Information processing method and device based on large language model and storage medium

Abstract

The invention discloses an information processing method and device based on a large language model, a storage medium and electronic equipment. The method comprises the steps of obtaining question expression information containing at least two pieces of sub-information; an attachment relationship between the at least two pieces of sub-information is obtained, the problem expression information is processed into N pieces of sub-problem information through the attachment relationship, the attachment relationship is used for describing connection or association between the sub-information, and the sub-problem information in the N pieces of sub-problem information is used for expressing the sub-problems in the N sub-problems; and respectively inputting each piece of sub-question information in the N pieces of sub-question information into M sub-models in the large language model to obtain K question information features, and obtaining question reply information based on the K question information features. The technical problem that information processing based on a large language model is low in accuracy is solved.

Description

Technical Field

[0001] This application relates to the field of computers, and in particular, to an information processing method, apparatus, storage medium, and electronic device based on a large language model. Background Art

[0002] In the information processing scenario based on a large language model, a large language model with a single structure performs poorly in expressing information for complex problems. For example, when the object of the problem expression information is unclear or there are multiple targets, directly processing such a problem may result in an unsatisfactory result, which may lead to a lower accuracy of information processing based on the large language model. Therefore, there is a problem of low accuracy in information processing based on the large language model.

[0003] In response to the above problems, no effective solution has been proposed yet. Summary of the Invention

[0004] Embodiments of this application provide an information processing method, apparatus, storage medium, and electronic device based on a large language model to at least solve the technical problem of low accuracy in information processing based on the large language model.

[0005] According to one aspect of the embodiments of this application, an information processing method based on a large language model is provided, including: obtaining problem expression information including at least two sub-information, where the problem expression information is used to instruct the large language model to output a problem response information that matches the problem expression information, the problem expression information includes at least one problem object and at least one problem target, the problem object is the object being queried mentioned in the problem expression information, the problem target is the target that the problem expression information expects to achieve by querying the problem object, the at least one problem object and the at least one problem target form N sub-problems, and N is an integer greater than or equal to 2; obtaining the dependency relationship between each sub-information in the at least two sub-information, and using the dependency relationship to process the problem expression information into N sub-problem information, where the dependency relationship is used to describe the connection or association between each sub-information, and the sub-problem information in the N sub-problem information is used to express the sub-problems in the N sub-problems; respectively inputting each sub-problem information in the N sub-problem information into M sub-models in the large language model to obtain K problem information features, and based on the K problem information features, obtaining the problem response information, where different sub-models are configured with different model parameters, the problem information features are used to characterize the problem content of the sub-problem information, M is an integer greater than or equal to 2, and K is less than or equal to the product of N and M.

[0006] According to another aspect of the embodiments of the present application, there is also provided an information processing device based on a large language model, including: a first acquisition unit, configured to acquire problem expression information including at least two sub-informations, where the problem expression information is used to instruct the large language model to output problem response information matching the problem expression information, the problem expression information includes at least one problem object and at least one problem target, the problem object is the object being queried mentioned in the problem expression information, the problem target is the target expected to be achieved by querying the problem object through the problem expression information, the at least one problem object and the at least one problem target form N sub-problems, and N is an integer greater than or equal to 2; a second acquisition unit, configured to acquire the dependency relationship between each of the at least two sub-informations, and process the problem expression information into N sub-problem informations by using the dependency relationship, where the dependency relationship is used to describe the connection or association between each of the sub-informations, and the sub-problem information in the N sub-problem informations is used to express the sub-problem in the N sub-problems; a first response unit, configured to respectively input each of the sub-problem informations in the N sub-problem informations into M sub-models in the large language model to obtain K problem information features, and based on the K problem information features, obtain the problem response information, where different ones of the sub-models are configured with different model parameters, the problem information features are used to characterize the problem content of the sub-problem information, M is an integer greater than or equal to 2, and K is less than or equal to the product of N and M.

[0007] As an optional solution, the second acquisition unit includes: a first acquisition module, configured to determine a first sub-information from the at least two sub-informations, and acquire the connection relationship between each of the first sub-informations, where the first sub-information is a noun or a phrase centered on the noun; a first determination module, configured to determine a second sub-information associated with the first sub-information from the at least two sub-informations, where the second sub-information is a verb; a second determination module, configured to determine the first sub-information having the connection relationship and the second sub-information associated with the first sub-information having the connection relationship in the at least two sub-informations as having the dependency relationship.

[0008] As an alternative solution, the above-mentioned first acquisition unit includes: a second acquisition module for acquiring the above-mentioned problem expression information input by the user account; the above-mentioned second acquisition unit includes: a third determination module for determining, from the above-mentioned at least two sub-information, a plurality of sub-information having the above-mentioned dependency relationship; a supplement module for, when the above-mentioned historical problem information input by the user account has been acquired and the above-mentioned plurality of sub-information lacks key sub-information, using the above-mentioned historical problem information to supplement the above-mentioned plurality of sub-information to obtain the above-mentioned N sub-problem information, where the above-mentioned key sub-information is the information of the above-mentioned first sub-information or the above-mentioned second sub-information.

[0009] As an alternative solution, the above-mentioned first reply unit includes: a third acquisition module for acquiring the first similarity between each problem information feature among the above-mentioned K problem information features and each problem reply feature in the knowledge base associated with the above-mentioned large language model; a fourth determination module for determining, from the above-mentioned knowledge base, the first P initial reply information corresponding to the first P problem reply features in descending order of the above-mentioned first similarity; a fourth acquisition module for obtaining the above-mentioned problem reply information based on the above-mentioned P initial reply information.

[0010] As an alternative solution, the above-mentioned fourth acquisition module includes: a word segmentation sub-module for performing word segmentation processing on each of the above-mentioned sub-problem information and each initial reply information in the above-mentioned P initial reply information to obtain an N sub-problem information set and an N initial reply information set, where the above-mentioned sub-problem information set includes a plurality of sub-problem words or short sentences, and the above-mentioned initial reply information set includes a plurality of reply words or short sentences; an acquisition sub-module for acquiring the second similarity between each sub-problem information set in the above-mentioned N sub-problem information sets and each initial reply information set in the above-mentioned N initial reply information sets; a determination sub-module for determining, from the above-mentioned P initial reply information, the initial reply information with the largest above-mentioned second similarity as the above-mentioned problem reply information.

[0011] As an alternative solution, the above-mentioned device further includes at least one of the following: a deletion module, configured to delete first processing information in each of the sub-question information and each of the initial response information in the above-mentioned P initial response information before performing word segmentation processing on each of the sub-question information and each of the initial response information to obtain N sub-question information sets and N initial response information sets, where the influence degree of the first processing information on the information semantics is less than or equal to a first preset threshold; a prohibition module, configured to set second processing information in each of the sub-question information and each of the initial response information to be prohibited from word segmentation before performing word segmentation processing on each of the sub-question information and each of the initial response information in the above-mentioned P initial response information to obtain N sub-question information sets and N initial response information sets, where the second processing information is a pre-specified fixed word or short sentence.

[0012] As an alternative solution, the above-mentioned acquisition sub-module includes: a first acquisition sub-unit, configured to acquire a first sub-similarity between each sub-question information set in the above-mentioned N sub-question information sets and keywords or short sentences in each of the initial response information sets; a second acquisition sub-unit, configured to acquire a second sub-similarity between each sub-question information set in the above-mentioned N sub-question information sets and non-keywords or short sentences in each of the initial response information sets; an allocation sub-unit, configured to allocate a first weight to the first sub-similarity and allocate a second weight to the second sub-similarity, where the first weight is greater than the second weight; an integration sub-unit, configured to integrate the first sub-similarity and the second sub-similarity according to the first weight and the second weight to obtain the second similarity.

[0013] As an alternative solution, the above-mentioned device further includes: a fifth acquisition module, configured to acquire knowledge text information to be input into the knowledge base before respectively inputting each of the N sub-question information into the M sub-models in the large language model to obtain K question information features; a splitting module, configured to, before respectively inputting each of the N sub-question information into the M sub-models in the large language model to obtain K question information features, split the knowledge text information into at least two first text paragraphs when the information length of the knowledge text information is greater than or equal to a second preset threshold; an extraction module, configured to, before respectively inputting each of the N sub-question information into the M sub-models in the large language model to obtain K question information features, extract respective corresponding first paragraph keywords from each of the at least two first text paragraphs; a subdivision module, configured to, before respectively inputting each of the N sub-question information into the M sub-models in the large language model to obtain K question information features, subdivide each of the first text paragraphs into at least two first text segments, and assign respective corresponding first paragraph keywords to the first text segments, wherein there is overlapping text information between the at least two first text segments.

[0014] As an alternative solution, the above-mentioned device further includes: a first display unit, configured to display at least two virtual character identifiers on the Q&A platform of the large language model before obtaining the question reply information based on the K question information features, wherein the question expression information is input into the Q&A platform, and the virtual character identifiers are used to indicate virtual characters for answering questions based on the large language model, and different virtual characters match different reply templates; the device further includes: a second reply unit, configured to, after obtaining the question reply information based on the K question information features, when the virtual character for answering questions based on the large language model is a specific virtual character, acquire a specific reply template matched by the specific virtual character, and process the question reply information using the specific reply template to obtain specific reply information; a second display unit, configured to display the specific reply information on the Q&A platform after obtaining the question reply information based on the K question information features.

[0015] As an alternative solution, the above-mentioned device further includes: a third acquisition unit, configured to acquire initial question information before acquiring the question expression information including at least two sub-information, wherein the initial question information is used to instruct the large language model to output a response information matching the initial question information; a splitting unit, configured to, before acquiring the question expression information including at least two sub-information, when the information length of the initial question information is greater than or equal to a third preset threshold, split the question expression information into at least two second text paragraphs; an extraction unit, configured to, before acquiring the question expression information including at least two sub-information, extract respective corresponding second paragraph keywords from each of the at least two second text paragraphs; a subdivision unit, configured to, before acquiring the question expression information including at least two sub-information, subdivide each of the at least two second text paragraphs into at least two second text segments, and assign respective corresponding second paragraph keywords to the second text segments, to obtain the at least two sub-information, wherein there is overlapping text information between the at least two second text segments.

