Entity recognition-based collaborative problem traceability and strategy matching method

Abstract

The invention discloses a collaborative problem traceability and strategy matching method based on entity recognition, and the method comprises the steps: carrying out the semantic expansion of a negative problem text through a T5 generation type pre-training model, and extracting time, a subject, a problem, a strategy and other entities and position information through a BERT-BiLSTM-CRF model; splicing the subject entity, the question and the strategy entity to generate a text with a prefix, correcting by using a BART generation type pre-training model, and classifying and mapping the question to a subsystem for storage through a BERT-DPCNN model; coarse screening is conducted on texts in the subsystem through a regular rule base, vector representation is generated in combination with BERT and RoBERTa models, cosine similarity is calculated, and matching results are determined by synthesizing sentence positions and keyword overlapping degrees; constructing a question-strategy keyword knowledge graph based on a matching result, and inputting the keywords into a DeepSeek-V3 big language model to generate an optimization strategy scheme; according to the method, multiple deep learning models and large language models are fused, so that semantic understanding, entity extraction and accurate matching of negative problem texts are realized, and a final optimization strategy scheme is generated.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of project management and natural language processing, and more specifically to a collaborative problem tracing and strategy matching method based on entity recognition. Background Technology

[0002] In the management of large-scale engineering projects, meeting minutes, weekly reports, and other documents are important carriers of feedback, opinions, suggestions, and decision-making information, containing a wealth of key content related to project collaboration and control. However, these documents are often unstructured, vaguely worded, and semantically complex, making it difficult to efficiently and accurately identify hidden collaboration issues and potential conflicts.

[0003] Currently, handling collaborative issues in engineering projects largely relies on manual screening and analysis of textual materials. This approach is not only inefficient but also prone to subjectivity and overlooking crucial information. Regarding problem-strategy matching, existing technologies often employ simple keyword comparison methods, which struggle to handle complex semantic relationships within the text, resulting in insufficient matching accuracy and failing to provide precise and effective support for project decision-making. Furthermore, since the number of problem and strategy samples in project textual materials is often limited, existing methods struggle to guarantee the stability of identification and matching with small sample sizes and lack the ability to systematically trace the root causes of problems, making it difficult to meet the refined requirements of collaborative management in large-scale engineering projects.

[0004] To address the shortcomings of existing technologies, this invention proposes a collaborative problem tracing and strategy matching method based on entity recognition. This mechanism integrates text generation, named entity recognition, semantic analysis, and knowledge graph technologies to accurately extract key entities from unstructured text, achieving effective tracing of collaborative problems. Furthermore, through multi-dimensional matching rules and intelligent optimization methods, it improves the matching accuracy between problems and strategies, providing an intelligent solution for collaborative management of engineering projects. Compared with existing technologies, this invention effectively addresses the complexity of unstructured text, improves the comprehensiveness of problem identification and the accuracy of strategy matching, reduces reliance on manual processing, and has significant practical application value. Summary of the Invention

[0005] The purpose of this invention is to provide a collaborative problem tracing and strategy matching method based on entity recognition. By integrating multiple deep learning models and large language models, it achieves end-to-end closed-loop processing from semantic understanding of negative problem texts, entity extraction, accurate matching to the generation of the final optimized strategy solution.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A collaborative problem tracing and policy matching method based on entity recognition includes: S1. Use the T5 generative pre-trained model to semantically expand the original question text with the sentiment label "negative"; S2. Entity extraction is performed on the expanded question text and the original strategy text using a BERT-BiLSTM-CRF three-layer fusion model, outputting entity label sequences and location information; wherein the entities include: time entity, subject entity, question entity, strategy entity, strategy verb entity, and strategy content entity; S3. Concatenate the main entity with the problem entity and the strategy entity respectively to generate problem text and strategy text with the main entity prefix; The BART generative pre-trained model is used to correct the question text and policy text with subject prefixes, and the BERT-DPCNN fusion model is used to classify the question text, map it to the subsystem, and store it in a structured manner. S4. Perform coarse filtering of problem texts and policy texts in the subsystem using regular expression rule base; The BERT model and the RoBERTa model are used to generate vector representations of the question text and the strategy text after coarse screening, respectively. The cosine similarity between the vectors generated by the two models is calculated and the average value is taken as the basic similarity score. The matching results are determined by combining the sentence position, keyword overlap, and basic similarity score of the question text and the strategy text after the initial screening. S5. Construct a question-strategy keyword knowledge graph based on the matching results, and input the keywords in the question-strategy keyword knowledge graph into the DeepSeek-V3 large language model to generate an optimized strategy solution.

