Question processing method and device, electronic equipment and storage medium

Abstract

The invention provides a problem processing method and device, electronic equipment and a storage medium, and relates to the field of artificial intelligence, in particular to the technical fields of NLP, large models, LLM, deep learning and the like. According to the specific implementation scheme, a first problem matched with a target processing demand is acquired; scheduling a plurality of agents in the distributed system to process the first question to obtain initial answers fed back by the plurality of agents; executing at least one round of feedback process based on the plurality of initial answers to obtain intermediate answers submitted by the plurality of agents in each round of feedback process; and determining a first answer for replying the first question from the intermediate answers submitted by the plurality of agents in each round of feedback process.

Description

Technical Field

[0001] The present disclosure relates to the field of AI (Artificial Intelligence), specifically to technical fields such as NLP (Natural Language Processing), large models, LLM (Large Language Model), deep learning, and especially to problem-solving methods, devices, electronic devices and storage media. Background Art

[0002] With the rapid development of NLP technology and the continuous expansion of its application fields, the demand for efficient and intelligent problem-solving systems is growing in all walks of life. Whether it is industrial automation, medical diagnosis, financial services, or smart homes and mobile devices in daily life, there is an urgent need for systems that can quickly and accurately understand and respond to user questions in order to solve complex and diverse user problems and meet the personalized query needs of different users. Summary of the invention

[0003] The present disclosure provides a method, device, electronic device and storage medium for problem handling.

[0004] According to one aspect of the present disclosure, a problem handling method is provided, comprising:

[0005] Obtaining a first problem that matches the target processing requirement;

[0006] Scheduling multiple agents in the distributed system to process the first problem, and obtaining initial answers fed back by the multiple agents;

[0007] Perform at least one round of feedback process based on the multiple initial answers to obtain intermediate answers submitted by the multiple agents in each round of the feedback process;

[0008] A first answer for responding to the first question is determined from the intermediate answers submitted by the multiple agents in each round of the feedback process.

[0009] According to another aspect of the present disclosure, there is provided a problem handling device, comprising:

[0010] An acquisition module, used for acquiring a first question matching a target processing requirement;

[0011] A scheduling module, used for scheduling multiple agents in the distributed system to process the first problem and obtain initial answers fed back by the multiple agents;

[0012] An execution module, configured to execute at least one round of feedback process based on the plurality of initial answers, so as to obtain intermediate answers submitted by the plurality of agents in each round of the feedback process;

[0013] A determination module is used to determine a first answer for responding to the first question from the intermediate answers submitted by the multiple agents in each round of the feedback process.

[0014] According to another aspect of the present disclosure, there is provided an electronic device, comprising:

[0015] at least one processor; and

[0016] a memory communicatively connected to the at least one processor; wherein,

[0017] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the problem handling method proposed in the above aspect of the present disclosure.

[0018] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium of computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the problem handling method proposed in the above aspect of the present disclosure.

[0019] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, wherein when the computer program is executed by a processor, the problem processing method proposed in the above aspect of the present disclosure is implemented.

[0020] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0022] Figure 1 A flowchart of a problem solving method provided in the first embodiment of the present disclosure;

[0023] Figure 2 A flowchart of a problem solving method provided in Embodiment 2 of the present disclosure;

[0024] Figure 3 A flowchart of a problem solving method provided in Embodiment 3 of the present disclosure;

[0025] Figure 4 A flowchart of a problem solving method provided in Embodiment 4 of the present disclosure;

[0026] Figure 5 A flowchart of a problem handling method provided in Embodiment 5 of the present disclosure;

[0027] Figure 6 A flowchart of a problem handling method provided in Embodiment 6 of the present disclosure;

[0028] Figure 7 This is a schematic diagram of the structure of a problem handling device provided in Embodiment 7 of the present disclosure;

[0029] Figure 8 A schematic block diagram of an example electronic device that can be used to implement embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0030] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0031] When faced with a large number of optional large models in a distributed system, it is very important to select the optimal solution to the problem from the output results of the optional large models in order to solve the problem of selecting the master of multiple large models.

[0032] The large model (agent) itself has strong fault tolerance, answer randomness, and model hallucination. In addition to the prompt itself, the output quality of the large model mainly depends on the following aspects:

[0033] First, model architecture and model parameter scale: A reasonable architecture (such as Transformer) helps capture semantic relationships. The more model parameters, the stronger the expressive power and the better it can process complex texts.

[0034] The second aspect is training data: the accuracy of training data is the basis. Incorrect training data will lead to incorrect output of the large model. The diversity of training data, including multiple dimensions such as text type, field, style, etc., can enable the large model to cope with different needs.

[0035] Third, training methods and optimization strategies: clear training goals are critical. Appropriate optimization algorithms (such as the Adam (adaptive moment estimation) algorithm) and hyperparameter adjustments can improve the performance of large models.

[0036] Fourthly, model fusion and integration technology: Model fusion can combine advantages, and integrated learning can reduce the errors of individual models and improve output quality.

[0037] Fifth, post-processing and evaluation mechanism: Post-processing the output content of the large model can improve grammar and other problems. The evaluation mechanism (a combination of automatic and manual evaluation) can judge the output quality of the large model and be used to improve the large model.

[0038] In related technologies, consensus algorithms are used to solve the leader selection problem in distributed systems. Consensus algorithms mainly include the following:

[0039] 1. Downtime error-tolerant algorithms. Among them, downtime error-tolerant algorithms include: Paxos (a distributed consistency algorithm that is widely used in distributed systems to ensure data consistency) and Raft (a consensus algorithm designed to understand consistency algorithms, which is easier to understand and implement, while providing the same fault tolerance guarantees as Paxos).

[0040] 2. Byzantine fault-tolerant algorithms. Among them, Byzantine fault-tolerant algorithms include: PBFT (Practical Byzantine Fault Tolerance, a consensus algorithm that can tolerate Byzantine faults and is widely used in distributed systems), PoW (Proof of Work, a most common consensus algorithm, widely used in blockchain systems), PoS (Proof of Stake, a consensus algorithm that distributes interest based on the amount and time of digital currency held; compared with PoW, PoS is more energy-efficient in terms of energy consumption), DPoS (Delegate Proof of Stake, an improved algorithm based on POS, which improves the efficiency and scalability of the system by electing a certain number of representatives to participate in the consensus process).

[0041] Based on the usage scenarios, consensus algorithms can be divided into the following categories:

[0042] 1', public chain consensus algorithm. Among them, public chain consensus algorithms include PoW, PoS, etc. These algorithms have been widely used in public blockchain systems to ensure that all nodes in the distributed system reach a consensus on the order and validity of transactions.

[0043] 2'. Alliance chain consensus algorithm. Among them, the alliance chain consensus algorithm includes PBFT, etc. These algorithms have been applied in the alliance blockchain system and are suitable for scenarios that require higher security and reliability.

[0044] In addition, other common consensus algorithms include: PoB (Proof of Burn, a burning proof mechanism, a voting mechanism that burns tokens in one's hands to determine who has a commitment to leading the network; the more tokens burned, the higher the probability of obtaining network leadership), dPoW (Delayed Proof-of-Work, a consensus algorithm that combines PoW and PoS, with the advantages of energy saving and increased security).

[0045] It should be noted that applying the leader selection algorithm (consensus algorithm) of the above distributed system to the leader selection of a large model will at least cause the following problems:

[0046] (1) Data consistency and model output characteristics: The output of the large model is uncertain and has "illusions", which is different from the traditional deterministic node data, making it difficult to determine the leader election criteria; semantic understanding is ambiguous, and different large models have different understandings and answers to the same question, making semantics difficult; large models are constantly updated, and version differences affect the leader election results, while traditional consensus algorithms do not take this dynamic factor into account;

[0047] (2) Algorithm complexity and performance: The master selection process increases computing resource consumption. Large model reasoning itself requires a lot of resources, which may lead to system performance degradation. It also introduces time delays, affecting scenarios with high real-time requirements because the output results of multiple large models need to be processed.

[0048] (3) Scalability and adaptability: It is difficult to adapt to the growth in the number and types of models. When a new model is added, the algorithm needs to be readjusted, which increases maintenance costs. It cannot adapt well to the different requirements of different application scenarios for large model results and lacks customization.

[0049] Therefore, in response to at least one of the above problems, in any embodiment of the present disclosure, the application scenario of the large model (intelligent agent) is: the accuracy requirement for the solution to the problem is relatively high, but the output performance (i.e., processing speed) of the solution is relatively low. In the face of a large number of optional large models, in the present disclosure, the optimal solution to the problem can be determined based on the capabilities of the large model itself and the Delphi Method to solve the problem of selecting the main model for multiple large models. For example, in the 0-1 stage of the R&D project, the system framework can be designed based on the capabilities of the large model. Usually, a long prompt is required to describe the functional items and specific implementation points. Since the initial system framework design determines the efficiency of subsequent iterations, it is often necessary to interact with multiple large models for multiple rounds to select the optimal solution. By extension, in similar critical and complex decision-making scenarios, the response speed requirements for large models are different, the quality requirements for the final decision product are extremely high, and there is a tendency to choose the only and optimal conclusion.

[0050] The problem solving method, device, electronic device and storage medium of the embodiment of the present disclosure are described below with reference to the accompanying drawings. Before describing the embodiment of the present disclosure in detail, for ease of understanding, common technical terms are first introduced:

[0051] A large model refers to a machine learning model with large-scale parameters and complex computational structures. It is usually built from deep neural networks and has billions or even hundreds of billions of parameters. The purpose of designing a large model is to improve the model's expressiveness and predictive performance, and to be able to handle more complex tasks and data. Large models are widely used in various fields, including natural language processing, computer vision, speech recognition, and recommendation systems.

