Association mining method based on large model

Abstract

The invention provides an association mining method based on a large model. The association mining method based on the large model comprises the following steps: constructing a standard expression according to a predetermined text format; cleaning input original text data according to the standard expression to obtain output text data conforming to the standard expression; outputting a preset sentence pattern text from the output text data based on a model according to the constraint of a preset prompt word template; and taking the predetermined sentence pattern text as the input of the model for training to obtain a strong correlation model so as to realize correlation mining. The question and answer matching degree of the model is improved through the predetermined cue word template, and the technical problem that the trained model lacks accuracy in the prior art is solved.

Description

Technical field:

[0001] The present invention relates to the field of artificial intelligence, and in particular to an association mining method based on a large model. Background technology:

[0002] The present invention is mainly based on large model technology and prompt word template design technology. In recent years, general artificial intelligence technology represented by large language models is developing and innovating rapidly, becoming the core technology engine to promote a new round of scientific and technological revolution and industrial transformation, and has been deployed in multiple industries and scenarios. In underwater command and decision-making scenarios, there are relatively few seats for underwater platform operators. Facing the dynamically changing hydrological environment and the sudden behavior of unknown targets, professionals are usually required to perform a large amount of query and calculation work, but lack the ability to respond quickly. Therefore, a large-model underwater domain knowledge template-based association mining method is studied to integrate underwater domain resources and provide an intelligent question-and-answer interface to assist operators in a series of decision-making behaviors.

[0003] Most of the existing underwater data governance methods in practical application are to clean and annotate the original data and then put it into the large model training process. However, in fact, after the data is pre-processed, there may still be redundant or low-domain-relevance information. After these low-quality data are fed into the large model in batches, the effect of improving the large model's domain professional knowledge question-and-answer capabilities is not significant, and it may result in the loss of the original general capabilities.

[0004] There is an urgent need for an association mining method based on a large model, which can help solve the technical problem of the lack of accuracy of the model after training in the existing technology. Summary of the invention:

[0005] In one embodiment, the present invention provides an association mining method based on a large model, which improves the question-answer matching degree of the model through a predetermined prompt word template, helping to solve the technical problem of the lack of accuracy of the model after training in the prior art.

[0006] The association mining method based on the large model includes:

[0007] Construct standard expressions based on predetermined text formats;

[0008] Cleaning the input original text data according to the standard expression to obtain output text data that conforms to the standard expression;

[0009] Outputting the output text data into a predetermined sentence text based on the model according to the constraints of the predetermined prompt word template;

[0010] The predetermined sentence text is used as the input of the model for training to obtain a strong association model to achieve association mining.

[0011] In one embodiment, the method for cleaning the original text data includes:

[0012] The original text data is cleaned and junk text is filtered, and text is deduplicated based on data evaluation of the model and user operations, and then output text data is obtained according to the standard expression.

[0013] In one embodiment, the predetermined prompt word template includes a prefix prompt word template and a cloze template.

[0014] In one embodiment, the predetermined prompt word template is constructed by task-driven and data-driven, as well as data of switch resources.

[0015] In one embodiment, the predetermined prompt word template is supervised prompt word data.

[0016] In one embodiment, before the step of cleaning the input original text data according to the standard expression to obtain output text data that conforms to the standard expression, the method further comprises:

[0017] The original resource data document is segmented and cut into multiple segments to generate original text data.

[0018] In one embodiment, the model collects prompt word data for training to obtain the strong association model to achieve iteration.

[0019] In one embodiment, the supervised prompt word data is used for training the model or for iteration of the strong association model.

[0020] In one embodiment, the predetermined sentence text is a typical question set.

[0021] In one embodiment, the raw data cleaning and the spam text filtering are modified based on the manual evaluation and the data model of the model. Description of the drawings:

[0022] Figure 1 A schematic diagram of a flow chart of an association mining method based on a large model in one embodiment of the present invention;

[0023] Figure 2 It is a schematic diagram of a flow chart of a method for constructing a dataset for fine-tuning underwater domain knowledge based on a large model in another embodiment of the present invention;

[0024] Figure 3 A schematic diagram of a data cleaning management process in another embodiment of the present invention;

[0025] Figure 4 A schematic diagram of a design scheme for constructing prompt words in another embodiment of the present invention;

[0026] Figure 5 A schematic diagram for generating a template-based question set in another embodiment of the present invention. Specific embodiment:

[0027] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0028] Various aspects and features of the present application are described herein with reference to the accompanying drawings.