[0007] Furthermore, in S1, the process of semantically expanding the original question text includes: Using the T5 generative pre-trained model, the input is question text with negative sentiment labels, the maximum output length is 256, Beam Search is used with a width of 5, and early stopping is enabled; The expanded question text was integrated with the original text and verified using ROUGE-L, BLEU, and BERTScore metrics.

[0008] Furthermore, in S2, the architecture of the BERT-BiLSTM-CRF three-layer fusion model includes: The BERT layer uses a bert-base-chinese pre-trained model and outputs 768-dimensional word embedding vectors. The BiLSTM layer takes the word embedding vectors output by the BERT layer as input and includes 128 hidden units each for the forward and backward directions, using the PReLU activation function. The CRF layer optimizes the entity label sequence through the label transition probability matrix to avoid logically contradictory labeling.

[0009] Furthermore, in S3, the BART generative pre-trained model corrects the question text and policy text with the subject prefix, specifically as follows: The BART generative pre-trained model generates 3-5 candidate corrected sentences through a text destruction-reconstruction mechanism, and retains the optimal expression after cosine similarity screening. The DPCNN in the BERT-DPCNN fusion model uses a 3-layer pyramidal convolution with a kernel size of 3×768 and 256 filters.

[0010] Furthermore, in S3, the subsystems include: an information coordination subsystem, a target coordination subsystem, a benefit coordination subsystem, and a resource coordination subsystem.

[0011] Furthermore, in S4, the regular expression rule base is divided by subsystem and covers keyword variations, which is used to quickly filter non-matching statements across subsystems; The problem text and policy text in the subsystem are coarsely filtered using a regular expression rule base, specifically as follows: During matching, based on the subsystem tag corresponding to the question statement in the question text, the corresponding rule base is called to perform keyword verification on the strategy statement: if the strategy statement includes at least one keyword of the subsystem, it is retained and proceeds to the next step; otherwise, it is directly eliminated to quickly narrow down the matching range.

[0012] Furthermore, the text content of the original question and original strategy text includes: meeting minutes, weekly reports, and is suitable for unstructured text processing in large-scale engineering EPC projects.

[0013] According to specific embodiments provided by the present invention, the present invention has the following technical effects compared to the prior art: (1) The problem text was expanded by using the T5 generative pre-training model, which effectively solved the problems of insufficient sample size and single expression, enriched the diversity of the corpus, provided data guarantee for the stability of subsequent entity recognition and matching tasks, and improved the mechanism's adaptability to different text expression forms.

[0014] (2) The BERT-BiLSTM-CRF three-layer fusion model is used for named entity recognition. It combines the context semantic capture capability of BERT, the sequence modeling capability of BiLSTM and the label constraint capability of CRF, which significantly improves the accuracy of key entity extraction, provides entity-level support for accurate source tracing of collaborative problems, and realizes efficient transformation from text to key information.

[0015] (3) The refined matching mechanism that integrates regular expression matching and semantic similarity takes into account both the structural features and semantic associations of the text. It achieves rapid coarse screening through regular expression matching and captures deep semantic associations through semantic similarity calculation, which greatly improves the accuracy of problem and strategy matching and overcomes the limitations of traditional keyword matching methods.

