Scientific problem value quantification method embedded with multi-dimensional mixed expert mechanism

Abstract

The invention provides a scientific problem value quantification method embedded with a multi-dimensional hybrid expert mechanism, and relates to the technical field of data processing, and the method comprises the steps: reading a predetermined quantification dimension; introducing a predetermined triple strategy to construct a target instruction fine tuning data set; performing parameter quantization based on a QLoRA large language model fine tuning principle; obtaining a scientific problem text, performing semantic feature analysis through a gating network, and obtaining a target weight coefficient of the scientific problem text according to text semantics; and analyzing the scientific problem text through the target identification large model, and obtaining a target quantification result in combination with the target weight coefficient. Through the method and the device, the technical problem of relatively low scientific problem value evaluation efficiency caused by different evaluation standards due to dependence on expert evaluation and subjective judgment in the prior art can be solved, internal evaluation of scientific problem values is realized by integrating multiple dimensions and combining a large-scale language model for fine adjustment, and the evaluation efficiency of the scientific problem values is improved. And the scientific problem value evaluation efficiency is improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism. Background Art

[0002] Scientific problems mainly focus on phenomena and laws in the field of natural sciences, usually involving the verification or expansion of hypotheses, theories and models. The quantification of the value of scientific problems refers to the process of quantitatively evaluating the value of scientific problems through evaluation and estimation based on certain standards and criteria. Existing methods for evaluating the value of scientific problems mainly rely on expert review and quantitative indicators. Expert review is one of the commonly used methods for evaluating the value of scientific problems. It relies on experts in the field to give subjective evaluations based on their own knowledge background and experience, combined with their understanding of scientific problems. Although it can provide profound insights and rich background knowledge, the knowledge structure and experience of different experts may lead to inconsistent evaluation standards, which often lack uniformity and standardization, thus affecting the objectivity and repeatability of the evaluation. External indicators such as citations and journal impact factors are also widely used to measure the academic influence of scientific research articles, journals or scientific problems. Citations reflect the attention of a scientific problem in the academic community, and journal impact factors measure the influence of journals in the academic circle. They are usually regarded as indirect standards for measuring the influence of scientific research problems. However, they have certain limitations in evaluating the value of scientific problems and cannot accurately reflect the actual innovation, cutting-edge nature and application potential of scientific problems.

[0003] In summary, the existing technology has technical problems such as low efficiency in the value assessment of scientific issues due to reliance on expert evaluation and subjective judgment and inconsistent evaluation standards. Summary of the invention

[0004] The purpose of this application is to provide a method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism, in order to solve the technical problem in the prior art that the efficiency of scientific problem value assessment is low due to reliance on expert evaluation and subjective judgment and inconsistent evaluation standards.

[0005] In view of the above problems, the present application provides a method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism, wherein the method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism includes: reading predetermined quantization dimensions, wherein the predetermined quantization dimensions include target quantization dimensions; introducing a predetermined triple strategy to construct a target instruction fine-tuning data set of the target quantization dimensions; performing parameter quantization on the target instruction fine-tuning data set based on the QLoRA large language model fine-tuning principle to obtain a target recognition large model; obtaining a scientific problem text, performing semantic feature analysis on the scientific problem text through a gated network to obtain text semantics, and obtaining a target weight coefficient of the scientific problem text on the target quantization dimension based on the text semantics; analyzing the scientific problem text through the target recognition large model, and obtaining a target quantization result in combination with the target weight coefficient.

[0006] The technical solution provided in this application has at least the following technical effects or advantages:

[0007] By reading the predetermined quantitative dimensions, the predetermined quantitative dimensions include the target quantitative dimensions; introducing the predetermined triple strategy to construct the target instruction fine-tuning data set of the target quantitative dimensions; quantizing the parameters of the target instruction fine-tuning data set based on the QLoRA large language model fine-tuning principle to obtain the target recognition large model; obtaining the scientific problem text, performing semantic feature analysis on the scientific problem text through the gating network to obtain the text semantics, and obtaining the target weight coefficient of the scientific problem text on the target quantitative dimensions according to the text semantics; analyzing the scientific problem text through the target recognition large model, and obtaining the target quantization result in combination with the target weight coefficient. In other words, through clear predetermined quantitative dimensions, including theoretical value, application value, frontier value and innovation value, a special data set is constructed using the triple strategy, and the parameter quantization is performed using the QLoRA large language model fine-tuning principle, and the gating network is introduced to automatically analyze the key information of the text, and the weight coefficient of the scientific problem text on each target quantitative dimension is determined to obtain a quantitative evaluation result, thereby improving the efficiency of the quantitative evaluation of the value of scientific problems.

[0008] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented according to the contents of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are specifically cited below. It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easy to understand through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the present application or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0010] Figure 1 A flowchart of the scientific problem value quantification method that embeds a multi-dimensional hybrid expert mechanism for this application;

[0011] Figure 2 Schematic diagram of the process of constructing a target instruction fine-tuning dataset in the scientific problem value quantification method embedded in a multi-dimensional hybrid expert mechanism for this application. DETAILED DESCRIPTION

[0012] This application solves the technical problem in the prior art that the value assessment of scientific problems is inefficient due to the reliance on expert evaluation and subjective judgment and the existence of different evaluation standards by providing a scientific problem value quantification method embedded in a multi-dimensional hybrid expert mechanism. Through clear predetermined quantitative dimensions, including theoretical value, application value, frontier value and innovation value, a special data set is constructed using a triple strategy, and the parameters are quantified using the QLoRA large language model fine-tuning principle. A gated network is introduced to automatically analyze the key information of the text, determine the weight coefficient of the scientific problem text in each target quantitative dimension, and obtain a quantitative evaluation result, thereby improving the efficiency of the quantitative evaluation of the value of scientific problems.

