Multi-language operation and maintenance question and answer method and system based on large model

Abstract

The invention relates to the technical field of operation and maintenance data processing, in particular to a multi-language operation and maintenance question and answer method and system based on a large model. Based on an operation and maintenance knowledge data set pair basic open source large model, an operation and maintenance user inputs questions into the large model, and an operation and maintenance knowledge database is constructed; performing iterative optimization on the large model by using data in the database regularly through a feedback iteration mechanism, and performing operation and maintenance knowledge multi-language question and answer by using the optimized model; the method has the beneficial effects that the system can quickly convert the operation and maintenance work order into multiple languages through a multi-language pre-training technology, so that the operation and maintenance service is not limited by the languages any more, global coverage is realized, and the customer satisfaction is improved.

Description

Technical Field

[0001] The present invention relates to the field of operation and maintenance data processing technology, and specifically to a multi-language operation and maintenance question-answering method and system based on a large model. Background Art

[0002] With the rapid evolution of artificial intelligence (AI), generative artificial intelligence (AIGC) has become a focal point in research and application, profoundly impacting both academic research and daily life. Large language models (LLMs), such as ChatGPT, simulate the interactive patterns of human conversations, enabling accurate parsing of user commands and the execution of diverse tasks, including but not limited to complex cognitive tasks such as question answering and text creation. This technological breakthrough not only demonstrates the enormous potential of AI in understanding natural language and generating high-quality content, but also opens up new research directions in the field of natural language processing (NLP). Leveraging their powerful natural language processing capabilities, large models can accurately understand complex questions or instructions posed by operations personnel, significantly reducing communication costs and improving work efficiency. Operations personnel no longer need to spend extensive time writing tedious scripts or query statements. Through simple conversations, they can obtain the required system status, troubleshooting suggestions, or operational guides.

[0003] However, as we expand into overseas markets, operations teams must not only handle technical requests and problem reports from around the world, but also ensure accurate communication and effective implementation of solutions. Traditional operations question-and-answer systems are often limited to a single language, making it difficult for non-native users to accurately express their questions, which in turn hinders the speed and efficiency of problem resolution. Therefore, a multilingual operations question-and-answer method based on a large model is needed. Summary of the Invention

[0004] The purpose of the present invention is to provide a multi-language operation and maintenance question-answering method and system based on a large model to solve the problems raised in the above background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a multi-language operation and maintenance question-answering method based on a large model, comprising the following steps:

[0006] Based on the operation and maintenance knowledge dataset, we pre-trained and fine-tuned the open-source large model using multilingual pre-training technology to build a large model capable of answering multilingual operation and maintenance knowledge questions. Pre-training uses a language-specific vocabulary approach, maintaining an independent vocabulary for each language. Through model learning, different languages are mapped to a shared semantic space, using a causal language model for training. Fine-tuning utilizes the operation and maintenance knowledge dataset to focus the model on operation and maintenance domain knowledge.

[0007] The operation and maintenance user inputs the problem into the large model. The input information includes long and short sentences, phrase expressions, and voice texts, and the results are returned to the user. When receiving user problems, multiple input forms are supported. Voice input is converted into text input through speech recognition technology. The input information is parsed and processed, named entity recognition is carried out using prompts, and retrieval-augmented generation (RAG) is adopted to solve the problems of hallucination, data timeliness, and data security during the generation of the large model. The parsed information is sent to the large model for prediction. The large model generates corresponding answers or solutions under the guidance of prompt engineering. The answers include answers in Chinese for operation and maintenance personnel and answers in the question language for directly replying to customers.

[0008] Build an operation and maintenance knowledge database, including the data used for training and the real question-and-answer information generated during the use of the large model.

[0009] Through the feedback iteration mechanism, the large model is iteratively optimized using the data in the database regularly. The scoring results of operation and maintenance personnel with rich operation and maintenance professional knowledge for the answers generated by the large model are collected and stored in the database. The large model answer data with wrong answers is analyzed to find the modification patterns, and the large model is intensively trained and updated using the corrected data set. After the training is completed, the improved large model is tested, and the scoring results are continuously collected to continuously update the data set and iteratively optimize the large model.

[0010] Use the optimized model for multi-language Q&A of operation and maintenance knowledge.

[0011] Preferably, in the step of building a large model with the ability to answer multi-language operation and maintenance knowledge, a data set containing multi-language operation and maintenance knowledge is sorted out and established, covering various operation and maintenance scenarios, problems, solutions, as well as relevant technical documents and discussions, and data cleaning and preprocessing are carried out.

