A plant document information retrieval method and device based on a large language model

Abstract

The application provides a factory document information retrieval method and device based on a large language model, and relates to the field of artificial intelligence.The method comprises the following steps: receiving question information about factory document information input by a user; preprocessing the question information; analyzing the preprocessed document information using a large language model to obtain processed text; constructing a query that captures key concepts and potential contradictions according to the processed text; sending the constructed and generated query to a factory document information retrieval model obtained by training a large language model based on a corpus to search a database, determining whether there is target answer text matching the question information, and outputting the target answer text.The application can improve the efficiency and accuracy of the information retrieval method, and further improve the information acquisition efficiency of factory technical research and development.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for retrieving factory document information based on a large language model. Background Technology

[0002] Information retrieval is crucial in the operation and management of modern factories. Modern factories, due to their numerous production stages, result in dispersed user terminals, a large volume of dynamic and timely information requiring low-cost, rapid integration and alignment, and diverse data presentation formats. Therefore, modern factories have a demand for rapid updates of dynamic and timely information and diverse presentation modalities in information retrieval. However, traditional factory information retrieval methods face challenges in addressing the increasing scale and complexity of factory operations, particularly in terms of information accumulation and updates, and information presentation modalities. Firstly, regarding information content accumulation, current common large language model-based information retrieval methods mostly rely on acquiring all document information from existing document databases or temporarily organizing manual input into the database to form a corpus. This method is inefficient, susceptible to subjective factors, and only suitable for static, systematic data. With the increasing frequency and wider distribution of dynamic and timely information, rapid information updates are essential for modern factories. Relying on manual updates is clearly insufficient. Therefore, it is necessary to develop and implement rapid information updates tailored to the characteristics of modern factories, building upon current large language model-based information retrieval technologies. Secondly, regarding information presentation modalities, most current methods employ keyword-based search. While this method can quickly locate documents containing specific keywords, the document-only search results still fall short of user needs. Therefore, it is necessary to develop diversified information presentation modalities that incorporate the characteristics of modern factories, building upon current information retrieval technologies based on large language models. For example, it could directly provide corresponding videos, images, tables, etc. In conclusion, current document retrieval methods based on large language models cannot meet the needs of modern factories for efficient and accurate document retrieval. Summary of the Invention

[0003] To overcome the shortcomings of the existing technology, this invention provides a method and apparatus for information retrieval in modern factories based on a large language model. This invention utilizes the powerful semantic understanding capabilities of a large language model to develop and implement rapid information updates and diversified information presentation modalities tailored to the characteristics of modern factories, thereby achieving efficient and accurate retrieval of factory document information.

[0004] To achieve the above objectives, in a first aspect, the present invention provides a factory document information retrieval method based on a large language model, comprising:

[0005] Receive user-input questions about factory documentation;

[0006] The query information is preprocessed;

[0007] The preprocessed document information is analyzed using a large language model to obtain the analyzed and processed text.

[0008] Based on the analyzed and processed text, construct a query that captures key concepts and potential contradictions;

[0009] The constructed and generated query is sent to the factory document information retrieval model obtained based on the corpus-trained large language model to search the database, determine whether there is a target answer text that matches the question information, and output it.

[0010] Furthermore, the preprocessing of the question information includes: removing noise from the question, and performing natural language processing such as word segmentation and part-of-speech tagging.

[0011] Furthermore, the step of sending the generated query to the factory document information retrieval model obtained based on the corpus-trained large language model for database retrieval, determining whether there is a target answer text matching the query information, and outputting it includes:

[0012] The factory document information retrieval model searches the database for documents or paragraphs related to the question information and accurately locates the documents related to the question, including: searching the database for relevant entries based on the generated query that captures key concepts and potential contradictions, and then extracting key facts or concepts related to the task from the retrieved entries;

[0013] After retrieving relevant documents, the semantic similarity between the question and the document content is calculated to determine whether there is a target answer text that matches the query information.

