A cross-cultural risk identification method and system for an overseas EPC project

Abstract

This invention, entitled "A Method and System for Cross-Cultural Risk Identification in Overseas EPC Projects," belongs to the field of engineering management technology. The technical problem it aims to solve is: to achieve accurate and efficient identification of cross-cultural risks in multilingual environments, outputting standardized and structured results with high interpretability and engineering applicability. The key technical solutions are: acquiring multi-source heterogeneous project texts, unifying the context through language identification and translation, and obtaining compliant compressed text using hybrid scoring and progressive multi-level compression; acquiring risk knowledge through a hybrid vector and full-text retrieval method, and using structured prompts to guide a large language model to generate structured results containing risk types, levels, basis, and recommendations; outputting documents and triggering alerts; and iteratively optimizing the knowledge base and model based on human feedback. The system consists of modules for data input, preprocessing, hybrid retrieval, model reasoning, structured output, and feedback iteration, achieving accurate and efficient identification of multilingual cross-cultural risks.

Description

Technical Field

[0001] This invention belongs to the field of engineering management technology, specifically relating to a method and system for identifying cross-cultural risks in overseas EPC projects. Background Technology

[0002] Overseas engineering general contracting (construction of overseas projects) The projects span multiple regions with diverse cultural backgrounds, including Asia, Africa, and Latin America. Significant differences exist among project stakeholders in terms of values, communication styles, power distance, and uncertainty avoidance, often leading to misunderstandings, lack of trust, and obstacles to collaboration. According to industry reports, over 30% of overseas engineering project delays or cost overruns are closely related to cross-cultural conflicts. Existing cross-cultural risk management methods largely rely on expert experience, cultural theory analysis such as the Hofstede cultural dimension, and qualitative research methods such as questionnaires and Delphi. These methods have significant limitations: (1) Lag: They often analyze problems after they occur, lacking real-time early warning capabilities; (2) Subjectivity: They rely on the experience of a few experts, making the results susceptible to bias; (3) Insufficient coverage: They struggle to handle massive amounts of unstructured multilingual text such as contracts, emails, and meeting minutes.

[0003] Text mining and deep learning models like BERT have been applied in contract risk analysis and safety incident detection, but they primarily focus on technical and financial risks, lacking quantitative extraction of soft risks such as "culture." Simple static template or keyword matching methods have poor universality and perform poorly on complex, long texts. Large Language Models (LLMs) have potential in multilingual understanding and knowledge reasoning, and can automatically extract risk clues from text. However, directly applying general-purpose LLMs to cross-cultural risk identification may lead to "illusions" or cultural biases, and has the following limitations: (1) No search enhancement generation ( It's easy to experience "illusions" or miss information when searching; (2) A single model struggles to acquire sufficient cultural background information when dealing with multilingual scenarios; (3) Even with an LLM with a 128K ultra-large context window, the computational cost is high, the response delay is large, and there is a lack of structured output guarantee.

[0004] In view of this, the present invention is hereby proposed. Summary of the Invention

[0005] To address the aforementioned technical problems in existing technologies, this invention provides a method and system for identifying cross-cultural risks in overseas EPC projects. This method and system solves the problem of accurate and efficient identification of cross-cultural risks in multilingual environments, outputs standardized and structured results, and possesses high interpretability and engineering applicability.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: Firstly, a method for identifying cross-cultural risks in overseas EPC projects includes: Acquire text data from diverse and heterogeneous overseas EPC projects; The text data is preprocessed to obtain compressed text with a unified context and a length that meets the input requirements of a large language model; Based on a hybrid retrieval strategy, cross-cultural risk knowledge related to the compressed text is retrieved from a preset knowledge base; The compressed text and the cross-cultural risk knowledge are input into the large language model; The large language model is guided by preset structured prompts to generate cross-cultural risk identification results; The system outputs the cross-cultural risk identification results and receives human feedback to iteratively optimize the preset knowledge base and model parameters.

[0007] Furthermore, the step of preprocessing the text data includes: Identify the document type and language category of the text data; Translate non-Chinese text data into Chinese to obtain text data with a unified context.

[0008] Furthermore, the step of preprocessing the text data to obtain compressed text includes: A hybrid scoring algorithm is used to calculate the importance score of each paragraph in the text data, and paragraphs with scores higher than a preset threshold are retained to complete the initial compression. If the number of text tokens after the initial compression exceeds the session length limit of the large language model, a progressive multi-level compression strategy will be used for dynamic compression until the number of text tokens meets the preset conditions.

