A scene-aware based knowledge query sentence intelligent generation method and system

Abstract

The present application relates to a kind of big model few sample context learning driven knowledge query sentence intelligent generation method and system based on scene perception.The steps of the method include: the logical skeleton analysis of the natural language question input by user, i.e.target question;According to target question, retrieve relevant knowledge components from knowledge base;Use big model to carry out few sample context learning, realize the selection of knowledge components;According to logical skeleton and selected knowledge components, construct logical expression, obtain knowledge query sentence.The present application combines scene learning and knowledge query sentence generation, not only improves the accuracy and execution efficiency of knowledge question and answer, but also significantly reduces the cost, and shows excellent generalization ability and wide application prospect.

Description

Technical Field

[0001] This invention belongs to the field of computer technology and artificial intelligence, with a particular focus on innovative applications combining knowledge graphs and natural language processing technologies. The core of this invention lies in addressing knowledge question-answering tasks, specifically leveraging the ability of large models to learn contextual information from few samples. It intelligently generates knowledge base query statements based on scene awareness for user-generated natural language questions. This invention proposes a unique "decomposition-then-discrimination" architecture, breaking down the complex knowledge question-answering task into three sub-tasks: logical skeleton parsing, knowledge component retrieval, and query statement generation. This method overcomes the limitation of traditional models relying on large amounts of labeled data, significantly reducing the cost of knowledge question-answering tasks. Furthermore, this invention possesses perception and learning capabilities in different scenarios, making it applicable to knowledge question-answering tasks in various contexts and possessing broad application potential. Background Technology

[0002] Knowledge-based question answering (KBQA) systems aim to transform natural language questions into queries to a knowledge base, providing accurate answers to non-expert users. This field has garnered widespread attention due to its importance in natural language processing. However, the sheer size of knowledge bases and the complexity of given questions often necessitate large amounts of labeled data and complex training architectures to build an efficient KBQA model, resulting in high costs. Furthermore, these supervised learning-based models are often designed specifically for particular knowledge bases, making them difficult to apply to diverse question-answering scenarios. Therefore, it is necessary to develop a scenario-based KBQA framework that can intelligently generate knowledge queries with limited or no labeled data and flexibly adapt to different knowledge-based question-answering scenarios.

[0003] Large Language Models (LLMs) are highly favored by academia and industry for their excellent generalization ability when handling various tasks. They have proven to perform well in converting natural language into structured executable code, which is the core of query generation in KBQA. Therefore, this invention attempts to leverage the context learning capabilities of LLMs to achieve efficient knowledge query generation based on a limited number of examples.

[0004] However, directly generating query statements for KBQA tasks is challenging because complete knowledge query statements typically involve complex logical structures and rich knowledge components, including information such as entities, relationships, and types. Research shows that decomposing complex tasks into simpler subtasks can significantly improve the performance of LLMs. Furthermore, decomposing coarse-grained query statements into fine-grained components can further improve the generalization ability and accuracy of KBQA. Based on these observations, this invention proposes an innovative method that decomposes the task of directly generating knowledge query statements into several subtasks, such as generating logical frameworks and filling expression components, while introducing the concept of scenario learning to improve the context sensitivity and accuracy of subtask processing. Through this method, this invention aims to improve the efficiency and accuracy of knowledge query statement generation while reducing dependence on large amounts of labeled data, thereby enabling more flexible and widespread KBQA applications.

[0005] First, logical frame generation is particularly challenging in existing large-scale language model (LLM) applications, especially in the KBQA domain. This is because LLMs have limited exposure to knowledge of logical reasoning and frame generation during the pre-training phase. When generating logical frames, LLMs may encounter two main problems. First, they may produce frames with numerous formatting errors, rendering the generated queries unexecutable. Second, LLMs typically perform poorly in logical reasoning, potentially leading to high error rates in logical types and reasoning structures. To address these challenges, this patent proposes a shift strategy: instead of relying on the generative capabilities of LLMs, it leverages their discriminative abilities to select appropriate candidates, thereby simplifying the logical reasoning process and improving the executability of queries.

