Intelligent Reasoning Methods and Systems Based on Large Language Models

Abstract

This invention discloses an intelligent reasoning method and system based on a large language model, belonging to the field of intelligent reasoning technology. The method specifically includes: acquiring input data and preprocessing it; identifying the current context and reasoning target based on the input data; combining a deep neural network with a large language model; employing a recursive reasoning mechanism to perform reasoning analysis in stages; and outputting the final reasoning result. This invention improves the accuracy and reliability of the reasoning results, especially performing excellently in complex application scenarios and tasks. Furthermore, it can be adjusted according to the specific needs of different users, exhibiting a high degree of personalization.

Description

Technical Field

[0001] This invention belongs to the field of intelligent reasoning technology, specifically to an intelligent reasoning method and system based on a large language model. Background Technology

[0002] With the rapid development of artificial intelligence technology, large language models (such as GPT and BERT) have made significant progress in the field of natural language processing (NLP). These models, through large-scale pre-training and fine-tuning, are able to understand and generate human language and are widely used in tasks such as text generation, machine translation, and question answering systems. However, traditional language models still have limitations in handling complex reasoning tasks, especially in scenarios that require logical reasoning, multi-step reasoning, and cross-domain knowledge fusion.

[0003] Existing reasoning methods typically rely on rule engines, knowledge graphs, or symbolic logic. While these methods perform well in certain specific domains, they lack flexibility and generalization ability. The emergence of large language models offers new possibilities for solving these problems. By combining deep learning and natural language understanding, large language models can perform reasoning without explicit rules, but their reasoning capabilities still need further optimization and expansion.

[0004] For example, Chinese Patent No. CN116992966B discloses a method and computing device for an artificial intelligence model inference platform. The method includes: acquiring first input data for a first inference task and executing the first inference task; acquiring second input data for a second inference task and executing the second inference task; if the second input data is the same as the first input data, and the second inference path is the same as the first inference path, then the second inference task reuses the output data of the first inference path. According to the technical solution of this invention, the computational load of the algorithm can be effectively reduced, the computing power of the chip can be rationally utilized, and resource waste can be avoided.

[0005] For example, Chinese patent CN118485150B discloses an artificial intelligence-based reasoning information generation method, including: acquiring user input text and splitting it into a knowledge graph; obtaining several subject nodes and several description nodes; obtaining the descriptive specificity of each description node for each subject node; obtaining the main descriptive degree of different hop neighborhoods of each subject node, thereby obtaining the effective neighborhood range of each subject node; obtaining the descriptive accuracy of each subject node; obtaining several important semantic windows based on the descriptive accuracy and effective neighborhood range of each subject node; updating the knowledge graph based on the important semantic windows, and completing text reasoning through local graph encoding. This invention aims to solve the problem of reasoning information bias caused by insufficient mining of connections between unconnected nodes in the knowledge graph.

[0006] The aforementioned patents have the following drawbacks: insufficient personalization and adaptability, and issues with real-time and dynamic aspects. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes an intelligent reasoning method and system based on a large language model. By integrating deep learning technology with a large language model, a multi-level, recursively optimized intelligent reasoning system is constructed to solve problems encountered in reasoning and analysis tasks in complex scenarios. By combining historical data, external knowledge bases, and contextual information, this invention can improve the accuracy, interpretability, and adaptability of reasoning results.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] The intelligent reasoning method based on large language models includes the following specific steps:

[0010] Obtain the input data and perform preprocessing;

[0011] Based on the input data, identify the current context and reasoning objective;

[0012] By combining deep neural networks with large language models, a recursive reasoning mechanism is adopted to perform reasoning analysis in stages and output the final reasoning result.

[0013] Specifically, the process of identifying the current context and reasoning objective based on the input data includes:

[0014] Analyze the preprocessed input data and perform scenario modeling based on the analysis results;

[0015] Based on context modeling and combined with preprocessed input data, the objective of the current task is derived.

