A method for constructing data training samples based on large language model

Abstract

The present invention discloses a method for constructing data training samples based on a large language model, which relates to the technical field of data training sample construction, including collecting raw data samples for training and performing cleaning operations; analyzing and processing the first data samples using a pre-trained language model to generate new data samples, and adjusting the pre-trained language model; performing key pattern recognition, extracting key features, and evaluating and tuning the newly generated data samples. The present invention analyzes raw data through a GPT‑4 pre-trained language model, identifies key patterns, and generates new data samples according to predefined rules and parameters, automatically generating high-quality training samples that support various machine learning tasks, especially in the field of natural language processing, significantly improving the diversity and coverage of data samples, while maintaining the quality of samples and improving the model performance of downstream tasks.

Description

Technical Field

[0001] The present invention relates to the technical field of data training sample construction, and in particular to a data training sample construction method based on a large language model. Background Art

[0002] In recent years, with the development of the field of natural language processing, new technologies and methods are constantly being used to improve the construction of data and the preparation of training samples. This includes the use of methods such as transfer learning, domain adaptation, and semi-supervised learning to improve model performance. When collecting and using data, ethical guidelines and privacy regulations must be followed to ensure the legality and security of the data, especially for the processing of sensitive information and personal identities. By effectively constructing and preparing training samples, language large models can better understand and generate natural language, thereby realizing their potential in various application scenarios.

[0003] At present, the Chinese invention patent with application number CN202311823749.1 discloses a training sample construction method and device for document question answering based on a large model. Although this method can automatically construct a batch of high-quality training samples by constructing prompt sets and chain reasoning processes, this process does not require manual sample writing, and can generate samples with quality comparable to manual writing, lowering the threshold for constructing training samples. However, the existing technology cannot automatically generate high-quality training samples that support various machine learning tasks, especially in the field of natural language processing, and cannot simultaneously maintain the quality of samples and improve the model performance of downstream tasks. Summary of the invention

[0004] S3: Using the adjusted pre-trained language model to perform key pattern recognition on the second data sample, obtain the key pattern, language structure and text pattern of the second data sample, extract key features of the key pattern, language structure and text pattern, generate a new data sample, define it as a third data sample, and evaluate and tune the third data sample;

[0005] S4: Use the adjusted pre-trained language model to randomly generate data, concatenate the randomly generated data with the third data sample to obtain a combined data set, which is defined as a fourth data sample, and increase the number and diversity of the fourth data sample through data enhancement technology to obtain a final data training sample.

[0006] As a preferred solution of the method for constructing data training samples based on a large language model described in the present invention, step S1 includes:

[0007] The sources of the raw data samples include text data on the Internet, news articles, social media content, and entertainment, gaming, education, business, reading, and map apps;

[0008] The cleaning operation on the first data sample includes:

[0009] Remove noise, repeated values ​​and non-text content in the first data sample;

[0010] Processing missing values ​​and outliers in the first data sample includes:

[0011] Using the code to identify missing values, the missing values ​​include null values ​​or NA, and performing a forward filling operation at the position of the missing value;

[0012] The statistical 3σ rule is used to identify outliers, and the interpolation method is used to correct outliers.

[0013] As a preferred solution of the method for constructing data training samples based on a large language model described in the present invention, step S2 includes:

[0014] A GPT-4 pre-trained language model is selected as a pre-trained language model, and the pre-trained language model is used to understand the textual meaning of the first data sample, and the long text is cut into paragraphs or sentences of appropriate lengths, and a context window of appropriate lengths of the paragraphs or sentences is selected as a new data sample, which is defined as a second data sample;

[0015] Input the second data sample into the pre-trained language model, use the Hugging FaceTransformers library to load the GPT-4 pre-trained language model, input the second data sample into the loaded pre-trained language model, define the training hyperparameter as the learning rate, combine the pre-trained GPT-4 pre-trained language model with the GRU layer, construct a GPT-4 pre-trained language model combined with the GRU layer, iterate and train the GPT-4 pre-trained language model combined with the GRU layer for multiple times, update the hyperparameters through the adaptive learning rate algorithm Adagrad in each iteration, and select the GPT-4 pre-trained language model combined with the GRU layer with the highest hyperparameter as the new pre-trained language model.