[0006] Furthermore, LLMs may encounter the "illusion" problem when processing knowledge queries, especially when accurate knowledge base support is lacking. To address this issue, this invention proposes that knowledge components should be rationally extracted and utilized from the knowledge base (KB) during the expression component population process. These components will be provided to LLMs as samples for few-shot context learning. This approach has two key considerations: first, ensuring that the number of knowledge components provided to LLMs is moderate, avoiding excessive components that could cause selection confusion for large models; and second, ensuring the relevance and accuracy of knowledge components to prevent misleading information. In addition, combining this with scene learning methods can further enhance the targeting of knowledge component selection and application, ensuring that the generated knowledge query statements are both accurate and efficient, thereby achieving a significant performance improvement in LLM applications.

[0007] To address the challenges encountered in traditional knowledge query generation, this invention proposes an innovative method that combines scene learning and few-shot context learning techniques from large-scale models. This method generates query statements through logical framework parsing, knowledge component retrieval, and logical expression construction. First, the logical skeleton of the user's natural language question is parsed by understanding the context. Next, a knowledge component retrieval tool selects knowledge components that meet the contextual conditions and are closely related to the question. Finally, these scene-related knowledge components are integrated into the logical framework, and the reasoning capabilities of a large-scale model are used to generate the final knowledge query statement, thereby effectively improving the adaptability of the question context and the accuracy of the query statement. Summary of the Invention

[0008] The technical problem this invention aims to solve is to propose an intelligent knowledge query statement generation method and system driven by scene-aware large-scale model few-shot context learning. Specifically, this method employs a scene-based learning strategy to enhance the scene awareness and logical reasoning efficiency of the large-scale model when handling knowledge question-answering tasks. First, through scene learning, this method decomposes the complex knowledge question-answering task into multiple sub-tasks (logical skeleton parsing, knowledge component retrieval, and query statement generation, etc.), improving the efficiency and accuracy of question parsing. Next, for the natural language question input by the user, this method performs efficient logical skeleton parsing to determine the core logical structure of the question. Then, based on the analyzed logical skeleton, a scene-aware mechanism is used to retrieve and filter knowledge components related to the target question, ensuring the relevance and accuracy of the selected components. Furthermore, leveraging the advantages of the large-scale model in few-shot context learning, candidate knowledge components are accurately distinguished, thereby generating highly targeted and logically sound knowledge query statements. This method not only improves the executability of the generated query statements but also significantly enhances the accuracy of the knowledge question-answering results, effectively reducing task costs.

[0009] The technical solution of this invention is as follows: A method for intelligently generating knowledge query statements based on scene-aware large-model few-shot context learning includes the following steps: (1) Perform logical skeleton analysis on the natural language question input by the user, i.e. the target question; (2) Based on the target question, retrieve relevant knowledge components from the knowledge base; (3) Use large models for few-shot context learning to achieve selection of knowledge components; (4) Construct logical expressions based on the logical skeleton and the selected knowledge components to obtain knowledge query statements.

[0010] Further, step (1) includes the following sub-steps: 1a) Based on the natural language question input by the user, use the BM25 retrieval technology to retrieve several most similar sample questions from the sample question database; 1b) Obtain the logical expression corresponding to the sample problem, and extract the logical skeleton from the logical expression; 1c) Count the number of occurrences of various types of logical skeletons, and select several logical skeletons that occur most frequently as preliminary logical skeleton candidates for the target problem; 1d) Utilize the semantic understanding of the target problem using a large model to determine the logical type of the target problem; 1e) Filter the preliminary logical skeleton candidates according to the logical type, and retain only the logical skeleton candidates with the same logical type as the final logical skeleton candidates for the target problem.