[0016] Break down the task objective into sub-objectives and adjust the sub-objectives in real time according to changes in the situation;

[0017] The relationship between task objectives and context is modeled by multi-level inference networks and graph neural networks, establishing the constraint relationship between contextual factors and task objectives, and the constraints of task objectives are perceived and adjusted in real time according to changes in context.

[0018] Based on historical data and external knowledge bases, the current task objectives are further refined and predicted.

[0019] Specifically, the context modeling includes: context modeling of text data, context modeling of speech data, and context modeling of image data.

[0020] Specifically, the method combines deep neural networks with a large language model, employs a recursive reasoning mechanism, performs reasoning analysis in stages, and outputs the final reasoning result, including:

[0021] Based on the identification of context and task objectives, deep neural networks and large language models are used to analyze existing data and generate preliminary reasoning results;

[0022] The preliminary reasoning results are analyzed and optimized by combining historical data, external knowledge bases, and contextual information.

[0023] Multi-path reasoning is performed through a recursive mechanism, which automatically selects the appropriate reasoning path according to different situations, performs local optimization on each path, and evaluates the reasoning result of each optimized path.

[0024] The evaluation results of each path are sorted, and the top N paths are aggregated to obtain the final inference result.

[0025] Specifically, the input data includes: text, voice, images, sensor data, structured table data, or log files.

[0026] Specifically, the preprocessing includes: cleaning, segmenting, removing stop words and stemming text data, and vectorizing the text using a pre-trained large language model;

[0027] Voice data, which is converted into text using a speech recognition system;

[0028] Image data is denoised and enhanced, and high-level feature vectors of the image are extracted using a convolutional neural network.

[0029] The text vector, speech vector, and image feature vector are fused.

[0030] The intelligent reasoning system based on a large language model is used to implement the intelligent reasoning method based on the large language model, including: a data processing module, a context and target analysis module, and a reasoning module;

[0031] The data processing module is used to acquire input data and perform preprocessing.

[0032] The context and target analysis module is used to identify the current context and reasoning target based on the input data;

[0033] The reasoning module is used to perform reasoning analysis in stages by combining deep neural networks with a large language model and adopting a recursive reasoning mechanism, and output the final reasoning result.

[0034] Specifically, the context and target analysis module includes: a context analysis unit and a target analysis unit;

[0035] The scenario analysis unit is used to analyze the preprocessed input data and perform scenario modeling based on the analysis results.

[0036] The target analysis unit is used to derive the target of the current task, break down the target of the task into sub-targets, adjust the sub-targets of the task in real time according to changes in the situation, establish the constraint relationship between situational factors and task targets, and adjust the task targets in real time based on historical data and external knowledge base.

[0037] Specifically, the reasoning module includes: a preliminary result unit and a final reasoning result unit;

[0038] The preliminary results unit is used to analyze existing data using deep neural networks and large language models based on context and task target recognition, and generate preliminary inference results.

[0039] The final reasoning result unit is used to optimize the preliminary reasoning result and aggregate the paths using a recursive multi-path mechanism to obtain the final reasoning result.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. This invention proposes an intelligent reasoning method based on a large language model, which can be adjusted according to the specific needs of different users and has a high degree of personalization.

[0042] 2. This invention proposes an intelligent reasoning method based on a large language model, which can process and update reasoning results in real time. It can dynamically adjust the reasoning path according to new data inputs (such as real-time examination results, patient feedback, etc.) to ensure that the reasoning results conform to the latest reality.

[0043] 3. This invention proposes an intelligent reasoning method based on a large language model, which can integrate knowledge from different fields, support cross-domain reasoning tasks, and achieve multi-step reasoning and logical inference in complex tasks. Attached Figure Description

[0044] Figure 1 A flowchart of the intelligent reasoning method based on a large language model provided by the present invention;

[0045] Figure 2 This is a diagram illustrating the architecture of an intelligent reasoning system based on a large language model, provided by the present invention. Detailed Implementation

[0046] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These all fall within the protection scope of the present application.