[0013] Below, the technical solutions in the present application will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments of the present application. It should be understood that the present application is not limited to the example embodiments described herein. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application. It should also be noted that, for the convenience of description, only the parts related to the present application are shown in the accompanying drawings, rather than all of them.

[0014] For examples, please see the attached Figure 1 The present application provides a method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism, wherein the method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism specifically includes the following steps:

[0015] Step 1: Read a predetermined quantization dimension, where the predetermined quantization dimension includes a target quantization dimension.

[0016] The predetermined quantitative dimensions include at least theoretical value dimension, application value dimension, frontier value dimension and innovation value dimension.

[0017] Specifically, the predetermined quantitative dimensions are pre-set evaluation indicators used to measure the value of scientific problems. Each dimension represents an aspect of the value of a scientific problem, such as theoretical value, application value, frontier value, and innovation value. Quantitative dimensions are essentially standardized, numerical indicators of certain specific aspects of the evaluation object (such as scientific problems, technical papers, etc.) to measure its different values ​​or characteristics. The target quantitative dimensions are a subset of the predetermined quantitative dimensions. They are a set of dimensions set for specific evaluation goals (such as the value assessment of scientific problems). They are set to solve specific problems and provide guidance on how to evaluate goals.

[0018] The predetermined quantitative dimensions include at least theoretical value dimension, application value dimension, frontier value dimension and innovation value dimension. The theoretical value dimension measures the contribution of scientific problems or research results to the theoretical field. Its evaluation criteria include developing new theories or theoretical frameworks (i.e. creating a completely new theoretical system to explain certain phenomena or fields), improving or expanding existing theoretical systems (i.e. revising and improving existing theories to make them more complete or applicable), discovering new scientific laws or principles (i.e. research reveals previously unknown natural laws or scientific principles), establishing new mathematical models (i.e. mathematical structures that describe or predict phenomena), and providing theoretical explanations for observed phenomena (i.e. using theories to explain data or phenomena obtained in experiments or observations).

[0019] The application value dimension measures the application potential of scientific problems or research results in practice. Its evaluation criteria include solving actual industrial or technological problems (that is, the technology or theory can solve specific problems in practical applications), demonstrating commercial potential or economic benefits (that is, the economic benefits that technology or products can bring in the market), improving existing technical processes or methods (that is, optimizing existing technical processes or methods to improve efficiency or effectiveness), demonstrating the feasibility of real-world applications (that is, whether the technology or theory can be successfully applied in a practical environment), and providing cost-effective or efficient solutions (that is, the proposed solution is superior to existing solutions in terms of cost and efficiency).

[0020] The frontier value dimension measures the role of scientific problems or research results in promoting the development of disciplinary frontiers. Its evaluation criteria include leading or opening up new research directions (that is, research work can guide the scientific community to explore new fields), promoting cutting-edge technologies (that is, promoting technology to be at the forefront of current scientific or engineering practice), promoting interdisciplinary breakthroughs (that is, research can cross the boundaries of different disciplines and generate new insights or technologies), answering unsolved scientific problems (that is, research can solve long-standing unresolved scientific problems), and creating new research paradigms (that is, research can establish a completely new research model or framework).

[0021] The innovation value dimension measures the innovativeness of scientific problems or research results, and evaluates whether the research has proposed original scientific discoveries, innovative methodologies, technologies or tools, or whether it has made unique innovations in existing methods. Its evaluation criteria include making original scientific discoveries (that is, the research can bring unprecedented scientific knowledge or understanding), developing new methodologies or technologies (that is, creating completely new technologies or research methods), creating new tools or technologies (that is, inventing new equipment or tools to support research or production), proposing unique solutions to existing problems (that is, providing novel and effective solutions to known problems), combining existing methods in novel ways (that is, combining existing methods or technologies in innovative ways to produce new effects), etc.

[0022] By pre-setting quantitative dimensions, we ensure that the scientific problems or research results being evaluated have clear evaluation standards in various key areas, making the entire evaluation process more objective and systematic. By quantifying the four dimensions of theory, application, frontier and innovation, we can comprehensively evaluate the value of scientific problems and fully understand the value of scientific problems from different angles, which will help decision makers or researchers to accurately judge the true value of technology or research.

[0023] Step 2: Introduce a predetermined triple strategy to construct a target instruction fine-tuning dataset of the target quantization dimension.

[0024] Specifically, the predetermined triple strategy is a structured data organization method that includes three core parts: instruction unit, input unit, and output unit, which generates a standardized data format for the task. The instruction unit contains a clear task definition and supplements the definition description of theoretical value, application value, frontier value, and innovation value. The input unit contains a large number of sentence samples from scientific and technological literature. The definitions of different types of value sentences provide clear task boundaries and assist the model in understanding the characteristics of different types of sentences. The output unit adopts the annotation method of "sentence type: sentence content" to achieve fine-grained speech step mapping and provide language features and context information for the model. Obtain the target definition description of the theoretical value dimension, application value dimension, frontier value dimension, and innovation value dimension in the target quantitative dimension, combine the task boundary definition, generate clear task instructions for each dimension, and together constitute the instruction unit.