[0011] Furthermore, step (2) includes the following sub-steps: 2a) Based on the natural language question input by the user, use the BM25 retrieval technology to retrieve the most similar knowledge components from the knowledge base, including knowledge components such as entities, relations, and types; 2b) Use an instance-level pattern checker to check the retrieved knowledge components and retain only those that meet the instance-level pattern specification; where instance level refers to a specific individual instance, such as a concrete thing or object. 2c) Use an ontology-level schema checker to check the retrieved knowledge components and retain only those that meet the ontology-level schema specification; where ontology level refers to abstract conceptual structures, such as general categories and relation schemas, used to describe the relationships and attributes between instances. 2d) Summarize the knowledge components that satisfy the instance-level and ontology-level pattern checks as candidates for knowledge components of the target problem.

[0012] Furthermore, the contextual learning in step (3) is a process of understanding and analyzing the task scenario and problem context. By learning from task examples, the performance of the model can be improved. Step (3) includes the following sub-steps: 3a) Based on the natural language question input by the user, use the BM25 retrieval technology to retrieve several most similar sample questions from the sample question database; 3b) Obtain the logical expression corresponding to the sample problem, and extract the sample knowledge component from the logical expression; 3c) Use the logical skeleton candidates from step 1e) to filter the samples, and only keep sample problems with the same logical skeleton; 3d) Use the knowledge component candidates from step 2d) to filter the samples, keeping only the knowledge components that appear in the candidates; 3e) Summarize the example questions and knowledge components that meet the logical skeleton and candidate selection criteria, and use them as examples for small-sample context learning of large models; 3f) Use a large model to perform few-shot context learning on sample problems and knowledge components to achieve the selection of candidate knowledge components.

[0013] Furthermore, step (4) includes the following sub-steps: 4a) Using the logical skeleton candidates obtained in step 1e) and the knowledge components obtained in step 3f), fill the knowledge components into the logical skeleton candidates to construct logical expression candidates for the target problem; 4b) Extract type knowledge components from logical expression candidates and use them as answer type candidates for the target question; 4c) Use a large model to select the answer type most likely to be the answer to the target question, and select the corresponding logical expression as the final query statement for the target question.

[0014] Furthermore, the above method also includes step (5): executing the query statement of the target question in the knowledge base, obtaining the query results, and then describing the query results using natural language, as the result of knowledge question answering to the user. A knowledge query statement intelligent generation system driven by scene-aware large model few-shot context learning includes: The logic skeleton parsing module is used to parse the logic skeleton of the natural language question input by the user, i.e., the target question; The knowledge component retrieval module is used to retrieve relevant knowledge components from the knowledge base based on the target question. The knowledge component selection module is used to select knowledge components for few-shot context learning using a large model. The query statement generation module is used to construct logical expressions based on the logical skeleton and the selected knowledge components to obtain knowledge query statements.

[0015] Furthermore, the system also includes a query execution module, which executes knowledge query statements in the knowledge base to obtain query results.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention optimizes the method for generating knowledge query statements using large models in few-shot context learning environments by combining scene learning strategies. Specifically, this method can understand and deconstruct complex knowledge question answering tasks into more manageable subtasks through scene learning, thereby reducing the difficulty for large language models to handle these tasks. Furthermore, this method can efficiently parse the logical skeleton and determine the logical type of user-input natural language questions, enhancing the executability of logical expressions and thus improving the quality and relevance of knowledge query statements. By selectively retrieving and pattern-checking relevant knowledge components, this invention further improves the matching degree and accuracy of knowledge components. Moreover, using a large model optimized for scene learning for few-shot context learning, employing a precise differentiation strategy rather than simple generation, greatly improves the accuracy and relevance of the target results. Overall, this method, through the combination of scene learning and knowledge query statement generation, not only improves the accuracy and execution efficiency of knowledge question answering but also significantly reduces costs, demonstrating excellent generalization ability and broad application prospects. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for intelligently generating knowledge query statements based on scene-aware large-model, few-shot context learning.

[0018] Figure 2 It is the logical skeleton and knowledge component of the logical expression corresponding to the knowledge query statement. Detailed Implementation

[0019] The method will be further explained below with reference to the accompanying drawings and specific embodiments.