[0016] According to another aspect of the embodiments of the present application, there is provided a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the information processing method based on a large language model as described above.

[0017] According to another aspect of the embodiments of the present application, there is further provided an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the above-mentioned information processing method based on a large language model through the computer program.

[0018] In an embodiment of the present application, problem expression information including at least two sub-information is obtained. The problem expression information is used to instruct a large language model to output problem response information matching the problem expression information. The problem expression information includes at least one problem object and at least one problem target. The problem object is the object being questioned mentioned in the problem expression information, and the problem target is the target that the problem expression information expects to achieve by questioning the problem object. The at least one problem object and the at least one problem target form N sub-problems, where N is an integer greater than or equal to 2; the dependency relationship between each sub-information in the at least two sub-information is obtained, and the problem expression information is processed into N sub-problem information by using the dependency relationship. The dependency relationship is used to describe the connection or association between each sub-information. The sub-problem information in the N sub-problem information is used to express the sub-problems in the N sub-problems; each sub-problem information in the N sub-problem information is respectively input into M sub-models in the large language model to obtain K problem information features, and based on the K problem information features, the problem response information is obtained. Different sub-models are configured with different model parameters. The problem information features are used to characterize the problem content of the sub-problem information. M is an integer greater than or equal to 2, and K is less than or equal to the product of N and M.

[0019] Facing the complex situation where the problem expression information consists of multiple sub-problems composed of problem objects and problem targets, in this embodiment, through the dependency relationship between each sub-information in the problem expression information, the problem expression information is split into sub-problem information corresponding to each sub-problem, so that the large language model faces sub-problem information with a simple situation instead of complex problem expression information; in addition, in this embodiment, the large language model is divided into multiple sub-models, and multiple sub-models are used to respectively perform information characterization on different dimensions of each sub-problem information. The configuration of multiple sub-problem information and multiple sub-models enables a larger number of problem information features to be obtained in information characterization, and the richness of information that can be referred to when outputting the problem response information is naturally higher. Furthermore, the purpose of improving the richness of information referred to when outputting the problem response information is achieved, thereby realizing the technical effect of improving the accuracy of information processing based on the large language model, and further solving the technical problem of low accuracy of information processing based on the large language model. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of the present application and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings:

[0021] Figure 1 is a schematic diagram of an application environment of an optional information processing method based on a large language model according to an embodiment of the present application;

[0022] Figure 2 It is a schematic diagram of the process of an alternative information processing method based on a large language model according to an embodiment of the present application;

[0023] Figure 3 It is a schematic diagram of an alternative information processing method based on a large language model according to an embodiment of the present application;

[0024] Figure 4 It is a schematic diagram of another alternative information processing method based on a large language model according to an embodiment of the present application;

[0025] Figure 5 It is a schematic diagram of another alternative information processing method based on a large language model according to an embodiment of the present application;

[0026] Figure 6 It is a schematic diagram of another alternative information processing method based on a large language model according to an embodiment of the present application;

[0027] Figure 7 It is a schematic diagram of another alternative information processing method based on a large language model according to an embodiment of the present application;

[0028] Figure 8 It is a schematic diagram of another alternative information processing method based on a large language model according to an embodiment of the present application;

[0029] Figure 9 It is a schematic diagram of an alternative information processing device based on a large language model according to an embodiment of the present application;

[0030] Figure 10 It is a schematic diagram of the structure of an alternative electronic device according to an embodiment of the present application. Detailed implementation manners

[0031] In order to enable those skilled in the art of the present technology to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0032] It should be noted that the terms "first", "second", etc. in the description, claims and the above-mentioned drawings of this application are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application described here can be implemented in an order other than those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0033] According to one aspect of the embodiments of the present application, an information processing method based on a large language model is provided. Optionally, as an alternative implementation, the above-mentioned information processing method based on a large language model can be, but is not limited to, applied to an environment such as Figure 1 shown. Among them, it may include, but is not limited to, a user device 102 and a server 112. The user device 102 may include, but is not limited to, a display 104, a processor 106 and a memory 108. The server 112 includes a database 114 and a processing engine 116.

[0034] The specific process can be as follows:

[0035] Step S102, the user device 102 obtains problem expression information containing at least two sub-information;

[0036] Step S104, send the problem expression information to the server 112 through the network 110;

[0037] Steps S106-S110, the server 112 obtains the dependency relationship between each sub-information among at least two sub-information through the processing engine 116, and processes the problem expression information into N sub-problem information by using the dependency relationship; further input each sub-problem information in the N sub-problem information into M sub-models in the large language model to obtain K problem information features, and based on the K problem information features, obtain problem response information;

[0038] Step S112, send the problem response information to the user device 102 through the network 110. The user device 102 displays the problem response information on the display 104 through the processor 106, and stores the position information of the above-mentioned rope points for pathfinding in the memory 108.

[0039] Except Figure 1In addition to the examples shown, the above steps can be completed independently by the user equipment or the server, or jointly by the user equipment and the server. For example, the above steps S106 - S110 and the like are executed by the user equipment 102, thereby reducing the processing pressure on the server 112. The user equipment 102 includes, but is not limited to, a handheld device (such as a mobile phone), a laptop computer, a tablet computer, a desktop computer, a vehicle-mounted device, a smart TV, etc. The present application does not limit the specific implementation manner of the user equipment 102. The server 112 can be a single server or a server cluster composed of multiple servers, or a cloud server.

[0040] Optionally, as an alternative implementation, as Figure 2 shown, the information processing method based on the large language model can be executed by an electronic device, such as Figure 1 the user equipment or the server shown, and the specific steps include:

[0041] S202, obtain problem expression information including at least two sub-informations, where the problem expression information is used to instruct the large language model to output problem response information matching the problem expression information. The problem expression information includes at least one problem object and at least one problem target. The problem object is the object being questioned mentioned in the problem expression information, and the problem target is the target that the problem expression information expects to achieve by asking the problem object. At least one problem object and at least one problem target form N sub-problems, and N is an integer greater than or equal to 2;

[0042] S204, obtain the dependency relationship between each sub-information among at least two sub-informations, and use the dependency relationship to process the problem expression information into N sub-problem information, where the dependency relationship is used to describe the connection or association between each sub-information, and the sub-problem information in the N sub-problem information is used to express the sub-problems in the N sub-problems;

[0043] S206, input each sub-problem information in the N sub-problem information into M sub-models in the large language model to obtain K problem information features, and based on the K problem information features, obtain the problem response information, where different sub-models are configured with different model parameters, the problem information features are used to characterize the problem content of the sub-problem information, M is an integer greater than or equal to 2, and K is less than or equal to the product of N and M.

[0044] Optionally, in this embodiment, the above information processing method based on the large language model can be but is not limited to being applied to an application platform built based on the large language model. Users can input questions for which they want to know the answers through this application platform, and then the application platform recalls the answers and displays the recalled answers on the application platform. Ideally, the questions are clear in object and single in goal, but in real scenarios, user questions are usually unclear in object or have multiple goals. Directly recalling answers based on such a question may, but is not limited to, result in unsatisfactory results. Therefore, in this embodiment, to address such situations, questions that are unclear in object or have multiple goals are first processed into sub-questions with clear objects and single goals, and then the application platform recalls answers for the sub-questions.

[0045] Optionally, in this embodiment, the large language model (LLM) is a model based on machine learning and natural language processing technologies. It can, but is not limited to, generate natural language text or understand the meaning of language text by training on a large amount of text data. Specifically, the large language model can handle various natural language tasks, such as text classification, question answering, dialogue, etc.

[0046] For further illustration, optionally, the model structure of the large language model can include, but is not limited to, at least one of the following: a preprocessing layer, a word embedding layer, an encoder, a decoder, an output layer, etc. Among them, the preprocessing layer can, but is not limited to, be responsible for preprocessing the input text data, including tasks such as word segmentation, part-of-speech tagging, named entity recognition, etc., to facilitate subsequent processing by the model; the word embedding layer can, but is not limited to, be responsible for mapping the input words or phrases into a high-dimensional space to enable the model to better understand the semantic relationships between words; the encoder can, but is not limited to, be the core of the large language model, which can transform the input text sequence into an intermediate semantic representation; the decoder can, but is not limited to, be responsible for transforming the intermediate semantic representation output by the encoder into a text output; the output layer can, but is not limited to, be responsible for post-processing the text output by the decoder, including tasks such as sentence-level semantic understanding, sentiment analysis, text classification, etc.

[0047] Optionally, in this embodiment, as Figure 3 shown, the question expression information is used to ask "how to make manual A and manual B", and then the large language model will, according to the indication of the question expression information 302, output question response information matching the question expression information 302, such as "The production steps of manual A and manual B are as follows..." etc.

[0048] Optionally, in this embodiment, the question object is the object being questioned mentioned in the question expression information, which can, but is not limited to, refer to entities, concepts, or themes clearly mentioned in the question expression information. For example Figure 3In the problem expression information 302 shown, the object being questioned can be, but is not limited to, understood as "Manual A" and "Manual B". The large language model can accept such an object being questioned and attempt to extract information related to "Manual A" and "Manual B" from the text data as the answer to the question.

[0049] Optionally, in this embodiment, the problem objective is the goal that the problem expression information expects to achieve by questioning the problem object, or rather, the problem objective is the questioning purpose or intention in the problem expression information. The problem objective can be diverse and depends on the content and background of the problem expression information. The problem objective can also be, but is not limited to, to obtain more information about the problem object, confirm a certain fact, solve a doubt, provide a suggestion, etc. For example, Figure 3 In the problem expression information 302 shown, the problem objective can be, but is not limited to, understood as "how to make". Furthermore, the large language model can accept the problem expression information and attempt to understand the problem objective and the problem object therein. Then, the large language model can search for and extract information related to the problem object to output an answer or response that conforms to the problem objective. This answer or response can be an explanation, a description, a proof, a recommendation, etc., depending on the specific circumstances of the problem objective and the problem object, such as the specific production steps.