[0025] Collect scientific problem samples related to the target quantitative dimension from scientific literature, such as scientific research articles and technical reports in different fields, and store them in input units. Specify the corresponding quantitative evaluation results and explanations for each input sample, and score the target dimension (such as high, medium, and low) with explanations. Combine the instruction unit, input unit, and output unit into a standardized triple format. Each triple will represent a complete training sample, providing structured data for fine-tuning the model.

[0026] Through the above, a target instruction fine-tuning dataset containing multiple triplets was constructed to identify different types of scientific value sentences. Through clear task definitions and detailed descriptions of value types (theoretical value, application value, frontier value, and innovation value), clear task boundaries are provided for the model. The structure of the target instruction fine-tuning dataset adopts a triple form, including task instructions, sentence input samples, and fine-grained output annotations, aiming to help the model accurately understand and identify different types of scientific value sentences in scientific literature. Through the triple strategy, the data format is standardized so that the model can consistently process input and output, reduce manual intervention and data bias, and conduct a comprehensive evaluation of scientific issues through four dimensions (theory, application, frontier, and innovation), ensuring that the model can more comprehensively understand and evaluate the value of research.

[0027] Step 3: Based on the QLoRA large language model fine-tuning principle, the target instruction fine-tuning dataset is parameterized to obtain a large target recognition model.

[0028] Specifically, in order to improve the quality of knowledge extraction, the fine-tuning method of QLoRA is used to fine-tune the large model. QLoRA is an efficient fine-tuning method for large language models that combines two techniques: quantization and low-rank adaptation. Its main purpose is to maintain the performance of large language models while reducing memory usage. By quantizing the model parameters to a low-precision format (such as 4-bit) and introducing low-rank matrices to adapt to new tasks, QLoRA effectively solves the application problem of large models in resource-constrained environments. QLoRA first quantizes the model's weight matrix to 4-bit to reduce memory and computing requirements; then introduces a low-rank adaptation matrix (usually with a smaller dimension) to correct the performance loss caused by quantization, thereby optimizing the performance of the model.

[0029] The constructed target instruction fine-tuning dataset is input into the pre-trained large language model. The weights of the pre-trained model are usually floating point numbers (such as 32-bit floating point numbers). When fine-tuning, they are quantized to 4-bit format through the QLoRA method, which significantly reduces memory usage while retaining the effectiveness of the model on new tasks. Parameter quantization is the conversion of high-precision floating point numbers (such as 32-bit floating point numbers) to low-precision formats (such as 4-bit integers) to reduce computational complexity and memory requirements. 4-bit quantization refers to compressing each parameter value to 4 bits of data, which is usually achieved by reducing the precision of the number.

[0030] The QLoRA method converts the parameters of the model (such as the weight matrix W of the i-th layer) i ) is converted from the original high-precision floating point format (such as 32-bit floating point numbers) to 4-bit integers, and a low-rank adaptation matrix (the first projection matrix A i and the second projection matrix Bi ), correcting the accuracy loss caused by quantization, while enabling the model to quickly adapt to new tasks. Through forward calculation correction and optimization, the QLoRA model can effectively adapt to new tasks, combining the quantized original model parameters with the low-rank adaptation matrix to correct the performance loss caused by quantization. In the forward calculation, the model not only uses the quantized weight matrix for calculation, but also uses the projection matrix for correction to ensure that the output of each layer accurately reflects the goal of the new task. After the quantization, correction and optimization process, the model generates a large object recognition model that has been fine-tuned and adapted to the target task, which can show high accuracy on task-related datasets and effectively perform tasks (such as value assessment of scientific problems). With QLoRA, although the memory usage of the model is greatly reduced, the performance is still guaranteed, especially when it is necessary to process large-scale datasets and large language models, the advantages of QLoRA are very obvious.

[0031] Step 4: Obtain the scientific question text, perform semantic feature analysis on the scientific question text through a gating network to obtain text semantics, and obtain the target weight coefficient of the scientific question text on the target quantitative dimension based on the text semantics.

[0032] Specifically, scientific problem texts are obtained from scientific research literature, technical reports, academic articles or others. Scientific problem texts refer to texts containing descriptions of scientific problems, such as abstracts, sentences or paragraphs describing a research topic, problem or challenge, which are usually the basis for evaluating its theoretical, applied, frontier and innovative value. The scientific problem text may be a paragraph or sentence in a scientific document that describes a specific scientific research problem. A gated network is a neural network structure used to control the flow of information. It uses a gating mechanism to decide which information needs to be passed through the network and which information needs to be blocked. Gated networks are often used to process complex input data and can apply different processing strategies to different parts of the network to improve the flexibility and effectiveness of the model.

[0033] The gated network receives the input scientific question text and performs semantic feature analysis on the text through a series of neural network layers, performs operations such as word segmentation, word embedding, and context modeling on the text, and extracts important semantic information from the text. Semantic feature analysis refers to analyzing the grammatical, semantic, and contextual information in the text to mine important features related to scientific problems. Text semantics is usually a vector of fixed size, called a text semantic vector. After processing by the gated network, a four-dimensional semantic feature vector (logit vector) is obtained, each dimension represents the potential score of the text on a certain target quantitative dimension, and the four dimensions represent the preliminary scores of theoretical value, application value, frontier value, and innovation value.