[0052] The LLM in the big model is a type of natural language processing model based on deep learning. Its main features are huge model parameters and complex neural network structure, strong language understanding, context perception and language generation capabilities. It can automatically learn useful feature representations from input data and generate relevant text.

[0053] Figure 1 A flowchart of a problem solving method provided in Embodiment 1 of the present disclosure.

[0054] The embodiment of the present disclosure is illustrated by an example in which the problem handling method is configured in a problem handling device. The problem handling device can be applied to any electronic device so that the electronic device can perform a problem handling function.

[0055] Among them, the electronic device can be any device with computing capabilities, such as a personal computer, a mobile terminal, a server, etc. The mobile terminal can be, for example, a mobile phone, a tablet computer, a personal digital assistant, a wearable device, and other hardware devices with various operating systems, touch screens and / or display screens.

[0056] like Figure 1 As shown, the problem handling method may include the following steps S101 to S104:

[0057] Step S101, obtaining a first question matching a target processing requirement.

[0058] Among them, the first question is the question to be processed, and the number of the first question can be one or more, which is not limited in the embodiment of the present disclosure.

[0059] Among them, the target processing requirement refers to the pre-set problem processing requirement. In any embodiment of the present disclosure, the target processing requirement may include the timeliness requirement for problem processing (referred to as the processing timeliness requirement in the present disclosure) and / or the accuracy requirement for problem processing (referred to as the processing accuracy requirement in the present disclosure); wherein, the processing timeliness requirement includes that the problem processing time limit is not less than the set time limit threshold, that is, the processing timeliness requirement is relatively low, and the processing accuracy requirement includes that the answer accuracy is higher than the set accuracy threshold, that is, the processing accuracy requirement is relatively high.

[0060] Therefore, it is not the case that for all problems, the method of selecting a master from multiple large models in a distributed system is used to decide the final solution to the problem. Instead, only for the first problem that matches the target processing requirement, the method of selecting a master from multiple large models in a distributed system is used to decide the final solution to the problem. This can take into account the processing timeliness requirements and processing accuracy requirements of different problems and meet the personalized application needs in different application scenarios.

[0061] There is no restriction on the method for obtaining the first question. For example, the first question may be a question input by a user, or the first question may be a question collected online, or the first question may be a question automatically generated by a generative model, or the first question may be a question obtained from a training set or a test set, etc. The embodiments of the present disclosure do not impose any restrictions on this.

[0062] Among them, the input method of the first question may include but is not limited to touch input (such as sliding, clicking, etc.), keyboard input, voice input, etc.

[0063] Step S102, scheduling multiple agents in the distributed system to process the first question and obtain initial answers fed back by the multiple agents.

[0064] The initial answer includes, but is not limited to, the answer to the question and the corresponding detailed reasons, basis, etc.

[0065] In the disclosed embodiment, multiple agents in a distributed system may be scheduled to process the first question, and initial answers fed back by the multiple agents may be obtained.

[0066] As an example, a POST (submit) request containing a first question can be sent to the HTTP interface of multiple agents through the HTTP (Hypertext Transfer Protocol) interface addresses of multiple agents, and initial answers returned by multiple agents can be obtained through the HTTP interface of multiple agents.

[0067] Exemplarily, taking the intelligent agent as the big model, the get_model_response function can be used to send a POST request containing the first question to the HTTP interface of multiple big models through the requests library of Python (a programming language), and obtain the initial answers returned by multiple big models (such as answers in JSON (JavaScript Object Notation, a lightweight data exchange format) format).

[0068] Among them, the parameters involved in the get_model_response function include: URL (uniform resource locator) address: the HTTP interface address of the big model; question: the first question entered by the user, sent to the big model in the question field in JSON format; return value: the initial answer in JSON format returned by the HTTP interface of the big model. If the request fails, None (null value) is returned.

[0069] Step S103, performing at least one round of feedback process based on the multiple initial answers to obtain intermediate answers submitted by the multiple agents in each round of feedback process.

[0070] In the disclosed embodiment, at least one round of feedback process may be performed based on the initial answers fed back by the multiple agents to obtain intermediate answers submitted by the multiple agents in each round of feedback process.

[0071] Exemplarily, the first-round feedback process may include: scheduling each agent to refer to multiple initial answers, adjusting the initial answer fed back by the agent, and obtaining an intermediate answer submitted by the agent in the first-round feedback process.

[0072] Exemplarily, the i-th (i is a positive integer greater than 1) round of feedback process may include: scheduling each agent to refer to the intermediate answers submitted by multiple agents in the i-1 round of feedback process, adjusting the intermediate answers submitted by the agent in the i-1 round of feedback process, and obtaining the intermediate answers submitted by the agent in the i-1 round of feedback process.

[0073] Step S104, determining a first answer for responding to the first question from the intermediate answers submitted by multiple agents in each round of feedback process.

[0074] In the embodiment of the present disclosure, a best answer can be selected from the intermediate answers submitted by multiple agents in each round of feedback process as the final answer to the first question, which is recorded as the first answer in the present disclosure.

[0075] In the problem-solving method of the disclosed embodiment, each intelligent agent has a certain degree of autonomy and intelligence, which enables the system to flexibly adapt to complex and changeable problems, select the most appropriate processing strategy and method according to the characteristics and requirements of the problem, thereby improving the efficiency and effect of problem processing, and, based on the initial answers to the problem feedback from multiple intelligent agents, at least one round of feedback process is performed, and the intelligent agents can learn from each other, draw lessons from each other and make corrections, and continuously optimize their respective answers. This feedback mechanism helps to share information and knowledge between intelligent agents, thereby improving the accuracy and reliability of the final answer generation. In addition, the intelligent agents in the distributed system can collaborate and interact with each other. Even if an intelligent agent fails or makes an error, other intelligent agents can still continue to work and provide effective answers. This fault-tolerant capability makes the system more robust and reliable, able to cope with various emergencies and challenges, and ensure the continuity and stability of problem processing.

[0076] It should be noted that in the technical solution of the present disclosure, the collection, storage, use, processing, transmission, provision and disclosure of user personal information are all carried out with the user's consent, and are in compliance with relevant laws and regulations and do not violate public order and good morals.

[0077] In order to clearly illustrate how to schedule multiple agents in a distributed system to process a first problem and obtain initial answers fed back by multiple agents in any embodiment of the present disclosure, the present disclosure also proposes a problem processing method.

[0078] Figure 2 A flowchart of a problem solving method provided in Embodiment 2 of the present disclosure.

[0079] like Figure 2 As shown, the problem handling method may include the following steps S201 to S206:

[0080] Step S201, obtaining a first question matching a target processing requirement and context information associated with the first question.

[0081] It should be noted that the explanations of the target processing requirements and the first question in the aforementioned embodiment are also applicable to this embodiment and will not be repeated here.

[0082] The context information includes, but is not limited to, detailed information such as background information, restrictions, relevant data and research materials of the first question.

[0083] For example, the first question is “Which new energy utilization solution is the best in terms of cost, environmental protection and sustainability?” The context information includes but is not limited to the following items:

[0084] 1. Background information: Known new energy types: solar energy, wind energy, biomass energy, geothermal energy, tidal energy, etc.; Current energy market trends: The global demand for renewable energy is growing, and all countries support the development of new energy.

[0085] 2. Restrictions: Cost requirements: The cost of new energy solutions must be within a reasonable range and be acceptable to the market; Environmental requirements: New energy solutions must have lower carbon emissions and environmental impacts; Sustainability requirements: New energy solutions must be able to be supplied continuously in the long term and not rely on scarce resources.

[0086] 3. Existing attempts and achievements: Existing research and practice: Currently, a variety of new energy technologies have achieved certain success in practical applications, but there are still problems such as high cost and immature technology; Case study: Some countries and regions have made breakthroughs in specific fields, such as wind power generation in country D and photovoltaic industry in country C.

[0087] 4. Relevant data and research materials: Cost data: cost data of different new energy sources, including initial investment, operation and maintenance costs, etc.; Environmental indicators: environmental indicators such as carbon emissions and pollutant emissions of different new energy sources; Sustainable development research: research on the sustainability of various new energy sources, including resource availability, technology maturity, policy support, etc.

[0088] Step S202: Generate a scheduling request according to the first question and context information.

[0089] In the embodiment of the present disclosure, a scheduling request may be generated according to the first question and the context information, wherein the scheduling request carries the first question and the context information. Taking an agent as a large model for example, the scheduling request may be a POST request.

[0090] Step S203, sending the scheduling request to multiple agents in the distributed system, so that the multiple agents process the first question based on the context information to obtain an initial answer.

[0091] In an embodiment of the present disclosure, a scheduling request may be sent to multiple agents in a distributed system so that the multiple agents can process the first question based on context information in the scheduling request to obtain an initial answer.

[0092] As an example, a scheduling request containing a first question and context information can be sent to the HTTP interfaces of multiple agents through the HTTP interface addresses of multiple agents, so that the multiple agents process the first question based on the context information in the scheduling request and obtain an initial answer.

[0093] Exemplarily, taking the intelligent agent as the big model, the get_model_response function can be used to send a scheduling request (such as a POST request) containing the first question and context information to the HTTP interfaces of multiple big models through Python's requests library, and obtain the initial answers (such as answers in JSON format) returned by multiple big models through the HTTP interfaces of multiple big models.

[0094] Among them, the parameters involved in the get_model_response function include: URL address: HTTP interface address of the big model; question: the first question and context information entered by the user, sent to the big model in the question field in JSON format; return value: the initial answer in JSON format returned by the HTTP interface of the big model. If the request fails, None (null value) is returned.

[0095] Step S204, receiving initial answers fed back by multiple agents.