[0016] As a preferred solution of the method for constructing data training samples based on a large language model described in the present invention, combining the pre-trained GPT-4 pre-trained language model with the GRU gating layer to construct a language model that can deeply understand natural language includes:

[0017] A GRU layer is added to the output end of the GPT-4 pre-trained language model, and the hidden state of the last layer is extracted from the output of the GPT-4 pre-trained language model. The hidden state is used as the input of the GRU layer, and the outputs at different times are obtained by the reset gate and the update gate in sequence. The outputs at different times are input into the Sigmoid activation function to obtain the activation representation, and the final output state is obtained. The final output state is the key information of the hidden state, and the key information captures the deep meaning of the second data sample. The deep meaning includes but is not limited to emotional emphasis and positive and negative images. The GRU layer optimizes internal parameters through the back propagation algorithm to obtain the GPT-4 pre-trained language model combined with the GRU layer; the adaptive learning rate algorithm Adagrad adaptively adjusts the learning rate of each parameter, and scales the learning rate by accumulating the square of the gradient.

[0018] As a preferred solution of the method for constructing data training samples based on a large language model described in the present invention, step S3 includes:

[0019] The key pattern recognition includes:

[0020] Cluster analysis methods using association rules are used to retrieve patterns, associated item sets and rules in data and identify differences between categories;

[0021] The association rule clustering analysis method comprises:

[0022] Use the association rule Apriori algorithm to identify frequently occurring item sets in the second data sample, generate association rules based on the frequent item sets, select the K-means clustering algorithm, divide the second data sample into different clusters, the data samples within the same cluster have high similarity, and the data samples between different clusters have low similarity, and associate the cluster to which each data sample belongs with the association rule.

[0023] As a preferred solution of the method for constructing data training samples based on a large language model described in the present invention, associating the cluster to which the data sample belongs with the association rule includes:

[0024] Filter out the association rules with high correlation with the data samples in each cluster, filter according to the confidence of the frequent itemsets, select the association rules corresponding to the frequent itemsets with the largest confidence, use code to analyze the association rules corresponding to the frequent itemsets with the largest confidence in the cluster, identify the differences between the data samples of the cluster, and extract the characteristics of the cluster;

[0025] According to the characteristics of the clusters, different characteristics of different clusters are identified to obtain the characteristics of the key pattern, language structure, and text pattern of the second data sample, and the characteristics of the cluster corresponding to the association rule corresponding to the frequent item set with the largest confidence in the cluster corresponding to the key pattern, language structure, and text pattern are defined as key characteristics. The key characteristics are input into the GPT-4 pre-trained language model combined with the GRU layer, and the obtained output is defined as the third data sample.

[0026] As a preferred solution of the method for constructing data training samples based on a large language model described in the present invention, evaluating the third data sample includes:

[0027] The quality of the third data sample is evaluated based on the smoothness and symmetry of the density histogram of the third data sample with respect to the key feature. If the third data sample has low smoothness or no symmetry, the third data sample is of low quality. If the third data sample has high smoothness and symmetry, the third data sample is of high quality.

[0028] As a preferred solution of the method for constructing data training samples based on a large language model described in the present invention, the third data sample optimization includes:

[0029] The confidence parameters corresponding to each cluster are adjusted to obtain the changed key features, and the changed third data samples are obtained according to the key features. The confidence parameters are adjusted repeatedly several times to obtain several corresponding changed third data samples. According to the evaluation rules and adding manual review, all the third data samples are compared to obtain the third data sample with the highest quality.

[0030] As a preferred solution of the method for constructing data training samples based on a large language model described in the present invention, step S4 includes:

[0031] The data enhancement techniques include:

[0032] Synonym replacement, replacing some words in the text with synonyms;

[0033] Random insertions or deletions, randomly inserting extra words or deleting parts of the words in the text;

[0034] Generate synonyms and use word vector method to generate synonyms of text.