[0014] Furthermore, the method also includes:

[0015] Establishing a corpus includes:

[0016] The extracted data is cleaned;

[0017] The cleaned data is then structured and organized.

[0018] The structured data is stored in the corpus; the structured data includes static systematic data and dynamic time-sensitive data.

[0019] Furthermore, the method also includes:

[0020] Based on the corpus, the underlying large language model is pre-trained to obtain candidate models;

[0021] The training instruction set was used to fine-tune some network parameters and enhance text in the candidate model, and then the expression alignment model was used to refine it, resulting in the factory document information retrieval model.

[0022] Furthermore, the pre-training of the base language model based on the corpus yields candidate models including:

[0023] The model is trained on a corpus using unsupervised learning methods; it learns linguistic patterns and semantic information in the text by predicting the next word in the text or filling in masked words.

[0024] The processed corpus is input into the encoder and decoder of the transformer architecture to predict the target sequence and calculate the loss function;

[0025] The base model is obtained by training and converging the neural network, which serves as a preliminary candidate model.

[0026] Furthermore, the step of fine-tuning some network parameters and text enhancement in the candidate model using the training instruction set includes:

[0027] The large language model, attention layer, and semantic fusion layer together constitute the model encoder;

[0028] The model decoder part adopts a Transformer structure, decodes the output vector of the model decoder, classifies it through a softmax layer to obtain the final output, and updates the model parameters through the final loss function;

[0029] Text augmentation techniques are employed to enhance the diversity of training data and the robustness of the model.

[0030] Furthermore, determining whether there exists a target answer text matching the question information and outputting it includes:

[0031] If a target answer text matching the question information is found, the matching target answer question is output.

[0032] If a target answer text matching the question is found, meaning no correct answer was obtained, the question is uploaded to the expert client via an interface for further analysis. The expert then adds the analysis results to the corpus for proactive information updates.

[0033] Furthermore, the method also includes:

[0034] Using large language models for key fact identification and association with input text, specifically including:

[0035] 1) Feature extraction: Generate context-sensitive word embeddings for each text word using a large language model to accurately capture the meaning of the text words in a specific context;

[0036] 2) Dereference resolution: Identifying and connecting pronouns or synonyms in text that refer to the same entity, used to clarify the relevance of different parts of the text and the attribution of viewpoints;

[0037] 3) Contrastive sentence embedding: A sentence transformer is used to construct a vector representation for each sentence, and the differences in meaning and perspective between different paragraphs are quantified by calculating the distance between these vectors;

[0038] 4) Utilize large language models to identify discourse relationships between sentences;

[0039] 5) Based on the analysis results of steps 1)-4) above, the large language model outputs the probability of conflicting viewpoints between paragraphs and generates integrated information to be updated.

[0040] Secondly, the present invention provides a factory document information retrieval device based on a large language model, comprising:

[0041] The input module is used to receive user-input questions about factory documentation.

[0042] The preprocessing module is used to preprocess the query information;

[0043] The analysis module is used to analyze the preprocessed document information using a large language model to obtain the analyzed and processed text.

[0044] A construction module is used to build queries that capture key concepts and potential contradictions based on the analyzed and processed text;

[0045] The retrieval module is used to send the constructed and generated query to the factory document information retrieval model obtained based on the large language model trained on the corpus to search the database, determine whether there is a target answer text that matches the question information, and output it.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] This invention leverages the powerful semantic understanding capabilities of a large language model to develop and implement rapid information updates and diversified information presentation modalities tailored to the characteristics of modern factories, thereby achieving efficient and accurate factory document information retrieval. By improving the efficiency and accuracy of information retrieval methods, it enhances the efficiency of information acquisition for factory technology research and development. The adoption of update timing detection and rapid update mechanisms meets the needs of modern factories for rapid corpus updates across multiple devices; the ability to trace the responsible parties for input information meets the need for clarifying responsibilities in multi-device scenarios of modern factories; the large language model processes tagged combined text transcriptions and potentially improves links by identifying keywords or semantic connections between text segments and their corresponding media sources, satisfying the needs of different device users for different presentation formats of search results. Attached Figure Description

[0048] Figure 1 A flowchart illustrating a factory document information retrieval method based on a large language model, provided in an embodiment of the present invention;

[0049] Figure 2 A schematic diagram of a retrieval model based on the characteristics of a modern factory, provided in an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram illustrating the basic situation of the large language model information corpus provided in the embodiments of the present invention;

[0051] Figure 4 This invention provides a schematic diagram of information interaction in a factory document information retrieval system based on a large language model, as an embodiment of the present invention.