[0096] As an example, initial answers returned by multiple agents can be received through the HTTP interfaces of multiple agents.

[0097] Step S205, performing at least one round of feedback process based on the multiple initial answers to obtain intermediate answers submitted by the multiple agents in each round of feedback process.

[0098] Step S206, determining a first answer for responding to the first question from the intermediate answers submitted by the multiple agents in each round of feedback process.

[0099] The explanation of step S205 to step S206 can refer to the relevant description in any embodiment of the present disclosure, and will not be repeated here.

[0100] The problem handling method of the embodiment of the present disclosure provides the first question and context information associated with the first question to the intelligent agent at the same time. The context information can help the intelligent agent better understand the specific background and details of the first question, thereby more accurately parsing the intention of the question and avoiding misunderstanding or ambiguity. By providing relevant background information, the intelligent agent can reduce erroneous judgments caused by lack of necessary information, thereby improving the accuracy of the intelligent agent's output.

[0101] In order to clearly illustrate the above embodiments of the present disclosure, the present disclosure also proposes a problem solving method.

[0102] Figure 3 This is a flowchart of the problem handling method provided in the third embodiment of the present disclosure.

[0103] like Figure 3As shown, the problem handling method may include the following steps S301 to S305:

[0104] Step S301, obtaining a first question matching a target processing requirement.

[0105] Step S302: Schedule multiple agents in the distributed system to process the first question and obtain initial answers fed back by the multiple agents.

[0106] For explanations of steps S301 to S302, reference may be made to the relevant descriptions in any embodiment of the present disclosure and will not be repeated here.

[0107] Step S303, performing at least one round of feedback process based on the multiple initial answers to obtain intermediate answers submitted by the multiple agents in each round of feedback process.

[0108] In any embodiment of the present disclosure, the first round of feedback process in at least one round of feedback process may include the following steps A to E:

[0109] Step A: Based on NLP technology, extract key response information from multiple initial answers, which is recorded as the first key response information in this disclosure.

[0110] Step B: Schedule multiple agents to refer to the first key reply information, adjust the initial answers fed back by the corresponding agents, and obtain the intermediate answers submitted by multiple agents in the first round of feedback process.

[0111] That is, for any agent, the agent can be scheduled to refer to the first key reply information, adjust the initial answer fed back by the agent, and obtain the intermediate answer submitted by the agent in the first round of feedback process.

[0112] Step C: Determine whether the intermediate answers submitted by multiple agents in the first round of feedback process have converged. If so, execute step D; if not, execute step E.

[0113] It should be noted that step D and step E are two parallel implementation methods. In actual application, only one needs to be executed.

[0114] As a first possible implementation method, the following steps C_1 to C_4 can be used to determine whether the intermediate answer submitted by any agent in the first round of feedback process converges:

[0115] Step C_1: For any one of the multiple agents, calculate the content similarity between the intermediate answer submitted by the agent in the first round of feedback process and the initial answer fed back by the agent, which is recorded as the first content similarity in this disclosure.

[0116] As an example, first, the intermediate answer submitted by the agent in the first round of feedback process can be converted into a text vector A through the bag-of-words model, TF-IDF (term frequency-inverse document frequency), etc., and the initial answer fed back by the agent can be converted into a text vector B. After that, the cosine similarity algorithm can be used to calculate the cosine similarity between the text vector A and the text vector B, so that the cosine similarity can be used as the first content similarity.

[0117] Among them, the value of cosine similarity is between -1 and 1. The closer the value of cosine similarity is to 1, the more similar the two answers are. That is, the higher the cosine similarity, the more similar the detailed text expressions of the two answers are.

[0118] As another example, the intermediate answer submitted by the agent in the first round of feedback process can be converted into word set A (ignoring word frequency and only focusing on whether the word appears), and the initial answer fed back by the agent can be converted into word set B. After that, the Jaccard similarity algorithm can be used to calculate the Jaccard similarity between word set A and word set B, so that the Jaccard similarity can be used as the first content similarity.

[0119] For example, when the initial answer or intermediate answer fed back by the agent mainly involves some key technical points (which can be regarded as a set of words), the Jaccard similarity can be used to determine the degree of overlap of these key technical points in different answers, and then determine whether the answers converge.

[0120] As another example, the edit distance between the intermediate answer submitted by the agent in the first round of feedback and the initial answer fed back by the agent can be calculated based on the edit distance (also known as Levenshtein distance) algorithm, so that the first content similarity can be determined based on the edit distance.

[0121] Among them, the edit distance algorithm is used to calculate the minimum number of editing operations (insertion, deletion, replacement of characters) required to convert one string into another string; the first content similarity is negatively correlated with the edit distance, that is, the smaller the edit distance, the greater the first content similarity, and vice versa, the larger the edit distance, the smaller the first content similarity.

[0122] That is, when judging whether the answer is converging, if the edit distance between answers in different rounds is small, it means that the answer has not changed much at the text level and may be converging. For example, after multiple rounds of feedback, the edit distance of the agent's answer to a question remains at a low level, such as changing from "adopting A technology and B strategy" to "adopting A technology and B strategy", then the edit distance is small, indicating that the answer is gradually converging.

[0123] Step C_2: Determine whether the first content similarity is higher than a set first similarity threshold, if so, execute step C_3, if not, execute step C_4.

[0124] It should be noted that step C_3 and step C_4 are two parallel implementation methods. In actual application, only one needs to be executed.

[0125] Step C_3: When the first content similarity is higher than the first similarity threshold, determine that the intermediate answer submitted by the agent in the first round of feedback process converges.

[0126] Step C_4: When the first content similarity is not higher than the first similarity threshold, it is determined that the intermediate answer submitted by the agent in the first round of feedback process has not converged.

[0127] In summary, content similarity provides a clear quantitative indicator to measure whether the output results of the agent in different rounds tend to be consistent. When the content similarity between the output results of the agent in two adjacent rounds reaches a preset threshold, the output results of the agent are considered to have converged, and the feedback process is stopped. This clear judgment standard helps to reduce the uncertainty caused by subjective judgment, thereby improving the accuracy and consistency of judgment.

[0128] As a second possible implementation, the following steps C_5 to C_7 can be used to determine whether the intermediate answer submitted by any agent in the first round of feedback process converges:

[0129] Step C_5, for any one of the multiple agents, use a semantic classification model to classify the text pair to obtain a classification result; wherein the classification result is used to indicate whether the text pair is semantically equivalent, and the text pair includes the intermediate answer submitted by the agent in the first round of feedback process and the initial answer fed back by the agent.

[0130] As an example, a pre-trained semantic classification model such as BERT (Bidirectional Encoder Representations from Transformers) can be used to classify text pairs to obtain a classification result indicating whether the text pairs are semantically equivalent.

[0131] Exemplarily, when the output probability of the semantic classification model is greater than the probability threshold, the classification result is used to indicate that the text pair is semantically equivalent, and when the output probability of the semantic classification model is less than or equal to the probability threshold, the classification result is used to indicate that the text pair is semantically unequal.

[0132] Among them, the training method of the semantic classification model is, for example: constructing a binary classification data set, wherein the data set contains answer text pairs from different rounds, and the annotated labels of the answer text pairs are 0 (used to indicate that the answer text pairs are semantically equivalent) and 1 (used to indicate that the answer text pairs are semantically equivalent). After that, the answer text pairs from different rounds can be used to train the semantic classification model, so that the trained semantic classification model can be used to determine whether new text pairs are semantically equivalent.

[0133] Step C_6: When the classification result indicates that the text pair is semantically equivalent, determine that the intermediate answers submitted by the agent in the first round of feedback process converge.

[0134] Step C_7: When the classification result indicates that the text pair is not semantically equivalent, it is determined that the intermediate answer submitted by the agent in the first round of feedback process has not converged.

[0135] It should be noted that the Sentence-BERT (SBERT) model can also be used to predict the similarity scores between text pairs (including the intermediate answers submitted by the agent in the first round of feedback and the initial answers fed back by the agent). When the similarity scores are relatively high, it is determined that the intermediate answers submitted by the agent in the first round of feedback are converged, and when the similarity scores are relatively low, it is determined that the intermediate answers submitted by the agent in the first round of feedback are not converged.

[0136] In summary, the semantic classification model can capture the deep connections between text pairs through in-depth understanding and analysis of text pairs, thereby more accurately judging whether the output results of the intelligent agent have converged. The accuracy of this judgment is crucial to ensuring the effectiveness and reliability of the intelligent agent's execution feedback process.

[0137] As a third possible implementation method, the following steps C_8 to C_12 can be used to determine whether the intermediate answer submitted by any agent in the first round of feedback process converges:

[0138] Step C_8: For any one of the multiple agents, annotate the initial answer fed back by the agent to obtain a first annotation result; wherein the first annotation result is used to indicate the semantic roles in the initial answer and the relationship between the semantic roles.

[0139] As an example, SRL (Semantic Role Labeling) may be performed on the predicate-argument structure in the initial answer to obtain a first labeling result.

[0140] Step C_9: Annotate the intermediate answer submitted by the agent in the first round of feedback process to obtain a second annotation result; wherein the second annotation result is used to indicate the semantic roles in the intermediate answer and the relationship between the semantic roles.

[0141] Step C_10: Calculate the semantic similarity between the first annotation result and the second annotation result, which is referred to as the first semantic similarity in the present disclosure, and determine whether the first semantic similarity is higher than the set second similarity threshold. If so, execute step C_11; if not, execute step C_12.

[0142] It should be noted that the embodiments of the present disclosure do not limit the size relationship between the first similarity threshold and the second similarity threshold. For example, the first similarity threshold may be greater than the second similarity threshold, or the first similarity threshold may be less than the second similarity threshold, or the first similarity threshold may be equal to the second similarity threshold.