[0029] These and other characteristics of the present application will become apparent from the following description of a preferred form of embodiment given as a non-limiting example with reference to the accompanying drawings.

[0030] It should also be understood that, although the present application has been described with reference to some specific examples, those skilled in the art will be able to realize many other equivalent forms of the present application that have the features described in the claims and are therefore within the scope of protection defined thereby.

[0031] The above and other aspects, features and advantages of the present application will become more apparent in view of the following detailed description when taken in conjunction with the accompanying drawings.

[0032] Specific embodiments of the present application are described hereinafter with reference to the accompanying drawings; however, it should be understood that the embodiments applied for are merely examples of the present application, which may be implemented in a variety of ways. Well-known and / or repeated functions and structures are not described in detail to determine the true intent based on the user's historical operations and to avoid unnecessary or redundant details that make the present application unclear. Therefore, the specific structural and functional details applied for herein are not intended to be limiting, but are merely used as the basis and representative basis for the claims to teach those skilled in the art to use the present application in a variety of ways with substantially any suitable detailed structure.

[0033] This specification may use the phrases "in one embodiment," "in another embodiment," "in yet another embodiment," or "in other embodiments," all of which may refer to one or more of the same or different embodiments according to the present application.

[0034] The core of knowledge extraction based on large model technology is the prompt word template design. By designing a structured hierarchical expression text template, the large model is guided to output underwater domain knowledge question and answer pairs according to a predetermined logic, generating high-quality underwater domain professional corpus. The purpose of the present invention is to overcome the technical defects of the existing data preprocessing technology that the data quality is low and needs further processing, and a template-based association mining method for underwater domain knowledge based on a large model is proposed.

[0035] Figure 1 A schematic diagram of a flow chart of an association mining method based on a large model in one embodiment of the present invention; Figure 2 It is a schematic diagram of a flow chart of a method for constructing a dataset for fine-tuning underwater domain knowledge based on a large model in another embodiment of the present invention; Figure 3 A schematic diagram of a data cleaning management process in another embodiment of the present invention; Figure 4 A schematic diagram of a design scheme for constructing prompt words in another embodiment of the present invention; Figure 5 A schematic diagram for generating a template-based question set in another embodiment of the present invention.

[0036] like Figures 1 to 5 As shown, in one embodiment, the present invention provides an association mining method based on a large model, and the association mining method based on a large model includes:

[0037] S101, constructing a standard expression according to a predetermined text format.

[0038] In this step, a specific step of constructing a standard expression according to a predetermined text format is provided.

[0039] S102: Clean the input raw text data according to the standard expression to obtain output text data that conforms to the standard expression. According to the format of the original underwater professional data resource, a regular expression is designed to complete the cleaning and management of the raw data resource to obtain semantically coherent text data.

[0040] S103, outputting the output text data according to the constraints of the preset prompt word template based on the model. The large model outputs underwater field professional question and answer pairs under specific constraints for specific text data, annotates the question and answer pairs, and obtains underwater field template-based high-quality associated data.

[0041] S104: The predetermined sentence text is used as the input of the model for training to obtain a strong correlation model to achieve correlation mining. The cleaned text data and underwater professional questions are used as the input of the big model, and an intelligent question and answer prompt word template is designed to constrain the big model to output: "You are an expert in the underwater field, familiar with the underwater sound propagation mechanism, underwater communication and detection calculation related knowledge. Please understand the input text and answer the questions asked."

[0042] In one embodiment, the method for cleaning the original text data includes:

[0043] The raw text data is cleaned and junk text is filtered based on the data evaluation of the model and the user operation, and the text is deduplicated, and then the output text data is obtained according to the standard expression. The large model outputs underwater professional problems under specific constraints for specific text data, manually annotates the problem set, and deletes low-quality data.

[0044] In one embodiment, the predetermined prompt word template includes a prefix prompt word template and a cloze template.

[0045] Prompt word templates can be divided into two categories: prefix prompt words and cloze prompt words. Prefix prompt words indicate filling in the subsequent text based on a prefix. By adding prefix prompt words, the model is given additional conditional information to guide the model to generate subsequent text; cloze prompt words indicate filling in the blanks in the sentence. Based on the language knowledge learned in the pre-trained language model, cloze prompt words are constructed using multiple data sets, and the pre-trained model predicts the missing words.