[0052] Figure 5 This is a schematic diagram of a factory document information retrieval device based on a large language model, provided as an embodiment of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] This invention first addresses the rapid updating of dynamic and timely information in modern factories. Based on understanding and parsing existing information uploaded from the client side, it achieves preprocessing and query generation, information retrieval and integration, and an update timing detection and rapid update mechanism. It uses a large language model to identify key facts and associate them with concepts derived from the input text, then checks for inconsistencies, omissions, or contradictions, and completes the tracing of input information. Secondly, to meet the diverse modalities required for information presentation in modern factories, in addition to text, it adds the conversion of voice, tables, physical images, and videos towards a text corpus, while establishing connections between modalities, thus making information presentation more diverse.

[0055] Figure 1 This is a flowchart illustrating a factory document information retrieval method based on a large language model, provided as an embodiment of the present invention. Figures 1 to 4 As shown, a factory document information retrieval method 100 based on a large language model includes:

[0056] S110: Receive user input regarding questions about factory documentation;

[0057] In step S110, in order to quickly update the dynamic and timely information of a modern factory, the system needs to receive and upload questions about factory document information entered by the user on the parsing end. These questions may involve various aspects such as equipment operation, maintenance guidelines, safety procedures, and product specifications.

[0058] S120: Preprocess the query information;

[0059] In step S120, before inputting the query into the factory document information retrieval model, the system preprocesses the query. Optionally, the preprocessing includes removing noise (such as punctuation marks, special characters, etc.) from the query, and performing natural language processing steps such as word segmentation and part-of-speech tagging, so that the model can better understand the intent and content of the query.

[0060] S130: Use a large language model to analyze the preprocessed document information to obtain the analyzed and processed text;

[0061] In step S130, a large language model is used to analyze the preprocessed text, remove toxic data, and remove annotations, links, images, HTML tags, and semantically irrelevant content; stop words in the collected text are also removed.

[0062] S140: Based on the analyzed and processed text, construct a query that captures key concepts and potential contradictions;

[0063] In step S140, a query is constructed to capture key concepts and potential contradictions. The generated query is sent to an information retrieval system to find relevant entries in the database and extract key facts or concepts related to the task from the retrieved entries.

[0064] S150: The constructed and generated query is sent to the factory document information retrieval model obtained based on the corpus-trained large language model to search the database, determine whether there is a target answer text that matches the question information, and output it.

[0065] In step S150, the system inputs the preprocessed and analyzed query information into the pre-trained factory document information retrieval model. The model searches the factory document information retrieval corpus for documents or paragraphs related to the query based on its content and intent. During this process, the model utilizes the language knowledge and factory document-related knowledge learned during pre-training and fine-tuning to quickly and accurately locate documents related to the query. After retrieving relevant documents, the system needs to determine if there is a target answer text matching the query information. This is achieved by calculating the semantic similarity between the query and the document content. If a document or paragraph has a high semantic similarity to the query and contains a clear answer to the query, then it can be considered the target answer text.

[0066] According to the above embodiments, by leveraging the powerful semantic understanding capabilities of large language models, we can develop and implement rapid information updates and diversified information presentation modalities that combine the characteristics of modern factories, thereby achieving efficient and accurate factory document information retrieval.

