CN117076607A - Method, device and query system for establishing logical expressions using large language models - Google Patents

Abstract

The present disclosure provides a method, device and query system for establishing logical expressions using a large language model, and relates to the technical field of natural language processing. It includes: detecting the training type of the large language model; if the training type is completed, continue training, Then input the natural language text describing the logical expression to the large language model, and obtain the logical expression corresponding to the natural language text output by the large language model; continued training includes strengthening learning of the fine-tuning model and reward model obtained based on the large language model, The fine-tuning model is used to generate predictive expressions, and the reward model is used to score the predictive expressions to update the parameters of the fine-tuning model. Using the method of establishing logical expressions in this disclosure, the logical expression corresponding to the natural language text can be automatically obtained, which is efficient and convenient, greatly improving the processing efficiency, and at the same time, no manual participation is required, which greatly reduces the labor cost.

Description

Method, device and query system for establishing logical expressions using large language models

Technical field

The present disclosure generally relates to the technical field of natural language processing, and specifically relates to a method, device and query system for establishing logical expressions using a large language model.

Background technique

With the continuous development of electronic informatization, massive recording data has been generated in fields such as banking and insurance. These recording data include structured information such as attributes, business dimensions, and data time, as well as unstructured information such as call recordings and speech recognition texts.

There are many business scenarios for enterprises to analyze unstructured voice and text content, such as complaint cause analysis, transaction order analysis, call reason analysis, and agent violation analysis. At present, related technologies manually establish logical expressions of semantic tags. For example, a greeting can be represented by a logical expression "hello or hello", which is then converted into a full-text search engine input, and the matching index is detected through the full-text search engine. , and output structured semantic tags.

However, the syntax of different logical expressions is complex and diverse, which brings great learning costs to business personnel. At the same time, manually establishing logical expressions requires a lot of manpower to read the text, summarize the rules, and then create logical expressions, especially In complex business scenarios, manually created logical expressions often cannot cover all business scenarios, resulting in a low recall rate.

Contents of the invention

In view of the above defects or shortcomings in related technologies, it is expected to provide a method, device and query system for establishing logical expressions using a large language model, which can automatically obtain logical expressions, improve processing efficiency, and reduce labor costs at the same time.

In a first aspect, the present disclosure provides a method for establishing logical expressions using a large language model. The method includes:

Detect the training type of large language models;

If the training type is Completed and Continued Training, then input the natural language text describing the logical expression to the large language model, and obtain the logical expression corresponding to the natural language text output by the large language model; Continuing training includes performing reinforcement learning on the fine-tuned model obtained according to the large language model and the reward model, wherein the fine-tuned model is used to generate a predictive expression, and the reward model is used to score the predicted expression to update The parameters of the fine-tuned model.

Optionally, in some embodiments of the present disclosure, the fine-tuning model is obtained by fine-tuning the large language model based on manual annotation data, where the artificial annotation data includes a mapping relationship between natural language and logical expressions.

Optionally, in some embodiments of the present disclosure, the reward model is obtained by training the results of manual scoring and sorting of the prediction expression.

Optionally, in some embodiments of the present disclosure, the method further includes:

If the training type is to continue training without completion, an example of converting natural language into a logical expression is provided to the large language model, and then the natural language text is input to obtain the logical expression corresponding to the natural language text.

Optionally, in some embodiments of the present disclosure, the output format of the large language model includes json format, xml format, yaml format or field names.

In a second aspect, the present disclosure provides a device for establishing logical expressions using a large language model. The device includes:

Detection module, used to detect the training type of large language models;

Establish a module, configured to input a natural language text describing a logical expression to the large language model if the training type is completed and continue training, and obtain the logic corresponding to the natural language text output by the large language model. Expression; the continued training includes performing reinforcement learning on the fine-tuning model and the reward model obtained according to the large language model, wherein the fine-tuning model is used to generate a predictive expression, and the reward model is used to generate the predictive expression. Scoring is performed to update the parameters of the fine-tuned model.

Optionally, in some embodiments of the present disclosure, the fine-tuning model is obtained by fine-tuning the large language model based on manual annotation data, where the artificial annotation data includes a mapping relationship between natural language and logical expressions.

Optionally, in some embodiments of the present disclosure, the reward model is obtained by training the results of manual scoring and sorting of the prediction expression.

Optionally, in some embodiments of the present disclosure, the establishment module is also configured to provide the large language model with an example of converting natural language into a logical expression if the training type is incomplete and continue training, and then input The natural language text is used to obtain the logical expression corresponding to the natural language text.