[0046] On this basis, we focus on optimizing and improving the application effect of large models in specific fields, design a customizable prompt word library to ensure that it is highly consistent with the business field, supports various military application scenarios, and can be flexibly adjusted and packaged. Through role setting, thought chain injection and prompt word compression, the original prompt words are comprehensively optimized to help the large model better capture the deep logic of the problem and reduce the computational cost in the reasoning process.

[0047] In one embodiment, the predetermined prompt word template is constructed by task-driven and data-driven, as well as data of switch resources.

[0048] In one embodiment, the predetermined prompt word template is supervised prompt word data.

[0049] In one embodiment, before the step of cleaning the input original text data according to the standard expression to obtain output text data that conforms to the standard expression, the method further comprises:

[0050] The original resource data document is segmented and cut into multiple segments to generate original text data.

[0051] Under the premise of ensuring the relevance of the previous and next contexts, the cleaned text data is sliced ​​into appropriate fields as the input of the big model, and a prompt word template for professional questions about underwater knowledge is designed to constrain the big model output: "You are an expert in the underwater field. Please understand the input text and raise questions related to the underwater sound propagation mechanism, underwater communication and detection calculation."

[0052] In one embodiment, the model collects prompt word data for training to obtain the strong correlation model to achieve iteration. Facing the problem of complex underwater environment and huge domain knowledge system, the training reasoning and application requirements of the auxiliary decision-making large model are combined with manual evaluation and model evaluation methods, and the data cleaning rules are iteratively evolved to integrate and build knowledge in the fields of underwater environment, target characteristics, communication and detection calculation, so as to solve the problem of scattered underwater data and unclear system classification.

[0053] In one embodiment, the supervised prompt word data is used for training the model or for iteration of the strong association model.

[0054] In one embodiment, the predetermined sentence text is a typical question set.

[0055] In one embodiment, the raw data cleaning and the spam text filtering are modified based on the manual evaluation and the data model of the model.

[0056] Beneficial effects:

[0057] Compared with the traditional data processing process, adding the prompt word template design process can generate domain-customized high-quality data sets, effectively improving the intelligent knowledge question-answering and decision-making support capabilities of the large decision-making support model.

[0058] The above embodiments are only exemplary embodiments of the present invention and are not intended to limit the present invention. The protection scope of the present invention is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present invention within the essence and protection scope of the present invention, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present invention.

Claims

1. A large model-based association mining method, characterized in that: The association mining method based on the large model includes: Construct standard expressions based on predetermined text formats; Cleaning the input original text data according to the standard expression to obtain output text data that conforms to the standard expression; Outputting the output text data into a predetermined sentence text based on the model according to the constraints of the predetermined prompt word template; The predetermined sentence text is used as the input of the model for training to obtain a strong association model to achieve association mining.

2. The large model-based association mining method according to claim 1, characterized in that: The method for cleaning the original text data comprises: The original text data is cleaned and junk text is filtered, and text is deduplicated based on data evaluation of the model and user operations, and then output text data is obtained according to the standard expression.

3. The large model-based association mining method according to claim 2, characterized in that: The predetermined prompt word templates include prefix prompt word templates and cloze templates.

4. The large model-based association mining method according to claim 3, characterized in that: The predetermined prompt word template is task-driven and data-driven, as well as constructed by the data of switch resources.

5. The large model-based association mining method according to claim 4, characterized in that: The predetermined prompt word template is supervised prompt word data.

6. The large model-based association mining method according to claim 5, characterized in that: Before the step of cleaning the input original text data according to the standard expression to obtain output text data that conforms to the standard expression, the method further comprises: The original resource data document is segmented and cut into multiple segments to generate original text data.

7. The large model-based association mining method according to claim 6, characterized in that: The model collects prompt word data for training to obtain the strong association model to achieve iteration.

8. The large model-based association mining method according to claim 7, characterized in that: The supervised prompt word data is used for training the model or for iteration of the strong association model.

9. The large model-based association mining method according to claim 8, characterized in that: The predetermined sentence text is a typical question set.

10. The large model-based association mining method according to claim 9, characterized in that: The raw data cleaning and the junk text filtering are modified according to the manual evaluation and the data model of the model.

Images