[0067] Furthermore, in some embodiments, the step of sending the generated query to a factory document information retrieval model based on a corpus-trained large language model for database retrieval, determining whether a target answer text matching the query information exists, and outputting it includes:

[0068] The factory document information retrieval model searches the database for documents or paragraphs related to the question information and accurately locates the documents related to the question, including: searching the database for relevant entries based on the generated query that captures key concepts and potential contradictions, and then extracting key facts or concepts related to the task from the retrieved entries;

[0069] After retrieving relevant documents, the semantic similarity between the question and the document content is calculated to determine whether there is a target answer text that matches the query information.

[0070] Furthermore, in some embodiments, the establishment of the corpus includes:

[0071] 1) Data Import and Storage: The extracted data is cleaned to remove redundant information, errors, and noise. The structured data is then imported into the corpus. This can be done by writing data migration scripts or using ETL tools. A distinction is made between static, systematic data and dynamic, time-sensitive data.

[0072] -Static, systematic data such as: CNC machining, inspection, general machining, fast spinning, slow spinning, blanking, etc.

[0073] - Dynamic and timely data, for example: φ60*30, rough machining allowance of 0.7mm, rough machining of the shape and cavity according to the programming drawing, leaving a finishing allowance of 0.6-0.8mm, checking the finishing allowance of each part according to the mold drawing, 0.6-0.9mm, attaching the corresponding order code to the top for easy customer identification, rough machining of the head shape according to the programming drawing (leaving an allowance of 0.6-0.8), leaving a process straight platform of Φ30x5 at the head, rough machining of the tail end, leaving a finishing allowance of 0.6-0.8mm, the flat platform on the 340 dimension side in the rough opening view has been cut to the correct shape, etc.

[0074] 2) Data storage: Design a reasonable database structure to store data, such as... Figure 3 As shown.

[0075] - For static, systematic data, use a relational database (such as MySQL or PostgreSQL);

[0076] - For dynamic, time-sensitive data, use NoSQL databases (such as MongoDB and Cassandra) to support high concurrency and flexible data structures.

[0077] 3) Design the upload interface

[0078] Interface Specification: Define the technical specifications of the upload interface, such as whether to use a RESTful API or SOAP protocol, and the data transmission format (e.g., JSON, XML).

[0079] Security: Ensure the interface uses HTTPS protocol for data transmission to achieve encrypted data transmission. Simultaneously, implement authentication and access control to ensure that only authorized users can upload data.

[0080] Furthermore, a factory document information retrieval model is established—a candidate model is obtained by pre-training the underlying large language model based on a corpus. Then, a training instruction set is used to fine-tune some network parameters and enhance the text in the candidate model. Finally, an expression alignment model is used for polishing to obtain the factory document information retrieval model. Afterwards, as... Figure 2 and Figure 4As shown, employees use the factory document information retrieval model to ask and answer questions about factory document information. The system receives user questions and preprocesses them to determine whether there is a target answer text that matches the query information. If there is an answer text, the system returns the matching target answer question.

[0081] Specifically, in some embodiments, the establishment of the factory document information retrieval model includes:

[0082] (1) Pre-training the base language model based on the corpus to obtain candidate models; more specifically, the pre-training process includes:

[0083] Unsupervised learning methods, such as autoregressive models (e.g., GPT series) or masked language models (e.g., BERT series), are trained on a corpus. By predicting the next word in the text or filling in masked words, the model learns linguistic patterns and semantic information within the text. The processed corpus is then input into the encoder and decoder of a transformer architecture to predict the target sequence and calculate the loss function. The model is trained through a neural network until convergence, resulting in a base model as a preliminary candidate model. However, at this stage, the model's performance is insufficient for direct application in document information retrieval.

[0084] (2) The training instruction set is used to fine-tune some network parameters and perform text enhancement on the candidate model. Then, the expression alignment model is used for polishing to obtain the factory document information retrieval model. More specifically, the fine-tuning and text enhancement process includes:

[0085] The fine-tuning process involves a set of instructions, including specific queries and responses for documents corresponding to the characteristics of modern factories, keyword annotation, and document classification labels. The model undergoes supervised learning for these specific tasks to optimize its performance. The adaptive fine-tuning training includes: constructing a model encoder by combining a large language model, an attention layer, and a semantic fusion layer; the decoder uses a Transformer structure to decode the decoder's output vector, classifies it through a softmax layer to obtain the final output, and updates the model parameters using the final loss function. To enhance the model's text processing capabilities, text enhancement techniques are employed, such as synonym replacement, random insertion, deletion, or swapping of words, to increase the diversity of training data and the model's robustness.