[0009] Furthermore, the calculation formula for the hybrid scoring algorithm is as follows:

[0010] in, For paragraph Importance score For paragraph Keyword coverage score For paragraph TF-IDF weighted score.

[0011] Furthermore, the compression ratio calculation formula for the progressive multi-stage compression strategy is as follows:

[0012] in, For the first The compression ratio of a rotary compressor. This is the initial compression ratio. This is the compression ratio step size; The preset conditions are: ,in For the compressed text number, The maximum conversation length for a large language model. This represents the maximum generation length of a large language model. To reserve Buffer.

[0013] Furthermore, the step of retrieving cross-cultural risk knowledge from a preset knowledge base based on a hybrid retrieval strategy includes: Perform new-bge-m3 vector search and BM25 full-text search in parallel; The two search results are fused with a weight ratio of 6:4, and a candidate set of knowledge entries is generated using the RRF algorithm.

[0014] Furthermore, it also includes: The knowledge item candidate set is re-ranked using a re-ranking model, and a preset number of knowledge items are taken as the final retrieval results. When the relevance of the search results from the local preset knowledge base is lower than the preset threshold, an online search is initiated to obtain the latest cross-cultural risk knowledge, and the filtered online search results are added to the final search results.

[0015] Furthermore, the preset structured prompts include: system prompts and task-level prompts; The system prompts are used to define the role of the large language model as an EPC cross-cultural risk identification expert. The task-level prompts are used to guide the large language model to output cross-cultural risk identification results according to a preset process.

[0016] Furthermore, the cross-cultural risk identification results are structured data, including risk type, risk level, cultural basis, mitigation suggestions, and original text basis; The large language model adopted is the Qwen3-30B-A3B model, with a generation temperature set to 0.1 and a maximum generation length limit of 2048 tokens.

[0017] Furthermore, the step of outputting the cross-cultural risk identification result includes: Export the cross-cultural risk identification results as a table or JSON document; The cross-cultural risk identification results are embedded into the project management platform to trigger automatic early warnings.

[0018] Secondly, a cross-cultural risk identification system for overseas EPC projects, applied to any of the cross-cultural risk identification methods for overseas EPC projects described above, includes: The multi-source text acquisition module is used to acquire text data from multi-source heterogeneous overseas EPC projects. The text preprocessing module is used to preprocess the text data to obtain compressed text with a unified context and a length that meets the input requirements of the large language model. A hybrid retrieval module is used to retrieve cross-cultural risk knowledge related to the compressed text from a preset knowledge base based on a hybrid retrieval strategy. The model input module is used to input the compressed text and the cross-cultural risk knowledge into the large language model; The structured risk identification module is used to guide the large language model to generate cross-cultural risk identification results through preset structured prompts; The results output and iteration module is used to output the cross-cultural risk identification results and receive human feedback to iteratively optimize the preset knowledge base and model parameters.

[0019] Compared with the prior art, the present invention has the following advantages: (1) It realizes full-process automation from multi-source multilingual text input to structured risk report output and feedback iteration, has modular design features, and ensures the reproducibility and engineering feasibility of the system; (2) A hybrid recall strategy of new-bge-m3 vector retrieval and BM25 full-text retrieval is adopted, combined with Top-K re-ranking and network completion mechanism, which significantly improves the recall comprehensiveness and accuracy of cross-cultural risk knowledge in multilingual contexts and effectively solves the illusion problem of general large model; (3) While fully preserving key risk information, significantly reduce the token consumption and response latency of long text processing, while balancing recognition accuracy and operating efficiency; (4) A two-layer structured Prompt mechanism is adopted to achieve standardized output of risk identification results, including clear risk types, levels, basis and recommendations, which has high interpretability and facilitates quick decision-making by project managers; (5) It is significantly better than traditional rule systems, BERT fine-tuning models and simple large window LLM solutions in core indicators such as F1-score, precision and recall. It can support cross-cultural risk identification in 6 languages ​​including Chinese, Arabic and Indonesian, and has broad engineering application value. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the cross-cultural risk identification method for overseas EPC projects provided in this embodiment of the invention; Figure 2This is an architecture diagram of a cross-cultural risk identification system for overseas EPC projects provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0022] It should be noted that, unless otherwise specifically stated, the relative arrangement and numerical expressions of the components and steps described in these embodiments should not be construed as limiting the scope of the invention.

[0023] The following description of exemplary embodiments is merely illustrative and is not intended to limit the invention or its application or use in any way. Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but where applicable, such techniques, methods, and apparatus should be considered part of this specification.