[0016] (4) By combining keyword knowledge graphs with large language models, strategy optimization is carried out. Knowledge graphs provide structured knowledge support for strategy generation, while large language models use their powerful generation capabilities to mine potential connections. The generated strategies are more targeted and innovative, which improves the scientificity and effectiveness of problem solving in project collaborative management and provides reliable technical support for engineering project decision-making. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] The following description, in conjunction with the accompanying drawings, further illustrates the collaborative problem tracing and strategy matching method based on entity recognition provided by this invention; Figure 1 This is a schematic diagram of the overall process of the collaborative problem tracing and strategy matching method based on entity recognition provided by the present invention; Figure 2 This is a schematic diagram of the BERT-BiLSTM-CRF named entity recognition model structure in this invention; Figure 3 This is a schematic diagram of the problem-strategy matching process in this invention; Figure 4 This is a partial schematic diagram of the problem and strategy knowledge graph in this invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0020] To better understand the purpose, structure, and function of this invention, the invention will be described in further detail below with reference to the accompanying drawings. Example 1

[0021] like Figure 1 As shown, this invention provides a method for tracing the source and matching the strategy of collaborative problems based on entity recognition, including the following steps: S1. Use the T5 generative pre-trained model to semantically expand the original question text with the sentiment label "negative"; In step S1, the process of semantically expanding the original question text includes: Using the T5 generative pre-trained model, the input is question text with negative sentiment labels, the maximum output length is 256, Beam Search is used with a width of 5, and early stopping is enabled; The expanded question text was integrated with the original text and verified using ROUGE-L, BLEU, and BERTScore metrics.

[0022] In this embodiment, the text generation and expansion part addresses the issue of a small number of question samples and their simplistic expressions in the project text by employing a T5 generative pre-trained model to expand the original question text. Representative question texts with a "negative" sentiment label are selected as input, and the model generates semantically similar but diverse question texts to enrich the corpus samples and provide a sufficient data foundation for subsequent entity recognition and matching tasks.

[0023] Among them, the T5 (Text-to-Text Transformer) model is a general-purpose pre-trained language model proposed by Google. This model transforms all natural language problems into text-to-text formats and solves them using a unified model. The core of the T5 model lies in its use of prefixed task declarations and text-based answer generation to standardize the input and output of all natural language processing tasks. Most pre-trained language models typically require additional non-linear layers during downstream task fine-tuning to convert the model's output to the format required for specific tasks. In contrast, the T5 model requires no adjustments to its architecture, only the fine-tuning data for downstream tasks, and no additional non-linear layers. Its only requirement is to add the appropriate task declaration prefix to the input data, a design that greatly simplifies the model's application process.

[0024] S2. Entity extraction is performed on the expanded question text and the original strategy text using a BERT-BiLSTM-CRF three-layer fusion model, outputting entity label sequences and location information; wherein the entities include: time entity, subject entity, question entity, strategy entity, strategy verb entity, and strategy content entity; This embodiment is specifically as follows: Figure 2 As shown, a three-layer BERT-BiLSTM-CRF fusion model is constructed to extract key entities from the augmented text and the original policy text.

[0025] The first layer is the BERT model (Bidirectional Encoder Representations from Transformers, a pre-trained language model based on a bidirectional Transformer architecture, whose core capability is to capture semantic relationships between words through bidirectional context encoding). It employs the BERT-base-chinese pre-trained model (a basic version of BERT trained on Chinese corpus, adapted to the semantic understanding needs of Chinese domain-specific terms such as "supervision unit" and "progress node" in EPC projects). First, the corpus text is tokenized—a fundamental preprocessing step in natural language processing. Continuous natural language text (e.g., "April 2023, General Contracting Project Department reported progress delay") is broken down into basic semantic units recognizable by the model (i.e., "tokens," which are "April 2023," "General Contracting Project Department," "feedback," and "progress delay"). Simultaneously, meaningless symbols such as punctuation and special characters are filtered out, forming a token sequence arranged according to the original text's word order (e.g., "[April 2023, General Contracting Project Department, Feedback, Progress Delay]"). Then, the token sequence is linked with attention... The mask (used to mark the positions of valid tokens and filter out invalid tokens) is input into the 12-layer bidirectional Transformer encoder of the BERT model. This encoder is based on the self-attention mechanism, which allows the model to automatically "pay attention" to all other tokens in the text when processing each token and calculate the association strength (e.g., when processing "progress delay", it prioritizes associating "general contracting project department" and "April 2023"). The final output is a word embedding vector with a dimension of 768, which provides semantic support for entity boundary recognition.

