Information processing method and device of large language model, electronic equipment and storage medium

Abstract

The invention provides an information processing method and device of a large language model, electronic equipment and a computer readable storage medium, and relates to the technical field of information processing. The information processing method comprises the following steps: acquiring a user problem and a network application portal identifier; according to the user problem and the network application portal identifier, obtaining a total intention and a total task in the demand scene; determining a target intention and a target task according to the user problem and the full intention and the full task in the demand scene; and according to the target intention and the target task, extracting index data from a database in the demand scene, and generating an answer based on the index data. The problems that the long text understanding ability and the intention recognition ability are weak, and the intention recognition and task matching accuracy and stability are low in the prior art are at least solved. The method is suitable for intention recognition and intention processing scenes.

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to an information processing method, device, electronic device and computer-readable storage medium for a large language model. Background Art

[0002] In recent years, big language model technology has made continuous breakthroughs, and more and more companies have begun to pay attention to how to apply the results of big language model technology to production practice in order to improve production efficiency, reduce production and operation costs, and improve service levels and product quality.

[0003] The scale of the communication operator's network is huge and the structure is complex. The pressure to ensure stable network operation and good network service quality is very high. Each link of network planning, construction, maintenance, optimization and operation is very complex and time-consuming, requiring a lot of manpower costs, which brings huge challenges and burdens to the daily production and operation of operators. At present, operators have tried to apply large language model technology to the links of network intelligent operation, such as: using large language models to realize intelligent question and answer of network problems, intelligent data collection of network indicators, intelligent generation of SQL (Structured Query Language), and network fault analysis (AI Agent).

[0004] However, the large language model technology applied to network intelligent operation scenarios is still not mature enough and has certain limitations. The reason is that the network intelligent operation scenarios of operators usually provide users with query functions for multiple network indicator data, and each indicator data corresponds to multiple different query dimensions. Therefore, there are many full-volume intents; and the complexity and diversity of user questions submitted by users when querying network indicator data in network intelligent operation scenarios are high, which puts high demands on the long text understanding ability of large language models. However, the large language model applied to the network intelligent operation link in the existing technology has weak long text understanding ability. When there are many full-volume intents, the large language model cannot accurately identify the intent, and there is a certain deviation in the intent identification process, which leads to entering the wrong task processing branch and obtaining output that does not meet expectations. In addition, large language models are usually generative models with randomness and unstable intent recognition capabilities. The same user question is sometimes recognized correctly and sometimes incorrectly during intent recognition. In the operator's network intelligent operation scenario, operators need to follow strict security and compliance standards. Even a 1% error may lead to the failure of important indicator data parsing, which in turn may affect the key parameter configuration of network equipment and cause network failures. In severe cases, it may have a large-scale negative social impact. Therefore, if there is a 1% error or uncertainty in the large language model technology, the large language model technology will become unavailable. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides an information processing method, device, electronic device and computer-readable storage medium for a large language model. The method can realize intelligent and accurate information processing of a large language model, improve the long text understanding and intent recognition capabilities of the large language model, improve the accuracy and stability of intent recognition and task matching, and thereby reduce the errors and uncertainties in the information processing process of the large language model.

[0006] In a first aspect, the present invention provides an information processing method for a large language model, comprising: obtaining a user question and a network application portal identifier; obtaining a full set of intentions and a full set of tasks in a demand scenario based on the user question and the network application portal identifier; determining a target intention and a target task based on the user question, the full set of intentions and the full set of tasks in the demand scenario; extracting indicator data from a database in the demand scenario based on the target intention and the target task, and generating an answer based on the indicator data.

[0007] Preferably, the method of obtaining the full amount of intentions and full amount of tasks under the demand scenario according to the user question and the network application portal identifier specifically includes: extracting indicator parameters from the user question, wherein the indicator parameters are used to characterize the field name of the indicator data; obtaining the full amount of scenarios under the network application portal according to the network application portal identifier; determining the user's demand scenario according to the indicator parameters and the full amount of scenarios under the network application portal; and obtaining the full amount of intentions and full amount of tasks according to the user's demand scenario.

[0008] Preferably, the target intent and target task are determined based on user questions, all intentions and all tasks in the demand scenario, specifically including: based on the semantic focus algorithm and user questions, screening the all intentions in the demand scenario to obtain the target intent; according to the preset mapping relationship between intentions and tasks, matching the target task corresponding to the target intent from the all tasks in the demand scenario.

[0009] Preferably, the method based on the semantic focus algorithm and user questions filters all the intentions in the demand scenario to obtain the target intention, specifically including: calculating the text similarity between each intention in the full amount of intentions in the demand scenario and the user question; based on the text similarity, filtering all the intentions in the demand scenario to obtain a list of similar intentions; calculating the semantic similarity between each similar intention in the list of similar intentions and the user question; taking the similar intention with the highest semantic similarity as the target intention.

[0010] Preferably, the method of extracting indicator data from a database under a demand scenario according to a target intent and a target task specifically includes: obtaining a task description of the target task, wherein the task description is used to characterize an output specification format; extracting intent supplementary parameters and indicator-related parameters from user questions according to the target intent and the task description, wherein the indicator-related parameters are used to characterize constraints on the indicator parameters; judging whether the indicator-related parameters comply with an interface specification corresponding to the task description; in response to the indicator-related parameters complying with an interface specification corresponding to the task description, combining the task description, the indicator-related parameters, the intent supplementary parameters and the indicator parameters to generate a parameter set in a json format; and extracting indicator data from the database under the demand scenario based on the indicator parameters in the json format parameter set.

[0011] Preferably, after determining whether the indicator-related parameters comply with the interface specifications corresponding to the task description, and before extracting the indicator data based on the indicator parameters in the json format parameter set, the information processing method of the large language model also includes: in response to the indicator-related parameters not complying with the interface specifications corresponding to the task description, adjusting the interface position of the indicator-related parameters in the task description to obtain an optimized task description, and combining the optimized task description, indicator-related parameters, intent supplementary parameters and indicator parameters to generate a json format parameter set.

[0012] Preferably, the generating of answers based on indicator data specifically includes: determining whether the intended supplementary parameters in the json format parameter set are for displaying data only; in response to the intended supplementary parameters in the json format parameter set being for displaying data only, based on a reply template, combining the json format parameter set and the indicator data to generate an answer; in response to the intended supplementary parameters in the json format parameter set being not for displaying data only, selecting an indicator data display style according to the intended supplementary parameters, generating an indicator data display result based on the indicator data display style, and generating an answer by combining the json format parameter set and the indicator data display result based on the reply template.

[0013] In a second aspect, the present invention also provides an information processing device for a large language model, comprising: a first acquisition module, a second acquisition module, a determination module and an extraction module, the first acquisition module being used to acquire user questions and network application portal identifiers, the second acquisition module being connected to the first acquisition module and being used to acquire the full amount of intentions and full amount of tasks in a demand scenario based on the user questions and the network application portal identifier, the determination module being connected to the second acquisition module and being used to determine the target intentions and target tasks based on the user questions and the full amount of intentions and full amount of tasks in the demand scenario, the extraction module being connected to the determination module and being used to extract indicator data from a database in the demand scenario based on the target intentions and target tasks, and generating answers based on the indicator data.

[0014] In a third aspect, the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the information processing method of the large language model provided in the first aspect.

[0015] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the information processing method of the large language model provided in the first aspect above is implemented.