[0086] (3) Information update function implementation - If the correct search answer is not obtained, you can choose to actively update. Specifically, the question is uploaded to the expert for subsequent analysis through the interface, and then the result is added to the information corpus for updating, thus completing the closed loop of the information retrieval model.

[0087] Furthermore, in some embodiments, determining whether there is a target answer text matching the question information and outputting it includes:

[0088] If a target answer text matching the question information is found, the matching target answer question is output.

[0089] If a target answer text matching the question is found, meaning no correct answer was obtained, the question is uploaded to the expert client via an interface for further analysis. The expert then adds the analysis results to the corpus for proactive information updates.

[0090] Furthermore, in embodiments of the present invention, the method further includes: using a large language model to identify key facts and associate them with the input text, specifically including:

[0091] 1) Feature Extraction: A large language model generates context-sensitive word embeddings for each text word. These embeddings accurately capture the meaning of words within their specific context. Specifically, based on the characteristics of tasks related to factory processes and combined with the features of discrete tabular data, when dealing with structured data such as part names and dimensions, structured data embeddings are generated, mapping discrete categorical variables to continuous numerical vectors. When dealing with unstructured data such as complex process descriptions, ordinary text embedding generation techniques are applied to generate corresponding word embeddings. Furthermore, user account information is incorporated into the word embedding generation process, which helps the large model, after training, to identify the responsible person for each information fragment by decoding the corresponding word embedding. These embeddings capture the meaning of words based on their surrounding context.

[0092] 2) Dereference Resolution: Identifying and connecting pronouns or synonyms in text that refer to the same entity helps clarify the connections and viewpoints between different parts of the text. Specifically, identifying and linking pronouns or synonyms in text that refer to the same entity (person, place). This helps in understanding who holds which viewpoint.

[0093] 3) Contrastive Sentence Embeddings: A sentence transformer is used to construct a vector representation for each sentence, and the differences in meaning and perspective between different paragraphs are quantified by calculating the distances between these vectors. Specifically, based on this, contrastive sentence embeddings are used, and a sentence transformer is used to generate a vector representation for each sentence. Calculating the distances between these vectors to quantify the differences in meaning and perspective between different paragraphs may require experimentation to find the optimal threshold for a specific task, thus enabling discourse relation classification.

[0094] 4) Next, we train a large language model to identify discourse relationships between sentences, such as causal relationships and contrastive relationships. In particular, we focus on identifying contrastive relationships between paragraphs within a shorter text span to reveal potential conflicts of opinion.

[0095] 5) Finally, based on the analysis results of steps 1)-4) above, the large language model can output the probability of conflicting viewpoints between paragraphs. Based on the generated and integrated information to be updated, this information may be uploaded in a form requiring further confirmation from experts or the team, or presented through questions, to trigger the corresponding update mechanism and ensure the accuracy and timeliness of the information. The entire process aims to provide timely and accurate decision support to the expert team through the deep analysis and precise output of the large language model. Therefore, the above-mentioned determination of whether to trigger the update mechanism serves as a real-time reminder to experts or the team for confirmation (e.g., ...). Figure 2 and Figure 4 This generates integrated information to be updated, which can be uploaded either as a form requiring only confirmation or as a question. In short, this will meet the needs of modern factories for rapid corpus updates across multiple devices.

[0096] According to the above embodiments of the present invention, an update timing detection and rapid update mechanism is established. A large language model is used to identify key facts and related concepts to the input text, while ensuring that the illusion of "pretending to know what you don't know" is not generated. Then, it is checked whether there are inconsistencies, omissions or contradictions.

