A multi-model fusion vertical domain knowledge question answering method and system

Abstract

The application provides a multi-model fusion vertical field knowledge question answering method and system, belonging to the field of artificial intelligence natural language processing. The scheme disassembles the traditional single large model into multiple specialized models for collaborative work: first, an enhanced small parameter model trained for vertical fields is used for document analysis, and fixed position information is extracted in combination with rule assistance; then a general language model with lower parameters is used for intent understanding, supplemented by a prompt word strategy, a rule model or a knowledge graph; then based on the analyzed document information and the understood user intent, a larger parameter model is used for knowledge retrieval and answer generation; finally, a special small parameter model is used for style optimization to generate answers that meet the characteristics of the vertical field. The application also designs an answer backflow mechanism to realize continuous optimization of the system. The scheme greatly reduces resource consumption, improves the accuracy and efficiency of vertical field knowledge question answering, and solves the "hallucination" problem existing in the prior art.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and specifically to a multi-model fusion vertical domain knowledge question answering method and system. Background Art

[0002] With the rapid development of artificial intelligence (AI), knowledge question-answering systems (QAs), as a key application in natural language processing, have been widely adopted across multiple fields. These systems understand users' natural language questions, retrieve relevant information from a knowledge base, and generate accurate and reasonable answers. These systems have become crucial tools for acquiring expertise and solving problems, significantly improving information acquisition efficiency.

[0003] Existing technical solutions for knowledge question answering often utilize a single large model to handle the entire process, from intent understanding to answer generation. This single large model architecture performs well when handling general domain problems, but it has significant shortcomings when dealing with highly specialized, vertical domain knowledge with complex organizational modalities. For example, knowledge question answering in the engineering field requires not only understanding complex multimodal information such as design drawings and technical specifications, but also a deep professional background.

[0004] Due to the limitations of pre-training data for large models, a single model solution struggles to achieve both comprehensive knowledge and accuracy. This is especially true in specialized fields, where complex problems requiring precise knowledge often generate seemingly plausible but in fact erroneous information, a phenomenon known as "hallucination." This phenomenon is particularly pronounced in high-precision fields like engineering design and medical diagnosis, potentially leading to serious decision-making biases.

[0005] To improve model accuracy in specialized domains, the current mainstream approach is to increase the model's parameter size or fine-tune it using large amounts of domain data. However, this approach requires significantly more training resources and takes longer to train, significantly increasing development and application costs. This solution is difficult to implement in many resource-limited scenarios. Furthermore, simply increasing parameters does not fundamentally address the model's inadequate understanding of specialized knowledge.

[0006] Furthermore, existing single, large model architectures struggle to parse complex documents (such as engineering design drawings), leading to inaccurate retrieval results and impacting the quality of question-answering. A single model lacks the ability to optimize for diverse tasks, such as understanding intent, generating answers, and adjusting answer styles, making it difficult to achieve optimal results across all stages. Summary of the Invention

[0007] The purpose of the present invention is to provide a multi-model fusion vertical field knowledge question answering method and system, which can not only ensure the accuracy and authority of professional field knowledge questions and answers, but also have advantages in resource consumption and response speed, and meet the actual needs of vertical field knowledge services.

[0008] To achieve the above object, the present invention is implemented through the following technical solutions:

[0009] A multi-model fusion vertical domain knowledge question answering method includes the following steps:

[0010] S1: Use the enhanced small-parameter model trained for vertical domains to parse documents and extract structured information from documents;

[0011] S2: Use a general language model with fewer parameters to understand the intent of user questions;

[0012] S3: Based on the document information parsed in step S1 and the intent understanding results in step S2, a model with larger parameters is used for knowledge retrieval and answer generation;

[0013] S4: Use a dedicated small-parameter model to optimize the style of the answer generated in step S3 to generate a final answer that conforms to the characteristics of the vertical field.

[0014] Furthermore: the step S1 also includes: using rules to assist the enhanced small model in document parsing, and the rules include positioning fixed position information according to the standard layout of vertical field documents.

[0015] Furthermore, step S2 further includes: using at least one of a prompt word strategy, a rule model, or a knowledge graph as an auxiliary means to supplement the general language model for intent understanding.

[0016] Furthermore: the prompt word strategy includes: rewriting the user's question to optimize and expand the question and provide a clear generation direction.

[0017] Furthermore: Step S3 also includes: for vertical field problems whose complexity exceeds a preset threshold, first use a small parameter model to quickly filter and roughly sort knowledge, and then call a large parameter model to generate in-depth answers.

[0018] Furthermore: it also includes: returning the final answer generated in step S4 to the enhanced small model in step S1, so as to continuously optimize the document parsing effect and enhance the generalization of the parsing model.