[0026] The second layer is a BiLSTM model (Bidirectional Long Short-Term Memory, which can simultaneously capture forward and backward dependencies of text sequences). It takes the word embedding vectors output by the BERT layer as input, with 128 hidden units in both the forward and backward directions. It uses the PReLU activation function (Parametric Rectified Linear Unit, which can adaptively adjust the activation threshold to solve the "dead neuron" problem of the traditional ReLU function). The forward LSTM captures the forward dependencies of the text sequence (such as "April 2023" → "General Contracting Project Department" → "Progress Delay"), and the backward LSTM captures the backward dependencies. The output is a sequence feature vector with a dimension of 256, which solves the problem that the single-direction model cannot fully capture the dependencies of long texts.

[0027] The third layer is a CRF model (Conditional Random Field, a probabilistic model for sequence labeling, which excels at learning global logical relationships between labels). It takes the sequence feature vector of the BiLSTM layer as input, constructs a label transition probability matrix, learns the reasonable association logic of entity labels such as "time-subject-problem" (e.g., the "problem" label cannot be directly followed by the "strategy verb" label), and optimizes the label sequence by minimizing the negative log-likelihood loss function to avoid logically contradictory labeling results.

[0028] Based on the above three-layer model, when extracting entities from the text of the question-policy corpus, the text is sequentially processed by BERT semantic encoding, BiLSTM sequence feature extraction, and CRF label optimization. The final output is a label sequence containing six types of entities: "time", "subject", "question", "policy", "policy verb", and "policy content". At the same time, the start and end positions of each entity in the text are recorded to ensure that the entity information is complete and traceable.

[0029] S3. Concatenate the main entity with the problem entity and the strategy entity respectively to generate problem text and strategy text with the main entity prefix; The BART generative pre-trained model is used to correct the question text and policy text with subject prefixes, and the BERT-DPCNN fusion model is used to classify the question text, map it to the subsystem, and store it in a structured manner. This embodiment specifically includes the entity processing and classification section; (1) The identified entities are concatenated to generate "problem" and "strategy" texts with subject prefixes; the concatenated texts are semantically corrected using the BART model to generate fluent, standardized, and semantically clear texts; the problem texts are classified using the BERT-DPCNN text classification model, mapping them to four categories: "information," "goal," "benefit," and "resource," and the processed problem and strategy texts are stored in a structured JSON format. The specific process is as follows: First, entity concatenation is performed. For the six entity categories extracted from named entity recognition—"Time," "Subject," "Issue," "Strategy," "Strategy Verb," ​​and "Strategy Content"—focusing on the "Issue-Subject" and "Strategy-Subject" relationships, fixed rules are used to combine entities: the "Subject" entity is used as a prefix and concatenated with the "Issue" and "Strategy" entities according to the semantic order of the text, forming structured statements. For example, in an EPC project text, "Subject: Supervision Unit" and "Issue: Inadequate edge protection in the construction area" are concatenated into the issue statement "[Subject: Supervision Unit] Inadequate edge protection in the construction area," and "Subject: General Contracting Project Department" and "Strategy: Supplement protective facilities within 24 hours" are concatenated into the strategy statement "[Subject: General Contracting Project Department] Supplement protective facilities within 24 hours." This concatenation rule clarifies the responsible subject of the issue and the implementing subject of the strategy, avoiding subject confusion caused by multiple participants in an EPC project (government, owner, construction party, etc.) and ensuring the consistency of the "Issue-Strategy" subject during subsequent matching.