[0020] The method flow of the present invention is as follows: Figure 1 As shown, the specific steps are as follows: Step 1: Based on the user's input natural language question, the BM25 retrieval technique is first used to retrieve several most similar sample questions from the sample question database. Next, the logical expressions corresponding to these sample questions are obtained, and the logical skeletons are extracted from these expressions. Then, the logical skeletons of various types are statistically analyzed, and the most frequently occurring logical skeletons are selected as preliminary logical skeleton candidates for the target question. Further, a large-scale model is used to perform semantic understanding on the target question to determine its logical type. Finally, the preliminary logical skeleton candidates are filtered based on their logical type, retaining only those of the same logical type. These candidates will serve as the final logical skeleton candidates for the target question, used for further processing and answering the user's question.

[0021] Step 2: Based on the user's input natural language question, the BM25 retrieval technique is first used to retrieve the most similar knowledge components from the knowledge base. These components include knowledge elements such as entities, relations, and types. Then, an instance-level pattern checker is used to examine the retrieved knowledge components, retaining only those that conform to the instance-level pattern specification. Next, an ontology-level pattern checker is used to examine the retrieved knowledge components, filtering out those that conform to the ontology-level pattern specification. Finally, the knowledge components that satisfy both instance-level and ontology-level pattern checks are summarized to form candidate knowledge components for the target question, enabling further processing and answering of the user's question.

[0022] Step 3: Based on the natural language question input by the user, firstly, the BM25 retrieval technique is used to retrieve several most similar sample questions from the sample question database. Next, the logical expressions corresponding to these sample questions are obtained, and sample knowledge components are extracted from the logical expressions. Subsequently, using the previously obtained logical skeleton, sample questions are filtered, retaining only those with the same logical skeleton. Simultaneously, using the previously obtained knowledge component candidates, sample knowledge components are filtered, retaining only those appearing in the candidate options. Then, the sample questions and knowledge components that meet the logical skeleton and candidate selection criteria are summarized as examples for few-shot context learning by the large model. Finally, the large model is used to learn from these sample questions and knowledge components to achieve the selection of knowledge component candidates.

[0023] Step 4: First, combine the previously obtained logical skeleton candidates and the selected knowledge components, then populate the knowledge components into the logical skeleton candidates to create logical expression candidates for the target problem, such as... Figure 2 As shown. Next, type knowledge components are extracted from the logical expression candidates; these components will serve as candidate answer types for the target question. Finally, leveraging the capabilities of a large model, the answer type most likely to be the answer to the target question is selected, and the corresponding logical expression is chosen accordingly, becoming the final logical expression result for the target question.

[0024] The above method may also include step 5: executing the logical expression of the target question in the knowledge base to obtain the query results, then describing the query results using natural language, and answering the user as the result of the knowledge question answering, thus completing the knowledge question answering process.

[0025] Specific examples: For example, when a user inputs the natural language question "How many administrative districts are there in Beijing?", as shown in step 1, this method first retrieves several most similar sample questions from the sample question database. For instance, sample question 1: "How many administrative districts are there in Shanghai?", corresponding to the logical expression "COUNT(AND Administrative District (JOIN Administrative Division) Shanghai)". Another example is sample question 2: "How many administrative districts are there in Guangzhou?", corresponding to the logical expression "COUNT(AND Administrative District (JOIN Administrative District) Guangzhou)". After statistical analysis of the logical skeleton, "COUNT(AND Type (JOIN Relationship) Entity)" is the most frequently occurring logical skeleton. Furthermore, based on the understanding of the question's semantics, this question is a counting question, which matches the type of logical skeleton. Therefore, this logical skeleton will be used as a candidate for subsequent processing steps.