[0019] A vertical domain knowledge question answering system that uses a multi-model fusion approach based on the above-mentioned question answering method includes:

[0020] The document parsing module is used to parse documents using an enhanced small-parameter model trained for vertical domains to extract structured information from documents.

[0021] An intent understanding module is used to understand the intent of user questions using a general language model with relatively low parameters; an answer generation module is used to perform knowledge retrieval and answer generation based on the document information parsed by the document parsing module and the intent understanding results of the intent understanding module using a model with relatively large parameters;

[0022] The style optimization module is used to optimize the style of the answer generated by the answer generation module using a dedicated small parameter model to generate a final answer that conforms to the characteristics of the vertical field.

[0023] Furthermore: the document parsing module also includes: a rule-assisted unit, which is used to use rules to assist the enhanced small model in document parsing, and the rules include positioning fixed position information according to the standard layout of vertical field documents.

[0024] Furthermore: the intention understanding module also includes: an auxiliary understanding unit, which is used to use at least one of the prompt word strategy, rule model or knowledge graph as an auxiliary means to supplement the general language model for intention understanding.

[0025] Furthermore: it also includes: a model optimization module, which is used to feed the final answer generated by the style optimization module back to the enhanced small model in the document parsing module, so as to continuously optimize the document parsing effect and enhance the generalization of the parsing model.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. This invention significantly improves the accuracy of knowledge question-answering in vertical domains. By using an enhanced small model trained for vertical domains to specifically parse documents, combined with rule-assisted extraction of fixed-position information, it can more accurately understand and process highly professional and complex documents, improving the quality of knowledge extraction from the source. At the same time, answer generation through a model with larger parameters ensures the fluency and logical coherence of the generated content, effectively reducing the occurrence of "hallucination" problems.

[0028] Second, this invention significantly reduces system resource consumption. By adopting a "division of labor and collaboration" model, models of various sizes are enabled to perform their respective roles. This transforms the pre-training process that originally required hundreds of computing cards into a training process that can be executed by just a few cards. This not only conserves computing resources but also reduces energy consumption, making the system more cost-effective and easier to deploy and apply.

[0029] Thirdly, the application significantly improves the response speed of the question and answer system. The intent understanding link uses a general language model with low parameters, which can quickly analyze the user's intent. For complex problems, the strategy of filtering by a small model and then generating in depth by a large model optimizes the processing efficiency while ensuring the quality of the answer. The style optimization link of the small parameter model is also more efficient than using a large model for style adjustment, effectively controlling the response time as a whole.

[0030] Fourthly, the application constructs a self-optimization mechanism for the system. By designing the final answer to flow back to the document parsing small model, the system performance is continuously optimized and the generalization ability is improved, so that the system can continuously adapt to new document formats and knowledge structures in practical applications, maintaining long-term effectiveness and applicability. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 A flowchart schematic diagram of a multi-model fusion vertical field knowledge question and answer method of the application in an embodiment;

[0032] Figure 2 A flowchart schematic diagram of a multi-model fusion vertical field knowledge question and answer method of the application in another embodiment. DETAILED DESCRIPTION

[0033] The technical solutions of the application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the application.

[0034] In the description of the application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore it cannot be understood as a limitation on the application. In addition, the terms "first", "second", "third" are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0035] The technical solutions provided by the application can be widely applied to the vertical field knowledge question and answer field with strong professionalism and complex organization modalities, such as but not limited to professional fields such as engineering design, medical health, legal consultation, etc. The following will take the engineering field as an example for detailed description.

[0036] In existing technologies, knowledge question-answering systems typically use a single large model to handle the entire process from intent understanding to answer generation. However, due to the limitations of the model's pre-training data, it is often difficult to achieve both comprehensiveness and accuracy when dealing with complex problems in specialized fields, and this can easily lead to "hallucination" problems. To address these issues, the present invention breaks down the answering process, which is handled by a large model, into multiple different models, each of which collaborates to complete the question-answering process. This approach balances the comprehensiveness of general knowledge understanding with the controllability and accuracy of specialized knowledge understanding and generation.

[0037] like Figure 1 As shown, the multi-model fusion vertical domain knowledge question answering method of the present invention includes the following steps:

[0038] S1: Use the enhanced small parameter model trained for vertical fields to parse the document and extract the structured information in the document, wherein the parameter amount of the small parameter model is at the level of 7M to 10M. As a pre-processing process for the application of large model knowledge, the result of document parsing is directly related to the accuracy of question and answer. Especially for documents with high professionalism and complex layout, the quality of the parsing effect is one of the most important factors affecting knowledge recall. For the engineering field in this embodiment, the model with smaller parameters can be fine-tuned to understand the format and content of engineering design drawings.