[0050] Optionally, in this embodiment, at least one problem object and at least one problem objective can, but are not limited to, form multiple sub-questions, such as Figure 4 In the problem expression information 402 shown, "What beneficial elements and non-beneficial elements to the body are contained in Vegetable A and Vegetable B", where "Vegetable A" is the object being questioned 404, "Vegetable B" is the object being questioned 406, "What beneficial elements to the body are contained" is the problem objective 408, "(What) non-beneficial elements to the body" is the problem objective 410, and multiple sub-questions can be formed by the object being questioned 404, the object being questioned 406, the problem objective 408, and the problem objective 410, such as sub-question 412 "What beneficial elements to the body are contained in Vegetable A", sub-question 414 "What non-beneficial elements to the body are contained in Vegetable A", sub-question 416 "What beneficial elements to the body are contained in Vegetable B", and sub-question 410 "What non-beneficial elements to the body are contained in Vegetable B".

[0051] Optionally, in this embodiment, the dependency relationship is used to describe the connection or association between each sub-information. In language processing, the dependency relationship usually refers to the syntactic relationship between words or phrases and is used to express their semantic relationship. When analyzing the sentence structure, this embodiment can, but is not limited to, consider the dependency relationship between words, such as the relationship between the subject and the predicate, the verb and the object, the attributive and the headword, etc.

[0052] Optionally, in this embodiment, before processing the problem expression information into N sub-problem information using the dependency relationship, it is possible but not limited to first split the problem expression information into independent words or phrases, and determine the semantic type of each word or phrase, such as word segmentation, part-of-speech tagging, and named entity recognition, etc. Then use the dependency relationship to connect these words or phrases to form N-tuple problem information.

[0053] To further illustrate with an example, optionally, for example, use the dependency relationship to process the problem expression information into N sub-problem information. According to the dependency relationship between the subject and the predicate, connect them to form a binary tuple. Similarly, extract other components, such as the object, attributive, etc., and add them to the binary tuple according to their dependency relationship with the core word. Continue to extract other components until all the words or phrases in the problem expression information are processed. Finally, connect each binary tuple to form an N-tuple problem information. This N-tuple can contain all the important information in the problem expression information and can be used for subsequent tasks such as problem understanding and answering.

[0054] Optionally, different model structures can model the problem from different perspectives and levels according to the characteristics of the problem, so as to better capture the internal features and laws of the problem, which helps to improve the reasoning ability and prediction accuracy of the model. Furthermore, in this embodiment, for the large language model, multiple sub-models with different model parameters are used to respectively perform content representation on the input sub-problems. Different model parameters can better represent the characteristics of the problem content, so that the large language model can more accurately understand and interpret the problem content, and then make accurate problem responses.

[0055] Optionally, in this embodiment, it is possible but not limited to first determine the semantic type and content of each sub-problem, and then, based on the semantic type and content, match different or the same sub-models for each sub-problem information. For example, the sub-problem information 1 of semantic type 1 is input into sub-model 1, and the sub-problem information 2 and sub-problem information 3 of semantic type 2 are input into sub-model 2;

[0056] Furthermore, in this embodiment, it is also possible but not limited to adopt an exhaustive method to sequentially input each sub-problem information into M sub-models, such as Figure 5 As shown, each sub-problem information (such as sub-problem information 1, sub-problem information 2, sub-problem information 3) in the 3 (N) sub-problem information is respectively input into the 2 (M) sub-models (such as sub-model 1, sub-model 2) in the large language model to obtain 6 (K) problem information features (such as problem information feature 1, problem information feature 2, problem information feature 3, problem information feature 4, problem information feature 5, problem information feature 6). In this embodiment, K is equal to the product of N and M.

[0057] Optionally, in this embodiment, based on the K question information features, question response information can be further obtained. For example, the K question information features can be aggregated and analyzed to obtain a comprehensive understanding and answer to the question. Here, aggregation and analysis can be understood as integrating and processing the question information features to obtain a more comprehensive and in-depth understanding of the question;

[0058] Taking it a step further for illustration, optionally, for example, multiple question information features can be concatenated or weighted and summed to obtain a comprehensive feature representation that contains content from multiple aspects of the question. Then, by analyzing and processing the features, an answer to the question can be further obtained. Or, by classifying or clustering the features, keywords or topics related to the question can be identified, thus providing support for subsequent tasks such as searching or recommendation.

[0059] It should be noted that in the complex situation where the question expression information consists of multiple sub-questions composed of a question object and a question target, in this embodiment, based on the dependency relationship between each sub-information in the question expression information, the question expression information is split into sub-question information corresponding to each sub-question, so that the large language model faces sub-question information with a simple situation instead of complex question expression information; in addition, this embodiment also divides the large language model into multiple sub-models and uses multiple sub-models to perform information characterization in different dimensions for each sub-question information. The configuration of multiple sub-question information and multiple sub-models results in a larger number of question information features obtained from the information characterization, and naturally, the richness of information that can be referred to when outputting the question response information is also higher. Furthermore, the purpose of improving the richness of information referred to when outputting the question response information is achieved, thereby realizing the technical effect of improving the accuracy of information processing based on the large language model.

[0060] Taking it a step further for illustration, optionally based on Figure 3 the scenario shown, continue for example Figure 6As shown, obtain the problem expression information 302, where the problem expression information 302 is used to instruct the large language model 604 to output the problem response information 606 that matches the problem expression information 302. The problem expression information 302 includes at least one problem object (such as "Manual A", "Manual B") and at least one problem target (such as "how to make"), and the at least one problem object and the at least one problem target form 2 sub-problems; obtain the dependency relationship between each sub-information in the problem expression information 302, and use the dependency relationship to process the problem expression information 302 into 2 sub-problem information 602 (such as "how to make Manual A", "how to make Manual B"), where the dependency relationship is used to describe the connection or association between each sub-information; input each sub-problem information 602 in the 2 sub-problem information 602 into 2 sub-models (such as sub-model 1, sub-model 2) in the large language model 604, obtain 4 problem information features (such as problem information feature 1, problem information feature 2, problem information feature 3, problem information feature 4), and based on the 4 problem information features, obtain the problem response information 606 (such as "The production steps of Manual A are as follows...", "The production steps of Manual B are as follows...").

[0061] Through the embodiments provided in this application, problem expression information including at least two sub-informations is obtained. Among them, the problem expression information is used to instruct the large language model to output problem response information matching the problem expression information. The problem expression information includes at least one problem object and at least one problem target. The problem object is the object being questioned mentioned in the problem expression information, and the problem target is the goal that the problem expression information expects to achieve by asking the problem object. At least one problem object and at least one problem target form N sub-problems, where N is an integer greater than or equal to 2; the dependency relationship between each of the at least two sub-informations is obtained, and the problem expression information is processed into N sub-problem informations by using the dependency relationship. The dependency relationship is used to describe the connection or association between each sub-information, and the sub-problem information in the N sub-problem informations is used to express the sub-problems in the N sub-problems; each of the N sub-problem informations in the N sub-problem informations is respectively input into M sub-models in the large language model to obtain K problem information features, and based on the K problem information features, problem response information is obtained. Among them, different sub-models are configured with different model parameters. The problem information feature is used to characterize the problem content of the sub-problem information. M is an integer greater than or equal to 2, and K is less than or equal to the product of N and M. Facing the complex situation where the problem expression information consists of multiple sub-problems composed of problem objects and problem targets, in this embodiment, through the dependency relationship between each sub-information in the problem expression information, the problem expression information is split into sub-problem informations corresponding to each sub-problem respectively, so that the large language model faces sub-problem informations with simple situations instead of problem expression informations with complex situations; in addition, in this embodiment, the large language model is also divided into multiple sub-models, and multiple sub-models are used to respectively perform information characterization on each sub-problem information in different dimensions. The configuration of multiple sub-problem informations and multiple sub-models makes the number of problem information features obtained from the information characterization more, and the richness of the information that can be referred to when outputting the problem response information is naturally higher. Furthermore, the purpose of improving the richness of the information referred to when outputting the problem response information is achieved, thereby realizing the technical effect of improving the accuracy of information processing based on the large language model.

[0062] As an optional solution, obtaining the dependency relationship between each of the at least two sub-informations includes:

[0063] S1-1, determining a first sub-information from the at least two sub-informations, and obtaining the connection relationship between each of the first sub-informations, where the first sub-information is a noun or a phrase centered on a noun;

[0064] S1-2, determining a second sub-information associated with the first sub-information from the at least two sub-informations, where the second sub-information is a verb;

[0065] S1-3. Determine that the first sub-information with a connection relationship and the second sub-information associated with the first sub-information with a connection relationship among at least two sub-informations have an attachment relationship.

[0066] Optionally, in this embodiment, the type of sub-information in a sentence can be determined, but not limited to, through part-of-speech tagging technology. For example, which are nouns or noun-centered phrases, and which are verbs. Furthermore, through attachment relationship analysis, the connection relationships between sub-informations in the sentence can be determined. For example, which sub-informations have a subject-predicate relationship, an object-verb relationship, etc.

[0067] Optionally, in this embodiment, the sub-informations and the connection relationships between them can be integrated, but not limited to, to construct an attachment relationship tree or graph to visually display the structure and semantics of the sentence. And by analyzing the attachment relationship of the sentence, the semantics and meaning of the sentence can be better understood, thus providing better support for subsequent natural language processing tasks.

[0068] It should be noted that by analyzing the attachment relationship of natural language, the accuracy of sentence understanding and semantic analysis can be improved, and then the problem expression information can be processed into more accurate N sub-problem information, improving the processing accuracy of sub-problem information.

[0069] For further illustration, optionally, for example, determining which sub-informations are nouns or noun-centered phrases, and which sub-informations are verbs can be achieved through part-of-speech tagging technology in natural language processing. Once the types of these sub-informations are determined, the attachment relationships between sub-informations can be determined. And through attachment relationship analysis, it can be determined which first sub-informations have connection relationships, and which first sub-informations have connection relationships with the second sub-informations. These connection relationships can be represented as an attachment relationship tree or an attachment relationship graph, where each node represents a sub-information and each edge represents an attachment relationship. Finally, it is necessary to determine which first sub-informations and second sub-informations have an attachment relationship. This can be achieved by checking the connection relationships between each first sub-information and its associated second sub-information. If the connection relationship exists, then it can be considered that there is an attachment relationship between these two sub-informations.