[0016] The present invention provides a large language model information processing method, device, electronic device and computer-readable storage medium, which identifies the user's demand scenario based on the user's question and the network application portal identifier, and directly obtains the preset full intention and full task under the demand scenario. The preset full intention is accurate and concise, which greatly reduces the difficulty of the large language model in understanding the user's question in the long text, and then screens the full intention and full intention under the demand scenario according to the user's question, accurately identifies the target intention and target task closest to the user's question, extracts indicator data based on the target intention and target task, and generates intelligent and accurate answers. Therefore, the present invention can realize intelligent and accurate information processing of the large language model, improve the long text understanding ability and intention recognition ability of the large language model, reduce the error and uncertainty of the information processing process of the large language model, and thus improve the accuracy and stability of intention recognition and task matching. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of an information processing method for a large language model according to Embodiment 1 of the present invention;

[0018] Figure 2 This is a flow chart of an information processing method for a large language model according to Embodiment 2 of the present invention;

[0019] Figure 3 Schematic diagram of the structure of an information processing device for a large language model in Example 3 of the present invention. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the technical solution of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0021] It should be understood that the specific embodiments and drawings described herein are only used to explain the present invention rather than to limit the present invention.

[0022] It can be understood that, in the absence of conflict, the various embodiments of the present invention and the various features in the embodiments can be combined with each other.

[0023] It can be understood that, for the convenience of description, the drawings of the present invention only show the parts related to the present invention, while the parts irrelevant to the present invention are not shown in the drawings.

[0024] It can be understood that each unit and module involved in the embodiments of the present invention may correspond to only one physical structure, or may be composed of multiple physical structures, or multiple units and modules may be integrated into one physical structure.

[0025] It can be understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of the present invention may occur in an order different from that marked in the drawings.

[0026] It is understood that the flowcharts and block diagrams of the present invention illustrate the possible architectures, functions, and operations of the systems, devices, equipment, and methods according to the various embodiments of the present invention. Each box in the flowchart or block diagram may represent a unit, module, program segment, or code, which contains executable instructions for implementing the specified functions. Moreover, each box or combination of boxes in the block diagram and flowchart may be implemented by a hardware-based system that implements the specified functions, or may be implemented by a combination of hardware and computer instructions.

[0027] It can be understood that the units and modules involved in the embodiments of the present invention can be implemented by software or hardware. For example, the units and modules can be located in a processor.

[0028] Embodiment 1:

[0029] like Figure 1 As shown, this embodiment provides an information processing method for a large language model. The information processing method for a large language model includes:

[0030] S101, obtaining a user question and a network application portal identifier.

[0031] In this embodiment, the large language model is applied to the network application APP. When the user submits a user question in the question and answer window of the network application APP portal, the large language model obtains the user question and the network application portal identifier.

[0032] S102, obtaining all intentions and all tasks in the demand scenario according to the user question and the network application portal identifier.

[0033] In this embodiment, a network application APP generally refers to an application that provides services through the Internet. The network application APP may include several portals or sub-applications, each of which may focus on a specific function or service, for example: the communication operator APP includes a base station information portal, a network optimization portal, and a real-time monitoring portal; wherein the base station information portal is used to query the number and type of base stations in a specific area; the network optimization portal is used to provide a user feedback function so that users can report signal problems, and the operator optimizes the base station layout based on feedback; the real-time monitoring portal is used to display the real-time operating status and maintenance information of the base station. The network application APP portal contains several scenario applications, for example: the base station information portal includes scenario applications for base station data query and computer room data query. Therefore, the demand scenario refers to the scenario application corresponding to the user problem identified by the large language model from the network application portal. The full amount of intent refers to all possible intents identified by the large language model from the user problem, that is, the set of user intents. The full amount of tasks refers to all tasks that the large language model needs to perform based on all possible intents, that is, the set of user tasks. Specifically, S102: according to the user question and the network application portal identifier, obtain the full amount of intentions and full amount of tasks in the demand scenario, including steps S1021-S1024:

[0034] S1021, extracting indicator parameters from the user question, wherein the indicator parameters are used to represent the field names of the indicator data.

[0035] In this embodiment, after the large language model obtains the user question and the network application portal logo, the large language model extracts the indicator parameters in the user question, such as: the user question query is "Please help me check the current provincial proportion of the number of 5G base stations", and the indicator parameter extracted from the user question query is the number of 5G base stations. Subsequently, the large language model of this embodiment can call the back-end interface through the indicator parameter "number of 5G base stations" to extract the indicator data corresponding to the number of 5G base stations, such as: "The number of 5G base stations in Province A is about 300,000, mainly concentrated in City a; the number of 5G base stations in Province B is about 250,000, and City b is the main construction area; the number of 5G base stations in Province C is about 200,000...", and then select the indicator data style of the intention supplementary parameter "provincial proportion" extracted later, such as: pie chart, to display the indicator data.

[0036] S1022, obtaining all scenarios under the network application portal according to the network application portal identifier.

[0037] In this embodiment, accordingly, the full set of scenarios refers to all scenario applications, that is, a set of scenario applications.

[0038] S1023, determining the user's demand scenario based on the indicator parameters and the full set of scenarios under the network application portal.

[0039] In this embodiment, by extracting indicator parameters from user questions, all scenarios under the network application portal are screened to obtain the user's demand scenarios. For example, if the indicator parameter is the number of 5G base stations and the network application portal is the base station information portal, the user's demand scenario can be screened to be base station data query.

[0040] S1024: Obtain all intents and tasks based on the user's demand scenarios.

[0041] In this embodiment, the question and answer window of the network application portal can provide users with query functions for multiple network indicator data, and each indicator data corresponds to multiple different query dimensions. Therefore, there are many types of intents that need to be processed in the same question and answer window. In this embodiment, the full amount of intents in each scenario that the large language model needs to identify is defined in advance based on the full amount of scenarios under the network application portal and the goals that the large language model needs to complete, and listed according to the number. As shown in Table 1, there are more than 40 full amount of intents for communication network related indicator queries, including network operation data, network work order data, network resource information, network alarm information related indicators and information queries, such as: querying monthly granularity network indicator data, daily granularity network indicator data, indicator changes compared with the previous month, indicator month-on-month, to-do work orders, completed work orders, professional to-do / completed work orders, application to-do, and completed work orders. Among them, there are more than 600 types of network indicators extracted for each intention, more than 150 types of scene application APPs corresponding to work orders, and 10 types of professional types corresponding to work orders, such as wireless networks, core networks, and broadband networks.

[0042] Table 1 Full set of intents in each scenario

[0043]

[0044] In addition, after defining the full set of intents in each scenario that the large language model needs to recognize, define the full set of tasks in each scenario based on the full set of intents in each scenario, that is, determine the corresponding required processing actions based on the full set of intents in each scenario. The full set of tasks in each scenario consists of several specific task prompts and are numbered. Therefore, after determining the user's demand scenario, the corresponding full set of intents and full set of tasks can be matched according to the demand scenario to obtain the full set of intents and full set of tasks in the demand scenario.

[0045] It should be noted that intents are numbered and saved in text form, and tasks are numbered and saved in list form. After the large language model answers the user's question, the intent identified by the large language model from the user's question and the task performed by the large language model on the user's question can be stored and updated to the full set of intents and full set of tasks in the corresponding scenario.

[0046] S103, determining the target intention and target task according to the full amount of intentions and full amount of tasks in the user problem and demand scenario.

[0047] In this embodiment, the prior art usually uses a large language model to directly perform intent recognition and corresponding task execution on user questions obtained from the question and answer window of the network application portal. Therefore, the length of the text that the large language model needs to understand in the prior art is too long. In the process of processing more than 40 intents and tasks, the number of tokens that the large language model needs to process is 7832, which will cause problems such as high error rate in intent understanding and poor task execution effect. However, according to the same user question, this embodiment directly obtains the preset streamlined full intent in the intent recognition process, that is, the full intent in the required scenario. Assuming that the number of full intents is still the same as the prior art, that is, more than 40, the number of text tokens corresponding to the full intent that the large language model needs to process is reduced to 1034. When the full intent in the required scenario is further screened and the target intent is obtained, the number of text tokens corresponding to the target intent that the large language model needs to process is reduced to 110. Therefore, this embodiment can significantly reduce the length of text processed by the large language model each time by directly obtaining the preset streamlined full intent during the intent recognition process of the large language model, and greatly improve the accuracy of model recognition and task execution effect.