[0039] To further improve parsing performance, the present invention also incorporates a rule-assisted parsing mechanism into the document parsing step. For example, when processing engineering design drawings, the title bar and fixed area of ​​the drawing frame can be located in the lower right corner according to standard drawing formats such as ISO / A3 frames, and key information can be extracted from these specific locations. This combination of small models and rules not only saves computing resources but also allows for the production of effective vertical classification models with less training time.

[0040] S2: Use a general language model to understand the intent of user questions, where the number of parameters of the general language model is at the level of 100M to 1B. Studies have shown that the increase in the scale of model parameters is not proportional to the improvement in the accuracy of natural language reasoning tasks. According to a 2020 study by Brown et al., for every 10-fold increase in parameters, the accuracy of natural language reasoning tasks only increases by an average of 3-5%. Therefore, a low-parameter model of level 7B can be used for intent understanding. When processing text input by the user, the amount of data that needs to be processed and calculated is small, and the analysis and understanding of the user's intent can be completed in a shorter time, and a response can be given quickly.

[0041] However, fewer parameters may result in the model's inability to understand complex and diverse user questions, especially for questions that appear less frequently in the training data or are expressed in novel and unique ways. To address this problem, the present invention introduces auxiliary methods in the intent understanding step, which mainly include three categories:

[0042] The first is the prompt strategy, which rewrites the user's question, optimizes the expanded question, and provides clear generation direction, making the question more targeted and operational. For example, "What should be paid attention to in the design of rockfill dam?" is rewritten as "Please query the matters needing attention of rockfill dam in dam site selection, material selection, structure design, construction process, etc."

[0043] The second is the rule model, which uses its accuracy in processing specific pattern text and the advantage of machine learning model in some local features as a supplement to general language model to improve the overall performance of intent understanding.

[0044] The third is the knowledge graph, which serves as an external knowledge base to help the model better understand the semantic information in the text. For example, when the user mentions certain professional entities, the model can use the knowledge graph to obtain the attributes and relationships of the related entities, thereby more accurately understanding the user's intent.

[0045] The above three auxiliary means are used in the following principles when used: the prompt strategy is used to expand the user's intent; the rule model generally handles scenarios with fixed document templates and fixed question and answer formats, and is a supplement to the algorithm; the knowledge graph is used to improve the accuracy of recall.

[0046] S3: Based on the document information parsed in S1 step and the intent understanding result in S2 step, a large parameter model is used for knowledge retrieval and answer generation, wherein the parameter quantity of the large parameter model is 1B level or more. The size of the model parameters is positively correlated with the fluency and logical coherence of the answer generation. Models with more parameters usually generate more fluent and logical content because they can remember more details and knowledge. Therefore, using a larger parameter model can generate better quality answers.

[0047] It is worth noting that for extremely complex engineering knowledge, the present application also provides an optimization strategy: first use a small model for rapid filtering and rough sorting to determine the general direction and key points of the answer; then call a large model for deep generation to perfect the details and logical structure. This hierarchical processing method ensures both processing efficiency and the quality of the final answer.

[0048] S4: A dedicated small parameter model is used to optimize the style of the answer generated in step S3, generating a final answer that meets the characteristics of the vertical field. The parameter amount of the dedicated small parameter model is 7M to 10M level. There are mainly two methods for style optimization, one is to optimize based on the existing large language model prompt, and the other is to fill in the corresponding style training set and learn from the model to generate. For example, for the engineering field, a specific style optimization model can be trained to make the generated answer more rigorous, accurate, and standardized, in line with the expression habits and professional standards of engineering professionals.

[0049] In another embodiment, to achieve continuous optimization of the system, the application also designs an answer backflow mechanism step S5: the finally optimized answer is backflowed to the document parsing small model for continuous optimization of parsing effect and enhancement of generalization of the parsing model. The specific backflow mechanism is to first create a data set, then backflow the question and answer pairs into the data set, and apply them to model training after manual correction. In this way, the system can learn from actual applications and continuously adapt to new document formats and question types, maintaining long-term effectiveness.

[0050] In another embodiment, the application also provides a multi-model fusion vertical field knowledge question and answer system to implement the above method. The system includes a document parsing module, an intent understanding module, an answer generation module, and a style optimization module, corresponding to each step in the above method. Among them, the document parsing module includes a rule-assisted unit for using rule-assisted enhanced small models for document parsing; the intent understanding module includes an auxiliary understanding unit for using prompt word strategies, rule models, or knowledge graphs as auxiliary means; the system also includes a model optimization module for implementing an answer backflow mechanism.