[0034] Using the text semantic vector, a mapping layer or regression model is used to predict the weight coefficient of the scientific question text on the target quantitative dimension, that is, it is obtained by processing the semantic features of the text (such as using the Softmax function). The role of the Softmax function is to convert these original logit values ​​into a probability distribution so that their sum is 1. The weight coefficient of each target dimension represents the relative importance of the dimension in the evaluation. The Softmax function is a commonly used normalization tool that can convert the logit vector into a probability distribution for multi-classification problem processing. Through the gating network to analyze the semantic features of the text, the meaning of the scientific question text can be deeply understood, and accurate evaluation can be performed from four dimensions (theory, application, frontier, innovation). Through Softmax normalization, it can ensure that the scores of each dimension are comparable, and the sum of the results in all dimensions is 1, so that the model output is more stable and unified. Using the gating network and Softmax normalization, the model can automatically extract the weight coefficient of the target quantitative dimension from the input scientific question text, avoiding manual intervention and subjective judgment, and improving the objectivity and efficiency of the evaluation process.

[0035] Step 5: Analyze the scientific question text through the target recognition model, and obtain the target quantification result in combination with the target weight coefficient.

[0036] Specifically, the input scientific question text is analyzed through the fine-tuned target recognition large model, and the target recognition function corresponding to the target quantitative dimension is matched in the target recognition large model, that is, the specific function used by the model to evaluate the input text (scientific question text) in each quantitative dimension. The input scientific question text will be processed by the target recognition large model, and a target evaluation result will be generated, which is the score of each quantitative dimension, indicating the performance of the text in this dimension. The evaluation results of each dimension are weighted according to the target weight coefficient obtained previously to obtain the target quantitative result, that is, the comprehensive value score of the scientific question text, which integrates the performance of the text in all dimensions and is adjusted according to the importance of each dimension (that is, weight). By using the target recognition large model to analyze the scientific question text and combining the target weight coefficient to obtain the target quantitative result, the importance and value of scientific problems are evaluated, which helps to identify high-value scientific problems.

[0037] Further, as attached Figure 2 As shown, step 2 of this application includes:

[0038] The predetermined triple strategy includes an instruction unit, an input unit and an output unit; the distinguishing characteristics of the theoretical value dimension, the application value dimension, the frontier value dimension and the innovation value dimension are analyzed to determine the task boundary definition; the target definition description of the target quantitative dimension is obtained, and the instruction unit is composed of the task boundary definition; a scientific problem sample set of the target quantitative dimension is assembled, and the scientific problem sample set is stored in the input unit; a predetermined labeling scheme is read, and the predetermined labeling scheme is stored in the output unit; the instruction unit, the input unit and the output unit are constructed according to the predetermined triple strategy to obtain the target instruction fine-tuning data set.

[0039] Specifically, the predetermined triple strategy is a data construction scheme in which each triple consists of three parts: instruction unit, input unit and output unit, providing a structured and standardized data format for model training. The instruction unit contains the clear instructions or description of the task, telling the model what operation should be performed, usually including the definition and requirements of the task; the input unit contains the actual input data or samples, which is the information that the model needs to process when performing the task; the output unit defines the output result that the model should produce after processing the input data, usually the expected prediction or label.

[0040] According to the predetermined triple strategy, a target instruction fine-tuning dataset with four quantitative dimensions (theoretical value, application value, frontier value, and innovation value) is constructed. First, by performing a distinguishing feature analysis on the four dimensions (theoretical, application, frontier, and innovation value), the evaluation criteria and requirements of each dimension are clarified, and the characteristics of each dimension and its role in the task are understood. Distinguishing feature analysis is a detailed analysis of each quantitative dimension to determine the unique characteristics and evaluation criteria, thereby clarifying the task boundaries. The task boundary definition is to clarify the scope and objectives of the task, and defines the input, output, and features or attributes that the model needs to focus on.

[0041] The target definition description is a detailed explanation of each dimension in the target quantitative dimension, clarifying the specific meaning and evaluation criteria of each dimension. On the basis of the definition of the task boundary, the definition of each quantitative dimension is further obtained and described to ensure that the model can accurately understand the task requirements, that is, the evaluation criteria of each value dimension explained above. In the instruction unit, a clear task description is included, such as "given a scientific abstract, evaluate its innovative value, identify original discoveries, methodological innovations or novel applications, and supplement the definition descriptions of theoretical value, application value, frontier value and innovative value."

[0042] Then, we collect text samples of scientific questions related to these target quantitative dimensions to form a scientific question sample set. The scientific question sample set contains sentences from various types of scientific and technological literature, all of which cover the different characteristics of the above four dimensions. Each sample needs to be labeled to indicate the corresponding quantitative dimension (theory, application, frontier, innovation) and its evaluation results. The scientific question sample set is stored in the input unit, and a variety of training data can be received during fine-tuning. The definition of different types of value sentences provides clear task boundaries and assists the model in understanding the characteristics of different types of sentences. In the input field, the data set contains a large number of sentence samples from scientific and technological literature, such as a description of a technical solution or research, covering various sentence features in scientific and technological literature.