[0048] Specifically, S103: determining the target intention and target task according to the user's questions and the full amount of intentions and full amount of tasks in the demand scenario, including steps S1031 and S1032:

[0049] S1031, based on the semantic focus algorithm and user questions, filter all the intents in the demand scenario to obtain the target intent.

[0050] In this embodiment, the semantic focus algorithm includes but is not limited to: topic model, word embedding, sentence and document embedding, clustering algorithm, information retrieval algorithm, deep learning model and graph model. Among them, topic models include but are not limited to: Latent Dirichlet Allocation (LDA) and Non-negative Matrix Factorization (NMF), word embeddings include but are not limited to: Word2Vec and GloVe, sentence and document embeddings include but are not limited to: Doc2Vec and Universal Sentence Encoder (USE), clustering algorithms include but are not limited to: K-means clustering and hierarchical clustering, information retrieval algorithms include but are not limited to: BM25 (Best Matching 25, a classic information retrieval algorithm) and TF-IDF (erm frequency-inverse document frequency), deep learning models include but are not limited to: BERT (Bidirectional Encoder Representations from Transformers, a natural language processing (NLP) model based on deep learning) and GPT (Generative Pre-trained Transformer), and graph models include but are not limited to: Graph Neural Networks (GNN).

[0051] Latent Dirichlet Allocation: A generative model for discovering topic distribution in documents; Non-negative Matrix Factorization: Extracting the topic of a document through matrix factorization. Word2Vec: Learn vector representations of words through context to capture semantic relationships between words; GloVe: A word embedding method based on global word frequency statistics that generates dense vector representations of words. Doc2Vec: An extension of Word2Vec for generating vector representations of documents; Universal Sentence Encoder: Generates fixed-length vector representations of sentences, suitable for a variety of NLP tasks. K-means clustering: Clusters text data into K clusters based on feature vectors of text; Hierarchical clustering: Clusters text by building a tree structure, suitable for discovering hierarchical relationships between texts. BM25: A retrieval model based on word frequency and document length, used to evaluate the relevance of documents to queries; TF-IDF: Evaluates the importance of words by calculating word frequency and inverse document frequency, commonly used in text classification and focusing. BERT: Learns word representations through bidirectional context, suitable for a variety of NLP tasks; GPT: A generative model that can understand and generate natural language text. Graph Neural Network: Modeling the relationship between texts through graph structures, suitable for complex semantic focus tasks.

[0052] Specifically, S1031: based on the semantic focus algorithm and user questions, all the intentions in the demand scenario are screened to obtain the target intention, including: calculating the text similarity between each intention in the demand scenario and the user question; based on the text similarity, all the intentions in the demand scenario are screened to obtain a list of similar intentions; calculating the semantic similarity between each similar intention in the list of similar intentions and the user question; taking the similar intention with the highest semantic similarity as the target intention.

[0053] In this embodiment, due to the high similarity between some intents in the prior art, the intent recognition error is caused, and the indicator data obtained by executing the corresponding task does not meet the user's expectations. For example, the two intents are "query the monthly granularity indicator data of a certain place" and "query the daily granularity indicator data of a certain place". The two intents are relatively similar, and the total number of intents is large. When the user's question is "query the number of 5G base stations in Beijing in July" and "query the number of 5G base stations in Beijing on July 1", the two intents are easily confused with each other. In addition, there are two intentions in the prior art that are quite different, but the large language model may also be confused when screening from many intents. For example, the two intents are "the regional distribution of indicators with monthly granularity" and "the provincial proportion of indicators with daily granularity". When the user's query is "please help me query the provincial proportion of the current number of 5G base stations", although the obvious intention "the provincial proportion of daily granularity indicators" is more matched, the large language model usually misidentifies it as "the regional distribution of indicators with monthly granularity".

[0054] After filtering out all the intents in the demand scenario based on the demand scenario, the text similarity between each intent in the demand scenario and the user's question is calculated, and based on the text similarity, all the intents in the demand scenario are filtered to obtain a list of similar intents. For example, if the user's query is "Please help me check the current provincial proportion of the number of 5G base stations", the large language model is asked to screen a list of similar intents similar to the user's question from the full amount of intents in the demand scenario, that is, more than 40 intents, and calculate and output the text similarity of the similar intents as the similar intent scores. The current large language model output out is "Intent recognition results are: 18, 11, 24; score list is: [0.9, 0.8, 0.1]", that is, the similar intent list includes 3 intents numbered 10, 11, and 24, and the text similarities of the 3 intents are 0.9, 0.8, and 0.1, respectively. It can be seen that intents 10 and 11 have higher scores, intent 24 is not related to the user's question, and intent 10 has the highest score. However, it is obvious that the time granularity of the user's question is inconsistent with intent 10, and the user's question is quite different from intent 10. Intent 11 is more in line with the user's question. If the large language model directly outputs the intent recognition result, that is, intent 10, and executes the corresponding numbered task, it will cause intent recognition errors, and the obtained indicator data will not meet user expectations. In this embodiment, the number of similar intents is 1-3 as an example. When the large language model filters out a list of similar intents based on user questions, the number of tokens processed by the large language model drops to 118.

[0055] Then, by calculating the semantic similarity between each similar intent in the similar intent list and the user's question, the similar intent list is filtered to obtain the target intent. For example: the user's query "Please help me check the current provincial proportion of the number of 5G base stations" is filtered through all the intents in the demand scenario, and the intents 10, 11, and 24 are obtained. Then, the intent closest to the user's question is selected from these three intents, and the output out of the current large language model is "the intent recognition result is 11, the daily granularity indicator provincial proportion, the provincial proportion, and the [provincial proportion] time parameter is "current", that is, the current provincial proportion of the number of 5G base stations is queried; the score list is [0.0, 1.0, 0.0], indicating that the question is completely matched with intent 11 and is not similar to other intents", that is, the target intent is intent 11, and the semantic similarity list of intents 10, 11, and 24 matching the user's question, that is, the score list is [0.0, 1.0, 0.0]. It can be seen that the cognition of the large language model meets the user's expectations. And the intent 11 can be output after executing the intent recognition several times, and the output result is accurate and stable. This embodiment combines the semantic focus algorithm to screen intents and tasks multiple times, reduce the possibility of high similarity between some intents, and improve the accuracy, stability and efficiency of intent recognition.

[0056] Algorithms for calculating text similarity include but are not limited to: word frequency-based methods, edit distance-based methods, TF-IDF-based methods, and N-gram models. Therefore, this embodiment can calculate the cosine value of each intent and the user question, or by calculating the Jaccard similarity of each intent and the user question (i.e., the ratio of the intersection to the union), or by converting the user question into the minimum number of editing operations (insertion, deletion, replacement) required for each intent, or by calculating the cosine similarity between each intent and the user question in combination with the TF-IDF weight, or by decomposing each intent and the user question into N-grams (continuous N words or characters), calculating the similarity of N-grams, and commonly using Jaccard or cosine similarity to obtain text similarity.