[0051] In actual application, taking the engineering design field as an example, when the user asks the question "What should be paid attention to in the design of reinforced concrete frame structure?", the system first parses relevant engineering design specifications, design drawings, and other documents using the trained enhanced small model; then analyzes the user's intent through the intent understanding module to understand that the user is asking about design considerations; then, the large parameter model retrieves relevant knowledge and generates an answer based on the parsed document information and understood user intent, covering key considerations in the design of reinforced concrete frame structure; finally, the style optimization module adjusts the expression of the answer to conform to the professional expression habits of the engineering field, forming the final answer returned to the user.

[0052] By the above multi-model fusion manner, the application effectively solves the problems of insufficient accuracy, large resource consumption and slow response speed of a single large model in vertical field knowledge question answering, and provides a new technical path for a vertical field intelligent question answering system. Compared with the traditional method, the technical scheme of the application can significantly reduce resource consumption while improving the knowledge accuracy and response efficiency of the question answering system in the professional field, and has a wide application prospect.

[0053] The above embodiments are only for illustrating the technical concept and characteristics of the application, and the purpose is to enable those skilled in the art to understand the content of the application and implement it, and cannot limit the protection scope of the application. Any equivalent transformation or modification made according to the spirit and essence of the application shall be covered within the protection scope of the application.

Claims

1. A multi-model fusion vertical domain knowledge question answering method, characterized by: The following steps are involved: S1: Use an enhanced small-parameter model trained for vertical domains to parse documents and extract structured information from them, where the number of parameters in the small-parameter model is in the range of 7M to 10M. S2: Use a general language model to understand the intent of the user question, where the number of parameters of the general language model is at the level of 100M to 1B; S3: Based on the document information parsed in step S1 and the intent understanding results in step S2, a large parameter model is used for knowledge retrieval and answer generation, where the parameter size of the large parameter model is at least 1B level; S4: Use a dedicated small-parameter model to optimize the style of the answer generated in step S3 to generate a final answer that conforms to the characteristics of the vertical field, where the parameter amount of the dedicated small-parameter model is at the level of 7M to 10M.

2. A multi-model fusion vertical domain knowledge question answering method according to claim 1, characterized in that: The step S1 also includes: using rules to assist the enhanced small parameter model in document parsing, and the rules include positioning fixed position information according to the standard layout of vertical field documents.

3. The multi-model fusion vertical domain knowledge question answering method according to claim 1 is characterized in that: The step S2 further includes: using at least one of a prompt word strategy, a rule model or a knowledge graph as an auxiliary means to supplement the general language model for intent understanding.

4. A multi-model fusion vertical domain knowledge question answering method according to claim 3, characterized in that: The prompt word strategy includes: rewriting the user's question to optimize and expand the question and provide a clear generation direction.

5. The multi-model fusion vertical domain knowledge question answering method according to claim 1 is characterized in that: The step S3 also includes: for complex vertical field problems, first use a small parameter model to quickly filter and roughly sort knowledge, and then call a large parameter model to generate in-depth answers.

6. A multi-model fusion vertical domain knowledge question answering method according to claim 1, characterized in that: The method further includes step S5: feeding the final answer generated in step S4 back to the enhanced small-parameter model in step S1 to continuously optimize the document parsing effect and enhance the generalization of the parsing model.

7. A vertical domain knowledge question answering system using a multi-model fusion of the question answering method according to any one of claims 1 to 6, characterized in that: include: The document parsing module is used to parse documents using an enhanced small-parameter model trained for vertical domains to extract structured information from documents. Intent understanding module, which is used to understand the intent of user questions using a general language model with low parameters; An answer generation module, configured to perform knowledge retrieval and answer generation using a model with larger parameters based on the document information parsed by the document parsing module and the intent understanding result of the intent understanding module; The style optimization module is used to optimize the style of the answer generated by the answer generation module using a dedicated small parameter model to generate a final answer that conforms to the characteristics of the vertical field.

8. A multi-model fusion vertical domain knowledge question answering system according to claim 7, characterized in that: The document parsing module also includes: a rule-assisted unit, which is used to use rules to assist the enhanced small-parameter model in document parsing, and the rules include positioning fixed position information according to the standard layout of documents in vertical categories.

9. A multi-model fusion vertical domain knowledge question answering system according to claim 7, characterized in that: The intention understanding module also includes: an auxiliary understanding unit, which is used to use at least one of a prompt word strategy, a rule model or a knowledge graph as an auxiliary means to supplement the general language model for intention understanding.

10. A multi-model fusion vertical domain knowledge question answering system according to claim 7, characterized in that: Also includes: The model optimization module is used to feed the final answer generated by the style optimization module back to the enhanced small-parameter model in the document parsing module, so as to continuously optimize the document parsing effect and enhance the generalization of the parsing model.

Images