[0043] Read the predetermined annotation scheme, that is, the predefined annotation rules or standards, for annotating the data. The annotation form of "sentence type: sentence content" is adopted to achieve fine-grained speech step mapping, providing language features and context information for the model. For example: Theoretical value: medium, reasons: 1. Verified existing theoretical predictions; 2. Contributed to quantum computing theory; 3. Provided mathematical verification.

[0044] According to the predetermined triple strategy, the instruction unit, input unit and output unit are combined to obtain the target instruction fine-tuning dataset. For example, a target instruction fine-tuning dataset is constructed and a scientific problem is evaluated. Assume that the scientific problem text used is: "This study proposes a new artificial intelligence algorithm that can effectively improve the accuracy of image recognition and can process large-scale data in real time." Data annotation is performed according to the predetermined triple strategy: Instruction unit: The task description is "to evaluate the theoretical value, application value, frontier value and innovation value of this research." Input unit: Scientific problem text: "This study proposes a new artificial intelligence algorithm that can effectively improve the accuracy of image recognition and can process large-scale data in real time." Output unit: Theoretical value: medium (provides a new method but does not involve a new theoretical framework); Application value: high (can solve practical problems in the field of image recognition and has practical application potential); Frontier value: medium (although a new algorithm is applied, it does not involve revolutionary technological breakthroughs); Innovation value: high (a new algorithm is proposed, which is innovative).

[0045] By using the triple strategy, all data are standardized into triplets with a clear structure, making data processing and model training more consistent and reproducible, providing a multi-dimensional value assessment for each scientific problem, not only focusing on a single dimension such as citation frequency or h-index, but comprehensively evaluating its theoretical, application, frontier and innovation value, combining instruction units, input units and output units, accurately understanding the characteristics of each dimension, and making reasonable evaluation results.

[0046] Further, step three of this application includes:

[0047] Get a pre-trained model, the pre-trained model includes the original weight matrix W of the i-th layer i ; Obtain a predetermined projection matrix, wherein the predetermined projection matrix includes a first projection matrix A i and the second projection matrix B i , and the first projection matrix A i The dimension is (d, r), the second projection matrix B i The dimension is (r, d), where d is the hidden layer dimension, r is the projection dimension, and r is much smaller than d; combined with the first projection matrix A i and the second projection matrix B i The original weight matrix W quantized to 4-bit format i Perform forward calculation correction and optimization to obtain the target recognition large model.

[0048] Specifically, obtain a pre-trained model, that is, a model that has been trained on a large-scale dataset and has valid parameters and weights. A pre-trained model refers to a model that has been trained on a large-scale dataset and has learned the underlying patterns and features in the data. During the fine-tuning process, the original weight matrix of the pre-trained model will be used as a basis for modification to suit a specific task. Extract the original weight matrix W of the i-th layer from the pre-trained model i , that is, the object that needs fine-tuning.

[0049] Get the predetermined projection matrix, including the first projection matrix A i and the second projection matrix B i , the first projection matrix A i The dimension is (d, r), which is used to project the high-dimensional weight matrix into a low-dimensional space; the second projection matrix B i The dimension is (r, d), which is used to reproject the representation of the low-dimensional space to the original high-dimensional space; where d is the model hidden layer dimension, which represents the size of the feature space of the neural network at this layer; r is the projection dimension, which is usually much smaller than d. For example, r may be only a few hundred or a few thousand, while d may be a few thousand or tens of thousands.

[0050] The role of the projection matrix is ​​to map the large-scale weight matrix to a low-dimensional space through the first projection matrix, and then map it back to the original space from the low-dimensional space through the second projection matrix, which can not only reduce the computational and storage costs, but also maintain sufficient expressiveness during the training process. In order to further reduce memory usage, the original weight matrix W iQuantize to 4-bit format, that is, compress each parameter value from 32-bit or 64-bit floating point number to 4-bit representation, significantly reducing memory usage. The quantized weight matrix storage is more compact and improves computational and storage efficiency without affecting model accuracy. i and the second projection matrix B i Combined, using the first projection matrix A i Projecting the quantized weight matrix into a low-dimensional space can be achieved through matrix multiplication, using the second projection matrix B i Restoring the representation of the low-dimensional space to the original dimension is also done through matrix multiplication.

[0051] Perform forward calculation, that is, use the modified weight matrix to perform forward propagation of the neural network. The quantized original weight matrix W i The high-precision floating point number (such as 32-bit) has been converted to a 4-bit format, which greatly reduces the storage space of the model, but also introduces quantization errors. In order to correct the accuracy loss caused by quantization, QLoRA introduces low-rank adaptation technology. By introducing the first projection matrix A i and the second projection matrix B i , QLoRA can effectively correct the quantized model weights, reduce errors and improve performance. The dimensions (d, r) and (r, d) of the projection matrix represent its adaptability and compressibility. r is much smaller than d, so the projection matrix introduces only a small amount of additional computational overhead. When performing forward calculations, the quantized weight matrix W i With the projection matrix A i and B i Combined with computational corrections, the output of the quantized model can recover some of the errors caused by quantization, ensuring that the model still performs well in new tasks. Combined with the correction results of the quantized weight matrix and projection matrix, after the forward calculation is completed, the output of the model will be used for target recognition tasks (such as text classification, question answering, etc.). By optimizing the training data, QLoRA can generate a large target recognition model that performs well on the target task, that is, a large model that can effectively perform recognition tasks after quantization and low-rank adaptation optimization.