[0070] Through the embodiments provided by the present application, the first sub-information is determined from at least two sub-informations, and the connection relationship between each first sub-information is obtained, where the first sub-information is a noun or a phrase centered on a noun; the second sub-information associated with the first sub-information is determined from at least two sub-informations, where the second sub-information is a verb; among at least two sub-informations, the first sub-information with a connection relationship and the second sub-information associated with the first sub-information with a connection relationship are determined to have an attachment relationship, thereby achieving the purpose of improving the understanding of sentences and the accuracy of semantic analysis through the analysis of the attachment relationship of natural language, and thus realizing the technical effect of improving the processing accuracy of sub-problem information.

[0071] As an alternative solution, obtaining the problem expression information including at least two sub-informations includes: obtaining the problem expression information input by the user account;

[0072] As an alternative solution, processing the problem expression information into N sub-problem informations by using the attachment relationship includes:

[0073] S2-1, determining multiple sub-informations with an attachment relationship from at least two sub-informations;

[0074] S2-2, when the historical problem information already input by the user account is obtained and the multiple sub-informations lack key sub-informations, using the historical problem information to supplement the multiple sub-informations to obtain N sub-problem informations, where the key sub-information is the information of the first sub-information or the second sub-information.

[0075] Optionally, in this embodiment, the historical problem information already input by the user account may but is not limited to refer to the problem information input by the user during the use of the product or service. These historical problem informations may include data such as text, voice, and images input by the user in aspects such as querying information, solving problems, purchasing products or services.

[0076] It should be noted that since splitting the problem expression information into multiple sub-informations may but is not limited to result in the lack of content of the sub-informations, thereby reducing the information integrity of the sub-problem informations, this embodiment can use the historical problem information already input by the user account to supplement the multiple sub-informations to obtain sub-problem informations with complete information. By analyzing the attachment relationship and complementing information of natural language, the structure and semantics of sentences can be better understood, thereby improving the accuracy of natural language processing tasks.

[0077] For further illustration, optionally in a question-and-answer system, through the analysis of the attachment relationship and information complementation of questions, the user's questions can be answered more accurately; in an intelligent customer service, by analyzing and processing the language input by the user, the user's needs can be understood more accurately and corresponding solutions can be provided.

[0078] Through the embodiments provided in this application, obtain the problem expression information input by the user account; determine multiple sub-information with dependent relationships from at least two sub-information; in the case where historical problem information already input by the user account is obtained and the multiple sub-information lacks key sub-information, use the historical problem information to supplement the multiple sub-information to obtain N sub-problem information, where the key sub-information is the information of the first sub-information or the second sub-information, thereby achieving the purpose of improving the information integrity of the sub-problem information, thus realizing better understanding of the sentence structure and semantics, and thus improving the accuracy of natural language processing tasks.

[0079] As an alternative solution, based on K problem information features, obtain problem response information, including:

[0080] S3-1, obtain the first similarity between each problem information feature in the K problem information features and each problem response feature in the knowledge base associated with the large language model;

[0081] S3-2, according to the first similarity from large to small, determine the P initial response information corresponding to the first P problem response features from the knowledge base;

[0082] S3-3, based on the P initial response information, obtain the problem response information.

[0083] It should be noted that the feature extraction of the input problem and the comparison with the problem response features in the knowledge base associated with the large language model in this embodiment can use, but are not limited to, natural language processing techniques, including methods such as word segmentation, part-of-speech tagging, dependency relationship analysis, and text similarity calculation, to determine the similarity and correlation between the problem and the answer, retrieve the response information most relevant to the problem from the knowledge base, and generate the final problem response, thereby realizing the matching and retrieval of the problem and the answer, and improving the accuracy and efficiency of the question-answering system.

[0084] For further illustration by example, optionally, for example, analyze the input K problems and extract the features of each problem. These features can be the semantics, context, keywords, etc. of the problem. Compare the features of the K problems with the problem response features in the knowledge base and calculate the similarity between them. This similarity can be, but is not limited to, obtained based on text similarity, semantic similarity, or other similarity calculation methods. Further compare the features of the K problems with the problem response features in the knowledge base and calculate the similarity between them. This similarity can be, but is not limited to, obtained based on text similarity, semantic similarity, or other similarity calculation methods. Analyze and process the P initial response information to generate the final problem response information, which can, but is not limited to, fuse multiple responses, screen or sort the answers, etc. during the analysis and processing

[0085] Through the embodiments provided in this application, each question information feature among the K question information features is obtained, and the first similarity between each question information feature and each question response feature in the knowledge base associated with the large language model is calculated; according to the first similarity from large to small, the first P question response features corresponding to the first P initial response messages are determined from the knowledge base; based on the first P initial response messages, the question response message is obtained, thereby achieving the purpose of matching and retrieving questions and answers, and thus realizing the technical effect of improving the accuracy and efficiency of question answering.

[0086] As an optional solution, obtaining the question response message based on the first P initial response messages includes:

[0087] S4-1, performing word segmentation processing on each sub-question information and each initial response message in the first P initial response messages to obtain N sub-question information sets and N initial response message sets, where the sub-question information set includes multiple sub-question words or short sentences, and the initial response message set includes multiple response words or short sentences;

[0088] S4-2, obtaining the second similarity between each sub-question information set in the N sub-question information sets and each initial response message set in the N initial response message sets;

[0089] S4-3, determining the initial response message with the largest second similarity from the first P initial response messages as the question response message.

[0090] Optionally, in this embodiment, by performing word segmentation processing on the text, the input sub-question information and initial response message are decomposed into smaller elements, such as single words or short sentences, which helps subsequent text similarity calculation and semantic understanding.

[0091] Optionally, in this embodiment, when determining the similarity between the sub-question information set and the initial response message set, a text similarity calculation method is adopted. By comparing the similarity between two texts, the response message most relevant to the sub-question can be selected. From the first P initial response messages, by comparing and sorting the similarities, the initial response message with the largest second similarity is selected as the question response message. This process can improve the accuracy and relevance of question answering.

[0092] It should be noted that, to further improve the accuracy of question answering, after the first screening of the response message based on the first similarity, a second screening of the response message is performed based on the second relevance.

[0093] For further illustration, optionally, for example, through word segmentation processing, the sub-question information and the initial response information can be decomposed into smaller elements, such as individual words or short phrases. After word segmentation processing, multiple sub-question information sets and multiple initial response information sets can be obtained. Each sub-question information set and each initial response information set contain multiple words or short phrases. Compare each sub-question information set with each initial response information set, and calculate the similarity between them. This similarity can be, but is not limited to, obtained based on text similarity, semantic similarity, or other similarity calculation methods. Compare each sub-question information set with each initial response information set, and calculate the similarity between them. This similarity can be, but is not limited to, obtained based on text similarity, semantic similarity, or other similarity calculation methods.

[0094] Through the embodiments provided in this application, perform word segmentation processing on each sub-question information and each initial response information among the P initial response information to obtain N sub-question information sets and N initial response information sets, where the sub-question information set includes multiple sub-question words or short phrases, and the initial response information set includes multiple response words or short phrases; obtain the second similarity between each sub-question information set in the N sub-question information sets and each initial response information set in the N initial response information sets; determine the initial response information with the largest second similarity from the P initial response information as the question response information, thereby achieving the purpose of performing a secondary screening of the response information based on the second relevance after a primary screening of the response information based on the first similarity, and thus realizing the technical effect of improving the accuracy of question answering.

[0095] As an optional solution, before performing word segmentation processing on each sub-question information and each initial response information among the P initial response information to obtain N sub-question information sets and N initial response information sets, the method further includes at least one of the following:

[0096] S5-1, delete the first processing information in each sub-question information and each initial response information, where the influence degree of the first processing information on the information semantics is less than or equal to the first preset threshold;

[0097] S5-2, set the second processing information in each sub-question information and each initial response information to be prohibited from word segmentation, where the second processing information is a predefined fixed word or short phrase.

[0098] Optionally, in this embodiment, the first processing information is a word or short phrase whose influence degree on the information semantics is less than or equal to the first preset threshold. By deleting this information, the task volume of subsequent processing can be reduced, and the complexity of text similarity calculation can be lowered. At the same time, deleting this information also helps to improve the accuracy and response speed of text similarity calculation.

[0099] Optionally, in this embodiment, the second processing information is a pre-specified fixed word or short sentence, and these information do not need to be segmented. When performing segmentation processing, setting these information as non-segmentable can ensure the integrity of these information and help improve the accuracy and efficiency of subsequent processing.

[0100] It should be noted that by deleting words or short sentences with less impact on information semantics, and setting pre-specified fixed words or short sentences as non-segmentable, the information processing flow is optimized, the accuracy and response speed of text similarity calculation are improved, and thus the performance of the entire natural language processing system is enhanced.

[0101] Through the embodiments provided by this application, the first processing information in each sub-question information and each initial reply information is deleted, where the influence degree of the first processing information on information semantics is less than or equal to the first preset threshold; the second processing information in each sub-question information and each initial reply information is set as non-segmentable, where the second processing information is a pre-specified fixed word or short sentence, thereby achieving the purpose of optimizing the information processing flow, and thus realizing the technical effect of improving the accuracy and response speed of text similarity calculation.

[0102] As an alternative solution, obtaining the second similarity between each sub-question information set in the N sub-question information sets and each initial reply information set in the N initial reply information sets includes:

[0103] S6-1, obtaining the first sub-similarity between each sub-question information set in the N sub-question information sets and the keywords or short sentences in each initial reply information set;

[0104] S6-2, obtaining the second sub-similarity between each sub-question information set in the N sub-question information sets and the non-keywords or short sentences in each initial reply information set;

[0105] S6-3, assigning a first weight to the first sub-similarity and a second weight to the second sub-similarity, where the first weight is greater than the second weight;

[0106] S6-4, integrating the first sub-similarity and the second sub-similarity according to the first weight and the second weight to obtain the second similarity.

[0107] It should be noted that to more comprehensively measure the similarity between the sub-question information set and the initial reply information set and improve the accuracy of information processing, the first sub-similarity is used to focus on the higher matching degree of keywords or short sentences, and the second sub-similarity is used to focus on the lower matching degree of non-keywords or short sentences. By assigning different weights to these two similarities and integrating them according to the weights, a more accurate and comprehensive similarity evaluation result can be obtained.