[0026] Then, as shown in step 2, based on the user's input question, the model retrieves several relevant knowledge components from the knowledge base, such as the entity component "Beijing," the relation component "administrative division," the relation component "provincial capital," the type component "administrative district," and the type component "street." When the entity component is determined to be "Beijing," since Beijing does not have a "provincial capital" relationship, the relation component "provincial capital" does not meet the instance-level requirements; and since the city's administrative division does not include the "street" type, the type component "street" does not meet the ontology-level requirements. Therefore, after instance-level and ontology-level pattern checks and filtering, components that meet the requirements, such as "Beijing," "administrative division," and "administrative district," will be used as candidates for subsequent steps.

[0027] Then, as shown in step 3, based on the user's input question, the model retrieves several most similar sample questions from the sample question library, such as "How many administrative districts are there in Shanghai?", "How many administrative districts are there in Guangzhou?", "How many administrative organs are there in Beijing?", and "Which is the largest administrative district in Beijing?". Based on the logical framework obtained in step 1, questions like "How many administrative organs are there in Beijing?" do not meet the requirements for an administrative district question, and "Which is the largest administrative district in Beijing?" does not meet the requirements for a counting question, and will be discarded. Therefore, questions like "How many administrative districts are there in Shanghai?" and "How many administrative districts are there in Guangzhou?" that meet the requirements, along with their corresponding logical expressions, will serve as examples for the large model's context learning, enabling the selection of subsequent knowledge components and the filling of the logical framework.

[0028] Then, as shown in step 4, based on the logical skeleton "COUNT(AND type(JOIN relationship) entity)" obtained in step 1, and the knowledge components "Beijing", "administrative division", and "administrative region" obtained in step 2, and by performing context learning on the sample question obtained in step 3, the final logical expression is obtained as "COUNT(AND administrative region(JOIN administrative division) Beijing)". This logical expression can be transformed into a knowledge base query statement to realize knowledge base question answering.

[0029] This method was evaluated on the GrailQA public knowledge base question-answering dataset, using exact match (EM) score and F1 score as evaluation metrics. After extensive experiments and evaluation using official evaluation scripts, the mainstream large-model-based knowledge question answering methods achieve an EM score of 53.2 and an F1 score of 58.5, while our method achieves an EM score of 56.5 and an F1 score of 59.1. This indicates that our method has higher accuracy in generating knowledge queries and provides more reliable answers. Especially for questions with unknown scenarios, the mainstream large-model-based knowledge question answering methods achieve an EM score of 45.4 and an F1 score of 50.7, while our method achieves an EM score of 57.8 and an F1 score of 60.5. This demonstrates that our method significantly improves query accuracy and answer reliability, possesses stronger scene awareness capabilities, and is superior in complex question reasoning and question context understanding.

[0030] Another embodiment of the present invention provides a knowledge query statement intelligent generation system driven by scene-aware large model few-shot context learning, comprising: The logic skeleton parsing module is used to parse the logic skeleton of the natural language question input by the user, i.e., the target question; The knowledge component retrieval module is used to retrieve relevant knowledge components from the knowledge base based on the target question. The knowledge component selection module is used to select knowledge components for few-shot context learning using a large model. The query statement generation module is used to construct logical expressions based on the logical skeleton and the selected knowledge components to obtain knowledge query statements.

[0031] The system may also include a query execution module, which executes knowledge query statements in the knowledge base to obtain query results.

[0032] For the specific implementation process of each module, please refer to the description of the method of the present invention above.

[0033] Another embodiment of the present invention provides a computer device (computer, server, smartphone, etc.) including a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the steps of the method of the present invention.

[0034] Another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disk) that stores a computer program, which, when executed by a computer, implements the various steps of the method of the present invention.

[0035] Although the specific details, implementation algorithms, and accompanying drawings of the present invention have been disclosed for illustrative purposes to aid in understanding and implementing the invention, those skilled in the art will understand that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the invention and the appended claims. For example, the logical skeleton and knowledge component analysis method used in these claims is one of many semantic parsing analysis methods, and applications of other similar methods also fall within the scope of this invention. This invention should not be limited to the content disclosed in the preferred embodiments and accompanying drawings; the scope of protection of this invention is defined by the claims.