[0057] Algorithms for calculating semantic similarity include but are not limited to: word embedding models, sentence and document embedding algorithms, deep learning models, and graph models. Therefore, in this embodiment, the word vectors of each similar intention and user question can be generated through the Word2Vec algorithm, and the similarity between the word vectors can be calculated, or the word embedding vector based on global word frequency statistics can be generated for each similar intention and user question through the GloVe algorithm, and the semantic similarity of the word embedding vector can be calculated, or the vector representation of each similar intention and user question can be generated through the Doc2Vec algorithm, and the similarity between the vector representation of each similar intention and user question can be calculated, or the fixed-length vector representation of each similar intention and user question can be generated through the USE algorithm, and the similarity between the fixed-length vector representation of each similar intention and user question can be calculated, or the vector representation of bidirectional context learning words can be generated for each similar intention and user question through the BERT model, and the semantic similarity between the vector representation of each similar intention and user question can be calculated, or the similarity between each similar intention and user question can be compared through the Siamese Network using a dual-tower structure, or the relationship between each similar intention and user question can be modeled through the graph structure of the graph neural network, and the semantic similarity can be calculated.

[0058] It should be noted that after verification of actual application scenarios, for user questions with high intent similarity, and user questions with large differences in user intent but considered to be highly similar by the large language model, the accuracy of processing all intents and all tasks only once by the large language model is close to 0; and after the full intent and full tasks are separated, the accuracy of the large language model filtering intents based on the demand scenario and then executing the corresponding tasks is about 20%; the full intent and full tasks are separated, and then the intents are filtered multiple times based on the demand scenario, text similarity, and semantic similarity, and the accuracy of executing the corresponding tasks is increased to 90%. The above experiments are based on an open source model quantized by 72b-int4, in which the question-answering calls the large language model capabilities multiple times, and the average question-answering reasoning response time is 0.15 seconds per question-answering.

[0059] The large language model output out can also include a simple explanation of the intent recognition results. The user query is: Please help me check the current provincial share of the number of 5G base stations. The corresponding large language model output out may also include "The reasons for selecting these intents are as follows: Intent 10. Regional distribution of indicators with monthly granularity, asking about regional distribution, time for example: this month, June, June 2024; although the question mentions "currently", "the provincial proportion of the number of 5G base stations" implies regional distribution, and the monthly granularity is a possible statistical granularity, so the score is higher. Intent 11. Provincial proportion of indicators with daily granularity, asking about the provincial proportion of network indicators, time for example: currently, today, June 1, June 1, 2024; the "currently" in the question matches the current intent, and the question asks about the provincial proportion, so the score is higher. Intent 24. [xxG / xxg] base station information with an address name of xx; although the question does not explicitly mention the specific address and the type of G, the "provincial proportion of the number of 5G base stations" is related to the base station information, so the score is lower, but there is still some correlation." From the analysis process of the above large language model output, it can be seen that the large language model first screens the intent. Figure 1-3 When selecting a list, the reason for selecting intent 10, intent 11, and intent 24. The large language model screens the intent that best matches the user's query for the second time, and selects the closest intent from intent 10, intent 11, and intent 24 output by the large language model. The large language model also outputs an explanation, such as: "The reasons for selecting the intent are as follows: The time parameter is "current", that is, to query the current provincial proportion of 5G base stations. The score list is [0.8, 1.0, 0.0], which means it fully matches intent 11 and is not similar to other intents. The intent recognition result is intent 11."

[0060] S1032, according to the preset mapping relationship between intention and task, match the target task corresponding to the target intention from all the tasks in the demand scenario.

[0061] In this embodiment, based on the full amount of scenarios under the network application portal and the goals that the large language model needs to complete, the full amount of intentions and full amount of tasks in each scenario are defined in advance, and listed according to the numbers, the numbers of the intentions and tasks correspond one by one, and the mapping relationship between the preset intentions and tasks is generated. The prior art usually puts all the intention numbers and tasks together to let the large language model select the intention and process the corresponding task, and only calls the large language model capability once to process the intention and task at the same time. Therefore, even if the large language model finds that the intention is misunderstood during the intention recognition process, it cannot stop the task that matches the error from being processed, and the task execution effect is poor. However, this embodiment separates all intentions and all tasks, and numbers them separately, the intention number and the task number correspond one by one, calls the large language model capability multiple times, performs intention recognition through the large language model, gets the intention number and then gets the task with the corresponding number, and then the large language model processes according to the task description of the corresponding number, which can timely avoid the execution of the wrong task, thereby improving the task execution effect, and thus ensuring the security and reliability of data information.

[0062] Therefore, after determining the target intent, the large language model clearly knows its purpose and can match the corresponding target task according to the target intent number to execute the target task.

[0063] S104, extracting indicator data from the database under the demand scenario according to the target intention and target task, and generating answers based on the indicator data.

[0064] Specifically, according to the target intent and target task, indicator data is extracted from the database under the demand scenario, including: obtaining a task description of the target task, wherein the task description is used to characterize the output specification format; according to the target intent and task description, extracting intent supplementary parameters and indicator-related parameters from user questions, wherein the indicator-related parameters are used to characterize the constraints of the indicator parameters; judging whether the indicator-related parameters comply with the interface specifications corresponding to the task description; in response to the indicator-related parameters complying with the interface specifications corresponding to the task description, combining the task description, indicator-related parameters, intent supplementary parameters and indicator parameters to generate a json format parameter set; based on the indicator parameters in the json format parameter set, extracting indicator data from the database under the demand scenario.

[0065] In this embodiment, after the large language model matches the corresponding target task according to the target intent number, the intent supplementary parameters and indicator-related parameters are extracted from the user question according to the specific task description Prompt and the target intent, and the indicator-related parameters include but are not limited to: time parameters and location parameters. For example: the user question is "How has the number of 5G base stations in Shanghai changed in the past 10 months?" The large language model intention recognition obtains the target intent as "Query the changes in the monthly granularity indicator data in the past month." According to the task description Prompt and the target intent, the location parameter extracted from the user question is "Shanghai", the intent supplementary parameter is "changes", and the time parameter is "monthly granularity" and "recent 10 months." The user question is "Please help me query the current provincial proportion of the number of 5G base stations." The large language model intention recognition obtains the target intent as "daily granularity indicator provincial proportion, query the provincial proportion, time for example: today, June 1, 2024." According to the task description Prompt and the target intent, the intent supplementary parameters extracted from the user question are "provincial proportion", and the time parameters are "daily granularity" and "currently".

[0066] Determine whether the indicator-related parameters conform to the interface specifications corresponding to the task description prompt. The large language model outputs text format. If it conforms to the interface specifications corresponding to the task description, it is necessary to combine the task description, indicator-related parameters, intent supplementary parameters and indicator parameters to generate a json format parameter set.

[0067] Specific task description prompt, such as:

[0068] f"""Query the indicator data at the monthly granularity;

[0069] Extraction of parameters for changes in the past month: The parameters include monthly granularity time, provincial address, and indicator parameters. Assuming the query is in the past 10 months, the end time is (now_month}, and the start time is 10 months before the end time;

[0070] Parameter format: Check time, location and indicator parameters. The time format is yyyymm, where yyyy represents year and mm represents month. For example, 202401, the current time is (now_month). Location is at the provincial level, including provinces, municipalities, autonomous regions, and special administrative regions, such as Anji and Beijing. The indicator parameters are xx number, xx quantity, and xx completion rate, such as the number of 5G medium-frequency base stations.

[0071] Please use the following json format to give the extracted parameters. Please strictly output in json format. Please fill in None with parameters not mentioned in the question: {{"type":"BI","params":[{{"start time":XX,"end time":XX,"province":XX,"indicator parameter":XX,"field":"value","method":"query"}}],"showway":"fold"}}.

