the node replacement cost, node insertion cost, and node deletion cost are respectively: (7) (8) in, Indicates the editing cost weight. Indicates the similarity of knowledge graphs. Represents node confidence; normalized structural similarity is calculated as follows: (9) in, Represents a subgraph. Indicates the edit distance of the subgraph; Table similarity is calculated using normalized distance for numeric cells, vector cosine for text cells, and tree-structured alignment for formula cells, ultimately generated by weighted average. .
5. The dual-modal intelligent plagiarism detection and similarity quantification evaluation system for science and technology project management according to claim 1, characterized in that, The specific steps of the uncertainty-aware adaptive fusion module are as follows: Calculate the similarity of each modality. With confidence level The confidence level is calculated based on anchor coverage, matching segment length, and time difference; and a modal weight vector is generated through a two-layer MLP. (10) (11) in, Represents the weight matrix. This represents the bias vector. Represents the similarity vector and confidence vector of modalities. Indicates the output of the hidden layer. Represents the modal weight vector; The final overall similarity is calculated by weighted summation and normalized using Sigmoid: (12) in, These represent the local similarity and total similarity of each modality, respectively.
6. The dual-modal intelligent deduplication and similarity quantification evaluation system for science and technology project management according to claim 1, characterized in that, The specific steps of the traceable evidence generation and scenario-based risk index module are as follows: First, evidence records are generated during the local alignment process: (13) in, Indicates the document number. Align text fragments, Indicates the anchor point number. Represents modal similarity, Indicates modal confidence. Indicates time difference, Indicates the document author; chained storage facilitates auditing. Next, the Project Similarity Risk Index (PSRI) is defined as follows: (14) in, Mapping time window effect Indicates the score for similar coverage area. Assign weights to structure reuse, visual redrawing, and table reuse. Configurable weights.
7. The dual-modal intelligent deduplication and similarity quantification evaluation system for science and technology project management according to claim 1, characterized in that, The specific steps of the human-machine closed-loop and online learning module are as follows: The results of manual review are fed back as a supervisory signal to fine-tune the learning-based structural editing cost model and fusion weights; The loss function for training the edit cost model is: (15) The fusion weight network is fine-tuned using regression loss; audit reports and training samples are generated periodically.