[0108] For further illustration, for example, keywords or short phrases in each sub-question information set and each initial response information set can be compared to calculate the similarity between them. This similarity can be, but is not limited to, based on text similarity, semantic similarity, or other similarity calculation methods. Further, non-keywords or short phrases in each sub-question information set and each initial response information set can be compared to calculate the similarity between them. This similarity can also be, but is not limited to, based on text similarity, semantic similarity, or other similarity calculation methods. When calculating the final similarity, the importance of the first sub-similarity is higher than that of the second sub-similarity. Among them, the greater the weight, the greater the impact of the corresponding similarity on the final similarity. Further, when calculating the final similarity, the importance of the first sub-similarity is higher than that of the second sub-similarity.

[0109] Through the embodiments provided in this application, the first sub-similarity between each sub-question information set in the N sub-question information sets and the keywords or short phrases in each initial response information set is obtained; the second sub-similarity between each sub-question information set in the N sub-question information sets and the non-keywords or short phrases in each initial response information set is obtained; a first weight is assigned to the first sub-similarity, and a second weight is assigned to the second sub-similarity, where the first weight is greater than the second weight; the first sub-similarity and the second sub-similarity are integrated according to the first weight and the second weight to obtain the second similarity, thereby achieving the purpose of more comprehensively measuring the similarity between the sub-question information set and the initial response information set, and thus realizing the technical effect of improving the accuracy of information processing.

[0110] As an optional solution, before each sub-question information in the N sub-question information is respectively input into M sub-models in the large language model to obtain K question information features, the method further includes:

[0111] S7-1, obtaining the knowledge text information to be input into the knowledge base;

[0112] S7-2, when the information length of the knowledge text information is greater than or equal to the second preset threshold, splitting the knowledge text information into at least two first text paragraphs;

[0113] S7-3, extracting the respective corresponding first paragraph keywords from each of the at least two first text paragraphs;

[0114] S7-4, dividing each of the first text paragraphs into at least two first text segments, and assigning the respective corresponding first paragraph keywords to the first text segments, where there is overlapping text information between the at least two first text segments.

[0115] It should be noted that in order to process and analyze knowledge text information in a more detailed and in-depth manner, splitting paragraphs and extracting keywords are used to better understand the structure and theme of the text information, while subdividing text fragments and assigning keywords can further explore the semantic information and context relationships of the text, thereby improving the accuracy of the text information in the knowledge base.

[0116] For further illustration, optionally, for example, more detailed and in-depth processing and analysis of knowledge text information is carried out. If the length of the knowledge text information is greater than or equal to the second preset threshold, then the text information is split into at least two paragraphs. Each first text paragraph is analyzed to extract the keywords in each paragraph. Each first text paragraph is analyzed to extract the keywords in each paragraph.

[0117] Through the embodiments provided by the present application, knowledge text information to be input into the knowledge base is obtained; when the information length of the knowledge text information is greater than or equal to the second preset threshold, the knowledge text information is split into at least two first text paragraphs; first paragraph keywords corresponding to each of the at least two first text paragraphs are extracted from each of the first text paragraphs; each of the first text paragraphs is subdivided into at least two first text fragments, and first paragraph keywords corresponding to each of the first text fragments are assigned, wherein there is overlapping text information between the at least two first text fragments, thereby achieving the purpose of processing and analyzing the knowledge text information in a more detailed and in-depth manner, and thus realizing the technical effect of improving the accuracy of the text information in the knowledge base.

[0118] As an optional solution, before obtaining the question reply information based on the K question information features, the method further includes: on the question and answer platform of the large language model, at least two virtual character identifiers are displayed, wherein the question expression information is input into the question and answer platform, and the virtual character identifiers are used to indicate the virtual characters that answer questions based on the large language model, and different virtual characters match different reply templates;

[0119] As an optional solution, after obtaining the question reply information based on the K question information features, the method further includes:

[0120] S8-1, when the virtual character that answers questions based on the large language model is a specific virtual character, obtain the specific reply template matched by the specific virtual character, and use the specific reply template to process the question reply information to obtain specific reply information;

[0121] S8-2, display the specific reply information on the question and answer platform.

[0122] It should be noted that different virtual characters and their corresponding reply templates can demonstrate different personalities and professional fields, making the replies more interesting and practical. At the same time, the application of large language models makes the answers more natural and fluent, enabling users to obtain more diverse and efficient answers on the Q&A platform, thereby improving the user experience.

[0123] For further illustration, for example, on a Q&A platform, there are at least two different virtual characters, and each virtual character has its own identifier. These virtual characters may represent different personalities, professional fields, or answering styles. Users can input questions to express information on the Q&A platform. Each virtual character has its corresponding reply template, and the large language model will select the matching virtual character and its corresponding reply template according to the input question for reply. Thus, each virtual character has its unique reply template for generating replies to questions related to that virtual character. For instance, one virtual character may correspond to the answering style of a historical figure, and another virtual character may correspond to the answering style of a science fiction character. When the question input by the user matches the reply template of a specific virtual character, the system will obtain that specific reply template. The system will use the specific reply template to generate a reply to the user's question, and this reply is generated based on the large language model and has the answering style of that specific virtual character.

[0124] Through the embodiments provided in this application, on the Q&A platform of the large language model, at least two virtual character identifiers are displayed. Among them, the question expression information is input into the Q&A platform, and the virtual character identifiers are used to indicate the virtual characters that reply to questions based on the large language model. Different virtual characters match different reply templates; when the virtual character that replies to questions based on the large language model is a specific virtual character, the specific reply template matched by the specific virtual character is obtained, and the specific reply template is used to process the question reply information to obtain specific reply information; the specific reply information is displayed on the Q&A platform, thereby achieving the purpose of enabling users to obtain more diverse and efficient answers on the Q&A platform, and thus realizing the technical effect of improving the user experience.

[0125] As an optional solution, before obtaining the question expression information containing at least two sub-information, the method further includes:

[0126] S9-1, obtaining initial question information, where the initial question information is used to indicate that the large language model outputs reply information matching the initial question information;

[0127] S9-2, when the information length of the initial question information is greater than or equal to the third preset threshold, splitting the question expression information into at least two second text paragraphs;

[0128] S9-3. Extract the respective corresponding second paragraph keywords from each of the at least two second text paragraphs.

[0129] S9-4. Subdivide each of the second text paragraphs into at least two second text segments, and assign the respective corresponding second paragraph keywords to the second text segments, obtaining at least two sub-informations, wherein there is overlapping text information between the at least two second text segments.

[0130] It should be noted that for more meticulous and in-depth processing and analysis of the initial problem information, combining paragraph splitting and keyword extraction can better understand the structure and theme of the text information, while subdividing the text segments and assigning keywords can further explore the semantic information and context relationship of the text.

[0131] For further illustration, optionally, for example, obtain the initial problem information to be further processed from a certain source (such as user input, external data source, etc.). If the length of the initial problem information is greater than or equal to the third preset threshold, then split the information into at least two paragraphs. Analyze each second text paragraph, and extract the keywords in each paragraph. Conduct a more meticulous analysis of each second text paragraph, and subdivide it into at least two text segments. Each text segment is assigned the corresponding paragraph keyword.

[0132] Through the embodiments provided in this application, obtain the initial problem information, wherein the initial problem information is used to instruct the large language model to output a reply information matching the initial problem information; when the information length of the initial problem information is greater than or equal to the third preset threshold, split the problem expression information into at least two second text paragraphs; extract the respective corresponding second paragraph keywords from each of the at least two second text paragraphs; subdivide each of the second text paragraphs into at least two second text segments, and assign the respective corresponding second paragraph keywords to the second text segments, obtaining at least two sub-informations, wherein there is overlapping text information between the at least two second text segments, thereby achieving the purpose of more meticulous and in-depth processing and analysis of the initial problem information, and thus realizing the technical effect of improving the accuracy of information processing.

[0133] As an alternative solution, for the convenience of understanding, the above information processing method based on large language models is applied to the domain knowledge enhanced natural language interaction system of virtual humans. It is an effective strategy to seamlessly integrate prompt engineering into natural language knowledge base interaction. When dealing with the knowledge base, related technologies first simply split the text based on full stops or line breaks, represent the obtained text fragments as vectors through a single embedding model, and then store them in the knowledge base. For the questions of users, the same embedding model is also used to represent them as vectors, and similarity retrieval is performed with the text fragments in the knowledge base to obtain the text fragments related to the questions. The found text fragments and the questions are input into the large language model together to obtain the answers required by the users;

[0134] However, the above related technologies divide text fragments too simply, often ignoring context information, and a single embedding model often has poor effects on texts from multiple data sources. At the same time, the connection between the question and the text title or keywords is easily ignored during similarity search. The knowledge base Q&A robot has no personality and cannot be directly used in the language interaction of virtual humans.

[0135] Optionally, in this embodiment, the virtual human (domain knowledge enhanced natural language interaction) system represents a highly complex platform, including key elements such as realistic visual rendering, emotional expression, and natural language interaction. The natural language interaction module is particularly important because the entire system needs to interact with users, but there are still problems in seamlessly integrating large language models into these systems.

[0136] Optionally, an important basis for natural language interaction is the need for a large amount of knowledge bases. Virtual humans must answer users' questions based on the knowledge they have mastered. However, it is also unlikely that large language models can promptly master sufficient knowledge, especially in professional fields. This knowledge gap in large language models can be attributed to two main reasons. First, the training data of these models may not comprehensively cover all knowledge domains. Second, large language models themselves may have difficulty recalling all details from their training data. Therefore, it is crucial to establish specialized knowledge bases, such as knowledge graphs or vector databases, to store and facilitate knowledge retrieval, enabling large language models to provide accurate and content-rich responses to users. And in this embodiment, as Figure 7 shown, in this embodiment, the above problems are solved through text splitting, retrieval query enhancement, and retrieval optimization to optimize the knowledge base module.