[0072] Since the large language model is a generative model, it has a certain degree of randomness. For example, the JSON format parameter set sometimes has spaces and comments. After the JSON format parameter set is generated, it needs to be parsed into format parameters that can be processed by the backend through post-processing methods so that it can be passed to the backend. The backend extracts and displays indicator data according to the indicator parameters of the JSON format parameter set through the specific application system interface.

[0073] It should be noted that when the intention supplement parameter showWay is extracted from the user question as an indicator trend, such as "change situation", it can be replaced according to the preset fixed parameters, such as "fold" or "single", where "fold" means a line chart and "single" means only displaying data. Since the prompt may contain the system parameter now_month, which is the month corresponding to the current time of the server system, in order to avoid errors caused by json format conflicts, {{ and}} are used here to replace { and} in json.

[0074] Optionally, after determining whether the indicator-related parameters comply with the interface specifications corresponding to the task description, and before extracting the indicator data based on the indicator parameters in the JSON format parameter set, the information processing method of the large language model also includes: in response to the indicator-related parameters not complying with the interface specifications corresponding to the task description, adjusting the interface position of the indicator-related parameters in the task description to obtain an optimized task description, and combining the optimized task description, indicator-related parameters, intent supplementary parameters and indicator parameters to generate a JSON format parameter set.

[0075] In this embodiment, if the task description does not conform to the interface specification corresponding to the task description, the task description prompt needs to be adjusted to make the large language model have an output format that conforms to the interface specification. For example: the user question query is "How has the number of 5G base stations in Shanghai changed in the past 10 months?", and the extracted time parameters are "monthly granularity" and "the past 10 months". Then, combining the time parameters with the task description is obviously not in line with the interface specification of the time parameters. Therefore, it is necessary to adjust the interface position of the start time and end time in the task prompt, so that the large language model can obtain the current time and the time parameter format "The current time is July 2024, the time format is yyyymm, yyyy represents year, mm represents month, sample 202401", and remind the large language model "The start time is before the end time" to generate indicator-related parameter data, that is, the start time "202309" and the end time "202407". Then, the adjusted task description Prompt is combined, that is, the task description, task description, indicator-related parameters, intent supplementary parameters and indicator parameters are optimized to generate a parameter set in json format, that is, the output out of the current large language model is: {"type":"BI", "params":[{"start time":"202309", "end time":"202407", "province":"Shanghai", "indicator parameter":"number of 5G base stations", "field":"value", "method":"query"}], "showway":"fold"}.

[0076] In addition, if the user's question query is "How many 5G base stations were there in Shanghai yesterday?", the large language model correctly understands the user's intention and performs the corresponding tasks according to the user's intention - extracting indicator parameters, indicator-related parameters and intention supplementary parameters. The extracted time parameter is "yesterday", the location parameter is "Shanghai", and the indicator parameter is "number of 5G base stations". After adjusting the task description Prompt, the current time is obtained as July 22, 2024, and the time is converted into a specific date in the format of YYYYMMDD. The final time parameter is "20240721", and the intention-related parameters type, field, method, and showWay are preset fixed parameters. The json format parameter set is obtained, that is, the data out of the current large language model is: {"type":"BI", "params":[{"time":"20240721", "province":"Shanghai", "indicator parameter":"number of 5G base stations", "field":"value", "method":"query"}], "showway":"single"}.

[0077] Specifically, generating an answer based on the indicator data includes: determining whether the intended supplementary parameters in the json format parameter set are for displaying data only; in response to the intended supplementary parameters in the json format parameter set being for displaying data only, based on a reply template, combining the json format parameter set and the indicator data to generate an answer; in response to the intended supplementary parameters in the json format parameter set being not for displaying data only, selecting an indicator data display style according to the intended supplementary parameters, generating an indicator data display result based on the indicator data display style, and combining the json format parameter set and the indicator data display result based on the reply template to generate an answer.

[0078] In this embodiment, since there are intention supplementary parameters in the json format parameter set, such as filed, method, and showWay shown in the task description prompt, most of them cannot be directly obtained from the user's question. Therefore, it is necessary to determine whether the intention supplementary parameters in the json format parameter set are only for displaying data according to the corresponding post-processing means. If only data is displayed, the json format parameter set and the indicator data are semantically spliced ​​through the reply template to obtain an answer. In this embodiment, the reply template is "s% (time parameter) s% (position parameter) s% (indicator parameter) is s% (indicator data)" as an example. The placeholder is replaced by the json format parameter set and the indicator data to obtain an answer, and the answer is forwarded to the front end. The front end displays the answer in the question and answer window of the network application APP portal, such as displaying text, pictures, jump links and other content. If it is not only for displaying data, the indicator data display style is selected according to the intention supplementary parameters to generate the indicator data display result. The json format parameter set and the indicator data display result are combined through the reply template to generate the answer.

[0079] The present embodiment provides a method for processing information of a large language model, which identifies the user's demand scenario according to the user's question and the web application portal identifier, and directly obtains the preset full intent and full task under the demand scenario. The preset full intent is accurate and concise, which greatly reduces the difficulty of the large language model in understanding the user's question in the long text. Then, according to the user's question, the full intent and full intent under the demand scenario are screened, and the target intent and target task closest to the user's question are accurately identified, so as to extract indicator data based on the target intent and target task, and generate intelligent and accurate answers, so as to realize intelligent and accurate information processing of the large language model, improve the long text understanding ability and intent recognition ability of the large language model, reduce the error and uncertainty of the information processing process of the large language model, and thus improve the accuracy and stability of intent recognition and task matching. In addition, by directly obtaining the preset streamlined full intent in the process of intent recognition of the large language model, and screening the full intent multiple times to obtain the most accurate intent, the length of the text processed by the large language model each time can be greatly reduced, and the accuracy of model recognition and task execution effect can be greatly improved. By combining the semantic focus algorithm, multiple screening of intents and tasks can reduce the possibility of high similarity between some intents, thereby improving the accuracy, stability and efficiency of intent recognition. By separating all intents and all tasks and numbering them separately, with intent numbers and task numbers corresponding one to one, calling the large language model capabilities multiple times, using the large language model to perform intent recognition, and then getting the intent number and then the task with the corresponding number, and then the large language model processes according to the task description with the corresponding number, it can timely avoid the execution of erroneous tasks, thereby improving the task execution effect, and thus ensuring the security and reliability of data information.

[0080] Embodiment 2:

[0081] like Figure 2 As shown, this embodiment provides an information processing method for a large language model. The information processing method for a large language model includes:

[0082] S201, obtaining a user question and a network application portal identifier, extracting an indicator parameter from the user question, and determining a demand scenario according to the network application portal identifier and the indicator parameter.

[0083] In this embodiment, the large language model is applied to the network application APP. When the user submits a user question in the question and answer window of the network application APP portal, the large language model obtains the user question and the network application portal identifier.

[0084] After the large language model obtains the user question and the network application portal logo, the large language model extracts the indicator parameters in the user question, such as: the user question query is "Please help me check the current provincial proportion of the number of 5G base stations", and the indicator parameter extracted from the user question query is the number of 5G base stations. Subsequently, the large language model of this embodiment can call the back-end interface through the indicator parameter "number of 5G base stations" to extract the indicator data corresponding to the number of 5G base stations, such as: "The number of 5G base stations in Province A is about 300,000, mainly concentrated in City a; the number of 5G base stations in Province B is about 250,000, and City b is the main construction area; the number of 5G base stations in Province C is about 200,000...", and then select the indicator data style of the intention supplementary parameter "provincial proportion" extracted later, such as: pie chart, to display the indicator data.