Claims

1. A method for intelligently generating knowledge query statements based on scene-aware large-model few-shot context learning, characterized in that, Includes the following steps: Perform logical skeleton parsing on the natural language question input by the user, i.e., the target question; Based on the target question, relevant knowledge components are retrieved from the knowledge base; Use large models for few-shot context learning to enable the selection of knowledge components; Based on the logical framework and the selected knowledge components, construct a logical expression to obtain a knowledge query statement; The logical skeleton parsing of the natural language question input by the user, i.e., the target question, includes: Based on the natural language question input by the user, retrieve the most similar sample questions from the sample question library; Obtain the logical expression corresponding to the sample problem, and extract the logical skeleton from the logical expression; Count the frequency of each type of logical skeleton, and select several logical skeletons that appear most frequently as preliminary logical skeleton candidates for the target problem. Utilize large-scale models to understand the semantics of the target problem and determine its logical type; The preliminary logical skeleton candidates are filtered based on the logical type, and only logical skeleton candidates of the same logical type are retained as the final logical skeleton candidates for the target problem.

2. The method according to claim 1, characterized in that, The step of retrieving relevant knowledge components from the knowledge base based on the target question includes: Based on the natural language question input by the user, retrieve the most similar knowledge components from the knowledge base, including entities, relations, and types; The retrieved knowledge components are inspected using an instance-level pattern inspector, and only those that meet the instance-level pattern specification are retained. Using an ontology-level schema inspector, the retrieved knowledge components are inspected, and only those that satisfy the ontology-level schema specification are retained. We compile knowledge components that satisfy both instance-level and ontology-level pattern checks and use them as candidates for knowledge components for the target problem.

3. The method according to claim 1, characterized in that, The use of a large model for few-shot context learning to achieve the selection of knowledge components includes: Based on the natural language question input by the user, retrieve the most similar sample questions from the sample question library; Obtain the logical expression corresponding to the sample question, and extract the sample knowledge component from the logical expression; Use logical skeleton candidates to filter samples, and only retain sample issues with the same logical skeleton; Use knowledge component candidates to filter the samples, and only keep the knowledge components that appear in the candidates; Summarize sample questions and knowledge components that meet the logical framework and candidate selection criteria, and use them as examples for small-sample context learning of large models; Using a large model, we can perform few-shot context learning on example problems and knowledge components to select candidates for knowledge components.

4. The method according to claim 1, characterized in that, The step of constructing a logical expression based on the logical skeleton and selected knowledge components to obtain a knowledge query statement includes: Using logical skeleton candidates and knowledge components, knowledge components are populated into logical skeleton candidates to construct logical expression candidates for the target problem; Extract type knowledge components from logical expression candidates and use them as answer type candidates for the target question; Use a large model to select the answer type most likely to be the answer to the target question, and select the corresponding logical expression as the final query statement for the target question.

5. The method according to claim 1, characterized in that, Also includes: The system executes a query for the target question in the knowledge base, obtains the query results, and then describes the query results using natural language, presenting them to the user as the answer to the knowledge question.

6. A knowledge query statement intelligent generation system based on scene-aware large model few-shot context learning driven by the method of any one of claims 1 to 5, characterized in that, include: The logic skeleton parsing module is used to parse the logic skeleton of the natural language question input by the user, i.e., the target question; The knowledge component retrieval module is used to retrieve relevant knowledge components from the knowledge base based on the target question. The knowledge component selection module is used to select knowledge components for few-shot context learning using a large model. The query statement generation module is used to construct logical expressions based on the logical skeleton and the selected knowledge components to obtain knowledge query statements.

7. The system according to claim 6, characterized in that, It also includes a query execution module, which is used to execute knowledge query statements in the knowledge base and obtain query results.

8. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the method of any one of claims 1 to 5.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a computer, implements the method according to any one of claims 1 to 5.

Images