[0035] As a preferred solution of the method for constructing data training samples based on a large language model described in the present invention, natural language processing tasks include but are not limited to text classification, named entity recognition, language model, information retrieval and question answering, machine translation, speech recognition, text generation, sentiment analysis, semantic understanding and language generation.

[0036] The beneficial effects of the present invention are: automatically generating high-quality training samples that support various machine learning tasks, especially in the field of natural language processing, significantly improving the diversity and coverage of data samples, while maintaining the quality of samples and improving the model performance of downstream tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A basic flow chart of a method for constructing data training samples based on a large language model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0039] Reference Figure 1 , as an embodiment of the present invention, provides a method for constructing data training samples based on a large language model, comprising the following steps:

[0040] S1: Collecting original data samples for training, defining the original data samples as first data samples, importing the first data samples into a computing environment, performing cleaning operations on the first data samples to ensure data quality and consistency, and analyzing the characteristics, distribution and correlation of the data through visual statistical methods to understand the basic characteristics of the data;

[0041] S2: Select a pre-trained language model, use the pre-trained language model to analyze and process the first data sample, generate a new data sample, define it as a second data sample, and adjust the pre-trained language model according to the second data sample to make it more suitable for specific data and generation requirements;

[0042] S3: Using the adjusted pre-trained language model to perform key pattern recognition on the second data sample, obtain the key pattern, language structure and text pattern of the second data sample, extract key features of the key pattern, language structure and text pattern, generate a new data sample, define it as a third data sample, and evaluate and tune the third data sample;

[0043] S4: Use the adjusted pre-trained language model to randomly generate data, concatenate the randomly generated data with the third data sample to obtain a combined data set, which is defined as a fourth data sample, and increase the number and diversity of the fourth data sample through data enhancement technology to obtain a final data training sample.

[0044] The advanced GPT-4 pre-trained language model is used to analyze the raw data, identify key patterns, and generate new data samples based on predefined rules and parameters to increase the quantity and diversity of data samples.

[0045] Step S1 includes:

[0046] The sources of the raw data samples include text data on the Internet, news articles, social media content, and entertainment, gaming, education, business, reading, and map apps;

[0047] The cleaning operation on the first data sample includes:

[0048] Remove noise, repeated values ​​and non-text content in the first data sample;

[0049] Processing missing values ​​and outliers in the first data sample includes:

[0050] Using the code to identify missing values, the missing values ​​include null values ​​or NA, and performing a forward filling operation at the position of the missing value;

[0051] The statistical 3σ rule is used to identify outliers, and the interpolation method is used to correct outliers.

[0052] Step S2 includes:

[0053] A GPT-4 pre-trained language model is selected as a pre-trained language model, and the pre-trained language model is used to understand the textual meaning of the first data sample, and the long text is cut into paragraphs or sentences of appropriate length, which helps the pre-trained language model to learn longer context information, and a context window of appropriate length of the paragraph or sentence is selected as a new data sample, which is defined as a second data sample; the context window is a paragraph of text or a complete document;

[0054] Input the second data sample into the pre-trained language model, use the Hugging FaceTransformers library to load the GPT-4 pre-trained language model, input the second data sample into the loaded pre-trained language model, define the training hyperparameter as the learning rate, combine the pre-trained GPT-4 pre-trained language model with the GRU layer, construct a GPT-4 pre-trained language model combined with the GRU layer that can deeply understand natural language, iterate the GPT-4 pre-trained language model combined with the GRU layer for multiple times, update the hyperparameters through the adaptive learning rate algorithm Adagrad in each iteration, and select the GPT-4 pre-trained language model combined with the GRU layer with the highest hyperparameter as the new pre-trained language model.

[0055] The GPT-4 pre-trained language model, as a language model pre-trained on large-scale text, can better understand and generate natural language text.