[0143] It should also be noted that step C_11 and step C_12 are two parallel implementation methods. In actual application, only one needs to be executed.

[0144] Step C_11: When the first semantic similarity is higher than the second similarity threshold, determine that the intermediate answer submitted by the agent in the first round of feedback process converges.

[0145] Step C_12: When the first semantic similarity is not higher than the second similarity threshold, it is determined that the intermediate answer submitted by the agent in the first round of feedback process has not converged.

[0146] For example, when the agent answers questions about the event processing flow, semantic role labeling is used to analyze each step (semantic role) and the order (relationship) between them. If these steps and the relationships between the steps remain consistent in multiple rounds of answers, it means that the agent's response is converging.

[0147] In summary, semantic role labeling can reveal the predicate-argument structure in a sentence, clearly mark the semantic roles in a sentence and the relationship between these semantic roles. In this way, the semantic content of the output results of the agent in different rounds can be more accurately understood, so as to more accurately judge whether the output results of the agent converge. Compared with methods that rely solely on literal matching or keyword comparison, semantic role labeling can capture deeper semantic information and improve the accuracy and reliability of convergence judgment.

[0148] As a fourth possible implementation method, the following steps C_13 to C_17 can be used to determine whether the intermediate answer submitted by any agent in the first round of feedback process converges:

[0149] Step C_13: For any one of the multiple agents, obtain the first directed graph corresponding to the initial answer fed back by the agent; wherein the nodes in the first directed graph are used to indicate the semantic roles in the initial answer, and the edges between the nodes are used to indicate the dependency relationship and / or sequence between the semantic roles.

[0150] As an example, for answers that contain logical content such as steps and causal relationships, a logical diagram can be constructed or the logical process can be parsed to obtain a directed graph corresponding to the answer. For example, when the agent answers a question about a project implementation plan, each step in the answer can be regarded as a node in a directed graph, and the sequence and dependency between the steps can be regarded as edges to construct a directed graph.

[0151] Step C_14: Obtain a second directed graph corresponding to the intermediate answer submitted by the agent in the first round of feedback process; wherein the nodes in the second directed graph are used to indicate the semantic roles in the intermediate answer, and the edges between the nodes are used to indicate the dependency relationship and / or sequence between the semantic roles.

[0152] Step C_15: Calculate the semantic similarity between the first directed graph and the second directed graph, which is referred to as the second semantic similarity in the present disclosure, and determine whether the second semantic similarity is higher than a set third similarity threshold. If so, execute step C_16; if not, execute step C_17.

[0153] It should be noted that the embodiment of the present disclosure does not limit the magnitude relationship among the first similarity threshold, the second similarity threshold, and the third similarity threshold.

[0154] It should also be noted that step C_16 and step C_17 are two parallel implementation methods. In actual application, only one needs to be executed.

[0155] Step C_16: When the second semantic similarity is higher than the third similarity threshold, determine that the intermediate answer submitted by the agent in the first round of feedback process converges.

[0156] Step C_17: When the second semantic similarity is not higher than the third similarity threshold, it is determined that the intermediate answer submitted by any agent in the first round of feedback process has not converged.

[0157] In summary, directed graphs can intuitively display the semantic roles in the output results of the agent and their dependencies and order. This graphical representation makes complex semantic relationships clear and easy to understand, and facilitates quick understanding of the response structure and content of the agent. Constructing directed graphs of the output results of the agent in different rounds and calculating the semantic similarity between directed graphs can accurately capture the semantic changes in the output results of the agent in different rounds. The semantic changes include not only changes in vocabulary and sentences, but also changes in semantic roles and dependencies. This precise capture method makes it more accurate and reliable to judge whether the output results of the agent have converged.

[0158] As a fifth possible implementation, the following steps C_18 to C_22 can be used to determine whether the intermediate answer submitted by any agent in the first round of feedback process converges:

[0159] Step C_18: For any one of the multiple agents, extract the first key element from the initial answer fed back by the agent.

[0160] Among them, the first key element includes but is not limited to: key technical terms, core ideas, entities, concepts, etc.

[0161] As an example, NLP technology (such as part-of-speech tagging, named entity recognition, keyword extraction algorithm, etc.) can be used to extract the first key element from the initial answer.

[0162] For example, when the intelligent agent answers questions about the application of scientific and technological innovation, NLP technology can be used to extract key elements in the question, such as the name of the key technology, application scenarios, and advantages. Subsequently, the consistency and overlap of the key elements in the answers in different rounds can be compared to determine whether the intelligent agent's responses have converged.

[0163] For example, when the intelligent agent answers questions about the industry's competitive landscape, NLP technology can be used to extract key elements such as entities (such as company names, product names, etc.) and concepts (such as theoretical models, business models, market concepts, etc.) in the question. Subsequently, the key elements in the answers of different rounds can be compared to determine whether they are the same to determine whether the intelligent agent's responses have converged.

[0164] Step C_19: Extract the second key element from the intermediate answer submitted by the agent in the first round of feedback process.

[0165] Step C_20: Calculate the content similarity (or semantic similarity) between the first key element and the second key element, which is referred to as the second content similarity in this disclosure, and determine whether the second content similarity is higher than the set fourth similarity threshold. If so, execute step C_21; if not, execute step C_22.

[0166] It should be noted that the embodiment of the present disclosure does not limit the magnitude relationship among the first similarity threshold, the second similarity threshold, the third similarity threshold, and the fourth similarity threshold.

[0167] It should also be noted that step C_21 and step C_22 are two parallel implementation methods. In actual application, only one needs to be executed.

[0168] Step C_21: When the second content similarity is higher than the fourth similarity threshold, determine that the intermediate answer submitted by the agent in the first round of feedback process converges.

[0169] Step C_22: When the second content similarity is not higher than the fourth similarity threshold, it is determined that the intermediate answer submitted by the agent in the first round of feedback process has not converged.

[0170] In summary, by extracting the key elements in the output results of the agent in different rounds and calculating the content similarity between them, we can quickly determine whether the agent's reply is convergent. This method avoids complex semantic analysis of the entire output result, thereby improving the judgment efficiency.

[0171] It should be noted that the above-mentioned method of judging whether the response or feedback of the intelligent agent converges is only an example. In actual application, other algorithms can also be used to judge whether the response of the intelligent agent converges. For example, a logical consistency check based on the rules can be used to judge whether the response of the intelligent agent converges. The embodiments of the present disclosure do not limit this.

[0172] Among them, rule-based logical consistency check: define some logical rules, such as whether the causal relationship is reasonable, whether the conditional judgment is complete, etc. By checking whether the answers in different rounds conform to these logical rules and whether the application of logical rules is consistent, it is judged whether the agent's response converges.

[0173] For example, when the agent answers questions about fault diagnosis, it checks whether the causal relationship between the cause and symptom of the fault in each answer conforms to the established professional rules, and whether the application of these logical rules is stable in multiple rounds of answers. If so, it is determined that the agent's reply has converged; if not, it is determined that the agent's reply has not converged.

[0174] Step D: If all converge, the feedback process ends.

[0175] Step E: If there is at least one agent whose intermediate answer submitted in the first round of feedback process has not converged, the next round of feedback process is executed.

[0176] In summary, each agent refers to the key response information in all initial answers to adjust the initial answer fed back by the corresponding agent, which can improve the rationality and reliability of the intermediate answers submitted by each agent in the first round of feedback, thereby improving the accuracy of the final answer generation.

[0177] In any embodiment of the present disclosure, at least one feedback process that is not the first feedback process (such as the i-th feedback process, where i is a positive integer greater than 1) may include the following steps F to J:

[0178] Step F: Based on NLP technology, extract the second key response information from the intermediate answers submitted by multiple agents in the previous round of feedback process.

[0179] Step G: Schedule multiple agents to refer to the second key reply information, adjust the intermediate answers submitted by the corresponding agents in the i-1th round of feedback process, and obtain the intermediate answers submitted by multiple agents in the i-th round of feedback process.

[0180] That is, for any agent, the agent can be scheduled to refer to the second key reply information, adjust the intermediate answer submitted by the agent in the i-1th round of feedback process, and obtain the intermediate answer submitted by the agent in the i-th round of feedback process.

[0181] Step H: Determine whether the intermediate answers submitted by multiple agents in the i-th round of feedback process have all converged, that is, determine whether the intermediate answers submitted by each agent in the i-th round of feedback process are similar to the intermediate answers submitted in the i-1th round (semantically similar or textually similar). If so, it is determined that the intermediate answers submitted by the agent in the i-th round of feedback process have converged; if not, it is determined that the intermediate answers submitted by the agent in the i-th round of feedback process have not converged.

[0182] In the present disclosure, if the intermediate answers submitted by multiple agents in the i-th round of feedback process are all converged, step I is executed; if there is at least one agent whose intermediate answer submitted in the i-th round of feedback process has not converged, step J is executed.

[0183] It should be noted that the implementation principle of step H is similar to that of step C, which will not be described in detail here.

[0184] It should also be noted that step I and step J are two parallel implementation methods. In actual application, only one needs to be executed.

[0185] Step I: If all converge, the feedback process ends.

[0186] Step J: If there is at least one agent whose intermediate answer submitted in the i-th round of feedback process has not converged, then determine whether i has reached the set round. If so, end the feedback process; if not, execute the next round of feedback process.

[0187] The set rounds are the preset maximum execution rounds. Exemplarily, the set rounds may be 3 rounds.

[0188] In summary, each agent refers to the key response information in the intermediate answers submitted by all agents in the previous round of feedback to adjust the intermediate answers submitted by the corresponding agent in the previous round of feedback, and obtains the intermediate answers submitted by each agent in the feedback process. This can improve the rationality and reliability of the intermediate answers submitted by each agent in each round of feedback, thereby improving the accuracy of the final answer generation.