[0012] Preferably, in the step of parsing and processing the input information, named entity recognition is carried out using prompts to identify the keywords in the problem, and the keywords are used as context information to make the large model focus on the information related to the keywords. When adopting retrieval-augmented generation (RAG), a vector database is added to store and process vector data. All fragments of operation and maintenance documents are vectorized and stored in the vector database. The question raised by the user is vectorized, and the cosine similarity algorithm is used to find the top k fragments most similar to the question vector in the vector database, which are used as context information and input into the large model.

[0013] Preferably, in the step of building an operation and maintenance knowledge database, the data set includes work order information over the years, operation and maintenance knowledge bases, question-and-answer generated during the use of the large model, and the scores of professional operation and maintenance personnel for the answers of the large model.

[0014] Preferably, in the step of iteratively optimizing the large model using the data in the database through a feedback iteration mechanism, after determining the modification mode, the corrected dataset is used to strengthen the training of the large model to update the model. After the training is completed, the improved large model is tested to ensure that while the large model improves its ability to answer incorrectly, it does not affect the questions that were answered correctly before the update.

[0015] A system for a multilingual operation and maintenance Q&A method based on a large model, comprising an information management module and a large model construction module;

[0016] The information management module is composed of an information input module, an information parsing module, an information Q&A module, and an information storage module;

[0017] The information input module is used to obtain multilingual customer operation and maintenance information input by the user;

[0018] The information parsing module is used to understand the user's information, extract the keywords in the question, and retrieve relevant information according to the keyword information. The information parsing module includes an information extraction module and an information retrieval module; the information extraction module uses named entity recognition technology and uses the large model to deeply understand the text, focusing on the key entities in the text; the information retrieval module is used to retrieve the text in the way of retrieval-enhanced generation, and send the retrieved results as context to the large model, so that the large model can give an answer based on the context;

[0019] The information Q&A module inputs the information related to the question into the large model, and generates an answer through the multilingual operation and maintenance large model and returns it to the user;

[0020] The information storage module is used to store the large model Q&A results in the database;

[0021] The large model construction module is composed of an operation and maintenance large model module and a feedback iteration module;

[0022] The operation and maintenance large model module is used for the deployment and use of the large model, and the processed question is input into this module to interact with the large model;

[0023] The feedback iteration module is used to optimize the multilingual operation and maintenance large model through a feedback iteration mechanism using the data in the database.

[0024] Preferably, the multilingual operation and maintenance large model in the large model construction module is constructed based on an operation and maintenance knowledge dataset for a basic open-source large model using multilingual pre-training technology and fine-tuning; the multilingual pre-training adopts a language-specific vocabulary method, maintains an independent vocabulary for each language, and maps different languages to a shared semantic space through model learning, and uses causal language model training; the fine-tuning uses the operation and maintenance knowledge dataset to make the model focus on operation and maintenance domain knowledge.

[0025] Preferably, the information input module supports multiple input forms, including long and short sentences, phrase expressions, and voice texts. Among them, voice input converts the user's voice into text input through voice recognition technology.

[0026] Preferably, the database stored in the information storage module includes the data used for training and the real question-and-answer information generated during the use of the large model. The data set covers the work order information over the years, the operation and maintenance knowledge base, the questions and answers generated during the use of the large model, and the scores given by professional operation and maintenance personnel to the answers of the large model.

[0027] Preferably, the feedback iteration module optimizes the large model in the following ways:

[0028] Collect the scoring results of operation and maintenance personnel with rich operation and maintenance professional knowledge on the answers generated by the multilingual large model and store them in the database. Conduct a detailed analysis of the large model answer data with wrong answers to find out the modification patterns. Use the corrected data set to perform intensive training on the large model to update the model. After the training is completed, test the improved large model to ensure that the large model can improve the ability to answer wrong questions without affecting the questions answered correctly before the update. Continuously collect the scoring results and continuously update the data set for iterative optimization.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] The multilingual operation and maintenance question-and-answer method and system based on a large model proposed by the present invention, through multilingual pre-training technology, the system can quickly convert operation and maintenance work orders into multiple languages, making operation and maintenance services no longer limited by language, achieving global coverage, and improving customer satisfaction. Improve the speed and accuracy of problem-solving. Based on the natural language processing and in-depth understanding capabilities of the large model, the system can accurately identify the core of the user's problem, provide accurate and detailed answers or operation guides, significantly accelerate the speed of problem-solving, and ensure the accuracy of the answers. For questions not in the knowledge base, the system can automatically generate natural and fluent operation and maintenance answers, reducing the workload of manual editing and improving work efficiency. Especially for foreign language work orders, there is no need to rely on the foreign language ability of customer service personnel, greatly reducing the operation difficulty. Users can obtain professional answers through simple questions without editing complex retrieval expressions, improving the user experience. At the same time, the intelligence and efficiency of the system also significantly improve the overall quality of operation and maintenance services. The system has high scalability and can easily adapt to the operation and maintenance needs of different fields. At the same time, through reinforcement learning technology, the system can continuously optimize the answer quality, keep up with the latest knowledge, and ensure long-term service effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flowchart of the method of the present invention;