[0052] Furthermore, the present application also includes the following steps:

[0053] Read the forward calculation correction function, wherein the expression of the forward calculation correction function is as follows: i =f i (x i ); Among them, h i refers to the original forward calculation function of the i-th layer, x irefers to the scientific question text i, Refers to the original weight matrix W quantized to 4-bit format i Dequantize the weights to floating point numbers, Δ i =A i B i f i (x i ) refers to the low-rank correction term introduced by QLoRA.

[0054] Specifically, forward calculation refers to the process of obtaining output after the input data is calculated through each layer of neurons. The forward calculation correction function refers to the function that corrects and optimizes the results of this calculation process. In the QLoRA method, the correction function is used to correct the quantized model weights to improve the accuracy and performance of the model. The expression of the forward calculation correction function is as follows: i =f i (x i ); The error caused by quantization is corrected by low-rank adaptation, thereby improving the accuracy of the fine-tuned model. i refers to the original forward calculation function of the i-th layer, through the forward calculation function f i (x i ) to calculate, that is, by entering the scientific question text x i The result obtained by calculating the i-th layer; x i It refers to the scientific question text i, that is, the input data, which is a certain text (such as a sentence in a scientific paper) here, which is processed by a neural network to obtain the corresponding output.

[0055] Refers to the original weight matrix W quantized to 4-bit format i Dequantize the weights to floating point numbers. The quantization operation compresses the values ​​of the original weight matrix into 4-bit representation (low precision), while dequantization restores these 4-bit quantized values ​​to the original floating point values. Since the QLoRA method performs 4-bit quantization on the weight matrix, it is necessary to restore the quantized weight matrix to floating point format in each calculation; Δ i =A i B i f i (x i ) refers to the low-rank correction term introduced by QLoRA, which is composed of two projection matrices A i and B i And the forward calculation function f i (x i ) are calculated together. The low-rank correction term Δ i Add to the original forward calculation result h i, and get the new output h i ′ represents the corrected output of the i-th layer, which includes the influence of the dequantized weights and the low-rank correction term. After the forward calculation correction, the new model will be better adapted to the specific task and can perform more accurate analysis of the input data. During the fine-tuning process, the forward calculation correction function helps the model maintain high calculation accuracy while quantizing.

[0056] Furthermore, the present application also includes the following steps:

[0057] Obtain a predetermined loss function; perform forward calculation correction optimization with the minimum loss value of the predetermined loss function as the optimization goal to obtain the target recognition large model; wherein the expression of the predetermined loss function is as follows: Among them, θ 4bit refers to the original model parameters after quantization, refers to all the projection matrices introduced by QLoRA, D is the training data set of the new task, l is the task-related loss function, where only While keeping θ 4bit constant.

[0058] Specifically, the loss function is used to measure the gap between the model's prediction results and the true label. The predetermined loss function refers to the loss function that is pre-set during the model training process, which is used to evaluate the performance of the model and guide the optimization process. The loss function in the QLoRA optimization process is used to minimize the prediction error of the quantized model. Pre-determined loss function: Among them, θ 4bit refers to the original model parameters after quantization, refers to all the projection matrices introduced by QLoRA, D is the training data set of the new task, and l is the task-related loss function, which is used to calculate the error between the model output and the true label; in the optimization process, only While keeping θ 4bit constant. is the target loss function, which measures the performance of the quantized model and the modified projection matrix on the new task. (x, y) is a pair of samples in the training dataset D, x is the input data (e.g., scientific question text), and y is the target output (e.g., text classification label or regression value). Represents the forward computation output of the quantized original model and all projection matrices on the input x.

[0059] In the optimization process of QLoRA, the goal is to minimize the loss function L, that is, to optimize the performance of the model on the new task. The core of the QLoRA method is to maintain the quantized weight matrix θ 4bit No change, only update The modified model can produce the smallest loss on the training dataset of the new task. For example, suppose a quantized BERT model is fine-tuned using QLoRA. The original weight matrix θ of the i-th layer of the model 4bit The training dataset D contains 1000 sample pairs (x, y), where x is a text paragraph of a scientific paper and y is the corresponding label (e.g., a classification label). The cross entropy loss function is used as the task-related loss function l. During the optimization process, the quantized model parameters θ 4bit Keep it unchanged, assuming that the dimensions of the projection matrix are 1024×128 and 128×1024 respectively, and gradually update the projection matrix through the gradient descent method to minimize the loss function and optimize the projection matrix. By optimizing all the introduced projection matrices, the QLoRA method can effectively correct the errors caused by quantization, so that the quantized model can still maintain a high accuracy on the new task.

[0060] Further, step 4 of this application includes:

[0061] The scientific question text is feed-forward mapped through the gating network to obtain a four-dimensional logit vector; the four-dimensional logit vector is normalized using Softmax to obtain the target weight coefficient.