[0189] Step S304, counting the number of occurrences of each intermediate answer.

[0190] As an example, similar intermediate answers can be determined from the intermediate answers based on the similarities between the intermediate answers (including semantic similarity, text similarity, and content similarity), and the similar intermediate answers can be regarded as the same intermediate answer, so that the number of occurrences of the same intermediate answer can be counted.

[0191] Step S305, determining a first answer from the intermediate answers according to the number of occurrences of each intermediate answer, and replying to the first question based on the first answer.

[0192] As an example, the middle answer with the highest number of occurrences may be used as the first answer, so that the first question may be replied to based on the first answer.

[0193] The question handling method of the embodiment of the present disclosure selects a first answer for replying to the first question from the intermediate answers submitted in each round of feedback process based on the number of occurrences of the intermediate answers submitted in each round of feedback process, thereby improving the reliability and rationality of the question reply.

[0194] In order to clearly illustrate how any embodiment of the present disclosure determines a first answer for responding to a first question from intermediate answers submitted by multiple agents in each round of feedback process, the present disclosure also proposes a question processing method.

[0195] Figure 4 This is a flowchart of the problem handling method provided in the fourth embodiment of the present disclosure.

[0196] like Figure 4 As shown, the problem handling method may include the following steps S401 to S406:

[0197] Step S401, obtaining a first question matching a target processing requirement.

[0198] Step S402: Schedule multiple agents in the distributed system to process the first question and obtain initial answers fed back by the multiple agents.

[0199] Step S403, performing at least one round of feedback process based on the multiple initial answers to obtain intermediate answers submitted by the multiple agents in each round of feedback process.

[0200] For explanations of steps S401 to S403, reference may be made to the relevant descriptions in any embodiment of the present disclosure, and will not be repeated here.

[0201] In any embodiment of the present disclosure, in order to improve the security of data, the data may be encrypted during the data transmission process.

[0202] In any embodiment of the present disclosure, the initial answers fed back by multiple agents and the intermediate answers submitted by multiple agents in each round of feedback process may also be stored. Exemplarily, the initial answers fed back by multiple agents and the intermediate answers submitted by multiple agents in each round of feedback process may be encrypted and stored in a database.

[0203] Step S404, evaluating each intermediate answer to obtain a score of each intermediate answer under multiple evaluation dimensions.

[0204] Among them, the evaluation dimensions include but are not limited to: feasibility, innovation, cost and other dimensions.

[0205] As an example, LLM can be called to evaluate each intermediate answer to obtain the score of each intermediate answer under multiple evaluation dimensions.

[0206] As another example, a pre-trained evaluation model may be called to evaluate each intermediate answer to obtain a score for each intermediate answer under multiple evaluation dimensions.

[0207] Step S405, determining a first score for each intermediate answer according to the scores of each intermediate answer under multiple evaluation dimensions.

[0208] As an example, based on the weights of the multiple evaluation dimensions, the scores of each intermediate answer under the multiple evaluation dimensions may be weighted and summed to obtain the first score of each intermediate answer.

[0209] As another example, the average or cumulative value of the scores of each intermediate answer under multiple evaluation dimensions may be used as the first score of each intermediate answer.

[0210] Step S406: Determine a first answer from the intermediate answers based on the first scores of the intermediate answers, and reply to the first question based on the first answer.

[0211] As an example, the intermediate answer with the highest first score may be used as the first answer for responding to the first question.

[0212] The question processing method of the embodiment of the present disclosure determines a first answer for responding to a first question from among the intermediate answers based on the scores of the intermediate answers under multiple evaluation dimensions, which can improve the rationality and accuracy of the response to the question.

[0213] In order to clearly illustrate how any embodiment of the present disclosure determines a first answer for responding to a first question from intermediate answers submitted by multiple agents in each round of feedback process, the present disclosure also proposes a question processing method.

[0214] Figure 5 This is a flowchart of the problem handling method provided in Embodiment 5 of the present disclosure.

[0215] like Figure 5 As shown, the problem handling method may include the following steps S501 to S506:

[0216] Step S501, obtaining a first question matching a target processing requirement.

[0217] Step S502: Schedule multiple agents in the distributed system to process the first question and obtain initial answers fed back by the multiple agents.

[0218] Step S503, performing at least one round of feedback process based on the multiple initial answers to obtain intermediate answers submitted by the multiple agents in each round of feedback process.

[0219] For explanations of steps S501 to S503, reference may be made to the relevant descriptions in any embodiment of the present disclosure, and will not be repeated here.

[0220] In any embodiment of the present disclosure, the number of rounds N of the feedback process is set, and the first round of the feedback process in the N rounds of feedback process may include the following steps a to c:

[0221] Step a: Classify multiple initial answers to obtain initial answers under multiple classification dimensions.

[0222] As an example, multiple initial answers can be classified and summarized according to different classification dimensions (such as the theoretical category on which the answer is based, the general direction of the answer or solution, etc.), so as to facilitate comparison, evaluation and analysis of initial answers of different classification dimensions.

[0223] Taking the example in step S201 as an example, multiple initial answers can be classified according to different energy types such as solar energy, wind energy, and hydropower to obtain initial answers for each energy type.

[0224] Step b: Schedule multiple agents to refer to the initial answers under multiple classification dimensions, adjust the initial answers fed back by the corresponding agents, and obtain the intermediate answers submitted by multiple agents in the first round of feedback process.

[0225] That is, for any agent, the agent can be scheduled to refer to the initial answers under multiple classification dimensions, adjust the initial answers fed back by the agent, and obtain the intermediate answers submitted by the agent in the i-th round of feedback process.

[0226] Step c: Schedule multiple agents to compare and evaluate the initial answers under multiple classification dimensions, and obtain the evaluation scores of the initial answers under multiple classification dimensions submitted by multiple agents in the first round of feedback process.

[0227] That is, for any agent, the agent can be scheduled to compare and evaluate the initial answers under multiple classification dimensions, and obtain the evaluation scores of the initial answers under multiple classification dimensions submitted by the agent in the first round of feedback process.

[0228] That is to say, each agent can not only modify the answer of the previous reply, but also compare and score the answers of the previous replies of each agent to obtain an evaluation score for each answer.

[0229] In summary, each agent refers to the initial answers under different classification dimensions to adjust the initial answers fed back by the corresponding agent, which can improve the rationality and reliability of the intermediate answers submitted by each agent in the first round of feedback, thereby improving the accuracy of the final answer generation. In addition, each agent can also compare and evaluate the initial answers under different classification dimensions to obtain the evaluation score of each initial answer, so that the first answer for responding to the first question can be selected based on the evaluation score of each answer, which can improve the reliability and accuracy of the response to the question.

[0230] In any embodiment of the present disclosure, the number of rounds N of the feedback process is set, and the non-first round of feedback process in the N rounds of feedback process (such as the i-th round of feedback process, where i is a positive integer greater than 1) may include the following steps e to g:

[0231] Step e: Classify the intermediate answers submitted by multiple agents in the i-1 round feedback process to obtain intermediate answers under multiple classification dimensions.

[0232] It should be noted that the implementation principle of step e is similar to that of step a, and will not be repeated here.

[0233] Step f: Schedule multiple agents to refer to the intermediate answers under multiple classification dimensions, adjust the intermediate answers submitted by the corresponding agents in the i-1th round of feedback process, and obtain the intermediate answers submitted by multiple agents in the i-th round of feedback process.

[0234] It should be noted that the implementation principle of step f is similar to that of step b, which will not be elaborated here.

[0235] Step g: Schedule multiple agents to compare and evaluate the intermediate answers under multiple classification dimensions, and obtain the evaluation scores of the intermediate answers under multiple classification dimensions submitted by multiple agents in the i-th round of feedback process.

[0236] It should be noted that the implementation principles of step g and step c are similar and will not be elaborated here.

[0237] In summary, each agent refers to the intermediate answers under different classification dimensions submitted by all agents in the previous round of feedback process to adjust the intermediate answers submitted by the corresponding agent in the previous round of feedback process, which can improve the rationality and reliability of the intermediate answers submitted by each agent in each round of feedback process, thereby improving the accuracy of the final answer generation. In addition, each agent can also compare and evaluate the intermediate answers under different classification dimensions submitted in the previous round of feedback process to obtain the evaluation score of each intermediate answer, so that the first answer for responding to the first question can be screened out based on the evaluation score of each answer, which can further improve the reliability and accuracy of the question response.

[0238] Step S504, determining the second scores of the multiple agents based on the evaluation scores of the intermediate answers submitted by the multiple agents in the N-round feedback process and the evaluation scores of the initial answers.

[0239] The second score is used to characterize the professionalism, accuracy and stability of the intelligent agent in handling the first problem.

[0240] In the disclosed embodiment, for any one of the multiple agents, the second score of the agent can be calculated according to the evaluation score of the intermediate answer submitted by the agent in the N-round feedback process and the evaluation score of the initial answer. The second score is positively correlated with the evaluation score of each answer.

[0241] In any embodiment of the present disclosure, the following steps X to Z may be used to calculate the second score of each agent:

[0242] Step X: For any one of the multiple agents, the first sub-score of the agent is calculated according to the evaluation score of the intermediate answer submitted by the agent in the N-round feedback process and the evaluation score of the initial answer.

[0243] As an example, the evaluation scores of the intermediate answers submitted by the agent in the N-round feedback process and the evaluation score of the initial answer can be weightedly summed to obtain the first sub-score of the agent.

[0244] As another example, the average or cumulative sum of the evaluation scores of the intermediate answers submitted by the agent in N rounds of feedback and the evaluation scores of the initial answers can be used as the first sub-score of the agent.