[0056] Combining the pre-trained GPT-4 pre-trained language model with the GRU gating layer to build a language model that can deeply understand natural language includes:

[0057] A GRU layer is added to the output end of the GPT-4 pre-trained language model, and the hidden state of the last layer is extracted from the output of the GPT-4 pre-trained language model. The hidden state is used as the input of the GRU layer. The GRU layer can further process the hidden state to obtain a deeper semantic understanding or perform a specific sequence modeling task. The output of the GRU layer can be used for various downstream tasks, including text classification, named entity recognition, and sentiment analysis. The output at different times is obtained by resetting the gate and updating the gate in sequence, and the output at different times is input into the Sigmoid activation function to obtain the activation representation, and the final output state is obtained. The final output state is the key information of the hidden state. The key information captures the deep meaning of the second data sample, and the deep meaning includes but is not limited to emotional emphasis and positive and negative images. The GRU layer optimizes the internal parameters through the back-propagation algorithm to obtain the GPT-4 pre-trained language model combined with the GRU layer; this method allows the GRU layer to further process the input sequence based on the semantic representation generated by the GPT-4 pre-trained language model, thereby enhancing the semantic understanding ability of the model;

[0058] The adaptive learning rate algorithm Adagrad adaptively adjusts the learning rate of each parameter and scales the learning rate by accumulating the square of the gradient.

[0059] Step S3 includes:

[0060] The key pattern recognition includes:

[0061] Cluster analysis methods using association rules are used to retrieve patterns, associated item sets and rules in data and identify differences between categories;

[0062] The association rule clustering analysis method comprises:

[0063] Use the association rule Apriori algorithm to identify frequently occurring item sets in the second data sample, generate association rules based on the frequent item sets, such as {A, B}->{C}, select the K-means clustering algorithm, divide the second data sample into different clusters, the data samples in the same cluster have high similarity, and the data samples between different clusters have low similarity, and associate the cluster to which each data sample belongs with the association rule.

[0064] Associating the cluster to which the data sample belongs with the association rule comprises:

[0065] Filter out the association rules with high correlation with the data samples in each cluster, filter according to the confidence of the frequent itemsets, select the association rules corresponding to the frequent itemsets with the largest confidence, ensure the validity and relevance of the association rules, use code to analyze the association rules corresponding to the frequent itemsets with the largest confidence in the cluster, identify the differences between the data samples of the cluster, and extract the characteristics of the cluster;

[0066] The association rule clustering analysis method can explain the characteristics and association rules of each cluster and understand the data patterns and relationships behind them;

[0067] According to the characteristics of the clusters, different characteristics of different clusters are identified to obtain the characteristics of the key pattern, language structure, and text pattern of the second data sample, and the characteristics of the cluster corresponding to the association rule corresponding to the frequent item set with the largest confidence in the cluster corresponding to the key pattern, language structure, and text pattern are defined as key characteristics. The key characteristics are input into the GPT-4 pre-trained language model combined with the GRU layer, and the obtained output is defined as the third data sample.

[0068] The evaluating of the third data sample comprises:

[0069] The quality of the third data sample is evaluated based on the smoothness and symmetry of the density histogram of the third data sample with respect to the key feature. If the third data sample has low smoothness or no symmetry, the third data sample is of low quality. If the third data sample has high smoothness and symmetry, the third data sample is of high quality.

[0070] Optimizing the third data sample includes:

[0071] The confidence parameters corresponding to each cluster are adjusted to obtain the changed key features, and the changed third data samples are obtained according to the key features. The confidence parameters are adjusted repeatedly several times to obtain several corresponding changed third data samples. According to the evaluation rules and adding manual review, all the third data samples are compared to obtain the third data sample with the highest quality.

[0072] Step S4 includes:

[0073] The data enhancement techniques include:

[0074] Synonym replacement, replacing some words in the text with synonyms;

[0075] Random insertions or deletions, randomly inserting extra words or deleting parts of the words in the text;

[0076] Generate synonyms and use word vector method to generate synonyms of text.