[0062] Specifically, the gating network processes and maps the scientific question text. Through the feedforward mapping process, the gating network extracts useful semantic information from the input text and converts it into a four-dimensional logit vector. In natural language processing tasks, the gating network is used to select and transmit useful information through a weighted mechanism to help the model focus on the most critical part of the input data. The structure of the gating network usually includes multiple gates to control different information flows, such as input gates, forget gates, output gates, etc. Each gate determines the transmission strength of information through learned weights. Logit refers to the original prediction value output by the neural network without activation. Logit is usually the original score of each category predicted by the model. The four-dimensional logit vector means that the network outputs a four-dimensional vector, in which each dimension corresponds to the score of the scientific question text on a certain dimension (such as theoretical value, application value, frontier value and innovation value). Feedforward mapping refers to the layer-by-layer calculation process from input data (such as scientific question text) to the output layer. In the feedforward network, information is transmitted through the network layer in a unidirectional manner, and the corresponding output is finally generated. Each layer of input vector will be weighted, nonlinearly activated and further processed to form the output.

[0063] After the input scientific question text is processed by feedforward mapping and weighted by the gating mechanism, the gating network can automatically adjust its information processing method according to different task requirements. The four-dimensional logit vector generated by the feedforward mapping reflects the preliminary evaluation of the scientific question text on four quantitative dimensions. These logit values ​​are not the final evaluation results, but the preliminary prediction scores of the text. Next, these logit values ​​can be further processed (such as Softmax normalization) to obtain the target weight coefficient for more accurate quantitative evaluation.

[0064] The Softmax function normalizes the above four-dimensional logit vector and converts it into a probability distribution. Through Softmax normalization, the logit value of each dimension will be converted to a value between 0 and 1, and the sum of the target weight coefficients of all four dimensions will be 1. Usually, the normalization calculation process includes: for each element in the logit vector, first calculate the exponent, then get the sum of the exponential values, calculate the probability value of each logit, and get the normalized target weight coefficient. The normalized target weight coefficient can convert the score of each dimension into a probability distribution, which is convenient for further analysis and decision-making, making the evaluation results of the text more standardized and unified, and avoiding the problem that the score values ​​of different dimensions are not comparable.

[0065] For example, if the four-dimensional logit vector is [2.3, 1.5, 0.7, 1.2], the result after Softmax processing may be [0.40, 0.30, 0.15, 0.15], that is, the target weight coefficients are 0.40, 0.30, 0.15 and 0.15 respectively, indicating that theoretical value is the most important, followed by application value, while frontier value and innovation value are relatively minor in this text. After Softmax normalization, the target weight coefficient obtained represents the relative importance of the scientific problem text in the four value dimensions. Through the feedforward mapping of the gating network, multi-dimensional information is extracted from the scientific problem text and converted into a logit vector. Through Softmax normalization, the original score of the scientific problem in the four dimensions can be converted into a target weight coefficient in the form of probability, thereby quantifying the relative importance of each dimension and more clearly identifying the contribution of the scientific problem in each value dimension in the evaluation task.

[0066] Further, step five of this application includes:

[0067] Matching the target recognition function corresponding to the target quantization dimension in the target recognition large model; analyzing the scientific question text through the target recognition function to obtain a target evaluation result; calculating the target evaluation result and the target weight coefficient to obtain the target quantification result.

[0068] Specifically, the target recognition large model (the model obtained by fine-tuning QLoRA) matches the corresponding target recognition function according to the target quantitative dimensions (such as theoretical value, application value, frontier value and innovation value). The target recognition function is a mathematical function used to calculate the score of a scientific problem text in a specific dimension based on the trained model parameters. For example, in the theoretical value dimension, the target recognition function may be evaluated based on the theoretical innovation, hypothesis verification, theoretical extension and other features in the text. The target recognition function is a text analysis function for different quantitative dimensions (such as theoretical value, application value, etc.). The target recognition function for each dimension is usually obtained by a deep learning model through a large amount of training, which can extract specific features in the text and give a score.

[0069] The input scientific question text is analyzed using the matched target recognition function to obtain the target evaluation result. Each target recognition function processes the key information in the text and outputs a corresponding score. The target recognition function processes the text to generate one or more evaluation scores, which represent the performance of the text in each quantitative dimension. The target evaluation result is a numerical score for each dimension, indicating the value of the text in that dimension. For example, suppose the target recognition function for application value gives an evaluation result of 3.8 points, and the evaluation result for innovation value is 4.5 points.

[0070] The target weight coefficient is a coefficient obtained by normalizing the Softmax function, which indicates the importance of each evaluation dimension in the target quantification. The target evaluation result is weighted and summed with the target weight coefficient, that is, the evaluation score on each dimension is multiplied by the target weight coefficient calculated previously, and the product results on all dimensions are weighted and summed to obtain the target quantification result, which reflects the comprehensive value of the scientific problem text on all target dimensions. Through the target recognition function, the text is analyzed in multiple dimensions, and combined with the weight coefficient, the scientific value of the text can be comprehensively evaluated, not limited to a single dimension, avoiding artificial subjectivity. Through weighted calculation, the information of different dimensions is effectively integrated to obtain a comprehensive target quantification result, making the evaluation more comprehensive and avoiding excessive reliance on or neglect of a certain dimension.