[0137] Optionally, in this embodiment, given the length limitation of large language models, a feasible solution is to split the input document into multiple text segments. However, such splitting has a significant impact on the accuracy of information retrieval. If the text segments are too long, the text differences during retrieval will decrease, and it will also be difficult for the large model to grasp the key points when answering. On the contrary, too short text segments will result in incomplete retrieved information. In addition, document splitting is prone to the problem of losing context during the question-and-answer process. Therefore, this embodiment designs a series of strategies to address these issues. To solve the context missing problem, this embodiment adopts a hierarchical method for document splitting. First, the document is divided into paragraphs, and then the large language model is used to extract keywords from each paragraph. Next, each paragraph is further subdivided into text segments, and the keywords extracted from their respective paragraphs are added to each text segment. At the same time, this embodiment also adopts a scheme of segment overlap, that is, the text in each text segment will overlap with the upper and lower text segments by a part. In summary, these strategies aim to bridge the context missing and improve the accuracy and efficiency of the document analysis process in this embodiment.

[0138] Optionally, in this embodiment, when the user submits an input question q that needs to be retrieved from the knowledge base of this embodiment, ideally the question is clear in object and single in target (e.g., "Where is the attack button in Game A?"). However, in real scenarios, user questions are usually unclear in object or have multiple targets (e.g., "Where are the attack buttons and item menus in these two games?"). If retrieved directly based on such a question, the results may be unsatisfactory because it is not specified which game, and due to multiple targets in the question, texts containing information about the attack button and item menu need to be retrieved separately.

[0139] Therefore, this embodiment introduces a method of enhancing retrieval questions in natural language interaction. As Figure 7 shown, the method of this embodiment includes two key strategies. First, to effectively handle multi-target problems, this embodiment performs dependency analysis to generate a dependency tree for the input query. In this dependency tree, this embodiment can identify noun phrase connections and the associated verbs, and then identify multiple targets in the question. As Figure 8 shown, this enables this embodiment to split the multi-target query into several different single-target queries;

[0140] Second, to solve unclear or ambiguous questions, this embodiment uses the summarization ability of the large language model to summarize the previous chat history, extract the details of the previous chat, and output the completed question. Figure 8 In this embodiment, "these two games" is restored to Game A and Game B based on the two-day record. This process ensures that the response of the large language model is not only more accurate but also highly relevant.

[0141] Optionally, in this embodiment, since the retrieval effect of text encoding by a single embedding model is lower than expected, this embodiment proposes a multi-embedding recall method. For an input question q, it will be represented as n different vectors by n different embedding models. At the same time, the text fragments in the knowledge base will also be represented by different embedding models and stored as a vector knowledge base D. When calculating the similarity, for the question and the knowledge base represented by the same embedding model, their similarity is calculated by the following formula (1), and then according to the similarity score in this embodiment, the top-k text fragments are selected from multiple embedding vectors as the recall result.

[0142] S(q,D) = cos(E(q),E(D))

[0143]

[0144] Optionally, in order to further improve the relevance between the recall result and the input question, the re-ranking strategy in this embodiment uses an N-gram based lexical Jaccard weight to calculate the re-ranking score. First, this embodiment constructs a stop word list and removes the symbols in the recalled text fragments at the same time. Then constructs an N-gram question set Q and a recall result set R, and calculates the Jaccard coefficient as shown in the following formula (2):

[0145]

[0146] At the same time, for the title T of each document, the Jaccard coefficient is also calculated with the question set as shown in the following formula (3):

[0147]

[0148] Then this embodiment defines the re-ranking weight W as shown in the following formula (4):

[0149]

[0150] Finally, for S ′ The score for re-ranking S

[0151] S ′ (Q,D) can be defined as shown in the following formula (5):

[0152] Optionally, in this embodiment, prompt engineering plays an important role in the performance and usability of large language models, involving designing effective prompts or inputs to guide the model to generate the desired accurate response. In the virtual person of this embodiment, this embodiment widely applies prompt engineering to improve its performance and user experience. By formulating effective prompts, this embodiment aims to guide the model to generate responses that are consistent with the unique characteristics of the character and accurately solve the user's query, ensuring a more attractive and interactive experience.

[0153] One of the most critical aspects for a virtual person is its unique personality, otherwise the virtual person will only be a static chatbot. However, developing these personalities can be challenging because the complete prompt becomes too large along with the retrieval results, while the character prompt only accounts for a small proportion. To solve this problem, this embodiment adopts two strategies.

[0154] First, in order to enhance the role prompting function in large language models, this embodiment provides a comprehensive solution that aims to make the role image more three-dimensional and vivid through multi-dimensional information injection. Specifically, this embodiment allows users to set a designated name for the role and answer based on this identity. In addition, in order to ensure that the AI can deeply understand and implement this role setting, this embodiment requires the AI to always keep in mind the role rules represented by the designated name during the interaction process.

[0155] These rules are not limited to the name itself, but also include a detailed description of the character's background, distinctive characteristics of the personality, the use of catchphrases, and subtle differences in tone. For example, a character set as a "seasoned detective" may have rich detective experience, a calm and calm personality, frequently used catchphrases such as "interesting clues", and a tone that is both mysterious and confident. When answering, the AI will follow these characteristics and construct a sentence structure that fits the character's image, such as using complex sentences, inserting professional terms, or simulating the detective reasoning process.

[0156] Inspired by the concept of contextual learning, this embodiment further introduces a comparative teaching method of high-quality and low-quality examples. By showing specific cases, it helps large language models to more intuitively understand the application scenarios of character features. High-quality examples can usually perfectly reflect the character characteristics, such as the detective character mentioned above who always remains calm and clear when answering, and shows his unique reasoning ability from time to time. On the contrary, low-quality examples may deviate from the role setting, such as the detective character suddenly showing an impetuous or uncertain attitude, or even revealing his identity as an AI model.

[0157] To enable users to define new character personalities more easily, this embodiment may also provide a series of practical tools and guidance materials. These include, but are not limited to, character setting templates, example libraries, and detailed operation guides. Through these resources, users can set characters in a more systematic manner, ensuring that every detail is properly handled, thus creating a more realistic and vivid character image.

[0158] Secondly, this embodiment requires the large language model to pay more attention to character prompts. Thus, this embodiment sets the character prompts as system_prompt to ensure that the chat model pays more attention to them and generates more vivid features. Among them, in the OpenAI Chat Completion API, there is a parameter called system_prompt. The system_prompt parameter is a string used to specify the system prompt. This parameter allows custom prompt text to be provided in the API call to guide the direction of the chat conversation. Through the system_prompt parameter, a piece of text can be input as a guide or prompt for the chat system to help the chat system better understand the user's intention and generate corresponding responses.

[0159] Through the embodiments provided in this application, a large amount of domain knowledge can be obtained through the knowledge base, thereby implementing functions such as relevant business introductions, game recommendations, game guides, and game companionship. At the same time, the virtual humans in this embodiment also have multiple different personalities that can be switched, and can be extended to systems that require natural language interaction such as intelligent customer service and tour guide robots.

[0160] It can be understood that in the specific implementation of this application, data related to user information, etc. When the above embodiments of this application are applied to specific products or technologies, user permission or consent needs to be obtained, and the collection, use, and processing of relevant data need to comply with relevant laws, regulations, and standards in relevant countries and regions.

[0161] It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that this application is not limited by the described action sequence, because according to this application, certain steps can be in other sequences or performed simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0162] According to another aspect of the embodiments of this application, there is also provided an information processing device based on a large language model for implementing the above information processing method based on a large language model. As Figure 9 shown, the device includes:

[0163] The first acquisition unit 902 is configured to acquire a question expression information including at least two sub-informations. The question expression information is used to instruct the large language model to output a question response information that matches the question expression information. The question expression information includes at least one question object and at least one question target. The question object is the object being questioned mentioned in the question expression information, and the question target is the target that the question expression information expects to achieve by asking the question object. At least one question object and at least one question target form N sub-questions, where N is an integer greater than or equal to 2;

[0164] The second acquisition unit 904 is configured to acquire the dependency relationships between the respective sub-informations in the at least two sub-informations, and process the question expression information into N sub-question informations by using the dependency relationships. The dependency relationship is used to describe the connection or association between the respective sub-informations, and the sub-question information in the N sub-question informations is used to express the sub-questions in the N sub-questions;

[0165] The first response unit 906 is configured to separately input each of the sub-question informations in the N sub-question informations into M sub-models in the large language model to obtain K question information features, and obtain the question response information based on the K question information features. Different sub-models are configured with different model parameters. The question information feature is used to characterize the question content of the sub-question information. M is an integer greater than or equal to 2, and K is less than or equal to the product of N and M.

[0166] For specific embodiments, reference may be made to the examples shown in the above information processing device based on the large language model. Such examples are not elaborated herein.

[0167] As an alternative solution, the second acquisition unit 904 includes:

[0168] The first acquisition module is configured to determine a first sub-information from the at least two sub-informations, and acquire the connection relationships between the respective first sub-informations, where the first sub-information is a noun or a noun-centered phrase;

[0169] The first determination module is configured to determine a second sub-information associated with the first sub-information from the at least two sub-informations, where the second sub-information is a verb;

[0170] The second determination module is configured to determine the first sub-information having a connection relationship and the second sub-information associated with the first sub-information having a connection relationship in the at least two sub-informations as having a dependency relationship.

[0171] For specific embodiments, reference may be made to the examples shown in the above information processing method based on the large language model. Such examples are not elaborated herein.

[0172] As an alternative solution, the first acquisition unit 902 includes: a second acquisition module for acquiring the question expression information input by the user account;

[0173] The second acquisition unit 904 includes:

[0174] A third determination module for determining multiple sub-information with an attachment relationship from at least two sub-information;

[0175] A supplement module for supplementing the multiple sub-information with the historical question information when the historical question information input by the user account has been acquired and the multiple sub-information lacks key sub-information to obtain N sub-question information, where the key sub-information is the information of the first sub-information or the second sub-information.

[0176] For specific embodiments, reference may be made to the examples shown in the above information processing method based on the large language model, and details are not described herein again.

[0177] As an alternative solution, the first reply unit 906 includes:

[0178] A third acquisition module for acquiring the first similarity between each question information feature among the K question information features and each question reply feature in the knowledge base associated with the large language model;

[0179] A fourth determination module for determining the first P initial reply information corresponding to the first P question reply features from the knowledge base in descending order of the first similarity;

[0180] A fourth acquisition module for obtaining the question reply information based on the P initial reply information.

[0181] For specific embodiments, reference may be made to the examples shown in the above information processing method based on the large language model, and details are not described herein again.