[0085] Network application APP usually refers to an application that provides services through the Internet. Network application APP can include several portals or sub-applications, each of which may focus on specific functions or services. For example, the communication operator APP includes a base station information portal, a network optimization portal, and a real-time monitoring portal; among them, the base station information portal is used to query the number and type of base stations in a specific area; the network optimization portal is used to provide user feedback functions so that users can report signal problems, and the operator optimizes the base station layout based on feedback; the real-time monitoring portal is used to display the real-time operating status and maintenance information of the base station. The network application APP portal contains several scenario applications, for example, the base station information portal includes scenario applications for base station data query and computer room data query.

[0086] By extracting indicator parameters from user questions, we can filter all scenarios under the network application portal to obtain the user's demand scenarios. For example, if the indicator parameter is the number of 5G base stations and the network application portal is the base station information portal, we can filter out the user's demand scenario as base station data query.

[0087] S202: Obtain all intents and tasks according to the required scenario.

[0088] In this embodiment, the Q&A window of the network application portal can provide users with query functions for multiple network indicator data. Each indicator data corresponds to multiple different query dimensions. Therefore, many types of intents need to be processed in the same Q&A window. This embodiment defines the full amount of intents in various scenarios that the large language model needs to identify based on the full amount of scenarios under the network application portal and the goals that the large language model needs to achieve, and lists them according to the numbers. Figure 2As shown in the figure, there are more than 40 full-volume intents for communication network-related indicator queries, including network operation data, network work order data, network resource information, and network alarm information-related indicators and information queries, such as: querying monthly granularity network indicator data, daily granularity network indicator data, indicator changes compared with the previous month, indicator month-on-month, pending work orders, completed work orders, professional pending / completed work orders, application pending, and completed work orders. Among them, each intent extracts more than 600 types of network indicators, more than 150 types of scenario application APPs corresponding to work orders, and 10 types of professional types corresponding to work orders, such as wireless networks, core networks, and broadband networks.

[0089] In addition, after defining the full set of intents in each scenario that the large language model needs to recognize, define the full set of tasks in each scenario based on the full set of intents in each scenario, that is, determine the corresponding required processing actions based on the full set of intents in each scenario. The full set of tasks in each scenario consists of several specific task prompts and are numbered. Therefore, after determining the user's demand scenario, the corresponding full set of intents and full set of tasks can be matched according to the demand scenario to obtain the full set of intents and full set of tasks in the demand scenario.

[0090] S203, calculate the text similarity between each intention in the full set of intentions in the demand scenario and the user's question; based on the text similarity, filter all the intentions in the demand scenario to obtain a list of similar intentions; calculate the semantic similarity between each similar intention in the list of similar intentions and the user's question; take the similar intention with the highest semantic similarity as the target intention.

[0091] In this embodiment, Figure 2 The selected intent in is the target intent. The prior art usually uses a large language model to directly perform intent recognition and corresponding task execution on user questions obtained in the question and answer window of the network application portal. Therefore, the length of the text that the large language model needs to understand in the prior art is too long. In the process of processing more than 40 intents and tasks, the number of tokens that the large language model needs to process is 7832, which will lead to problems such as high error rate in intent understanding and poor task execution effect. However, according to the same user question, this embodiment directly obtains the preset streamlined full intent in the intent recognition process, that is, the full intent in the required scenario. The number of text tokens corresponding to the full intent that the large language model needs to process is 1034. When the full intent in the required scenario is further screened to obtain the target intent, the number of text tokens corresponding to the target intent that the large language model needs to process is reduced to 118.

[0092] Because the existing technology has a high similarity between some intents, it leads to intent recognition errors, and the indicator data obtained by executing the corresponding task does not meet the user's expectations. For example, two intents are "query the monthly granularity indicator data of a certain place" and "query the daily granularity indicator data of a certain place". The two intents are relatively similar, and the total number of intents is large. When the user's questions are "query the number of 5G base stations in Beijing in July" and "query the number of 5G base stations in Beijing on July 1", the two intents are easily confused with each other. In addition, the existing technology also has two intents that are quite different, but the large language model may also be confused when screening from many intents. For example, the two intents are "the regional distribution of indicators with monthly granularity" and "the provincial proportion of indicators with daily granularity". When the user's query is "please help me query the provincial proportion of the current number of 5G base stations", although the obvious intent "the provincial proportion of indicators with daily granularity" is more matched, the large language model usually misidentifies it as "the regional distribution of indicators with monthly granularity".

[0093] After filtering out all the intents in the demand scenario based on the demand scenario, the text similarity between each intent in the demand scenario and the user's question is calculated, and based on the text similarity, all the intents in the demand scenario are filtered to obtain a list of similar intents. For example, if the user's query is "Please help me check the current provincial proportion of the number of 5G base stations", the large language model is asked to screen a list of similar intents similar to the user's question from the full amount of intents in the demand scenario, that is, more than 40 intents, and calculate and output the text similarity of the similar intents as the similar intent scores. The current large language model output out is "Intent recognition results are: 18, 11, 24; score list is: [0.9, 0.8, 0.1]", that is, the similar intent list includes 3 intents numbered 10, 11, and 24, and the text similarities of the 3 intents are 0.9, 0.8, and 0.1, respectively. It can be seen that intents 10 and 11 have higher scores, intent 24 is not related to the user's question, and intent 10 has the highest score. However, it is obvious that the time granularity of the user's question is inconsistent with intent 10, and the user's question is quite different from intent 10. Intent 11 is more in line with the user's question. If the large language model directly outputs the intent recognition result, that is, intent 10, and executes the corresponding numbered task, it will cause intent recognition errors, and the obtained indicator data will not meet user expectations. In this embodiment, the number of similar intents is 1-3 as an example. When the large language model filters out a list of similar intents based on user questions, the number of tokens processed by the large language model drops to 118.

[0094] Then, by calculating the semantic similarity between each similar intent in the similar intent list and the user's question, the similar intent list is filtered to obtain the target intent. For example: the user's query "Please help me check the current provincial proportion of the number of 5G base stations" is filtered through all the intents in the demand scenario, and the intents 10, 11, and 24 are obtained. Then, the intent closest to the user's question is selected from these three intents, and the output out of the current large language model is "the intent recognition result is 11, the daily granularity indicator provincial proportion, the provincial proportion, and the [provincial proportion] time parameter is "current", that is, the current provincial proportion of the number of 5G base stations is queried; the score list is [0.0, 1.0, 0.0], indicating that the question is completely matched with intent 11 and is not similar to other intents", that is, the target intent is intent 11, and the semantic similarity list of intents 10, 11, and 24 matching the user's question, that is, the score list is [0.0, 1.0, 0.0]. It can be seen that the cognition of the large language model meets the user's expectations. And the intention recognition results can be executed several times to output intention 11, and the output results are accurate and stable.

[0095] S204, according to the preset mapping relationship between intent and task, match the target task corresponding to the target intent from all tasks in the demand scenario; obtain the task description of the target task, wherein the task description is used to characterize the output specification format; according to the task description and the target intent, extract the intent supplementary parameters and indicator-related parameters from the user question.

[0096] In this embodiment, based on the full amount of scenarios under the network application portal and the goals that the large language model needs to complete, the full amount of intents and tasks in each scenario are defined in advance, and listed according to the numbers respectively, the numbers of the intents and tasks correspond one by one, and a mapping relationship between preset intents and tasks is generated. The prior art usually puts all intent numbers and tasks together to let the large language model select the intent and process the corresponding tasks. The large language model capability is only called once to process the intent and tasks at the same time. Therefore, even if the large language model finds that the intent is misunderstood during the intent recognition process, it cannot stop the task that matches the error from being processed, resulting in poor task execution.

[0097] Therefore, after determining the target intent, the large language model knows its purpose clearly and can match the corresponding target task according to the target intent number to execute the target task. The number of tokens that the large language model needs to process is about 110.