In a third aspect, the present disclosure provides a query system. The logical expression of the query system is obtained by the method of establishing a logical expression described in any one of the first aspects.

It can be seen from the above technical solutions that the embodiments of the present disclosure have the following advantages:

Embodiments of the present disclosure provide a method, device and query system for using a large language model to create logical expressions. By continuing to train the large language model, its prediction results can be more accurate, thereby converting the natural language describing the logical expression into When text is input into this large language model, the logical expression corresponding to the natural language text can be automatically obtained, which is efficient and convenient, greatly improving the processing efficiency, and at the same time, no manual participation is required, which greatly reduces labor costs.

Description of the drawings

Other features, objects and advantages of the present disclosure will become more apparent upon reading the detailed description of the non-limiting embodiments with reference to the following drawings:

Figure 1 is a schematic flowchart of a method for establishing logical expressions using a large language model according to an embodiment of the present disclosure;

Figure 2 is a schematic flow chart of continued training of a large language model provided by an embodiment of the present disclosure;

FIG. 3 is a structural block diagram of a device for establishing logical expressions in a large language model provided by an embodiment of the present disclosure.

Detailed ways

In order to enable those skilled in the art to better understand the present disclosure, the following will clearly and completely describe the technical solutions in the present disclosure embodiments in conjunction with the accompanying drawings. Obviously, the described embodiments are only These are some embodiments of the present disclosure, but not all embodiments. Based on the embodiments in this disclosure, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the scope of protection of this disclosure.

The terms "first", "second", "third", "fourth", etc. (if present) in the description and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects without necessarily using Used to describe a specific order or sequence. It is to be understood that the figures so used are interchangeable under appropriate circumstances so that the described embodiments of the present disclosure can be practiced in sequences other than those illustrated or described herein.

In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or modules and need not be limited to those explicitly listed. Those steps or modules may instead include other steps or modules not expressly listed or inherent to the processes, methods, products or devices.

It should be noted that, as long as there is no conflict, the embodiments and features in the embodiments of the present disclosure can be combined with each other. In order to facilitate a better understanding of the embodiments of the present disclosure, the method, device and query system for establishing logical expressions using the large language model provided by the embodiments of the present disclosure are explained in detail below through FIGS. 1 to 3 .

Please refer to FIG. 1 , which is a schematic flowchart of a method for establishing logical expressions using a large language model according to an embodiment of the present disclosure. The method specifically includes the following steps:

S101, detect the training type of the large language model.

It should be noted that in the embodiment of the present disclosure, large language models (Large Language Models, LLM) refer to language models that are trained on large-scale text corpus and contain tens of billions of parameters (or more). For example, large language models can Including but not limited to GPT-3, PaLM, LLaMA, etc. Although existing large language models have strong text generation capabilities in general fields, their performance is still lacking for specific fields and specific tasks, especially when there is little data in the general corpus. Therefore, the embodiments of the present disclosure use the collected domain data and manual annotation data to continue training the large language model, which can enhance the ability of the large language model to convert natural language into logical expressions in relevant business fields, where the logical expression An expression refers to an expression that has a fixed syntax and is generally composed of keywords, reserved words and logical relationships (for example, AND, OR, NOT). It can be used to describe the structured and unstructured features of a document.

S102, if the training type is Completed and Continued Training, input the natural language text describing the logical expression to the large language model, and obtain the logical expression corresponding to the natural language text output by the large language model.

It should be noted that continued training includes but is not limited to reinforcement learning (Reinforcement Learning, RL) on fine-tuned models and reward models obtained based on large language models. Reinforcement learning is also called reinforcement learning, evaluation learning or reinforcement learning. It is a machine learning It is one of the paradigms and methodologies used to describe and solve problems in which an agent learns strategies to maximize returns or achieve specific goals during its interaction with the environment. Among them, the fine-tuning model is used to generate prediction expressions, and the reward model is used to score the prediction expressions to update the parameters of the fine-tuning model.

Illustratively, as shown in Figure 2, the embodiment of the present disclosure uses manual annotation data to perform fine-tuning training on a large language model to obtain fine-tuning model Model 1. The artificial annotation data may include the mapping relationship between natural language and logical expressions, that is, ( Natural language-logical expression) pairs, so that the fine-tuning model has the preliminary ability to convert natural language into logical expressions.

For example, natural language [prompt corresponding to the prediction stage]: Query the text of Wuhan workplace agents who did not say hello to customers from March 1, 2022 to May 1, 2023.

Logical expression [corresponding to the expected output of the prediction phase]:

It is understandable that the output format of the large language model here is json format only for example. In fact, according to the needs of different logical expression systems, the output format of the large language model can also be xml format, yaml format or field name, etc.