[0182] As an alternative solution, the fourth acquisition module includes:

[0183] A word segmentation sub-module for performing word segmentation on each of the sub-question information and each of the initial reply information in the P initial reply information to obtain an N sub-question information set and an N initial reply information set, where the sub-question information set includes multiple sub-question words or short sentences, and the initial reply information set includes multiple reply words or short sentences;

[0184] An acquisition sub-module for acquiring the second similarity between each sub-question information set in the N sub-question information sets and each initial reply information set in the N initial reply information sets;

[0185] A determining sub-module, configured to determine the initial response message with the second largest similarity from the P initial response messages as the question response message.

[0186] For specific embodiments, reference may be made to the examples shown in the above information processing method based on a large language model, which will not be elaborated herein.

[0187] As an alternative solution, the apparatus further includes at least one of the following:

[0188] A deletion module, configured to delete the first processing information in each sub-question information and each initial response message before performing word segmentation processing on each sub-question information and each initial response message in the P initial response messages to obtain N sub-question information sets and N initial response message sets, where the influence degree of the first processing information on the information semantics is less than or equal to a first preset threshold;

[0189] A prohibition module, configured to set the second processing information in each sub-question information and each initial response message to be prohibited from word segmentation before performing word segmentation processing on each sub-question information and each initial response message in the P initial response messages to obtain N sub-question information sets and N initial response message sets, where the second processing information is a predetermined fixed word or short sentence.

[0190] For specific embodiments, reference may be made to the examples shown in the above information processing method based on a large language model, which will not be elaborated herein.

[0191] As an alternative solution, the obtaining sub-module includes:

[0192] A first obtaining sub-unit, configured to obtain the first sub-similarity between each sub-question information set in the N sub-question information sets and the keywords or short sentences in each initial response message set;

[0193] A second obtaining sub-unit, configured to obtain the second sub-similarity between each sub-question information set in the N sub-question information sets and the non-keywords or short sentences in each initial response message set;

[0194] An assignment sub-unit, configured to assign a first weight to the first sub-similarity and a second weight to the second sub-similarity, where the first weight is greater than the second weight;

[0195] An integration sub-unit, configured to integrate the first sub-similarity and the second sub-similarity according to the first weight and the second weight to obtain the second similarity.

[0196] For specific embodiments, reference may be made to the examples shown in the above information processing method based on a large language model, which will not be elaborated herein.

[0197] As an alternative solution, the apparatus further includes:

[0198] A fifth acquisition module, configured to acquire knowledge text information to be input into a knowledge base before respectively inputting each sub-problem information in N pieces of sub-problem information into M sub-models in a large language model to obtain K pieces of problem information features;

[0199] A splitting module, configured to, before respectively inputting each sub-problem information in N pieces of sub-problem information into M sub-models in a large language model to obtain K pieces of problem information features, split the knowledge text information into at least two first text paragraphs when the information length of the knowledge text information is greater than or equal to a second preset threshold;

[0200] An extraction module, configured to, before respectively inputting each sub-problem information in N pieces of sub-problem information into M sub-models in a large language model to obtain K pieces of problem information features, extract respective corresponding first paragraph keywords from each of the at least two first text paragraphs;

[0201] A subdivision module, configured to, before respectively inputting each sub-problem information in N pieces of sub-problem information into M sub-models in a large language model to obtain K pieces of problem information features, subdivide each of the first text paragraphs into at least two first text segments, and assign respective corresponding first paragraph keywords to the first text segments, wherein there is overlapping text information between the at least two first text segments.

[0202] For specific embodiments, reference may be made to the examples shown in the above information processing method based on a large language model, and details are not described herein again.

[0203] As an optional solution, the apparatus further includes: a first display unit, configured to display at least two virtual character identifiers on a question and answer platform of the large language model before obtaining question reply information based on the K pieces of problem information features, wherein question expression information is input into the question and answer platform, and the virtual character identifiers are used to indicate virtual characters for replying questions based on the large language model, and different virtual characters match different reply templates;

[0204] The apparatus further includes:

[0205] A second reply unit, configured to, after obtaining question reply information based on the K pieces of problem information features, when the virtual character for replying questions based on the large language model is a specific virtual character, obtain a specific reply template matched by the specific virtual character, and process the question reply information using the specific reply template to obtain specific reply information;

[0206] A second display unit, configured to display the specific reply information on the question and answer platform after obtaining question reply information based on the K pieces of problem information features.

[0207] For specific embodiments, reference may be made to the examples shown in the above information processing method based on a large language model, and details will not be elaborated herein in this example.

[0208] As an alternative solution, the apparatus further includes:

[0209] A third acquisition unit, configured to acquire initial question information before acquiring question expression information including at least two sub-information, where the initial question information is used to instruct the large language model to output a response information matching the initial question information;

[0210] A splitting unit, configured to split the question expression information into at least two second text paragraphs when the information length of the initial question information is greater than or equal to a third preset threshold before acquiring the question expression information including at least two sub-information;

[0211] An extraction unit, configured to extract respective second paragraph keywords from each of the at least two second text paragraphs before acquiring the question expression information including at least two sub-information;

[0212] A subdivision unit, configured to subdivide each of the second text paragraphs into at least two second text fragments and assign respective second paragraph keywords to the second text fragments to obtain at least two sub-information before acquiring the question expression information including at least two sub-information, where there is overlapping text information between the at least two second text fragments.

[0213] For specific embodiments, reference may be made to the examples shown in the above information processing method based on a large language model, and details will not be elaborated herein in this example.

[0214] According to another aspect of the embodiments of the present application, there is also provided an electronic device for implementing the above information processing method based on a large language model. The electronic device may be, but is not limited to, Figure 1 the user device 102 or the server 112 shown in Figure 10 As shown, the electronic device includes a memory 1002 and a processor 1004. A computer program is stored in the memory 1002, and the processor 1004 is configured to execute the steps in any of the above method embodiments through the computer program.

[0215] Optionally, in this embodiment, the above electronic device may be at least one of multiple network devices in a computer network.

[0216] Optionally, in this embodiment, the above processor may be configured to execute the following steps through the computer program:

[0217] S1. Obtain problem expression information containing at least two sub-informations. The problem expression information is used to instruct the large language model to output problem response information that matches the problem expression information. The problem expression information contains at least one problem object and at least one problem target. The problem object is the object being questioned mentioned in the problem expression information, and the problem target is the goal that the problem expression information expects to achieve by asking the problem object. At least one problem object and at least one problem target form N sub-problems, where N is an integer greater than or equal to 2.

[0218] S2. Obtain the dependency relationships between the respective sub-informations among the at least two sub-informations, and use the dependency relationships to process the problem expression information into N sub-problem information. The dependency relationships are used to describe the connections or associations between the respective sub-informations. The sub-problem information in the N sub-problem information is used to express the sub-problems in the N sub-problems.

[0219] S3. Input each of the sub-problem information in the N sub-problem information into M sub-models in the large language model to obtain K problem information features, and based on the K problem information features, obtain the problem response information. Different sub-models are configured with different model parameters. The problem information features are used to characterize the problem content of the sub-problem information. M is an integer greater than or equal to 2, and K is less than or equal to the product of N and M.

[0220] Optionally, those of ordinary skill in the art can understand that Figure 10 the structure shown is only schematic Figure 10 and does not limit the structure of the above electronic device. For example, the electronic device may further include more or fewer components (such as network interfaces, etc.) than those shown Figure 10 or have a different configuration from that shown. Figure 10

[0221] ​Among them, the memory 1002 can be used to store software programs and modules, such as the program instructions / modules corresponding to the information processing method and device based on the large language model in the embodiments of the present application. The processor 1004 executes various functional applications and data processing by running the software programs and modules stored in the memory 1002, that is, implements the above-mentioned information processing method based on the large language model. The memory 1002 may include a high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memories. In some instances, the memory 1002 may further include a memory remotely disposed relative to the processor 1004, and these remote memories can be connected to the electronic device through a network. Examples of the above network include but are not limited to the Internet, enterprise intranet, local area network, mobile communication network, and combinations thereof. Among them, the memory 1002 can specifically but not limitedly be used to store information such as problem expression information, problem information features, and problem reply information. As an example, as Figure 10 shown, the above memory 1002 may include, but is not limited to, the first acquisition unit 902, the second acquisition unit 904, and the first reply unit 906 in the above information processing device based on the large language model. In addition, it may also include, but is not limited to, other module units in the above information processing device based on the large language model, which will not be elaborated in this example.

[0222] Optionally, the above transmission device 1006 is used to receive or send data via a network. Specific examples of the above network may include a wired network and a wireless network. In one instance, the transmission device 1006 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices and routers through a network cable, thereby enabling communication with the Internet or a local area network. In one instance, the transmission device 1006 is a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0223] In addition, the above electronic device further includes: a display 1008, which is used to display information such as the above problem expression information, problem information features, and problem reply information; and a connection bus 1010, which is used to connect each module component in the above electronic device.

[0224] In other embodiments, the above user equipment or server may be a node in a distributed system. Among them, the distributed system may be a blockchain system, and the blockchain system may be a distributed system formed by connecting the multiple nodes through network communication. Among them, the nodes can form a peer-to-peer network, and any form of computing device, such as an electronic device such as a server or a user equipment, can become a node in the blockchain system by joining the peer-to-peer network.

[0225] According to one aspect of the present application, a computer program product is provided, which includes a computer program / instructions that contain program code for performing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through the communication part, and / or installed from a removable medium. When the computer program is executed by the central processing unit, various functions provided by the embodiments of the present application are performed.

[0226] The serial numbers of the embodiments of the present application above are only for description and do not represent the superiority or inferiority of the embodiments.

[0227] It should be noted that the computer system of the electronic device is only an example and should not impose any limitations on the functions and usage scope of the embodiments of the present application.

[0228] The computer system includes a central processing unit (CPU), which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) or the program loaded from the storage part into the random access memory (RAM). In the random access memory, various programs and data required for system operation are also stored. The central processing unit, the read-only memory, and the random access memory are connected to each other through a bus. The input / output interface (Input / Output interface, i.e., I / O interface) is also connected to the bus.