[0030] (2) Next, semantic correction is carried out. Since EPC project meeting minutes, weekly reports and other texts are mostly recorded in real time, they are prone to problems such as disordered word order and ambiguous wording (such as "insufficient protection" and "slow data transmission"). It is necessary to introduce the BART generative pre-trained model (Bidirectional and Auto-Regressive Transformers, a sequence reconstruction model based on the Transformer architecture, which is good at text semantic optimization and standardized expression) for processing. The bart-base-chinese pre-trained weights (adapted to Chinese engineering text semantic understanding) are used. The concatenated sentences are input into the model. The model optimizes the non-standard expressions in the original sentences through the "text destruction-reconstruction" mechanism. For example, "[Main body: Ministry of Construction] Data transmission is slow" is corrected to "[Main body: Ministry of Construction] The data upload rate of each participating unit's system is lower than the standard agreed in the contract", and "[Main body: Construction unit] Insufficient protection" is corrected to "[Main body: Construction unit] The number of edge protection facilities in the construction area does not meet the safety specifications". The model simultaneously generates 3-5 semantically similar candidate corrected sentences. By calculating cosine similarity (comparing the vector similarity between the candidate sentences and the original concatenated sentences), duplicate expressions are eliminated, and the sentences with the most accurate semantics and the most standardized expression are retained to ensure semantic consistency in subsequent classification and matching.

[0031] Subsequently, subsystem classification is implemented, requiring the corrected problem statements to be assigned to the four subsystems of EPC project collaborative management: "Information Collaboration," "Goal Collaboration," "Interest Collaboration," and "Resource Collaboration." The BERT-DPCNN fusion model is used to complete the classification task: the BERT model (Bidirectional Encoder Representations from Transformers) is responsible for extracting deep semantic features of the problem statements, such as extracting "information transmission" related semantics from "low data upload rate" and "goal deviation" related semantics from "20 days behind schedule." The DPCNN model (Deep Pyramid Convolutional Neural Networks, which captures global structural features of text through multi-layer convolution) further processes the semantic features output by BERT, strengthening subsystem-related features (such as the feature weights of keywords like "data" and "transmission" in "information collaboration" statements) through 3 layers of pyramidal convolution (3×768 kernel size, 256 filters). Finally, the probability values ​​of the problem statements corresponding to the four subsystems are output. The category with the highest probability is selected as the final subsystem label. For example, the probability of the category "[Main Body: Ministry of Construction] The data upload rate of each participating unit's system is lower than the standard agreed in the contract" is 0.94. Therefore, it is determined that it belongs to the "Information Collaboration" subsystem, which realizes the accurate classification of the problematic statement and provides a basis for subsequent cross-subsystem statement filtering.

[0032] (3) Finally, structured storage is performed. To facilitate data retrieval and management in the fine matching stage, the processed information is stored in a standardized format: JSON is used as the data format, and five core fields are designed: "Issue ID", "Issue Statement", "Subsystem Tag", "Strategy Statement" and "Entity Details". The "Entity Details" field contains the original extracted six types of entity information (such as "Time: March 15, 2024", "Subject: Ministry of Construction", "Issue: Low data upload rate" etc.). Example stored content is as follows: "{"Problem ID":"Q2024031501","Problem Statement":"[Subject: Ministry of Construction] The data upload rate of each participating unit's system is lower than the contractually agreed standard","Subsystem Tag":"Information Collaboration","Strategy Statement":"[Subject: Ministry of Construction] Designate a specific person to be responsible for data upload review and establish an anomaly warning mechanism","Entity Details":{"Time":"March 15, 2024","Subject":"Ministry of Construction","Problem":"System data upload rate is lower than the contractually agreed standard","Strategy":"Designate a specific person to be responsible for data upload review and establish an anomaly warning mechanism","Strategy Verb":"Designate, Establish","Strategy Content":"Data upload review, anomaly warning mechanism"}}". This storage method can realize the associated management of problem, strategy, and entity information, which facilitates the quick retrieval of corresponding data during subsequent matching and improves the efficiency of mechanism operation.