[0245] As another example, when the evaluation score includes scores under multiple evaluation dimensions (such as set dimensions such as rationality, innovation, feasibility, etc.), the calculation method of the first sub-score of the agent is, for example: weighted sum of the scores of the intermediate answers and initial answers submitted by the agent in N rounds of feedback process under the same evaluation dimension to obtain the weighted score of the agent under the same evaluation dimension, so that in the present disclosure, the weighted scores of the agent under each evaluation dimension can be weighted summed to obtain the first sub-score of the agent.

[0246] Therefore, the scores of all answers replied by the same intelligent agent in different evaluation dimensions are processed in a targeted manner to obtain the score of the same intelligent agent in each evaluation dimension, and then the scores of the same intelligent agent in each evaluation dimension are weightedly summed to obtain the first sub-score of the intelligent agent, which can improve the targetedness, rationality and reliability of the calculation of the first sub-score.

[0247] Step Y: Calculate the second sub-score of the agent according to the degree of change in the content of the intermediate answers submitted by the agent in the N-round feedback process.

[0248] Among them, the second sub-score is used to indicate the content stability of the intermediate answers submitted by the agent in each round of feedback process.

[0249] Among them, the second sub-score is negatively correlated with the degree of content change between the intermediate answers submitted in the N-round feedback process, that is, the smaller the degree of content change, the higher the second sub-score, and conversely, the greater the degree of content change, the lower the second sub-score.

[0250] Step Z: Calculate the second score of the agent according to the first sub-score and the second sub-score.

[0251] As an example, the first sub-score and the second sub-score may be weightedly summed to obtain the second score of the agent.

[0252] As another example, the average or cumulative sum of the first sub-score and the second sub-score may be used as the second score of the agent.

[0253] In summary, by comprehensively evaluating the scores of all answers given by the same agent in different rounds, as well as the degree of content change of all answers in different rounds (i.e., answer stability), the second score of the same agent (used to characterize the professionalism, accuracy, and stability of the same agent in handling problems) is calculated, which can improve the rationality and reliability of the second score calculation.

[0254] Step S505: determining a first agent from the multiple agents based on the second scores of the multiple agents.

[0255] As an example, the agent with the second highest score can be used as the first agent.

[0256] Step S506: Use the intermediate answer submitted by the first agent in the last round of feedback process as the first answer, and reply to the first question based on the first answer.

[0257] In the disclosed embodiment, the intermediate answer submitted by the first agent in the last round of feedback process may be used as the first answer for responding to the first question.

[0258] The problem-solving method of the disclosed embodiment obtains evaluation scores of all answers replied by the same agent in different rounds through multiple rounds of feedback and evaluation, and calculates the second score of the same agent based on the evaluation score (used to characterize the professionalism, accuracy and stability of the same agent in handling the problem). This can improve the rationality of the calculation of the second score, and then based on the reasonable second score, can screen out the best performing agent from multiple agents to obtain the final answer, which can improve the generation quality of the final answer.

[0259] In order to clearly illustrate the above embodiments of the present disclosure, the present disclosure also proposes a problem solving method.

[0260] Figure 6 This is a flowchart of the problem handling method provided in Embodiment 6 of the present disclosure.

[0261] like Figure 6 As shown, the problem handling method may include the following steps S601 to S604:

[0262] Step S601, obtaining a second problem that does not match the target processing requirement.

[0263] It should be noted that the explanation of the target processing requirements in the aforementioned embodiment is also applicable to this embodiment and will not be repeated here.

[0264] In the embodiments of the present disclosure, there is no limitation on the method for obtaining the second question. For example, the second question may be a question input by a user, or the second question may be a question collected online, or the second question may be a question automatically generated by a generative model, or the second question may be a question obtained from a training set or a test set, etc. The embodiments of the present disclosure do not limit this.

[0265] Among them, the input method of the second question may include but is not limited to touch input (such as sliding, clicking, etc.), keyboard input, voice input, etc.

[0266] Step S602, determining a second agent from multiple agents.

[0267] As an example, an agent may be randomly selected from a plurality of agents as the second agent.

[0268] As another example, the agent with the most model parameters can be used as the second agent to improve the accuracy of problem processing.

[0269] As another example, based on the current load information of multiple agents, the second agent with the lowest load can be determined from the multiple agents. This not only balances the load of each agent, but also ensures that each problem can receive an efficient and timely response, thereby improving problem handling performance and user experience.

[0270] Step S603: dispatch a second agent to process the second question and obtain a second answer.

[0271] In the disclosed embodiment, a second agent may be scheduled to process the second question and obtain a second answer fed back by the second agent.

[0272] As an example, a POST request containing the second question can be sent to the HTTP interface of the second agent through the HTTP interface address of the second agent, and the second answer returned by the second agent can be obtained through the HTTP interface of the second agent.

[0273] Step S604: reply to the second question based on the second answer.

[0274] In the disclosed embodiment, the second answer may be used to respond to the second question.

[0275] The problem handling method of the embodiment of the present disclosure only schedules one intelligent agent to provide one round of replies to the second question with relatively high processing time requirements, which can ensure that the second question can receive an efficient and timely response, thereby meeting the user's real-time query needs.

[0276] In any embodiment of the present disclosure, the output results of multiple agents can be integrated based on the Delphi method. Taking the agent as a large model for example, the implementation process mainly includes the following steps:

[0277] 1. First, collect the initial response solutions (referred to as initial answers in this disclosure) of each large model to the complex questions input by the user.

[0278] 2. Then, multiple rounds of Delphi process are carried out. In each round, the key information of the response plans of all the current large models are summarized and fed back to each large model, so that they can adjust their own response plans after reference. Then, check whether the response plans of the large models converge. If converged, the process ends, otherwise continue to the next round.

[0279] 3. Finally, select one response plan from all the response plans obtained after multiple rounds of feedback as the final response plan (here we use a simple method of selecting the plan with the highest frequency).

[0280] Among them, the Delphi method mainly involves the following functions:

[0281] First, the get_model_response function:

[0282] Function: Send a POST request containing a question to the HTTP interface of the specified big model through Python's requests library, and obtain the response plan returned by the big model (assuming the response plan is in JSON format). If an error occurs in the request (such as network problems, the interface returns an error status code, etc.), print the corresponding error information and return None.

[0283] Parameters: URL: HTTP interface address of the big model; question: complex question entered by the user, sent to the big model in the question field in JSON format; return value: reply solution in JSON format returned by the HTTP interface of the big model, if the request fails, None is returned.

[0284] It should be noted that, in the above POST request parameters received only by the HTTP interface of the big model, the question is sent in the JSON format of the question field, and the response plan returned by the big model is also in JSON format and contains relevant fields such as the plan and the key content of the plan, which is convenient for extraction and analysis for example. In actual use, it is necessary to accurately adjust the construction of the request parameters and the parsing method of the response plan according to the specific requirements of the HTTP interface of each big model.

[0285] Second, the delphi_method function:

[0286] Function: Implement the main process of the Delphi method. First, collect the response plans of all large models in the first round, then conduct multiple rounds of feedback loops in the specified rounds, summarize the response plans of all large models in the current round in each round, obtain the new response plan of the large model based on the summarized feedback, determine whether the response plan of the large model has converged, and finally select the final response plan.

[0287] Parameters: question: complex question input by the user, passed to the HTTP interface of each large model; num_rounds: the number of rounds of the Delphi method, the default is 3 rounds, which can be adjusted according to actual conditions; return value: the output plan of the large model finally selected.

[0288] Third, the are_responses_converged function:

[0289] Function: Simply judge whether the key contents of the response plans of all large models in the previous and next rounds are the same, so as to determine whether the response plans of the large models have converged. If the key contents of the response plans of all large models in the previous and next rounds are the same, it is considered to be converged and True is returned; as long as the key contents of one large model change, False is returned.

[0290] Parameters: old_responses: the dictionary of response information of each large model in the previous round, the key is the HTTP interface address of the large model, and the value is the JSON data of the response plan; new_responses: the dictionary of response information of each large model in the current round, the structure is the same as old_responses; return value: Boolean value, indicating whether it converges.

[0291] It should be noted that in the are_responses_converged function, the method of simply comparing the key contents of the response plans to determine whether they have converged is relatively rough. In actual complex scenarios, we can compare the response plans in more depth in multiple dimensions (such as the core logic of the implementation details, the expected effects, etc.), and even set a certain similarity threshold to judge the convergence more scientifically.

[0292] Fourth, select_final_scheme function:

[0293] Function: Simply count the number of times each response plan appears in all model responses, and then select the response plan with the largest number of appearances as the final response plan; Parameter: all_responses is the response information dictionary of each large model in the last round, including the complete response plans given by each large model; Return value: the response plan of the large model finally selected.

[0294] It should be noted that in the select_final_scheme function, the method of selecting the final response plan according to the number of times the response plan appears is only a basic approach. In actual application, you can choose the best response plan based on the specific application scenario, for example, you can comprehensively consider the feasibility, innovation, cost and other factors of the response plan for weighted scoring.

[0295] It should also be noted that although the above simple processing is performed for network anomalies when requesting the HTTP interface of the large model and the interface returning an error status code (such as printing error information and returning None), the overall system can be further enhanced in robustness, such as setting up a retry mechanism and more detailed error logging, etc., to cope with more complex problems in the actual operating environment.

[0296] In summary, the problem handling method provided by the present disclosure can be applied to different scenarios such as initialization framework code generation for software engineering projects and sales forecasting, which can improve the applicability of the method.

[0297] With the above Figures 1 to 6 Corresponding to the problem handling method provided in the embodiment, the present disclosure also provides a problem handling device. Figures 1 to 6 The problem handling method provided in the embodiment corresponds to the problem handling method, so the implementation of the problem handling method is also applicable to the problem handling device provided in the embodiment of the present disclosure, and will not be described in detail in the embodiment of the present disclosure.