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] It should be noted that, unless there is a conflict, the various features in the embodiments of this application can be combined with each other, all of which are within the protection scope of this application. Furthermore, although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than the module division in the device or the order in the flowchart. In addition, the terms "first," "second," and "third" used in this application do not limit the data or execution order, but only distinguish identical or similar items with essentially the same function and effect.

[0049] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0050] Example 1

[0051] Please see Figure 1 The present invention provides an embodiment of an intelligent reasoning method based on a large language model, comprising the following specific steps:

[0052] Step S1: Obtain input data and perform preprocessing;

[0053] The input data in step S1 includes, but is not limited to, text, voice, images, sensor data, structured table data, or log files. In this step, the data sources are automatically labeled, and corresponding preprocessing procedures are performed according to the type of data source. For example, for text data, the system performs preliminary text cleaning using natural language processing tools (such as word segmentation and part-of-speech tagging); for voice data, a speech recognition system is used to convert speech into text; for image data, image recognition technology is used to extract key features or labels from the images. At the same time, data that does not meet the quality standards (such as parts with excessive noise or lack of effective information) is initially screened at this stage to remove unnecessary noise and erroneous data, so as to ensure the accuracy and efficiency of subsequent processing.

[0054] Preprocessing also includes multimodal fusion and standardization. Text data standardization involves transforming the original text into a unified semantic representation through word segmentation, stop word removal, and stemming. Furthermore, a pre-trained large language model is used to vectorize the text to capture deep semantic relationships between words. Image data standardization employs image enhancement techniques (such as rotation, scaling, and flipping) to increase data diversity and reduce bias. Convolutional neural networks (CNNs) are used to extract high-level feature vectors from the images, and these features are then integrated with other modal data. Speech data processing converts speech data into text using automatic speech recognition (ASR) technology. Next, natural language processing (NLP) techniques are used to analyze the text content. If the speech data comes from a dialogue system, sentiment analysis and tone recognition techniques are used to extract the emotional or contextual information of the speech. Structured data processing: First, the tabular data is standardized by filling in missing values, removing outliers, and performing normalization and standardization. Then, graph neural networks (GNNs) are used to model the relationships between structured data, forming feature vectors that are fused with unstructured data. Unstructured data processing: Unstructured data (such as images and text) undergoes feature extraction using convolutional neural networks (CNNs), recurrent neural networks (RNNs), etc. These features are then fused with the features of structured data in the same feature space.

[0055] Finally, multimodal fusion is performed. The goal of fusion is to construct a multidimensional feature space so that data from different modalities can be effectively integrated within the same inference framework.

[0056] Step S2: Based on the input data, identify the current context and reasoning objective;

[0057] The specific steps of step S2 are as follows:

[0058] Step S201: Analyze the preprocessed input data and perform scenario modeling based on the analysis results;

[0059] In this embodiment, context modeling includes three categories: 1) Context modeling of text data: For input text data, semantic understanding is performed using large language models (such as BERT, GPT, etc.). These models not only identify the relationship between keywords and sentences, but also analyze the context and potential emotions behind the text. For example, in a medical scenario, the system will identify the symptoms, medications, and disease names mentioned by the patient, and combine them with the context to understand the patient's medical background and form a complete context description; 2) Context modeling of speech data: For speech data, automatic speech recognition (ASR) technology is first used to convert speech into text. Then, sentiment analysis technology is used to identify tone, emotion, and implied meaning to ensure that the potential emotions and intentions in the dialogue are accurately captured. For example, if the user's tone appears anxious, the system will adjust the reasoning path according to this emotional change; 3) Context modeling of image data: For image data, especially medical images and satellite images, deep analysis is performed using convolutional neural networks (CNN) to extract key features in the image (such as lesion areas, object locations, etc.). Furthermore, the system combines the image content with an external knowledge base to generate a context description corresponding to the image, and combines it with text data to form a multimodal context understanding.