Furthermore, embodiments of the present disclosure can also collect a batch of natural languages and use fine-tuning models to generate corresponding prediction expressions. Afterwards, manual scoring is performed. The higher the score, the higher the quality of the prediction expression generated by the model. At this time, a reward model (Reward Model) is trained using the score ranking results. The reward model is a separate classification model, that is, using To determine which predictive expression generated by two models is better, it can be trained based on a fine-tuned model or a new large language model.

Furthermore, the embodiments of the present disclosure use reinforcement learning to enhance the fine-tuning model. This stage does not require manual labeling of data, that is, the reward model learned in the previous stage is used, and the parameters of the fine-tuning model are updated through the reinforcement learning method. Finally, the scoring process and reinforcement learning process are repeated until the model generation effect reaches the expectation, and a large language model that has completed continued training is obtained, which is called the Text2Query model.

In actual use, if the large language model continues to be trained, the natural language text describing the logical expression is directly input to the large language model according to the business needs. For example, the input is "You are a natural language-to-logical expression robot that helps you." I converted the following text into a logical expression: Query the text "Wuhan workplace agents did not say hello to customers from March 1, 2022 to May 1, 2023". At this time, the model output is the corresponding logical expression.

If the large language model has not been continuously trained, that is, S103, and the training type is incomplete continued training, then the large language model will be provided with examples of natural language converted into logical expressions. The advantage of this setting is that the accuracy requirements are not high and the labeling In scenarios where data is difficult to obtain, the In Context Learning (ICL) capability of large language models can be used for rapid learning, while saving computing power and meeting diversified usage needs. For example, the example is: You are a natural language into logical expression robot, which can convert natural language text like "Query the text that Wuhan workplace agents did not say hello to customers from March 1, 2022 to May 1, 2023" For the following logical expression: "{

Afterwards, the natural language text is input into the large language model to obtain the logical expression corresponding to the natural language text. For example, the input natural language text is: When the natural language description is "Query the text of Beijing workplace agents who did not say goodbye to customers from March 1, 2022 to May 1, 2023", what is the logical expression that should be output?

At this time, the embodiment of the present disclosure expects the output of the large language model to be:

The large language model provided by the embodiments of the present disclosure is used to create a method for logical expressions. By continuing to train the large language model, its prediction results are more accurate, and then the natural language text describing the logical expression is input into the large language model. , can automatically obtain the logical expression corresponding to the natural language text, which is efficient and convenient, greatly improving the processing efficiency, and at the same time, no manual participation is required, which greatly reduces the labor cost.

Based on the foregoing embodiments, embodiments of the present disclosure provide a device for establishing logical expressions using a large language model. The apparatus 100 for establishing a logical expression using a large language model can be applied to the method for establishing a logical expression using a large language model in the corresponding embodiment of FIGS. 1 to 2 . Please refer to Figure 3. The device 100 of the large language model for establishing logical expressions includes:

The detection module 101 is used to detect the training type of the large language model;

The establishment module 102 is used to input natural language text describing logical expressions to the large language model if the training type is completed and continue training, and obtain the logical expression corresponding to the natural language text output by the large language model; continuing training includes Reinforcement learning is performed based on the fine-tuning model and reward model obtained from the large language model, where the fine-tuning model is used to generate predictive expressions, and the reward model is used to score the predictive expression to update the parameters of the fine-tuning model.

Optionally, in some embodiments of the present disclosure, the fine-tuning model is obtained by fine-tuning and training the large language model based on manual annotation data, where the artificial annotation data includes the mapping relationship between natural language and logical expressions.

Optionally, in some embodiments of the present disclosure, the reward model is obtained by training the results of manual scoring and sorting of predictive expressions.

Optionally, in some embodiments of the present disclosure, the establishment module 102 is also used to provide the large language model with examples of converting natural language into logical expressions if the training type is to continue training without completion, and then input the natural language text to obtain the natural language model. The logical expression corresponding to the language text.

Optionally, the output format of the large language model in some embodiments of the present disclosure includes json format, xml format, yaml format or field names.

It should be noted that for descriptions of the same steps and content in this embodiment as in other embodiments, reference may be made to the descriptions in other embodiments, and will not be described again here.

The large language model provided by the embodiment of the present disclosure is a device for establishing logical expressions. By continuing to train the large language model, its prediction results are more accurate, and then the natural language text describing the logical expression is input into the large language model. , can automatically obtain the logical expression corresponding to the natural language text, which is efficient and convenient, greatly improving the processing efficiency, and at the same time, no manual participation is required, which greatly reduces the labor cost.