[0077] Natural language processing tasks include but are not limited to text classification, named entity recognition, language modeling, information retrieval and question answering, machine translation, speech recognition, text generation, sentiment analysis, semantic understanding, and language generation.

[0078] The text classification: classifying text into predefined categories or labels;

[0079] Named entity recognition: identifying named entities with specific meanings in text, such as names of people, places, dates, etc.

[0080] The language model: understands and generates text through a pre-trained model, such as GPT;

[0081] Information retrieval and question answering: Retrieve relevant information from large-scale text or directly answer user questions, such as search engines and question answering systems;

[0082] The machine translation: automatically translates text from one language into another language;

[0083] Speech recognition: technology that converts speech signals into text form, such as voice assistants;

[0084] The text generation: generating text that complies with grammatical and semantic rules, such as automatic replies in a dialogue system;

[0085] Sentiment analysis: analyzing the sentiment tendency of a text, such as determining whether the sentiment of a review is positive, negative, or neutral;

[0086] Semantic understanding: understanding the semantic meaning of text, including the polysemy of word meanings and the influence of context;

[0087] The language generation described herein generates natural and fluent text according to semantic and grammatical rules.

[0088] The present invention analyzes the original data through the GPT-4 pre-trained language model, identifies key patterns, and generates new data samples according to predefined rules and parameters, automatically generating high-quality training samples that support various machine learning tasks, especially in the field of natural language processing, significantly improving the diversity and coverage of data samples while maintaining the quality of the samples and improving the model performance of downstream tasks.

[0089] It should be understood by those skilled in the art that the embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. Among them, the storage medium may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red Only Memory, referred to as PROM), read-only memory (Read Only Memory, referred to as ROM), magnetic memory, flash memory, magnetic disk or optical disk. These computer program instructions may also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which is implemented in the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0090] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for constructing data training samples based on a large language model, characterized in that: The following steps are involved: S1: Collecting original data samples for training, defining the original data samples as first data samples, importing the first data samples into a computing environment, performing a cleaning operation on the first data samples, and analyzing the characteristics, distribution, and correlation of the data by a visual statistical method; S2: Select a pre-trained language model, use the pre-trained language model to analyze and process the first data sample, generate a new data sample, define it as a second data sample, and adjust the pre-trained language model according to the second data sample; S3: Using the adjusted pre-trained language model to perform key pattern recognition on the second data sample, obtain the key pattern, language structure and text pattern of the second data sample, extract key features of the key pattern, language structure and text pattern, generate a new data sample, define it as a third data sample, and evaluate and tune the third data sample; S4: randomly generate data using the adjusted pre-trained language model, concatenate the randomly generated data with the third data sample to obtain a combined data set, which is defined as a fourth data sample, and increase the number and diversity of the fourth data sample by data enhancement technology to obtain a final data training sample; Step S2 includes: A GPT-4 pre-trained language model is selected as a pre-trained language model, and the pre-trained language model is used to understand the textual meaning of the first data sample, and the long text is cut into paragraphs or sentences of appropriate lengths, and a context window of appropriate lengths of the paragraphs or sentences is selected as a new data sample, which is defined as a second data sample; Input the second data sample into the pre-trained language model, use the Hugging Face Transformers library to load the GPT-4 pre-trained language model, input the second data sample into the loaded pre-trained language model, define the training hyperparameter as the learning rate, combine the pre-trained GPT-4 pre-trained language model with the GRU layer, construct a GPT-4 pre-trained language model combined with the GRU layer, iterate and train the GPT-4 pre-trained language model combined with the GRU layer for multiple times, update the hyperparameters by the adaptive learning rate algorithm Adagrad in each iteration, and select the GPT-4 pre-trained language model combined with the GRU layer with the highest hyperparameter as the new pre-trained language model; Combining the pre-trained GPT-4 pre-trained language model with the GRU gating layer to build a language model that can deeply understand natural language includes: Add a GRU layer at the output end of the GPT-4 pre-trained language model, extract the hidden state of the last layer from the output of the GPT-4 pre-trained language model, use the hidden state as the input of the GRU layer through the reset gate and the update gate in sequence to obtain outputs at different times, input the outputs at different times into the Sigmoid activation function to obtain activation representation, and obtain the final output state, the final output state is the key information of the hidden state, the key information captures the deep meaning of the second data sample, the deep meaning includes but is not limited to emotional emphasis and positive and negative images, the GRU layer optimizes internal parameters through the back propagation algorithm, and obtains the GPT-4 pre-trained language model combined with the GRU layer; The adaptive learning rate algorithm Adagrad adaptively adjusts the learning rate of each parameter and scales the learning rate by accumulating the square of the gradient.