[0060] Step S202: Based on context modeling and combined with the preprocessed input data, deduce the objective of the current task;

[0061] For example, in an enterprise decision support scenario, the task objective might be "increase product sales" or "optimize inventory management," while in a medical scenario, the task objective might be "diagnose a disease" or "determine a treatment plan." Natural Language Processing (NLP) technology is used to identify the user's intent and abstract the objective.

[0062] Step S203: Break down the task objective into several sub-objectives and adjust the task sub-objectives in real time according to changes in the situation;

[0063] For example, in a medical task, if the task objective is "diagnosing a disease," it can be further refined into sub-objectives such as "analyzing patient symptoms," "retrieving relevant cases," and "recommending possible treatment plans." When the situation changes, such as when the patient mentions relevant cases, the task objective is further refined into sub-objectives such as "analyzing patient symptoms" and "recommending possible treatment plans."

[0064] Step S204: Model the relationship between task objectives and context through multi-level inference networks and graph neural networks (GNNs), establish the constraint relationship between contextual factors and task objectives, and perceive and adjust the constraints of task objectives in real time according to changes in context.

[0065] In this embodiment, in financial decision-making, the task objective may be "predicting stock market trends," while contextual factors may include "economic data," "market sentiment," etc. By connecting these different dimensions of information through a graph neural network, a constraint relationship between contextual factors and task objectives is established. If certain contextual factors change (such as an economic crisis, sudden changes in customer demand, etc.), these changes are automatically identified, and the task objectives are optimized or adjusted to ensure the effectiveness and rationality of the final decision.

[0066] Step S205: Based on historical data and external knowledge base, further refine and predict the current task objectives.

[0067] In this embodiment, historical data analysis involves pattern recognition and trend inference based on similar scenarios and task execution in historical data. For example, in a customer service scenario, historical dialogue records can be analyzed to infer the current user's needs and identify targets related to those needs. By learning from common problems, solutions, and customer feedback in historical data, potential task targets in the current context can be quickly located. External knowledge base integration and reasoning involves connecting with external knowledge bases (such as professional knowledge bases in fields like medicine, law, and finance) to ensure the comprehensiveness and accuracy of the target when identifying the current task target. For example, in a medical reasoning task, an external knowledge base can provide the latest disease diagnosis and treatment guidelines, drug treatment plans, and other information to further refine the task target and ensure it meets domain requirements.

[0068] Step S3: By combining deep neural networks with a large language model, a recursive reasoning mechanism is adopted to perform reasoning analysis in stages and output the final reasoning result.

[0069] The specific steps of step S3 are as follows:

[0070] Step S301: Based on the identification of context and task objectives, analyze existing data (such as literature databases) using deep neural networks and large language models to generate preliminary reasoning results;

[0071] In this embodiment, this step is the second stage of recursive reasoning (the first stage is data collection and data processing). The reasoning results output in this stage are only preliminary diagnoses or suggestions, but they provide an important basis for subsequent reasoning.

[0072] Step S302: Analyze the preliminary reasoning results, and optimize the preliminary reasoning results by combining historical data, external knowledge base and contextual information;

[0073] In this embodiment, historical case review: Data sources include historical case databases, patient health records, known disease patterns, etc.; the analysis goal is to identify historical cases similar to the current patient's condition, and by reviewing the treatment outcomes and disease progression of these cases, the preliminary inference results can be optimized; application method: for example, for the early diagnosis of a disease, if the preliminary inference result suggests a certain disease, cases with similar symptoms and disease courses can be retrieved from historical cases, and the symptom manifestations, treatment methods, and treatment effects in these cases can be analyzed to help rule out mismatched diseases or adjust the diagnosis; disease course and treatment effect data: Data sources include the patient's treatment history, medication records, treatment response, etc.; the analysis goal is to judge the patient's treatment response and drug efficacy, and to revise the preliminary inference results by combining historical data. For example, the inference of the treatment effect of a certain drug can be revised by the patient's response to the drug (improvement, ineffectiveness, or side effects, etc.) in historical cases;

[0074] Integrate external knowledge bases, including professional knowledge graphs, academic literature, and research databases in the medical field. These data sources can provide the latest medical research results, drug information, disease classifications, and other information for preliminary reasoning.