[0229] The following components are connected to the input / output interface: an input part including a keyboard, a mouse, etc.; an output part including a cathode ray tube (CRT), a liquid crystal display (LCD), etc. and a speaker, etc.; a storage part including a hard disk, etc.; and a communication part including a network interface card such as a local area network card, a modem, etc. The communication part performs communication processing via a network such as the Internet. A drive is also connected to the input / output interface as needed. A removable medium, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive as needed so that the computer program read from it can be installed into the storage part as needed.

[0230] In particular, according to an embodiment of the present application, the processes described in each method flow chart can be implemented as computer software programs. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains program codes for executing the methods shown in the flow charts. In such an embodiment, the computer program can be downloaded and installed from a network through a communication part, and / or installed from a removable medium. When the computer program is executed by a central processing unit, various functions defined in the system of the present application are executed.

[0231] According to one aspect of the present application, there is provided a computer-readable storage medium, and a processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in the above various optional implementation manners.

[0232] Optionally, in this embodiment, the above computer-readable storage medium may be set to store a computer program for executing the following steps:

[0233] S1. Obtain problem expression information containing at least two sub-information, where the problem expression information is used to instruct a large language model to output problem response information matching the problem expression information. The problem expression information contains at least one problem object and at least one problem target. The problem object is the object being queried mentioned in the problem expression information, and the problem target is the target that the problem expression information expects to achieve by querying the problem object. At least one problem object and at least one problem target form N sub-problems, and N is an integer greater than or equal to 2.

[0234] S2. Obtain the dependency relationship between each sub-information among at least two sub-information, and use the dependency relationship to process the problem expression information into N sub-problem information, where the dependency relationship is used to describe the connection or association between each sub-information, and the sub-problem information in the N sub-problem information is used to express the sub-problems in the N sub-problems.

[0235] S3. Input each sub-problem information in the N sub-problem information into M sub-models in the large language model to obtain K problem information features, and based on the K problem information features, obtain the problem response information. Different sub-models are configured with different model parameters. The problem information features are used to characterize the problem content of the sub-problem information. M is an integer greater than or equal to 2, and K is less than or equal to the product of N and M.

[0236] Optionally, in this embodiment, those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware of the electronic device through a program. This program can be stored in a computer-readable storage medium, and the storage medium can include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disc, etc.

[0237] The serial numbers of the embodiments of the present application above are only for description and do not represent the advantages or disadvantages of the embodiments.

[0238] If the integrated unit in the above embodiments is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in the above computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing one or more computer devices (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application.

[0239] In the above embodiments of the present application, the descriptions of the various embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0240] In the several embodiments provided by the present application, it should be understood that the disclosed user equipment can be implemented in other ways. Among them, the device embodiments described above are only illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling, direct coupling, or communication connection can be through some interfaces. The indirect coupling or communication connection of the units or modules can be in an electrical or other form.

[0241] The unit described as a separated component may or may not be physically separated. The component displayed as a unit may or may not be a physical unit, that is, it can be located in one place, or it can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0242] In addition, each functional unit in various embodiments of the present application may be integrated into one processing unit, may exist separately as individual physical units, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware or in the form of software functional units.

[0243] The above are only the preferred embodiments of the present application. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. An information processing method based on a large language model, characterized in that, Including: Obtain problem expression information including at least two sub-informations, where the problem expression information is used to instruct a large language model to output problem response information matching the problem expression information. The problem expression information includes at least one problem object and at least one problem target. The problem object is the object being questioned mentioned in the problem expression information, and the problem target is the target expected to be achieved by asking the problem object in the problem expression information. The at least one problem object and the at least one problem target form N sub-problems, where N is an integer greater than or equal to 2; Obtain the dependency relationship between each of the at least two sub-informations, and use the dependency relationship to process the problem expression information into N sub-problem informations, where the dependency relationship is used to describe the connection or association between the sub-informations, and the sub-problem information in the N sub-problem informations is used to express the sub-problems in the N sub-problems; Input each of the N sub-problem informations in the N sub-problem informations into M sub-models in the large language model to obtain K problem information features, and based on the K problem information features, obtain the problem response information. Different sub-models are configured with different model parameters. The problem information feature is used to characterize the problem content of the sub-problem information. M is an integer greater than or equal to 2, and K is less than or equal to the product of N and M.

2. The method according to claim 1, wherein The obtaining the dependency relationship between each of the at least two sub-informations includes: Determine a first sub-information from the at least two sub-informations, and obtain the connection relationship between each of the first sub-informations, where the first sub-information is a noun or a phrase centered on the noun; Determine a second sub-information associated with the first sub-information from the at least two sub-informations, where the second sub-information is a verb; Determine that the first sub-information having the connection relationship and the second sub-information associated with the first sub-information having the connection relationship in the at least two sub-informations have the dependency relationship.

3. The method according to claim 2, wherein: The obtaining the problem expression information including at least two sub-informations includes: obtaining the problem expression information input by the user account; The using the dependency relationship to process the problem expression information into N sub-problem informations includes: Determine multiple sub-informations having the dependency relationship from the at least two sub-informations; In the case where the historical problem information input by the user account has been obtained and the multiple sub-informations lack key sub-informations, use the historical problem information to supplement the multiple sub-informations to obtain the N sub-problem informations, where the key sub-information is the information of the first sub-information or the second sub-information.

4. The method according to claim 1, characterized in that The obtaining the problem response information based on the K problem information features includes: Obtain the first similarity between each of the K problem information features and each of the problem response features in the knowledge base associated with the large language model; Determine the top P initial response messages corresponding to the P question response features from the knowledge base in descending order according to the first similarity; Based on the P initial response messages, obtain the question response message.

5. The method according to claim 4, wherein The obtaining the question response message based on the P initial response messages includes: Perform word segmentation on each sub-question message and each initial response message in the P initial response messages to obtain N sub-question message sets and N initial response message sets, where the sub-question message set includes multiple sub-question words or short sentences, and the initial response message set includes multiple response words or short sentences; Obtain the second similarity between each sub-question message set in the N sub-question message sets and each initial response message set in the N initial response message sets; Determine the initial response message with the largest second similarity from the P initial response messages as the question response message.

6. The method according to claim 5, wherein Before performing word segmentation on each sub-question message and each initial response message in the P initial response messages to obtain N sub-question message sets and N initial response message sets, the method further includes at least one of the following: Delete the first processing information in each sub-question message and each initial response message, where the influence degree of the first processing information on the information semantics is less than or equal to the first preset threshold; Set the second processing information in each sub-question message and each initial response message as non-segmentable, where the second processing information is a predetermined fixed word or short sentence.

7. The method according to claim 5, wherein The obtaining the second similarity between each sub-question message set in the N sub-question message sets and each initial response message set in the N initial response message sets includes: Obtain the first sub-similarity between each sub-question message set in the N sub-question message sets and the keyword or short sentence of each initial response message set; Obtain the second sub-similarity between each sub-question message set in the N sub-question message sets and the non-keyword or short sentence of each initial response message set; Assign a first weight to the first sub-similarity and a second weight to the second sub-similarity, where the first weight is greater than the second weight; Integrate the first sub-similarity and the second sub-similarity according to the first weight and the second weight to obtain the second similarity.

8. The method according to claim 4, characterized in that, Before inputting each sub-question message in the N sub-question messages into M sub-models in the large language model to obtain K question message features, the method further includes: Obtain the knowledge text information to be input into the knowledge base; In the case where the information length of the knowledge text information is greater than or equal to the second preset threshold, split the knowledge text information into at least two first text paragraphs; Extract the corresponding first paragraph keywords from each of the at least two first text paragraphs; Each of the first text paragraphs is subdivided into at least two first text segments, and respective corresponding first paragraph keywords are assigned to the first text segments, wherein there is overlapping text information between the at least two first text segments.

9. The method according to any one of claims 1 to 8, characterized in that Before obtaining the question response information based on the K question information features, the method further includes: on the question and answer platform of the large language model, at least two virtual character identifiers are displayed, wherein the question expression information is input into the question and answer platform, and the virtual character identifiers are used to indicate virtual characters that answer questions based on the large language model, and different virtual characters match different response templates; After obtaining the question response information based on the K question information features, the method further includes: In the case where the virtual character that answers questions based on the large language model is a specific virtual character, obtain the specific response template matched by the specific virtual character, and use the specific response template to process the question response information to obtain specific response information; Display the specific response information on the question and answer platform.

10. The method according to any one of claims 1 to 8, characterized in that Before obtaining the question expression information including at least two sub-information, the method further includes: Obtain initial question information, wherein the initial question information is used to indicate that the large language model outputs response information matching the initial question information; In the case where the information length of the initial question information is greater than or equal to a third preset threshold, split the question expression information into at least two second text paragraphs; Extract respective corresponding second paragraph keywords from each of the at least two second text paragraphs; Each of the second text paragraphs is subdivided into at least two second text segments, and respective corresponding second paragraph keywords are assigned to the second text segments to obtain the at least two sub-information, wherein there is overlapping text information between the at least two second text segments.

11. An information processing device based on a large language model, characterized in that, Including: A first obtaining unit, configured to obtain question expression information including at least two sub-information, wherein the question expression information is used to indicate that the large language model outputs question response information matching the question expression information, the question expression information includes at least one question object and at least one question target, the question object is the object being questioned mentioned in the question expression information, the question target is the target that the question expression information expects to achieve by asking the question object, and the at least one question object and the at least one question target form N sub-questions, and N is an integer greater than or equal to 2; A second obtaining unit, configured to obtain the dependency relationship between each of the at least two sub-information, and use the dependency relationship to process the question expression information into N sub-question information, wherein the dependency relationship is used to describe the connection or association between each of the sub-information, and the sub-question information in the N sub-question information is used to express the sub-questions in the N sub-questions; The first response unit is configured to separately input each of the N sub-question information into M sub-models in the large language model, obtain K question information features, and obtain the question response information based on the K question information features, wherein different ones of the sub-models are configured with different model parameters, the question information features are used to characterize the question content of the sub-question information, M is an integer greater than or equal to 2, and K is less than or equal to the product of N and M.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program, when run by an electronic device, executes the method described in any one of claims 1 to 10.

13. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by a processor, the steps of the method described in any one of claims 1 to 10 are implemented.

14. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is configured to execute the method described in any one of claims 1 to 10 through the computer program.

Images