[0033] S4. Perform coarse filtering of problem texts and policy texts in the subsystem using regular expression rule base; The BERT model and the RoBERTa model are used to generate vector representations of the question text and the strategy text after coarse screening, respectively. The cosine similarity between the vectors generated by the two models is calculated and the average value is taken as the basic similarity score. The matching result is determined by combining the sentence position, keyword overlap, and basic similarity score of the question text and the strategy text after the initial screening.

[0034] This embodiment specifically refers to: the refined matching part; This paper employs a combination of regular expression matching and semantic similarity calculation to match questions and strategies. First, regular expression matching defines text patterns or rules to perform a coarse screening of structured questions and strategies, extracting potential matching entities. Then, BERT and RoBERTa models are used to generate vector representations of the question and strategy texts, respectively, and the semantic similarity between the two is calculated using the cosine similarity formula. Finally, the final matching result is obtained by combining sentence position distance, keyword overlap, and average semantic similarity score. The specific details are as follows: Among them, the regular expression matching coarse filter is used to quickly filter out non-matching statements across subsystems. For example... Figure 3As shown, a dedicated keyword rule base is constructed for the four subsystems of "Information Collaboration," "Goal Collaboration," "Interest Collaboration," and "Resource Collaboration." "Information Collaboration" includes keywords such as "data upload" and "information transmission"; "Goal Collaboration" includes keywords such as "schedule delay" and "node delay"; "Interest Collaboration" includes keywords such as "resident complaints" and "cost disputes"; and "Resource Collaboration" includes keywords such as "staff allocation" and "equipment shortage." The rules also cover common variations of expressions found in engineering texts (e.g., "data upload" encompasses "data reporting" and "upload data"). During matching, based on the subsystem tag of the problem statement, the corresponding rule base is invoked to perform keyword validation on the strategy statement: if the strategy statement contains at least one keyword from that subsystem, it is retained and proceeds to the next step; otherwise, it is directly eliminated, quickly narrowing down the matching scope.

[0035] Semantic similarity matching is used to accurately locate the optimal strategy from the initial screening results. A dual-model fusion calculation is employed, using BERT (Bidirectional Transformer Encoder Representation Model) and RoBERTa (an optimized version of BERT): the question and strategy statements after initial screening are input into the two models respectively, generating vectors containing semantic features; the similarity between the vectors is calculated using the cosine similarity formula (the closer the value is to 1, the more semantically similar), and the average of the two model results is taken as the base similarity score. For example, the question statement "low data upload rate" and the strategy statement "optimize data upload channels" have a base similarity score of 0.83, indicating a high semantic correlation between them. Through this two-level matching, irrelevant strategies are first quickly filtered out using regularization rules, and then precise association is achieved through dual-model semantic calculation, ultimately determining the strategy that best matches the question, ensuring both efficiency and improved accuracy.

[0036] S5. Construct a question-strategy keyword knowledge graph based on the matching results, such as Figure 4 As shown, keywords from the problem-strategy keyword knowledge graph are input into the DeepSeek-V3 large language model to generate optimized strategy solutions.