[0032] Figure 2 This is the system block diagram of the present invention. Specific embodiments

[0033] In order to clearly and completely describe the purpose, technical solution of the present invention, and make the advantages more clearly understood, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are some embodiments of the present invention, rather than all embodiments, and are only used to explain the embodiments of the present invention, not to limit the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0034] Embodiment 1, please refer to Figure 1 , the present invention provides a technical solution: a multi-language operation and maintenance answering method based on a large model, mainly for operation and maintenance personnel involved in cross-border business, facilitating the daily use of operation and maintenance personnel. In order to enable those of ordinary skill in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be completely described below with reference to the accompanying drawings. Figure 1 This is the flowchart of the present invention.

[0035] As Figure 1 shown, the flowchart of the multi-language operation and maintenance question and answer method based on a large model proposed by the present invention includes the following steps:

[0036] Step 1: Based on the operation and maintenance knowledge data set, pre-train and fine-tune the basic open-source large model using multi-language pre-training technology to construct a large model with the ability to answer multi-language operation and maintenance knowledge.

[0037] Step 2: The operation and maintenance user inputs the question into the large model. The information input by the user includes but is not limited to long and short sentences, phrase expressions, voice texts, etc., and the result is returned to the user.

[0038] Step 3: Construct an operation and maintenance knowledge database, including the data used for training and the real question and answer information generated during the use of the large model.

[0039] Step 4: Regularly use the data in the database to iteratively optimize the large model through a feedback iteration mechanism.

[0040] Step 5: Use the optimized model to perform multi-language question and answer of operation and maintenance knowledge.

[0041] In step 1, first, a dataset containing multilingual operation and maintenance knowledge is organized and established, covering various operation and maintenance scenarios, problems, solutions, as well as relevant technical documents and discussions, and data cleaning and preprocessing are carried out. To achieve multilingual pre-training, the method of language-specific vocabulary tables is adopted, maintaining an independent vocabulary table for each language, and mapping different languages to a shared semantic space through the learning of the model. During the pre-training process, a causal language model is used to train the model, requiring the model to predict the next word based on the previous word, which helps the model learn the statistical patterns and semantic knowledge of the language. On the basis of pre-training, the model is fine-tuned using the operation and maintenance knowledge dataset to make the model more focused on the knowledge in the operation and maintenance field and improve the performance of the model in actual applications.

[0042] In step 2, first, the questions raised by users are received, supporting various input forms, including but not limited to long and short sentences, phrase expressions, and speech texts. Among them, for speech input, through speech recognition technology, the user's speech is converted into text input. After the user submits the input, the input information needs to be parsed and processed so that it can be accurately understood by the large model. Prompt is used for named entity recognition (NER), and the large model is used to deeply understand the text, identify the keywords in the question, and use the keywords as context information to make the large model more focused on the information related to the keywords, improve the large model's understanding ability of the question, and accurately control the detail level of the generated text content. To address the issues of hallucination during the generation of the large model, data timeliness, and data security, retrieval-augmented generation (RAG) is adopted to solve this problem. A vector database is added for storing and processing vector data. The data is stored in vector form and can be retrieved and queried through the similarity between vectors. All operation and maintenance document fragments are vectorized and stored in the vector database. The question raised by the user is vectorized, and the cosine similarity algorithm is used to find the top k most similar fragments in the vector database to the question vector, which are used as context information and input into the large model. The parsed information is sent to the constructed large model for prediction. The large model will, under the guidance of the prompt engineering prompt, reason based on its own knowledge base and the provided context information to generate corresponding answers or solutions. For the answers given by the large model, there are answers in Chinese for operation and maintenance personnel and answers in the question language for directly replying to customers.