[0097] Additionally, some embodiments of the present invention include: tracing the person in charge of the input information, including:

[0098] For any information in a corpus, the large language model can be used at any time to trace the information based on its understanding of natural language. For structured information, the input text is parsed in a structured manner, and the uploader's account is included. For content that is difficult to express linearly, such as processes or techniques, the large language model integrates updated information into the original information while retaining the source tag for later accountability. The implementation process is as follows: The large language model first generates an embedding that captures the semantics of each sentence in the original and merged paragraphs. Given the embedding of a specific sentence, the large language model performs a similarity search on the embeddings of all sentences in the original paragraph, identifying the sentence in the original paragraph that is closest to the specific sentence's embedding. Then, by comparing the embeddings of smaller segments (phrases or keywords) within the sentence, the parts in the merged paragraph that closely match the specific sentence are highlighted. If an attention mechanism is used for training, analyzing the attention weights assigned during the merging process reveals which parts of the original paragraph contribute most to each part of the merged paragraph. In summary, this will meet the need for clarifying responsibilities in multi-terminal scenarios in modern factories.

[0099] Preferably, in some embodiments, to meet the need for diversified information presentation modalities, in addition to text, the conversion of audio, tables, images, and videos towards a text corpus is added. For example, for audio, Automatic Speech Recognition (ASR) is used to convert audio into text; for tables and images, Optical Character Recognition (OCR) is used to extract text from tables; for videos, ASR is combined for speech recognition and OCR for on-screen text recognition. Simultaneously, connections are established between modalities, making information presentation more diverse. Tags or timestamps corresponding to the original sources (audio, tables, OCR, video) are inserted into the generated text transcription. A large language model processes the tagged combined text transcription and potentially improves links by identifying keywords or semantic connections between text segments and their corresponding media sources. This will meet the needs of different users on different devices for different presentation formats of search results.

[0100] Furthermore, some embodiments of the present invention also include: test deployment and maintenance, specifically including:

[0101] 1) Testing and Deployment

[0102] Unit testing: Perform unit tests on the upload interface and data structuring tool to ensure that each module works properly.

[0103] Integration testing: Perform integration testing on the entire system to verify whether the collaboration between the various modules is normal.

[0104] Performance testing: Test the system's performance, including response time, throughput, and concurrent processing capabilities. Ensure the system meets business requirements.

[0105] Deployment and Go-live: Deploy the system to the production environment and monitor and maintain it to ensure stable operation.

[0106] 2) User Support and Maintenance

[0107] User training: Provide users with the necessary training and support to ensure they can use the upload interface and query the corpus proficiently.

[0108] Regular maintenance: Regularly maintain and upgrade the system to fix potential problems and add new features.

[0109] Data backup: Regularly back up the data in the corpus to prevent data loss or corruption.

[0110] The above is an introduction to the method embodiments. The following system embodiments will further illustrate the solution of the present invention.

[0111] Figure 5 A schematic diagram of a factory document information retrieval device based on a large language model is provided for an embodiment of the present invention, as shown below. Figure 5 As shown, a factory document information retrieval device 200 based on a large language model includes:

[0112] Input module 210 is used to receive user input of questions about factory documentation information;

[0113] Preprocessing module 220 is used to preprocess the query information;

[0114] Analysis module 230 is used to analyze the preprocessed document information using a large language model to obtain the analyzed text;

[0115] Module 240 is used to construct a query that captures key concepts and potential contradictions based on the analyzed and processed text;

[0116] The retrieval module 250 is used to send the constructed and generated query to the factory document information retrieval model obtained based on the large language model trained on the corpus to retrieve the database, determine whether there is a target answer text that matches the question information, and output it.