Based on the foregoing embodiments, embodiments of the present disclosure provide a query system. The logical expression of the query system is obtained through the method of establishing logical expressions in the corresponding embodiments of Figures 1 to 2. For example, the query system can be used for full-text retrieval, which means that a computer program scans each word in an article and creates an index for the necessary words, indicating the number and position of the word in the article. When users query text that contains certain words, they can search based on the established index, similar to the process of looking up words through a dictionary's search list.

As another aspect, embodiments of the present disclosure provide an electronic device including a processor and a memory. At least one instruction, at least one program, code set or instruction set is stored in the memory. The instruction, program, code set or instruction set is loaded and executed by the processor to implement the method of establishing a logical expression according to the embodiment corresponding to Figures 1 to 2 A step of.

As yet another aspect, an embodiment of the present disclosure provides a computer-readable storage medium for storing program code. The program code is used to execute any one of the methods of establishing logical expressions in the corresponding embodiments of FIGS. 1 to 2 . implementation.

Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be described again here.

In the several embodiments provided in this disclosure, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of modules is only a logical function division. In actual implementation, there may be other division methods. For example, multiple modules or components may be combined or can be integrated into another system, or some features can be ignored, or not implemented. On the other hand, the coupling or direct coupling or communication connection between each other shown or discussed may be through some interfaces, indirect coupling or communication connection of devices or modules, and may be in electrical, mechanical or other forms. Modules described as separate components may or may not be physically separated, and components shown as modules may or may not be physical units, that is, they may be located in one place, or they may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional module in various embodiments of the present disclosure may be integrated into one processing unit, or each module may exist physically alone, or two or more units may be integrated into one module. The above integrated units can be implemented in the form of hardware or software functional units. If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.

Based on this understanding, the technical solution of the present disclosure is essentially or contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , including several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method for establishing a logical expression in various embodiments of the present disclosure. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk and other media that can store program code. .

It should be noted that the above embodiments are only used to illustrate the technical solution of the present disclosure, but not to limit it; although the present disclosure has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still The technical solutions described in the foregoing embodiments may be modified, or some of the technical features may be equivalently replaced; however, these modifications or substitutions shall not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims (10)

1. A method for establishing logical expressions in a large language model, characterized in that the method includes: Detect the training type of large language models; If the training type is Completed and Continued Training, then input the natural language text describing the logical expression to the large language model, and obtain the logical expression corresponding to the natural language text output by the large language model; Continuing training includes performing reinforcement learning on the fine-tuned model obtained according to the large language model and the reward model, wherein the fine-tuned model is used to generate a predictive expression, and the reward model is used to score the predicted expression to update The parameters of the fine-tuned model. 2. The method according to claim 1, characterized in that the fine-tuning model is obtained by fine-tuning the large language model based on manual annotation data, and the artificial annotation data includes a mapping relationship between natural language and logical expressions. . 3. The method according to claim 1, characterized in that the reward model is obtained by training the artificial scoring and sorting results of the prediction expression. 4. The method according to any one of claims 1 to 3, characterized in that the method further includes: If the training type is to continue training without completion, an example of converting natural language into a logical expression is provided to the large language model, and then the natural language text is input to obtain the logical expression corresponding to the natural language text. 5. The method according to claim 4, characterized in that the output format of the large language model includes json format, xml format, yaml format or field name. 6. A device for establishing logical expressions in a large language model, characterized in that the device includes: Detection module, used to detect the training type of large language models; Establish a module, configured to input a natural language text describing a logical expression to the large language model if the training type is completed and continue training, and obtain the logic corresponding to the natural language text output by the large language model. Expression; the continued training includes performing reinforcement learning on the fine-tuning model and the reward model obtained according to the large language model, wherein the fine-tuning model is used to generate a predictive expression, and the reward model is used to generate the predictive expression. Scoring is performed to update the parameters of the fine-tuned model. 7. The device according to claim 6, wherein the fine-tuning model is obtained by fine-tuning the large language model based on manual annotation data, and the artificial annotation data includes a mapping relationship between natural language and logical expressions. . 8. The device according to claim 6, wherein the reward model is obtained by training the artificial scoring and sorting results of the prediction expression. 9. The device according to any one of claims 6 to 8, characterized in that the establishment module is also used to provide natural language conversion to the large language model if the training type is incomplete and continued training. is an example of a logical expression, and then the natural language text is input to obtain the logical expression corresponding to the natural language text. 10. A query system, characterized in that the logical expression of the query system is obtained through the method of establishing a logical expression using a large language model according to any one of claims 1 to 5.