[0043] In step 3, building an operation and maintenance knowledge base is a key process. It aims to integrate training data and the real Q&A information generated during the use of the large model to form a comprehensive and dynamic operation and maintenance knowledge resource pool, supporting the smooth progress of operation and maintenance work and the continuous optimization of the large model. The dataset includes work order information over the years, the operation and maintenance knowledge base, Q&A generated during the use of the large model, and the scores given by professional operation and maintenance personnel to the answers of the large model.

[0044] In step 4, iteratively training the large model regularly using the data in the database through a feedback iteration mechanism is a crucial step to ensure that the large model continuously adapts to operation and maintenance requirements and improves the accuracy and practicality of answers. First, collect the scoring results. Operation and maintenance personnel with rich operation and maintenance professional knowledge score the answers generated by the multilingual large model and store the scoring results in the database. Second, determine the modification mode. For a certain type of question, if the large model gives incorrect answers, it is necessary to conduct a detailed analysis of the incorrect answer data of the large model to find a suitable modification mode. Use the corrected dataset to intensively train and update the large model to improve its answering ability. After training, it is necessary to test the improved large model to ensure that while improving the ability to answer incorrectly, it does not affect the questions that were answered correctly before the update. Finally, it is necessary to continuously collect the scoring results, continuously update the dataset, and continuously iterate to optimize the answering ability of the large model.

[0045] In step 5, use the iterated multilingual operation and maintenance large model as the final application model to generate the operation and maintenance knowledge required for operation and maintenance and foreign language answers that can be directly used to reply to customers according to the operation and maintenance questions raised by users, achieving effective answers for multilingual operation and maintenance.

[0046] Example 2, based on Example 1, proposes as Figure 2As shown in the figure, the multi-language operation and maintenance Q&A system based on large models proposed by the present invention is composed of an information management module and a large model construction module. The information management module consists of an information input module, an information parsing module, an information Q&A module, and an information storage module. The information parsing module includes an information extraction module and an information retrieval module; the large model construction module is composed of an operation and maintenance large model module and a feedback iteration module. The information input module is used to obtain multi-language customer operation and maintenance information input by users; the information parsing module is used to understand the information of users, extract the keywords in the questions, and retrieve relevant information according to the keyword information; the information Q&A module inputs the information related to the questions into the large model, and generates answers through the multi-language operation and maintenance large model and returns them to users; the information storage module is used to store the large model Q&A results in the database; the information extraction module utilizes the named entity recognition technology, deeply understands the text by using the large model, and focuses on the key entities in the text; the information retrieval module is used to retrieve and generate the text in an enhanced manner, and send the retrieved results as context to the large model, so that the large model can give answers based on the context; the operation and maintenance large model module is used for the deployment and use of the large model, and the questions after information processing are input into this module to interact with the large model; the feedback iteration module is used to optimize the multi-language operation and maintenance large model through the feedback iteration mechanism by using the data in the database.

[0047] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi - language operation and maintenance Q&A method based on large models, characterized in that: It includes the following steps: Based on the operation and maintenance knowledge dataset, pre-train and fine-tune the basic open-source large model using multi-language pre-training technology to build a large model with the ability to answer multi-language operation and maintenance knowledge; For pre-training, use the language-specific vocabulary method, maintain independent vocabularies for each language, and map different languages to a shared semantic space through model learning, and use causal language model training; for fine-tuning, use the operation and maintenance knowledge dataset to make the model focus on operation and maintenance domain knowledge; The operation and maintenance user inputs the question into the large model. The input information includes long and short sentences, phrase expressions, and voice texts, and the result is returned to the user; when receiving the user's question, multiple input forms are supported. The voice input is converted into text input through speech recognition technology, the input information is parsed and processed, named entity recognition is performed using prompt, and retrieval-augmented generation (RAG) is used to solve the problems of hallucination, data timeliness, and data security during the generation of the large model. The parsed information is sent to the large model for prediction. Under the guidance of the prompt engineering prompt, the large model generates corresponding answers or solutions. The answers include answers in Chinese for operation and maintenance personnel and answers in the question language for directly replying to customers; Build an operation and maintenance knowledge database, including the data used for training and the real question-and-answer information generated during the use of the large model; Through the feedback iteration mechanism, regularly use the data in the database to iteratively optimize the large model; Collect the scoring results of operation and maintenance personnel with rich operation and maintenance professional knowledge on the answers generated by the large model and store them in the database. Analyze the large model answer data with wrong answers to find the modification patterns, use the corrected dataset to strengthen the training of the large model to update the model. After the training is completed, test the improved large model, continuously collect the scoring results, and continuously update the dataset to iteratively optimize the large model; Use the optimized model for multi-language Q&A of operation and maintenance knowledge.