[0298] Figure 7 This is a schematic diagram of the structure of the problem handling device provided in Embodiment 7 of the present disclosure.

[0299] like Figure 7 As shown, the problem handling device 700 may include: an acquisition module 710 , a scheduling module 720 , an execution module 730 and a determination module 740 .

[0300] The acquisition module 710 is used to acquire a first problem matching the target processing requirement;

[0301] A scheduling module 720 is used to schedule multiple agents in the distributed system to process the first question and obtain initial answers fed back by the multiple agents;

[0302] An execution module 730, configured to execute at least one round of feedback process based on the multiple initial answers to obtain intermediate answers submitted by the multiple agents in each round of feedback process;

[0303] The determination module 740 is used to determine a first answer for responding to the first question from the intermediate answers submitted by multiple agents in each round of feedback process.

[0304] In a possible implementation of the embodiments of the present disclosure, the scheduling module 720 is used to: generate a scheduling request based on a first question and context information associated with the first question; send the scheduling request to multiple agents in a distributed system so that the first question can be processed by multiple agents based on the context information to obtain an initial answer; and receive the initial answers fed back by multiple agents.

[0305] In a possible implementation of the embodiment of the present disclosure, the execution module 730 executes the first round of feedback process in at least one round of feedback process, specifically: extracting the first key reply information from multiple initial answers; scheduling multiple intelligent agents to refer to the first key reply information, adjusting the initial answers fed back by the corresponding intelligent agents, and obtaining the intermediate answers submitted by multiple intelligent agents in the first round of feedback process; determining whether the intermediate answers submitted by multiple intelligent agents in the first round of feedback process are all converged; if they are all converged, ending the feedback process.

[0306] In a possible implementation of the embodiment of the present disclosure, the execution module 730 executes a non-first-round feedback process in at least one round of feedback process, specifically: extracting the second key reply information from the intermediate answers submitted by multiple agents in the previous round of feedback process; scheduling multiple agents to refer to the second key reply information, and adjusting the intermediate answers submitted by the corresponding agents in the previous round of feedback process to obtain the intermediate answers submitted by multiple agents in this round of feedback process; judging whether the intermediate answers submitted by multiple agents in this round of feedback process are all converged; if they are all converged, ending the feedback process.

[0307] In a possible implementation of the embodiments of the present disclosure, the execution module 730 is used to: determine, for any agent among the multiple agents, a first content similarity between an intermediate answer submitted by any agent in a first round of feedback and an initial answer; when the first content similarity is higher than a first similarity threshold, determine that the intermediate answer submitted by any agent in the first round of feedback has converged; when the first content similarity is not higher than the first similarity threshold, determine that the intermediate answer submitted by any agent in the first round of feedback has not converged.

[0308] In a possible implementation of the embodiments of the present disclosure, the execution module 730 is used to: for any agent among multiple agents, classify the text pairs using a semantic classification model to obtain a classification result; wherein the classification result is used to indicate whether the text pairs are semantically equivalent, and the text pairs include an intermediate answer and an initial answer submitted by any agent in the first round of feedback process; when the classification result indicates that the text pairs are semantically equivalent, determine that the intermediate answers submitted by any agent in the first round of feedback process have converged; when the classification result indicates that the text pairs are not semantically equivalent, determine that the intermediate answers submitted by any agent in the first round of feedback process have not converged.

[0309] In a possible implementation of the embodiments of the present disclosure, the execution module 730 is used to: for any one of the multiple agents, annotate an initial answer fed back by any one of the multiple agents to obtain a first annotation result; wherein the first annotation result is used to indicate the semantic roles and the relationship between the semantic roles in the initial answer; annotate an intermediate answer submitted by any one of the agents in the first round of feedback process to obtain a second annotation result; when the first semantic similarity between the first annotation result and the second annotation result is higher than the second similarity threshold, determine that the intermediate answer submitted by any one of the agents in the first round of feedback process has converged; when the first semantic similarity is not higher than the second similarity threshold, determine that the intermediate answer submitted by any one of the agents in the first round of feedback process has not converged.

[0310] In a possible implementation of the embodiments of the present disclosure, the execution module 730 is used to: for any agent among multiple agents, obtain a first directed graph corresponding to an initial answer fed back by any agent; wherein the nodes in the first directed graph are used to indicate semantic roles in the initial answer, and the edges between the nodes are used to indicate dependencies and / or sequences between the semantic roles; obtain a second directed graph corresponding to an intermediate answer submitted by any agent in a first round of feedback; when the second semantic similarity between the first directed graph and the second directed graph is higher than a third similarity threshold, determine that the intermediate answer submitted by any agent in the first round of feedback has converged; when the second semantic similarity is not higher than the third similarity threshold, determine that the intermediate answer submitted by any agent in the first round of feedback has not converged.

[0311] In a possible implementation of the embodiments of the present disclosure, the execution module 730 is used to: for any intelligent agent among multiple intelligent agents, extract the first key element from the initial answer fed back by any intelligent agent; extract the second key element from the intermediate answer submitted by any intelligent agent in the first round of feedback process; if the second content similarity between the first key element and the second key element is higher than a fourth similarity threshold, determine that the intermediate answer submitted by any intelligent agent in the first round of feedback process has converged; if the second content similarity is not higher than the fourth similarity threshold, determine that the intermediate answer submitted by any intelligent agent in the first round of feedback process has not converged.

[0312] In a possible implementation of the embodiment of the present disclosure, the determination module 740 is used to: count the number of occurrences of each intermediate answer; and determine the first answer from each intermediate answer according to the number of occurrences of each intermediate answer.

[0313] In a possible implementation of the embodiments of the present disclosure, the determination module 740 is used to: evaluate each intermediate answer to obtain a score of each intermediate answer under multiple evaluation dimensions; determine a first score of each intermediate answer based on the score of each intermediate answer under multiple evaluation dimensions; and determine a first answer from each intermediate answer based on the first score of each intermediate answer.

[0314] In a possible implementation of the embodiment of the present disclosure, the number of rounds N of the feedback process is set, and the execution module 730 executes the first round of feedback process in the N rounds of feedback process, specifically: classifying multiple initial answers to obtain initial answers under multiple classification dimensions; scheduling multiple intelligent agents to refer to the initial answers under multiple classification dimensions, adjusting the initial answers fed back by the corresponding intelligent agents, and obtaining intermediate answers submitted by multiple intelligent agents in the first round of feedback process; scheduling multiple intelligent agents to compare and evaluate the initial answers under multiple classification dimensions, and obtaining evaluation scores of the initial answers submitted by multiple intelligent agents in the first round of feedback process under multiple classification dimensions.

[0315] In a possible implementation of the embodiment of the present disclosure, the execution module 730 executes a non-first-round feedback process in the N-round feedback process, specifically: classifying the intermediate answers submitted by multiple agents in the previous round of feedback process to obtain intermediate answers under multiple classification dimensions; scheduling multiple agents to refer to the intermediate answers under multiple classification dimensions, adjusting the intermediate answers submitted by the corresponding agents in the previous round of feedback process to obtain the intermediate answers submitted by multiple agents in this round of feedback process; scheduling multiple agents to compare and evaluate the intermediate answers under multiple classification dimensions to obtain evaluation scores of the intermediate answers submitted by multiple agents in multiple classification dimensions in this round of feedback process.

[0316] In a possible implementation of the embodiments of the present disclosure, the determination module 740 is used to: determine the second scores of the multiple agents based on the evaluation scores of the intermediate answers submitted by the multiple agents in N rounds of feedback process and the evaluation scores of the initial answers; determine the first agent from the multiple agents based on the second scores of the multiple agents; and use the intermediate answer submitted by the first agent in the last round of feedback process as the first answer.

[0317] In a possible implementation of the embodiments of the present disclosure, the determination module 740 is used to: determine, for any one of the multiple agents, a first sub-score of any one of the agents based on the evaluation score of the intermediate answer submitted by any one of the agents in the N-round feedback process and the evaluation score of the initial answer; determine a second sub-score of any one of the agents based on the degree of content change of the intermediate answer submitted by any one of the agents in the N-round feedback process; determine a second score of any one of the agents based on the first sub-score and the second sub-score.

[0318] In a possible implementation of the embodiments of the present disclosure, the evaluation score includes scores under multiple evaluation dimensions, and the determination module 740 is used to: for any intelligent agent among the multiple intelligent agents, perform a weighted sum of the scores of the intermediate answers and initial answers submitted by any intelligent agent in the N rounds of feedback process under the same evaluation dimension to obtain the weighted score of any intelligent agent under the same evaluation dimension; perform a weighted sum of the weighted scores of any intelligent agent under each evaluation dimension to obtain the first sub-score of any intelligent agent.

[0319] In a possible implementation of the embodiment of the present disclosure, the target processing requirement includes at least one of the following:

[0320] Processing time limit requirements; among which, the processing time limit requirements include that the problem handling time limit is not less than the time limit threshold;

[0321] Processing accuracy requirements; wherein, the processing accuracy requirements include the answer accuracy being higher than the accuracy threshold.

[0322] In a possible implementation of the embodiment of the present disclosure, the acquisition module 710 is further used to acquire a second problem that does not match the target processing requirement;

[0323] The determination module 740 is further used to determine a second agent from the plurality of agents;

[0324] The scheduling module 720 is further used to schedule the second agent to process the second question and obtain a second answer;

[0325] The problem handling device 700 further includes:

[0326] The reply module is used to reply to the second question based on the second answer.