[0071] In summary, the scientific problem value quantification method embedded with a multi-dimensional hybrid expert mechanism provided in this application has the following technical effects:

[0072] By reading the predetermined quantitative dimensions, the predetermined quantitative dimensions include the target quantitative dimensions; introducing the predetermined triple strategy to construct the target instruction fine-tuning data set of the target quantitative dimensions; quantizing the parameters of the target instruction fine-tuning data set based on the QLoRA large language model fine-tuning principle to obtain the target recognition large model; obtaining the scientific problem text, performing semantic feature analysis on the scientific problem text through the gating network to obtain the text semantics, and obtaining the target weight coefficient of the scientific problem text on the target quantitative dimensions according to the text semantics; analyzing the scientific problem text through the target recognition large model, and obtaining the target quantization result in combination with the target weight coefficient. In other words, through clear predetermined quantitative dimensions, including theoretical value, application value, frontier value and innovation value, a special data set is constructed using the triple strategy, and the parameter quantization is performed using the QLoRA large language model fine-tuning principle, and the gating network is introduced to automatically analyze the key information of the text, and the weight coefficient of the scientific problem text on each target quantitative dimension is determined to obtain a quantitative evaluation result, thereby improving the efficiency of the quantitative evaluation of the value of scientific problems.

[0073] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

[0074] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalent technology, the present application is also intended to include these modifications and variations.

Claims

1. A scientific problem value quantification method embedded with a multi-dimensional hybrid expert mechanism, characterized by: include: Reading a predetermined quantization dimension, wherein the predetermined quantization dimension includes a target quantization dimension; Introducing a predetermined triple strategy to construct a target instruction fine-tuning dataset of the target quantization dimension; Based on the QLoRA large language model fine-tuning principle, the target instruction fine-tuning data set is parameterized to obtain a target recognition large model; Acquire a scientific question text, perform semantic feature analysis on the scientific question text through a gating network to obtain text semantics, and obtain a target weight coefficient of the scientific question text on the target quantitative dimension according to the text semantics; The scientific question text is analyzed by the target recognition model, and the target quantification result is obtained by combining the target weight coefficient.

2. The method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism according to claim 1 is characterized in that: The predetermined quantitative dimensions include at least theoretical value dimension, application value dimension, frontier value dimension and innovation value dimension.

3. The method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism according to claim 2 is characterized in that: The predetermined triple strategy includes an instruction unit, an input unit, and an output unit. The predetermined triple strategy is introduced to construct a target instruction fine-tuning data set of the target quantization dimension, including: Conduct distinguishing characteristic analysis on the theoretical value dimension, the applied value dimension, the frontier value dimension and the innovative value dimension to determine the definition of task boundaries; Obtaining a target definition description of the target quantization dimension and combining it with the task boundary definition to form the instruction unit; assembling a sample set of scientific questions of the target quantitative dimension, and storing the sample set of scientific questions in the input unit; Reading a predetermined labeling scheme, and storing the predetermined labeling scheme in the output unit; The instruction unit, the input unit and the output unit are constructed according to the predetermined triple strategy to obtain the target instruction fine-tuning data set.

4. The method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism according to claim 1 is characterized in that: Based on the QLoRA large language model fine-tuning principle, the target instruction fine-tuning dataset is parameterized to obtain a target recognition large model, including: Get a pre-trained model, the pre-trained model includes the original weight matrix W of the i-th layer i ; Obtain a predetermined projection matrix, wherein the predetermined projection matrix includes a first projection matrix A i and the second projection matrix B i , and the first projection matrix A i The dimension is (d, r), the second projection matrix B i The dimension is (r, d), where d is the hidden layer dimension, r is the projection dimension, and r is much smaller than d; Combined with the first projection matrix A i and the second projection matrix B i The original weight matrix W quantized to 4-bit format i Perform forward calculation correction and optimization to obtain the target recognition large model.

5. The method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism according to claim 4 is characterized in that: Read the forward calculation correction function, wherein the expression of the forward calculation correction function is as follows: h i =f i (x i ); Among them, h i refers to the original forward calculation function of the i-th layer, x i refers to the scientific question text i, Refers to the original weight matrix W quantized to 4-bit format i Dequantize the weights to floating point numbers, Δ i =A i B i f i (x i ) refers to the low-rank correction term introduced by QLoRA.

6. The method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism according to claim 5 is characterized in that: Based on the QLoRA large language model fine-tuning principle, the target instruction fine-tuning dataset is parameterized to obtain a target recognition large model, including: Obtain a predetermined loss function; Taking the minimum loss value of the predetermined loss function as the optimization goal, forward calculation correction optimization is performed to obtain the target recognition large model; Wherein, the expression of the predetermined loss function is as follows: Among them, θ 4bit refers to the original model parameters after quantization, refers to all the projection matrices introduced by QLoRA, D is the training data set of the new task, l is the task-related loss function, where only While keeping θ 4bit constant.

7. The method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism according to claim 1 is characterized in that: Obtaining a scientific question text, performing semantic feature analysis on the scientific question text through a gating network to obtain text semantics, and obtaining a target weight coefficient of the scientific question text on the target quantitative dimension according to the text semantics, including: Perform feed-forward mapping on the scientific question text through the gating network to obtain a four-dimensional logit vector; The four-dimensional logit vector is normalized using Softmax to obtain the target weight coefficient.

8. The method for quantifying the value of scientific problems embedded in a multi-dimensional hybrid expert mechanism according to claim 1 is characterized in that: The scientific question text is analyzed by the target recognition model, and the target quantification result is obtained by combining the target weight coefficient, including: Matching the target recognition function corresponding to the target quantization dimension in the target recognition large model; Analyze the scientific question text by using the target identification function to obtain a target evaluation result; The target evaluation result and the target weight coefficient are calculated to obtain the target quantification result.

Images