[0024] Example 1 See Figure 1 , Figure 1 This is a flowchart of a cross-cultural risk identification method for overseas EPC projects proposed in this invention. Specific steps may include: S1. Multi-source text acquisition: Acquire text data of overseas EPC projects from multiple heterogeneous sources; The text data comes from all stages of the overseas EPC project lifecycle, including but not limited to project contract terms, technical specifications, meeting minutes, project emails, on-site construction logs, supplier communications, and local government regulations. All input data has been anonymized to ensure project information security and business confidentiality.

[0025] S2. Text preprocessing and compression: The text data is preprocessed to obtain compressed text with a unified context and a length that meets the input requirements of the large language model; The preprocessing process first identifies the document type and language category of the text data. For non-Chinese text data, the Qwen3-30B-A3B translation model is called for high-quality translation. The translation temperature is set to 0.0 to ensure the consistency and accuracy of the translation results. All texts are uniformly converted into Chinese context. Subsequently, a hybrid scoring algorithm is used to calculate the importance score of each paragraph in the text data, retaining paragraphs with scores higher than a preset threshold to complete the initial compression. The calculation formula for the hybrid scoring algorithm is as follows:

[0026] in, For paragraph Importance score For paragraph The keyword coverage score takes into account the weight of cross-cultural risk-related keywords contained in the paragraph, document titles, list symbols, and special tags. For paragraph The TF-IDF weighted score is used to statistically characterize the importance of a paragraph.

[0027] If the number of text tokens after initial compression exceeds the session length limit of the large language model, a progressive multi-level compression strategy is adopted for dynamic compression until the number of text tokens meets the preset conditions. The compression ratio calculation formula of the progressive multi-level compression strategy is as follows:

[0028] in, For the first The compression ratio of a rotary compressor. This is the initial compression ratio. Compression ratio step size; initial compression ratio Set to 0.5, compression ratio step size The value is set to 0.15, which means the compression ratios are 0.5, 0.65, and 0.8 respectively. The preset conditions are: ,in For the compressed text number, The maximum conversation length for a large language model. This represents the maximum generation length of a large language model. To reserve Buffer; In this embodiment, the reserved token buffer is set to 400 tokens. After each round of compression, the system automatically verifies whether the number of text tokens meets the preset conditions, and terminates the compression loop when the conditions are met.

[0029] S3. Hybrid retrieval of risk knowledge: Based on a hybrid retrieval strategy, retrieve cross-cultural risk knowledge related to the compressed text from a preset knowledge base; The pre-stored knowledge base contains Hofstede's cultural dimension theory, guidelines on cultural customs of various countries, a case study library of cross-cultural risks in overseas EPC projects, local laws and regulations, and industry standards and norms.

[0030] The hybrid retrieval process executes new-bge-m3 vector retrieval and BM25 full-text retrieval in parallel. New-bge-m3 vector retrieval is used to capture the semantic similarity of texts, while BM25 full-text retrieval is used to accurately match keywords. The system merges the two retrieval results with a weight of 6:4 and uses the RRF algorithm to generate a candidate set of knowledge entries.

[0031] Subsequently, a re-ranking model is used to perform Top-K re-ranking on the candidate set of knowledge items. In this embodiment, the K value is set to 3, and the top 3 knowledge items are taken as the basic search results. When the relevance of the search results in the local preset knowledge base is lower than the preset threshold of 0.6, an online search is automatically initiated to obtain the latest cross-cultural risk-related news, industry guidelines, government announcements and other information. After filtering and abstract extraction, these are added to the final search results.

[0032] S4. Model Input: Input the compressed text and the cross-cultural risk knowledge into the large language model; The large language model adopts the Qwen3-30B-A3B model, with the generation temperature set to 0.1 to ensure the stability and accuracy of the output results, and the maximum generation length is limited to 2048 tokens.

[0033] S5. Structured Risk Identification: The large language model is guided to generate cross-cultural risk identification results through preset structured prompts; The preset structured prompts include system prompts and task-level prompts. System prompts define the role of the large language model as an EPC cross-cultural risk identification expert, requiring the model to strictly analyze the input text and retrieved cross-cultural risk knowledge, and not to fabricate information. Task-level prompts guide the model to output structured cross-cultural risk identification results according to the preset process, specifically including five parts: risk type, risk level, cultural basis, mitigation suggestions, and original text basis.