2. The multilingual operation and maintenance Q&A method based on a large model according to claim 1, wherein: In the step of building a large model with the ability to answer multi-language operation and maintenance knowledge, organize and establish a dataset containing multi-language operation and maintenance knowledge, covering various operation and maintenance scenarios, questions, solutions, and related technical documents and discussions, and perform data cleaning and preprocessing.

3. A multi-language operation and maintenance Q&A method based on a large model according to claim 2, characterized in that: In the step of parsing and processing the input information, use prompt for named entity recognition to identify the keywords in the question, and use the keywords as context information to make the large model focus on the information related to the keywords; When using retrieval-augmented generation (RAG), add a vector database for storing and processing vector data. Vectorize all operation and maintenance document fragments and store them in the vector database. Vectorize the question raised by the user, and use the cosine similarity algorithm to find the top k fragments most similar to the question vector in the vector database, and use them as context information to input into the large model.

4. The multilingual operation and maintenance Q&A method based on a large model according to claim 3, wherein: In the step of building the operation and maintenance knowledge database, the dataset includes work order information over the years, operation and maintenance knowledge bases, question-and-answer generated during the use of the large model, and the scores of professional operation and maintenance personnel on the answers of the large model.

5. The multi - language operation and maintenance Q&A method based on a large - model according to claim 4, wherein: In the step of iteratively optimizing the large model regularly using the data in the database through a feedback iteration mechanism, after determining the modification mode, the corrected dataset is used to perform reinforcement training on the large model to update the model. After the training is completed, the improved large model is tested to ensure that while the large model improves its ability to answer incorrectly, it does not affect the questions that were answered correctly before the update.

6. A system for the multi - language operation and maintenance Q&A method based on a large - model according to claim 5, characterized in that: It includes an information management module and a large model construction module; The information management module consists of an information input module, an information parsing module, an information Q&A module, and an information storage module; The information input module is used to obtain multilingual customer operation and maintenance information input by the user; The information parsing module is used to understand the user's information, extract the keywords in the question, and retrieve relevant information according to the keyword information. The information parsing module includes an information extraction module and an information retrieval module; the information extraction module uses named entity recognition technology to deeply understand the text using the large model, focusing on the key entities in the text; the information retrieval module is used to retrieve the text in the way of retrieval-augmented generation, and send the retrieved results as context into the large model, so that the large model can give answers based on the context; The information Q&A module inputs the information related to the question into the large model, and generates an answer through the multilingual operation and maintenance large model and returns it to the user; The information storage module is used to store the large model Q&A results in the database; The large model construction module consists of an operation and maintenance large model module and a feedback iteration module; The operation and maintenance large model module is used for the deployment and use of the large model, and the question after information processing is input into this module to interact with the large model; The feedback iteration module is used to optimize the multilingual operation and maintenance large model through the feedback iteration mechanism using the data in the database.

7. A system according to claim 6, characterized in that: The multilingual operation and maintenance large model in the large model construction module is constructed based on the operation and maintenance knowledge dataset for the basic open-source large model using multilingual pre-training technology and fine-tuning; multilingual pre-training adopts the language-specific vocabulary method, maintains an independent vocabulary for each language, and maps different languages to a shared semantic space through model learning, and uses causal language model training; fine-tuning uses the operation and maintenance knowledge dataset to make the model focus on operation and maintenance domain knowledge.

8. A system according to claim 7, wherein: The information input module supports multiple input forms, including long and short sentences, phrase expressions, and speech texts, where the speech input converts the user's speech into text input through speech recognition technology.

9. A system according to claim 8, wherein: The database stored in the information storage module includes the data used for training and the real Q&A information generated during the use of the large model. The dataset covers the work order information over the years, the operation and maintenance knowledge base, the Q&A generated during the use of the large model, and the scores given by professional operation and maintenance personnel to the large model answers.

10. A system according to claim 9, characterized in that: The feedback iteration module optimizes the large model in the following ways: Collect the scoring results of operation and maintenance personnel with rich operation and maintenance expertise on the answers generated by the multilingual large model and store them in the database. Conduct a detailed analysis of the large model's answer data with incorrect answers to find modification patterns. Use the corrected dataset to perform reinforcement training on the large model to update the model. After the training is completed, test the improved large model to ensure that while the large model improves its ability to answer incorrectly, it does not affect the questions that were answered correctly before the update. Continuously collect the scoring results, continuously update the dataset, and perform iterative optimization.

Images