2. The method for constructing data training samples based on a large language model according to claim 1, characterized in that: Step S1 includes: The sources of the raw data samples include text data on the Internet, news articles, social media content, and entertainment, gaming, education, business, reading, and map apps; The cleaning operation on the first data sample includes: Remove noise, repeated values ​​and non-text content in the first data sample; Processing missing values ​​and outliers in the first data sample includes: Using the code to identify missing values, the missing values ​​include null values ​​or NA, and performing a forward filling operation at the position of the missing value; The statistical 3σ rule is used to identify outliers, and the interpolation method is used to correct outliers.

3. The method for constructing data training samples based on a large language model according to claim 2, characterized in that: Step S3 includes: The key pattern recognition includes: Cluster analysis methods using association rules are used to retrieve patterns, associated item sets and rules in data and identify differences between categories; The association rule clustering analysis method comprises: Use the association rule Apriori algorithm to identify frequently occurring item sets in the second data sample, generate association rules based on the frequent item sets, select the K-means clustering algorithm, divide the second data sample into different clusters, the data samples within the same cluster have high similarity, and the data samples between different clusters have low similarity, and associate the cluster to which each data sample belongs with the association rule.

4. The method for constructing data training samples based on a large language model according to claim 3, characterized in that: Associating the cluster to which the data sample belongs with the association rule comprises: Filter out the association rules with high correlation with the data samples in each cluster, filter according to the confidence of the frequent itemsets, select the association rules corresponding to the frequent itemsets with the largest confidence, use code to analyze the association rules corresponding to the frequent itemsets with the largest confidence in the cluster, identify the differences between the data samples of the cluster, and extract the characteristics of the cluster; According to the characteristics of the clusters, different characteristics of different clusters are identified to obtain the characteristics of the key pattern, language structure, and text pattern of the second data sample, and the characteristics of the cluster corresponding to the association rule corresponding to the frequent item set with the largest confidence in the cluster corresponding to the key pattern, language structure, and text pattern are defined as key characteristics. The key characteristics are input into the GPT-4 pre-trained language model combined with the GRU layer, and the obtained output is defined as the third data sample.

5. The method for constructing data training samples based on a large language model according to claim 4, characterized in that: The evaluating of the third data sample comprises: The quality of the third data sample is evaluated based on the smoothness and symmetry of the density histogram of the third data sample with respect to the key feature. If the third data sample has low smoothness or no symmetry, the third data sample is of low quality. If the third data sample has high smoothness and symmetry, the third data sample is of high quality.

6. The method for constructing data training samples based on a large language model according to claim 5, characterized in that: Tuning the third data sample includes: The confidence parameters corresponding to each cluster are adjusted to obtain the changed key features, and the changed third data samples are obtained according to the key features. The confidence parameters are adjusted repeatedly several times to obtain several corresponding changed third data samples. According to the evaluation rules and adding manual review, all the third data samples are compared to obtain the third data sample with the highest quality.

7. The method for constructing data training samples based on a large language model according to claim 6, characterized in that: Step S4 includes: The data enhancement techniques include: Synonym replacement, replacing some words in the text with synonyms; Random insertions or deletions, randomly inserting extra words or deleting parts of the words in the text; Generate synonyms and use word vector method to generate synonyms of text.

Images