[0034] S6. Output and Iteration: Output the cross-cultural risk identification results and receive human feedback to iteratively optimize the preset knowledge base and model parameters; The output process exports the cross-cultural risk identification results as a table or JSON document and embeds the identification results into the project management platform. When a high risk level of 4 or above is identified, an early warning is automatically triggered, and SMS and email notifications are sent to project managers. The feedback iteration process receives manual feedback from project managers and industry experts. Based on the feedback results, the cross-cultural risk knowledge entries in the preset knowledge base are updated. At the same time, the weights of the hybrid scoring algorithm, the fusion ratio of the hybrid retrieval, and the parameters of the large language model are adjusted to achieve continuous iterative optimization of the system.

[0035] Example 2 This embodiment uses a meeting minutes document from a hydropower station EPC project in Brazil as an example to illustrate the cross-cultural risk identification method for overseas EPC projects proposed in this invention, and details the specific application process of this method; specifically including: S1. Obtain the meeting minutes of this project. The content is as follows: The project manager emphasized at the morning meeting: We must complete all construction approvals within this week, otherwise you will have to bear the costs yourself.

[0036] S2. The document type is identified as meeting minutes, and the language is Chinese. No translation is required. A hybrid scoring algorithm is used to calculate the importance score of each paragraph of the text. Since the text length is short, the number of tokens after the initial compression is 32, which is far below the maximum session length limit of the large language model. Therefore, there is no need to start the dynamic compression process.

[0037] S3. Input the compressed text into the new-bge-m3 vector search engine and the BM25 full-text search engine respectively, and perform the search operation in parallel. Merge the two search results with a weight of 6:4 and generate a candidate set of knowledge items using the RRF algorithm. Then, use the re-ranking model to perform Top-3 re-ranking of the candidate set. The resulting knowledge items include the collectivist characteristics of Brazilian culture, the euphemistic requirements of Brazilian business communication, and typical cases of communication conflicts in overseas EPC projects. The relevance of the local search results is 0.82, which is higher than the preset threshold of 0.6, so there is no need to initiate an online search.

[0038] S4. Input the compressed text and retrieved cross-cultural risk knowledge into the Qwen3-30B-A3B model.

[0039] S5. The system guides the analysis using structured prompts. The prompt is: "You are an EPC cross-cultural risk identification expert. As an international engineering risk management expert, identify potential risks based on the input text and retrieved cross-cultural risk knowledge. Strictly follow the provided information for analysis. Do not fabricate content. The output must include five parts: risk type, risk level, cultural basis, mitigation suggestions, and original text basis." The task-level prompt is: "Please analyze the cross-cultural risks present in the meeting minutes of the above-mentioned EPC project for a hydropower station in Brazil, and output structured results as required." The model-generated cross-cultural risk identification results are as follows: the risk type is communication conflict, the risk level is level 3, and the cultural basis is that Brazilian culture emphasizes collectivism and humility. Publicly emphasizing personal responsibility and using a tough tone may cause resentment and communication awkwardness among local subcontractors, affecting the efficiency of subsequent collaboration. Mitigation suggestions include communicating on behalf of the team, such as using the expression "Let's work together to complete the construction approvals this week," using a tactful tone to remind project progress, and providing subcontractors with necessary support resources to help them complete the task on time. The original basis is the statement in the meeting minutes, "We must complete all construction approvals this week, otherwise you will bear the costs yourself."

[0040] S6. Export the identification results as an Excel spreadsheet document and embed it into the project management platform. Since the risk level is level 3, no automatic warning is triggered. After receiving the identification results, the project manager confirms the accuracy of the risk analysis based on their own experience, makes a feasible suggestion, and submits manual feedback. The system adds the case to the cross-cultural risk case library in the preset knowledge base, and at the same time fine-tunes the parameters of the reordering model to further improve the identification accuracy of Brazilian culture-related risks.

[0041] Example 3 See Figure 2 , Figure 2 This invention presents an architecture diagram of a cross-cultural risk identification system for overseas EPC projects, comprising: M1, Multi-source text acquisition module, is used to acquire text data of overseas EPC projects from multiple heterogeneous sources; M2, the text preprocessing module, is used to preprocess the text data to obtain compressed text with a unified context and a length that meets the input requirements of the large language model; M3, a hybrid retrieval module, is used to retrieve cross-cultural risk knowledge related to the compressed text from a preset knowledge base based on a hybrid retrieval strategy; M4, Model Input Module, is used to input the compressed text and the cross-cultural risk knowledge into the large language model; M5, a structured risk identification module, is used to guide the large language model to generate cross-cultural risk identification results through preset structured prompts; M6, Result Output and Iteration Module, is used to output the cross-cultural risk identification results and receive human feedback to iteratively optimize the preset knowledge base and model parameters.