[0037] This embodiment specifically refers to the strategy optimization section; Based on the matching results, a keyword knowledge graph of questions and strategies is constructed to organize the knowledge structure and accurately locate key information. The DeepSeek-V3 large language model is selected, using keywords from the knowledge graph as support, to optimize and generate the matched strategies. Leveraging its powerful language understanding and generation capabilities, the large language model improves and expands upon the original strategies, generating more targeted and diversified solutions. Example 2

[0038] The present invention also provides a specific embodiment of the collaborative problem tracing and strategy matching method based on entity recognition in Embodiment 1; Taking the Yichang Two-Network Phase II EPC Project for the Yangtze River Protection as an example: The text generation augmentation part 1 involved collecting meeting minutes and weekly report texts from the project between 2020 and 2024, and selecting 147 texts (average length 96.5 characters) with a "negative" sentiment label and relevant to the objectives. The T5 model was used to augment these texts, with a maximum generated text length of 256 characters, a Beam Search width of 5, and early termination enabled. This generated 500 semantically similar question texts, which were then integrated with the original text for subsequent processing. Validation using ROUGE-L, BLEU, and BERTScore metrics showed that the T5 model generated the best performance in terms of semantic consistency and expressive diversity.

[0039] Part 2 of the named entity recognition component involved using Label Studio to annotate the expanded question text and collected strategy text, categorizing them into six entity types: "time," "subject," "question," "strategy," "strategy verb," ​​and "strategy content." After processing into BIO format, the data was divided into training, validation, and test sets (756 data points in total) in an 8:1:1 ratio. A BERT-BiLSTM-CRF model was trained with a batch size of 16, an epoch of 50, a learning rate of 5e-5, and a max seq length of 64. Experimental results showed that the model achieved an F1 score, precision, and recall of 0.82, accurately extracting various entities, such as identifying "April 7, 2023" (time), "general contracting project department" (subject), and "insufficient safety management personnel" (question) from the text.

[0040] Part 3, Entity Processing and Classification, involves concatenating identified entities. For example, "General Contracting Project Department + Insufficient Safety Management Personnel" generates problem text, and "General Contracting Project Department + Full Coverage of Investigation" generates strategy text. The concatenated text is semantically corrected using the BART model to remove redundant information and standardize the expression. The problem text is then classified using the BERT-DPCNN model, mapped to "Information," "Objective," "Benefit," and "Resource" tags, and stored in JSON format. An example is shown below: {"Problem":"The progress of each participating unit is far from the planned target","Type":"Objective","Strategy Candidate":["Sorting the progress plan and arranging it reasonably, with supervision ensuring the completion of node targets"]}.

[0041] The refined matching part 4 firstly uses regular expression matching to coarsely screen the structured storage questions and strategies. For example, for questions related to "information", strategies containing keywords such as "data", "transmission", and "channel" are matched. Then, text vectors are generated using BERT and RoBERTa models, and semantic similarity is calculated using the cosine similarity formula. For example, the similarity score between "the data upload rate of each unit's system is extremely low" and "each unit ensures that the information transmission channels are unobstructed" is 0.78. Finally, the final matching results are obtained by combining sentence position distance (distance within the same paragraph has higher weight) and keyword overlap (such as both containing "data" and "unit"). Some results are as follows: "complaints from surrounding residents" (interest category) matches "strengthen communication, and the supervision unit issues a fine notice".

[0042] Among them, strategy optimization part 5: construct a knowledge graph of problems and strategies based on the matching results, such as associating the "information problem" node with keywords such as "data upload" and "transmission channel", inputting the keywords in the knowledge graph into the DeepSeek-V3 large language model to generate optimization strategies.

[0043] For example, regarding the issue of "extremely low data upload rates in various unit systems," the optimization strategy is as follows: 1. Define clear responsibilities and processes: Each unit designates a data upload specialist and establishes a standardized 'data submission-review-upload' process; 2. Optimize information transmission channels: Deploy a unified data upload platform and establish an anomaly warning mechanism; 3. Establish a supervision and feedback mechanism: Regularly check and report upload status, collect issues, and optimize processes. Practical application verification shows that this mechanism improves issue identification accuracy by 40%, strategy matching efficiency by 35%, and significantly enhances the feasibility of optimized strategies.