[0117] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A factory document information retrieval method based on a large language model, characterized in that, include: Receive user-input questions about factory documentation; The query information is preprocessed; The preprocessed document information is analyzed using a large language model to obtain the analyzed and processed text. Based on the analyzed and processed text, construct a query that captures key concepts and potential contradictions; The constructed and generated query is sent to the factory document information retrieval model obtained based on the corpus-trained big language model to search the database, determine whether there is a target answer text that matches the question information, and output it. The method further includes: Using large language models for key fact identification and association with input text, specifically including: 1) Feature extraction: Generate context-sensitive word embeddings for each text word using a large language model to accurately capture the meaning of the text words in a specific context; 2) Dereference resolution: Identifying and connecting pronouns or synonyms in text that refer to the same entity, used to clarify the relevance of different parts of the text and the attribution of viewpoints; 3) Contrastive sentence embedding: Using a sentence transformer to construct a vector representation for each sentence, and quantifying the differences in meaning and perspective between different paragraphs by calculating the distance between these vectors; 4) Utilize large language models to identify discourse relationships between sentences; 5) Based on the analysis results of steps 1)-4) above, the large language model outputs the probability of conflicting viewpoints between paragraphs and generates integrated information to be updated.

2. The method according to claim 1, characterized in that, in, Preprocessing the query information includes removing noise from the query and performing natural language processing such as word segmentation and part-of-speech tagging.

3. The method according to claim 1, characterized in that, The generated query is sent to the factory document information retrieval model obtained based on the corpus-trained large language model to search the database, determine whether there is a target answer text that matches the query information, and output it, including: The factory document information retrieval model searches the database for documents or paragraphs related to the question information and accurately locates the documents related to the question, including: searching the database for relevant entries based on the generated query that captures key concepts and potential contradictions, and then extracting key facts or concepts related to the task from the retrieved entries; After retrieving relevant documents, the semantic similarity between the question and the document content is calculated to determine whether there is a target answer text that matches the query information.

4. The method according to claim 1, characterized in that, The method further includes: Establishing a corpus includes: The extracted data is cleaned; The cleaned data is then structured and organized. The structured data is stored in the corpus; the structured data includes static systematic data and dynamic time-sensitive data.

5. The method according to claim 4, characterized in that, The method further includes: Based on the corpus, the underlying large language model is pre-trained to obtain candidate models; The training instruction set was used to fine-tune some network parameters and enhance text in the candidate model, and then the expression alignment model was used to refine it, resulting in the factory document information retrieval model.

6. The method according to claim 5, characterized in that, The pre-training of the base language model based on the corpus yields candidate models including: The model is trained on a corpus using unsupervised learning methods; it learns linguistic patterns and semantic information in the text by predicting the next word in the text or filling in masked words. The processed corpus is input into the encoder and decoder of the transformer architecture to predict the target sequence and calculate the loss function; The base model is obtained by training and converging the neural network, which serves as a preliminary candidate model.

7. The method according to claim 6, characterized in that, The step of fine-tuning some network parameters and text enhancement in the candidate model using a training instruction set includes: The large language model, attention layer, and semantic fusion layer together constitute the model encoder; The model decoder part adopts a Transformer structure to decode the output vector of the model encoder, classifies it through a softmax layer to obtain the final output, and updates the model parameters through the final loss function; Text augmentation techniques are employed to enhance the diversity of training data and the robustness of the model.

8. The method according to claim 7, characterized in that, The step of determining whether there exists a target answer text that matches the question information and outputting it includes: If a target answer text matching the question information is found, the matching target answer question is output. If it is determined that there is no target answer text matching the question information, that is, no correct search answer is obtained, the question information is uploaded to the expert client through the interface for further analysis, and then the expert adds the analysis results to the corpus for active information update.

9. A factory document information retrieval device based on a large language model, used to implement the factory document information retrieval method based on a large language model as described in any one of claims 1-8, characterized in that, The device includes: The input module is used to receive user-input questions about factory documentation. The preprocessing module is used to preprocess the query information; The analysis module is used to analyze the preprocessed document information using a large language model to obtain the analyzed and processed text. A construction module is used to build queries that capture key concepts and potential contradictions based on the analyzed and processed text; The retrieval module is used to send the constructed and generated query to the factory document information retrieval model obtained based on the large language model trained on the corpus to search the database, determine whether there is a target answer text that matches the question information, and output it.

Images