[0098] After the large language model matches the corresponding target task according to the target intent number, the intent supplementary parameters and indicator-related parameters are extracted from the user question according to the specific task description prompt and target intent. The indicator-related parameters include but are not limited to: time parameters and location parameters. For example: the user question is "How has the number of 5G base stations in Shanghai changed in the past 10 months?" The large language model intent recognition obtains the target intent as "Query the changes in the monthly granularity indicator data in the past month." According to the task description prompt and target intent, the location parameter "Shanghai" is extracted from the user question, the intent supplementary parameter is "changes," and the time parameters are "monthly granularity" and "recent 10 months." The user question is "Please help me query the current provincial proportion of the number of 5G base stations." The large language model intent recognition obtains the target intent as "daily granularity indicator provincial proportion, query provincial proportion, time for example: today, June 1, 2024." According to the task description prompt and target intent, the intent supplementary parameters extracted from the user question are "provincial proportion," and the time parameters are "daily granularity" and "current."

[0099] S205, determining whether the indicator-related parameters conform to the interface specifications corresponding to the task description.

[0100] In this embodiment, Figure 2 The task prompt is the task description. Figure 2 The parameters in are the indicator related parameters. Figure 2 The expectation in the task description corresponds to the interface specification. Determine whether the indicator-related parameters meet the interface specification corresponding to the task description prompt.

[0101] S206, in response to the indicator-related parameters complying with the interface specifications corresponding to the task description, the task description, the indicator-related parameters, the intent supplementary parameters and the indicator parameters are combined to generate a parameter set in json format.

[0102] In this embodiment, Figure 2 The json in the text is a parameter set in json format. The large language model outputs text format. If it meets the interface specifications corresponding to the task description, it is necessary to combine the task description, indicator-related parameters, intent supplementary parameters, and indicator parameters to generate a parameter set in json format.

[0103] S207, in response to the indicator-related parameters not complying with the interface specifications corresponding to the task description, adjusting the interface position of the indicator-related parameters in the task description to obtain an optimized task description, and combining the optimized task description, indicator-related parameters, intent supplementary parameters and indicator parameters to generate a parameter set in json format.

[0104] In this embodiment, if the task description does not conform to the interface specification corresponding to the task description, the task description prompt needs to be adjusted to make the large language model have an output format that conforms to the interface specification. For example: the extracted time parameters are "monthly granularity" and "last 10 months", then the combination of the time parameters and the task description is obviously not in line with the interface specification of the time parameters. Therefore, it is necessary to adjust the interface position of the start time and end time in the task prompt, so that the large language model can obtain the current time and the time parameter format "the current time is July 2024, the time format is yyyymm, yyyy represents year, mm represents month, sample 202401", and remind the large language model that "the start time is before the end time" to generate indicator-related parameter data, that is, the start time "202309" and the end time "202407". Then combine the adjusted task description prompt, that is, optimize the task description, task description, indicator-related parameters, intention supplementary parameters and indicator parameters, and generate a json format parameter set. For example, if the user's question is "How many 5G base stations were there in Shanghai yesterday?", the large language model correctly understands the user's intent and performs the corresponding task based on the user's intent - extracting indicator parameters, indicator-related parameters and intent supplementary parameters. The extracted time parameter is "yesterday", the location parameter is "Shanghai", and the indicator parameter is "number of 5G base stations". After adjusting the task description Prompt, the current time is obtained as July 22, 2024, and the time is converted into a specific date in the format of YYYYMMDD. The final time parameter is "20240721", and the intent-related parameters type, field, method, and showWay are preset fixed parameters, and a json format parameter set is obtained.

[0105] S208, based on the indicator parameters in the json format parameter set, extract indicator data from the database under the demand scenario; determine whether the intended supplementary parameters in the json format parameter set are for displaying data only; in response to the intended supplementary parameters in the json format parameter set being for displaying data only, based on the reply template, combine the json format parameter set and the indicator data to generate an answer; in response to the intended supplementary parameters in the json format parameter set being not for displaying data only, select an indicator data display style according to the intended supplementary parameters, generate an indicator data display result based on the indicator data display style, and combine the json format parameter set and the indicator data display result based on the reply template to generate an answer.

[0106] In this embodiment, since the large language model is a generative model, it has a certain degree of randomness. For example, the json format parameter set sometimes has spaces and comments. After the json format parameter set is generated, it needs to be parsed into format parameters that can be processed by the back end through post-processing means so as to be passed to the back end. The back end extracts and displays indicator data according to the indicator parameters of the json format parameter set through a specific application system interface.

[0107] Since there are intention supplementary parameters in the json format parameter set, such as filed, method, and showWay shown in the task description prompt, most of them cannot be directly obtained from the user's question. Therefore, it is necessary to determine whether the intention supplementary parameters in the json format parameter set are only for displaying data according to the corresponding post-processing means. If only data is displayed, the json format parameter set and the indicator data are semantically spliced ​​through the reply template to get the answer. In this embodiment, the reply template is "s% (time parameter) s% (position parameter) s% (indicator parameter) is s% (indicator data)" as an example. The placeholder is replaced by the json format parameter set and the indicator data to get the answer, and the answer is forwarded to the front end. The front end displays the answer in the question and answer window of the network application APP portal, such as displaying text, pictures, jump links and other content. If it is not only for displaying data, the indicator data display style is selected according to the intention supplementary parameters to generate the indicator data display result. The json format parameter set and the indicator data display result are combined through the reply template to generate the answer.

[0108] The present embodiment provides an information processing method for a large language model, which identifies the user's demand scenario based on the user's question and the network application portal identifier, and directly obtains the preset full intent and full task under the demand scenario. The preset full intent is accurate and concise, which greatly reduces the difficulty of the large language model in understanding the user's question in a long text. Then, based on the user's question, the full intent and full intent in the demand scenario are screened, and the target intent and target task closest to the user's question are accurately identified, so as to extract indicator data based on the target intent and target task, and generate intelligent and accurate answers, so as to realize intelligent and accurate information processing of the large language model, improve the long text understanding ability and intent recognition ability of the large language model, reduce the error and uncertainty of the information processing process of the large language model, and thus improve the accuracy and stability of intent recognition and task matching.

[0109] Embodiment 3:

[0110] like Figure 3As shown, the present embodiment also provides an information processing device of a large language model, including: a first acquisition module 31, a second acquisition module 32, a determination module 33 and an extraction module 34, the first acquisition module 31 is used to obtain user questions and network application portal identifiers, the second acquisition module 32 is connected to the first acquisition module 31, and is used to obtain the full amount of intentions and full amount of tasks in the demand scenario according to the user questions and the network application portal identifier, the determination module 33 is connected to the second acquisition module 32, and is used to determine the target intention and target task according to the user questions and the full amount of intentions and full amount of tasks in the demand scenario, and the extraction module 34 is connected to the determination module 33, and is used to extract indicator data from the database in the demand scenario according to the target intention and target task, and generate answers based on the indicator data.

[0111] Specifically, the second acquisition module 32 includes: a first extraction unit 321, a first acquisition unit 322, a first determination unit 323 and a second acquisition unit 324, the first extraction unit 321 is used to extract indicator parameters from user questions, wherein the indicator parameters are used to characterize the field names of the indicator data, the first acquisition unit 322 is used to obtain the full amount of scenarios under the network application portal according to the network application portal identifier, the first determination unit 323 is used to determine the user's demand scenario according to the indicator parameters and the full amount of scenarios under the network application portal, and the second acquisition unit 324 is used to obtain the full amount of intentions and the full amount of tasks according to the user's demand scenario.