[0042] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying cross-cultural risks in overseas EPC projects, characterized in that, include: Acquire text data from diverse and heterogeneous overseas EPC projects; The text data is preprocessed to obtain compressed text with a unified context and a length that meets the input requirements of a large language model; Based on a hybrid retrieval strategy, cross-cultural risk knowledge related to the compressed text is retrieved from a preset knowledge base; The compressed text and the cross-cultural risk knowledge are input into the large language model; The large language model is guided by preset structured prompts to generate cross-cultural risk identification results; The system outputs the cross-cultural risk identification results and receives human feedback to iteratively optimize the preset knowledge base and model parameters.

2. The method for identifying cross-cultural risks in overseas EPC projects according to claim 1, characterized in that, The preprocessing steps for the text data include: Identify the document type and language category of the text data; Translate non-Chinese text data into Chinese to obtain text data with a unified context.

3. The method for identifying cross-cultural risks in overseas EPC projects according to claim 1, characterized in that, The step of preprocessing the text data to obtain compressed text includes: A hybrid scoring algorithm is used to calculate the importance score of each paragraph in the text data, and paragraphs with scores higher than a preset threshold are retained to complete the initial compression. If the number of text tokens after the initial compression exceeds the session length limit of the large language model, a progressive multi-level compression strategy will be used for dynamic compression until the number of text tokens meets the preset conditions.

4. The method for identifying cross-cultural risks in overseas EPC projects according to claim 3, characterized in that, The calculation formula for the hybrid scoring algorithm is as follows: in, For paragraph Importance score For paragraph Keyword coverage score For paragraph TF-IDF weighted score.

5. The method for identifying cross-cultural risks in overseas EPC projects according to claim 3, characterized in that, The compression ratio calculation formula for the progressive multi-stage compression strategy is as follows: in, For the first The compression ratio of a rotary compressor. This is the initial compression ratio. This is the compression ratio step size; The preset conditions are: ,in For the compressed text number, This represents the maximum conversation length for a large language model. This represents the maximum generation length of a large language model. To reserve Buffer.

6. The method for identifying cross-cultural risks in overseas EPC projects according to claim 1, characterized in that, The steps for retrieving cross-cultural risk knowledge from a pre-set knowledge base based on a hybrid retrieval strategy include: Perform new-bge-m3 vector search and BM25 full-text search in parallel; The two search results are fused with a weight ratio of 6:4, and a candidate set of knowledge entries is generated using the RRF algorithm.

7. The method for identifying cross-cultural risks in overseas EPC projects according to claim 6, characterized in that, Also includes: The knowledge item candidate set is re-ranked using a re-ranking model, and a preset number of knowledge items are taken as the final retrieval results. When the relevance of the search results from the local preset knowledge base is lower than the preset threshold, an online search is initiated to obtain the latest cross-cultural risk knowledge, and the filtered online search results are added to the final search results.

8. The method for identifying cross-cultural risks in overseas EPC projects according to claim 1, characterized in that, The preset structured prompts include: system prompts and task-level prompts; The system prompts are used to define the role of the large language model as an EPC cross-cultural risk identification expert. The task-level prompts are used to guide the large language model to output cross-cultural risk identification results according to a preset process.

9. The method for identifying cross-cultural risks in overseas EPC projects according to claim 1, characterized in that, The cross-cultural risk identification results are structured data, including risk type, risk level, cultural basis, mitigation suggestions, and original text basis; The large language model adopted is the Qwen3-30B-A3B model, with a generation temperature set to 0.1 and a maximum generation length limit of 2048 tokens.

10. The method for identifying cross-cultural risks in overseas EPC projects according to claim 1, characterized in that, The step of outputting the cross-cultural risk identification result includes: Export the cross-cultural risk identification results as a table or JSON document; The cross-cultural risk identification results are embedded into the project management platform to trigger automatic early warnings.

11. A cross-cultural risk identification system for overseas EPC projects, characterized in that, A method for identifying cross-cultural risks in overseas EPC projects according to any one of claims 1-10, comprising: The multi-source text acquisition module is used to acquire text data from multi-source heterogeneous overseas EPC projects. The text preprocessing module is used to preprocess the text data to obtain compressed text with a unified context and a length that meets the input requirements of the large language model. A hybrid retrieval module is used to retrieve cross-cultural risk knowledge related to the compressed text from a preset knowledge base based on a hybrid retrieval strategy. The model input module is used to input the compressed text and the cross-cultural risk knowledge into the large language model; The structured risk identification module is used to guide the large language model to generate cross-cultural risk identification results through preset structured prompts; The results output and iteration module is used to output the cross-cultural risk identification results and receive human feedback to iteratively optimize the preset knowledge base and model parameters.

Images