[0075] Dynamic integration of contextual information, which refers to the patient's current living environment, health status, clinical symptoms and their changes, etc., significantly influences disease inference and treatment decisions. This includes: real-time health status analysis (data sources: real-time patient health data, such as body temperature, blood pressure, laboratory test results, imaging data, etc.); analysis objectives: real-time monitoring of changes in the patient's health status and comparison with initial inference results for dynamic adjustment. For example, if the patient's temperature rises and is accompanied by new symptoms, the inference result can be adjusted to increase the probability of possible infectious diseases (such as influenza, pneumonia, etc.); dynamic changes in clinical symptoms (data sources: symptoms...). The analysis includes recording changes in symptoms, patient-reported feelings, etc.; analyzing objectives by monitoring symptom trends (e.g., worsening or relief) to adjust the inference path and dynamically updating symptom descriptions to make the inference results more consistent with the patient's current clinical presentation; application methods, for example, if a patient's symptoms change during treatment (e.g., worsening headache or cough), the likelihood of the disease is reassessed based on the symptom changes, and corresponding adjustments are made to the inference path; environmental and lifestyle considerations, combining the patient's lifestyle and environmental factors to optimize the initial inference results, for example, if the patient has recently been exposed to an epidemic area or has a history of occupational exposure, the probability of diseases related to these risk factors will increase.

[0076] For example, medication recommendations can be adjusted based on the patient's specific medical history, drug side effects, and other medical literature.

[0077] Step S303: Perform multi-path reasoning through a recursive mechanism, that is, explore multiple possible reasoning paths, automatically select the appropriate reasoning path according to different situations, perform local optimization on each path, and evaluate the reasoning result of each optimized path.

[0078] In this embodiment, the generated path can be inferred using a deep neural network model and a large language model respectively. If the results of the two models are consistent, the credibility of the result is enhanced. If there are differences, the system will further analyze and adjust the final result based on contextual information and historical data. After local optimization of each path, it will also provide dynamic feedback and adjust the inference result based on real-time input information. For example, if the user reports that their symptoms have worsened or new medical data is input, the inference result will be dynamically updated and recursive optimization will continue.

[0079] Step S304: Sort the evaluation results of each path, select the top N paths by evaluation results and aggregate them to obtain the final inference result.

[0080] For example, in a medical setting, the system may aggregate multiple possible drug recommendations and output a comprehensive recommendation result.

[0081] Example 2

[0082] Please see Figure 2 Another embodiment of the present invention provides: an intelligent reasoning system based on a large language model, comprising: a data processing module, a context and target analysis module, and a reasoning module;

[0083] The data processing module is used to acquire input data and perform preprocessing.

[0084] The context and target analysis module is used to identify the current context and reasoning target based on the input data;

[0085] The reasoning module is used to perform reasoning analysis in stages by combining deep neural networks with a large language model and adopting a recursive reasoning mechanism, and output the final reasoning result.

[0086] The context and objective analysis module includes: a context analysis unit and an objective analysis unit;

[0087] The scenario analysis unit is used to analyze the preprocessed input data and perform scenario modeling based on the analysis results.

[0088] The target analysis unit is used to derive the target of the current task, break down the target of the task into sub-targets, adjust the sub-targets of the task in real time according to changes in the situation, establish the constraint relationship between situational factors and task targets, and adjust the task targets in real time based on historical data and external knowledge base.

[0089] The reasoning module includes: a preliminary result unit and a final reasoning result unit;

[0090] The preliminary results unit is used to analyze existing data using deep neural networks and large language models based on context and task target recognition, and generate preliminary inference results.