[0044] The above description of the disclosed embodiments enables those skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A collaborative problem tracing and strategy matching method based on entity recognition, characterized in that, include: S1. Use the T5 generative pre-trained model to semantically expand the original question text with the sentiment label "negative"; S2. Entity extraction is performed on the expanded question text and the original strategy text using a BERT-BiLSTM-CRF three-layer fusion model, outputting entity label sequences and location information; among which, entities include: time entity, subject entity, question entity, strategy entity, strategy verb entity, and strategy content entity; S3. Concatenate the main entity with the problem entity and the strategy entity respectively to generate problem text and strategy text with the main entity prefix; The BART generative pre-trained model is used to correct the question text and policy text with subject prefixes, and the BERT-DPCNN fusion model is used to classify the question text, map it to the subsystem, and store it in a structured manner. S4. Perform coarse filtering of problem texts and policy texts in the subsystem using regular expression rule base; The BERT model and the RoBERTa model are used to generate vector representations of the question text and the strategy text after coarse screening, respectively. The cosine similarity between the vectors generated by the two models is calculated and the average value is taken as the basic similarity score. The matching results are determined by combining the sentence position, keyword overlap, and basic similarity score of the question text and the strategy text after the initial screening. S5. Construct a question-strategy keyword knowledge graph based on the matching results, and input the keywords in the question-strategy keyword knowledge graph into the DeepSeek-V3 large language model to generate an optimized strategy solution.

2. The method for tracing the source and matching the strategy of collaborative problems based on entity recognition according to claim 1, characterized in that, In S1, the process of semantically expanding the original question text includes: Using the T5 generative pre-trained model, the input is question text with negative sentiment labels, the maximum output length is 256, Beam Search is used with a width of 5, and early stopping is enabled; The expanded question text was integrated with the original text and verified using ROUGE-L, BLEU, and BERTScore metrics.

3. The method for tracing the source and matching the strategy of collaborative problems based on entity recognition according to claim 1, characterized in that, In S2, the architecture of the BERT-BiLSTM-CRF three-layer fusion model includes: The BERT layer uses a bert-base-chinese pre-trained model and outputs 768-dimensional word embedding vectors. The BiLSTM layer takes the word embedding vectors output by the BERT layer as input and includes 128 hidden units each for the forward and backward directions, using the PReLU activation function. The CRF layer optimizes the entity label sequence through the label transition probability matrix to avoid logically contradictory labeling.

4. The collaborative problem tracing and strategy matching method based on entity recognition according to claim 1, characterized in that, In S3, the BART generative pre-trained model corrects the question text and policy text with subject prefixes, specifically as follows: The BART generative pre-trained model generates 3-5 candidate corrected sentences through a text destruction-reconstruction mechanism, and retains the optimal expression after cosine similarity screening. The DPCNN in the BERT-DPCNN fusion model uses a 3-layer pyramidal convolution with a kernel size of 3×768 and 256 filters.

5. The method for tracing the source and matching the strategy of collaborative problems based on entity recognition according to claim 1, characterized in that, In S3, the subsystems include: information coordination subsystem, target coordination subsystem, benefit coordination subsystem, and resource coordination subsystem.

6. The method for tracing the source and matching the strategy of collaborative problems based on entity recognition according to claim 1, characterized in that, In S4, the regular expression rule base is divided by subsystem and covers keyword variations, which is used to quickly filter non-matching statements across subsystems; The problem text and policy text in the subsystem are coarsely filtered using a regular expression rule base, specifically as follows: During matching, based on the subsystem tag corresponding to the question statement in the question text, the corresponding rule base is called to perform keyword verification on the strategy statement: if the strategy statement includes at least one keyword of the subsystem, it is retained and proceeds to the next step; otherwise, it is directly eliminated to quickly narrow down the matching range.

7. The method for tracing the source and matching the strategy of collaborative problems based on entity recognition according to claim 1, characterized in that, The text content of the original questions and original strategies includes: meeting minutes, weekly reports, and is suitable for unstructured text processing in large-scale engineering EPC projects.

Images