[0112] Specifically, the determination module 33 includes: a screening unit 331 and a matching unit 332. The screening unit 331 is used to screen all intentions in the demand scenario based on the semantic focus algorithm and user questions to obtain the target intention. The matching unit 332 is used to match the target task corresponding to the target intention from all tasks in the demand scenario according to the preset mapping relationship between the intention and the task.

[0113] Specifically, the screening unit 331 includes: a first calculation subunit, a first screening subunit, a second calculation subunit and a second screening subunit. The first calculation subunit is used to calculate the text similarity between each intention in the full amount of intentions in the demand scenario and the user's question. The first screening subunit is used to screen the full amount of intentions in the demand scenario based on the text similarity to obtain a list of similar intentions. The second calculation subunit is used to calculate the semantic similarity between each similar intention in the list of similar intentions and the user's question. The second screening subunit takes the similar intention with the highest semantic similarity as the target intention.

[0114] Specifically, the extraction module 34 includes: a third acquisition unit 341, a second extraction unit 342, a first judgment unit 343, a first combination unit 344 and a third extraction unit 345, the third acquisition unit 341 is used to obtain the task description of the target task, wherein the task description is used to characterize the output specification format, the second extraction unit 342 is used to extract intent supplementary parameters and indicator-related parameters from user questions according to the task description and the target intent, wherein the indicator-related parameters are used to characterize the constraints of the indicator parameters, the first judgment unit 343 is used to judge whether the indicator-related parameters comply with the interface specifications corresponding to the task description, the first combination unit 344 is used to combine the task description, the indicator-related parameters, the intent supplementary parameters and the indicator parameters in response to the indicator-related parameters complying with the interface specifications corresponding to the task description to generate a json format parameter set, and the third extraction unit 345 is used to extract indicator data from the database under the demand scenario based on the indicator parameters in the json format parameter set.

[0115] Optionally, the extraction module 34 also includes: a second combination unit 346, which is used to adjust the interface position of the indicator-related parameters in the task description in response to the indicator-related parameters not complying with the corresponding interface specifications of the task description, to obtain an optimized task description, and to combine the optimized task description, indicator-related parameters, intent supplementary parameters and indicator parameters to generate a parameter set in JSON format.

[0116] Optionally, the extraction module 34 also includes: a second judgment unit 347, a first generation unit 348 and a second generation unit 349, the second judgment unit 347 is used to judge whether the intended supplementary parameters in the json format parameter set are for displaying data only, the third generation unit 348, in response to the intended supplementary parameters in the json format parameter set being for displaying data only, based on the reply template, combines the json format parameter set and the indicator data to generate an answer, the second generation unit 349 is used to, in response to the intended supplementary parameters in the json format parameter set being not for displaying data only, select an indicator data display style according to the intended supplementary parameters, generate an indicator data display result based on the indicator data display style, and generate an answer by combining the json format parameter set and the indicator data display result based on the reply template.

[0117] It can be understood that the above-mentioned information processing device for a large language model executes the information processing method for a large language model corresponding to the embodiment 1 provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects of the scheme corresponding to the information processing method for a large language model in the embodiment 1 above, and will not be repeated here.

[0118] Embodiment 4:

[0119] This embodiment also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the information processing method of the large language model in the above-mentioned embodiment 1.

[0120] Embodiment 5:

[0121] This embodiment also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the information processing method of the large language model in the above-mentioned embodiment 1 is implemented.

[0122] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for processing information of a large language model, characterized in that: include: Get user question and web application portal ID; Based on user questions and web application portal identifiers, obtain all intents and tasks in the required scenarios; Determine the target intent and target task based on the full set of intent and tasks in the user's problem and demand scenario; According to the target intention and target task, the indicator data is extracted from the database under the demand scenario, and the answer is generated based on the indicator data.

2. The information processing method of a large language model according to claim 1, characterized in that: The method of obtaining the full amount of intents and tasks in the required scenario according to the user question and the network application portal identifier specifically includes: Extract indicator parameters from user questions, where the indicator parameters are used to represent field names of indicator data; According to the network application portal identifier, obtain all scenarios under the network application portal; Determine the user's demand scenarios based on the indicator parameters and the full range of scenarios under the network application portal; Obtain all intents and tasks based on user demand scenarios.

3. The information processing method of a large language model according to claim 1, characterized in that: Determining the target intent and target task based on the full amount of intent and full amount of tasks in the user problem and demand scenario specifically includes: Based on the semantic focus algorithm and user questions, filter all the intents in the demand scenario and obtain the target intent; According to the preset mapping relationship between intentions and tasks, the target tasks corresponding to the target intentions are matched from all the tasks in the demand scenario.

4. The information processing method of a large language model according to claim 3, characterized in that: Based on the semantic focus algorithm and user questions, the full amount of intent in the demand scenario is screened to obtain the target intent, which specifically includes: Calculate the text similarity between each intent and the user's question in all intents under the demand scenario; Based on text similarity, filter all intents in the required scenario and obtain a list of similar intents; Calculate the semantic similarity between each similar intent in the similar intent list and the user's question; The similar intent with the highest semantic similarity is taken as the target intent.

5. The information processing method of a large language model according to claim 2, characterized in that: According to the target intention and target task, the indicator data is extracted from the database under the demand scenario, specifically including: Obtaining a task description of a target task, wherein the task description is used to characterize an output specification format; According to the target intention and task description, the intention supplement parameters and indicator-related parameters are extracted from the user questions, where the indicator-related parameters are used to characterize the constraints of the indicator parameters; Determine whether the parameters related to the indicator meet the interface specifications corresponding to the task description; In response to the indicator-related parameters complying with the interface specification corresponding to the task description, the task description, the indicator-related parameters, the intent supplementary parameters and the indicator parameters are combined to generate a parameter set in JSON format; Based on the indicator parameters in the JSON format parameter set, extract indicator data from the database in the required scenario.

6. The information processing method of a large language model according to claim 5, characterized in that: After determining whether the indicator-related parameters conform to the interface specification corresponding to the task description, and before extracting the indicator data based on the indicator parameters in the json format parameter set, the method further includes: In response to the indicator-related parameters not complying with the corresponding interface specifications of the task description, the interface positions of the indicator-related parameters in the task description are adjusted to obtain an optimized task description, and the optimized task description, indicator-related parameters, intent supplementary parameters and indicator parameters are combined to generate a parameter set in JSON format.

7. The information processing method of a large language model according to claim 5, characterized in that: The generating of answers based on the indicator data specifically includes: Determine whether the intended supplementary parameter in the json format parameter set is to display data only; In response to the intention to supplement the parameter in the json format parameter set to only display data, based on the reply template, the json format parameter set and the indicator data are combined to generate an answer; In response to the intention to supplement the parameters in the json format parameter set for not just displaying data, select the indicator data display style based on the intention to supplement the parameters, generate the indicator data display results based on the indicator data display style, and generate an answer based on the reply template by combining the json format parameter set and the indicator data display results.

8. An information processing device for a large language model, characterized in that: include: a first acquisition module, a second acquisition module, a determination module and an extraction module, The first acquisition module is used to obtain the user question and the network application portal identifier. The second acquisition module is connected to the first acquisition module and is used to acquire the full amount of intentions and full amount of tasks in the demand scenario according to the user question and the network application portal identifier. The determination module is connected to the second acquisition module and is used to determine the target intention and target task according to the user's questions and the full amount of intentions and full amount of tasks in the demand scenario. The extraction module is connected to the determination module and is used to extract indicator data from the database under the demand scenario according to the target intention and target task, and generate answers based on the indicator data.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement an information processing method for a large language model as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the information processing method of a large language model as described in any one of claims 1 to 7 is implemented.

Images