[0091] The final reasoning result unit is used to optimize the preliminary reasoning result and aggregate the paths using a recursive multi-path mechanism to obtain the final reasoning result.

[0092] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0093] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent reasoning method based on a large language model, characterized in that, include: Obtain the input data and perform preprocessing; Based on the input data, identify the current context and reasoning objective; By combining deep neural networks with large language models, a recursive reasoning mechanism is adopted to perform reasoning analysis in stages and output the final reasoning result. The process of identifying the current context and reasoning objective based on the input data includes: Analyze the preprocessed input data and perform scenario modeling based on the analysis results; Based on context modeling and combined with preprocessed input data, the objective of the current task is derived. Break down the task objective into sub-objectives and adjust the sub-objectives in real time according to changes in the situation; The relationship between task objectives and context is modeled by multi-level inference networks and graph neural networks, establishing the constraint relationship between contextual factors and task objectives, and the constraints of task objectives are perceived and adjusted in real time according to changes in context. Based on historical data and external knowledge bases, the current task objectives are further refined and predicted; The method combines deep neural networks with a large language model, employs a recursive reasoning mechanism, performs reasoning analysis in stages, and outputs the final reasoning result, including: Based on the identification of context and task objectives, deep neural networks and large language models are used to analyze existing data and generate preliminary reasoning results; The preliminary reasoning results are analyzed and optimized by combining historical data, external knowledge bases, and contextual information. Multi-path reasoning is performed through a recursive mechanism, which automatically selects the appropriate reasoning path according to different situations, performs local optimization on each path, and evaluates the reasoning result of each optimized path. The evaluation results of each path are sorted, and the top N paths are aggregated to obtain the final inference result.

2. The intelligent reasoning method based on a large language model as described in claim 1, characterized in that, The context modeling includes: context modeling of text data, context modeling of voice data, and context modeling of image data.

3. The intelligent reasoning method based on a large language model as described in claim 1, characterized in that, The input data includes: text, voice, images, sensor data, structured table data, or log files.

4. The intelligent reasoning method based on a large language model as described in claim 1, characterized in that, The preprocessing includes: cleaning, segmenting, removing stop words and stemming text data, and vectorizing the text using a pre-trained large language model; Voice data, which is converted into text using a speech recognition system; Image data is denoised and enhanced, and high-level feature vectors of the image are extracted using a convolutional neural network. The text vector, speech vector, and image feature vector are fused.

5. An intelligent reasoning system based on a large language model, used to implement the intelligent reasoning method based on a large language model as described in any one of claims 1-4, characterized in that, include: Data processing module, context and goal analysis module, and reasoning module; The data processing module is used to acquire input data and perform preprocessing. The context and target analysis module is used to identify the current context and reasoning target based on the input data; The reasoning module is used to perform reasoning analysis in stages by combining deep neural networks with a large language model and adopting a recursive reasoning mechanism, and output the final reasoning result.

6. The intelligent reasoning system based on a large language model as described in claim 5, characterized in that, The context and objective analysis module includes: a context analysis unit and an objective analysis unit; The scenario analysis unit is used to analyze the preprocessed input data and perform scenario modeling based on the analysis results. The target analysis unit is used to derive the target of the current task, break down the target of the task into sub-targets, adjust the sub-targets of the task in real time according to changes in the situation, establish the constraint relationship between situational factors and task targets, and adjust the task targets in real time based on historical data and external knowledge base.

7. The intelligent reasoning system based on a large language model as described in claim 6, characterized in that, The reasoning module includes: a preliminary result unit and a final reasoning result unit; The preliminary results unit is used to analyze existing data using deep neural networks and large language models based on context and task target recognition, and generate preliminary inference results. The final reasoning result unit is used to optimize the preliminary reasoning result and aggregate the paths using a recursive multi-path mechanism to obtain the final reasoning result.

Images