[0327] In the problem handling device of the disclosed embodiment, each intelligent agent has a certain degree of autonomy and intelligence, which can enable the system to flexibly adapt to complex and changeable problems, select the most appropriate processing strategy and method according to the characteristics and requirements of the problem, thereby improving the efficiency and effect of problem handling, and based on the initial answers to the problem feedback from multiple intelligent agents, at least one round of feedback process is performed, and the intelligent agents can learn from each other, draw lessons from each other and make corrections, and continuously optimize their respective answers. This feedback mechanism helps to share information and knowledge between intelligent agents, thereby improving the accuracy and reliability of the final answer generation. In addition, the intelligent agents in the distributed system can collaborate and interact with each other. Even if an intelligent agent fails or makes an error, other intelligent agents can still continue to work and provide effective answers. This fault-tolerant capability makes the system more robust and reliable, able to cope with various emergencies and challenges, and ensure the continuity and stability of problem handling.

[0328] In order to implement the above embodiments, the present disclosure also provides an electronic device, which may include at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the problem handling method proposed in any of the above embodiments of the present disclosure.

[0329] In order to implement the above embodiments, the present disclosure further provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute the problem handling method proposed in any of the above embodiments of the present disclosure.

[0330] In order to implement the above embodiments, the present disclosure further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the problem processing method proposed in any of the above embodiments of the present disclosure.

[0331] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0332] Figure 8 A schematic block diagram of an example electronic device that can be used to implement embodiments of the present disclosure is shown.

[0333] Among them, the electronic device may include the server and the client in the above-mentioned embodiment. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices and other similar computing devices. The components shown herein, their connections and relationships, and their functions are only examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0334] like Figure 8 As shown, the electronic device 800 includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a ROM (Read-Only Memory) 802 or a computer program loaded from a storage unit 807 to a RAM (Random Access Memory) 803. In the RAM 803, various programs and data required for the operation of the device 800 can also be stored. The computing unit 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An I / O (Input / Output) interface 805 is also connected to the bus 804.

[0335] A number of components in the device 800 are connected to the I / O interface 805, including: an input unit 806, such as a keyboard, a mouse, etc.; an output unit 807, such as various types of displays, speakers, etc.; a storage unit 808, such as a disk, an optical disk, etc.; and a communication unit 809, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 809 allows the device 800 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0336] The computing unit 801 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a CPU (Central Processing Unit), a GPU (Graphic Processing Units), various dedicated AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, a DSP (Digital Signal Processor), and any appropriate processor, controller, microcontroller, etc. The computing unit 801 performs the various methods and processes described above, such as the above-mentioned problem handling method. For example, in some embodiments, the above-mentioned problem handling method may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 808. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 800 via the ROM 802 and / or the communication unit 809. When the computer program is loaded into the RAM 803 and executed by the computing unit 801, one or more steps of the problem handling method described above may be performed. Alternatively, in other embodiments, the computing unit 801 may be configured to execute the above problem handling method in any other appropriate manner (for example, by means of firmware).

[0337] Various embodiments of the systems and techniques described above herein may be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application Specific Standard Products), SOCs (System On Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include: being implemented in one or more computer programs that may be executed and / or interpreted on a programmable system including at least one programmable processor that may be a dedicated or general-purpose programmable processor that may receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0338] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0339] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include electrical connections based on one or more lines, portable computer disks, hard disks, RAM, ROM, EPROM (Electrically Programmable Read-Only-Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0340] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0341] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: LAN (Local Area Network), WAN (Wide Area Network), the Internet, and blockchain networks.

[0342] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or a cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services (Virtual Private Server). The server may also be a server of a distributed system, or a server combined with a blockchain.

[0343] It should be noted that artificial intelligence is a discipline that studies how computers can simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, planning, etc.), and includes both hardware-level and software-level technologies. Artificial intelligence hardware technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, and big data processing; artificial intelligence software technologies mainly include computer vision technology, speech recognition technology, natural language processing technology, as well as machine learning / deep learning, big data processing technology, knowledge graph technology, and other major directions.

[0344] According to the technical solution of the embodiment of the present disclosure, each intelligent agent has a certain degree of autonomy and intelligence, which can enable the system to flexibly adapt to complex and changeable problems, select the most appropriate processing strategy and method according to the characteristics and requirements of the problem, thereby improving the efficiency and effect of problem processing, and based on the initial answers to the problem feedback from multiple intelligent agents, at least one round of feedback process is performed, and the intelligent agents can learn from each other, draw lessons from each other and make corrections, and continuously optimize their respective answers. This feedback mechanism helps to share information and knowledge between intelligent agents, thereby improving the accuracy and reliability of the final answer generation. In addition, the intelligent agents in the distributed system can collaborate and interact with each other. Even if an intelligent agent fails or fails, other intelligent agents can still continue to work and provide effective answers. This fault-tolerant capability makes the system more robust and reliable, able to cope with various emergencies and challenges, and ensure the continuity and stability of problem processing.

[0345] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0346] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A problem solving method, characterized in that: The method comprises: Obtaining a first problem that matches the target processing requirement; Scheduling multiple agents in the distributed system to process the first problem, and obtaining initial answers fed back by the multiple agents; Perform at least one round of feedback process based on the multiple initial answers to obtain intermediate answers submitted by the multiple agents in each round of the feedback process; A first answer for responding to the first question is determined from the intermediate answers submitted by the multiple agents in each round of the feedback process.

2. The method according to claim 1, wherein: The multiple agents in the scheduling distributed system process the first problem to obtain initial answers fed back by the multiple agents, including: generating a scheduling request based on the first question and context information associated with the first question; Sending the scheduling request to multiple agents in the distributed system, so that the multiple agents process the first question based on the context information to obtain an initial answer; Receive initial answers fed back by the multiple agents.

3. The method according to claim 1, wherein: The first round of feedback process in the at least one round of feedback process includes: extracting first key response information from the plurality of initial answers; Scheduling the multiple agents to refer to the first key reply information, adjusting the initial answers fed back by the corresponding agents, and obtaining the intermediate answers submitted by the multiple agents in the first round of feedback process; Determining whether the intermediate answers submitted by the multiple agents in the first round of feedback process are all converged; If both converge, the feedback process ends.

4. The method according to claim 3, wherein: The non-first round feedback process in the at least one round of feedback process includes: Extracting second key reply information from the intermediate answers submitted by the multiple agents in the previous round of feedback process; Scheduling the multiple agents to refer to the second key reply information, adjusting the intermediate answers submitted by the corresponding agents in the previous round of feedback process, and obtaining the intermediate answers submitted by the multiple agents in the current round of feedback process; Determine whether the intermediate answers submitted by the multiple agents in the current round of feedback process are all converged; If both converge, the feedback process ends.

5. The method according to claim 3, wherein: The determining whether the intermediate answers submitted by the multiple agents in the first round of feedback process are all converged includes: For any one of the multiple agents, determining a first content similarity between an intermediate answer submitted by the any one of the agents in the first round of feedback process and an initial answer; In a case where the first content similarity is higher than a first similarity threshold, determining that the intermediate answer submitted by any of the agents in the first round of feedback process converges; When the first content similarity is not higher than the first similarity threshold, it is determined that the intermediate answer submitted by any of the agents in the first round of feedback process has not converged.

6. The method according to claim 3, wherein: The determining whether the intermediate answers submitted by the multiple agents in the first round of feedback process are all converged includes: For any of the multiple agents, a semantic classification model is used to classify the text pair to obtain a classification result; wherein the classification result is used to indicate whether the text pair is semantically equivalent, and the text pair includes an intermediate answer and an initial answer submitted by any of the agents in the first round of feedback process; In the case where the classification result indicates that the text pair is semantically equivalent, determining that the intermediate answers submitted by any of the agents in the first round of feedback process converge; When the classification result indicates that the text pairs are not semantically equivalent, it is determined that the intermediate answers submitted by any of the agents in the first round of feedback process have not converged.

7. The method according to claim 3, wherein: The determining whether the intermediate answers submitted by the multiple agents in the first round of feedback process are all converged includes: For any one of the multiple agents, annotate an initial answer fed back by the any one of the multiple agents to obtain a first annotation result; wherein the first annotation result is used to indicate the semantic roles in the initial answer and the relationship between the semantic roles; Annotating the intermediate answer submitted by any of the agents in the first round of feedback process to obtain a second annotation result; When a first semantic similarity between the first annotation result and the second annotation result is higher than a second similarity threshold, determining that an intermediate answer submitted by any agent in the first round of feedback process converges; When the first semantic similarity is not higher than the second similarity threshold, it is determined that the intermediate answers submitted by any of the agents in the first round of feedback process have not converged.

8. The method according to claim 3, wherein: The determining whether the intermediate answers submitted by the multiple agents in the first round of feedback process are all converged includes: For any one of the multiple agents, a first directed graph corresponding to an initial answer fed back by the any one of the agents is obtained; wherein the nodes in the first directed graph are used to indicate semantic roles in the initial answer, and the edges between the nodes are used to indicate dependency relationships and / or a sequence between the semantic roles; Obtaining a second directed graph corresponding to an intermediate answer submitted by any of the agents in the first round of feedback process; When the second semantic similarity between the first directed graph and the second directed graph is higher than a third similarity threshold, determining that the intermediate answer submitted by any of the agents in the first round of feedback process converges; When the second semantic similarity is not higher than the third similarity threshold, it is determined that the intermediate answers submitted by any of the agents in the first round of feedback process have not converged.

9. The method according to claim 3, wherein: The determining whether the intermediate answers submitted by the multiple agents in the first round of feedback process are all converged includes: For any one of the multiple agents, extract a first key element from an initial answer fed back by the any one of the agents; Extracting a second key element from an intermediate answer submitted by any of the intelligent agents in the first round of feedback process; In a case where the second content similarity between the first key element and the second key element is higher than a fourth similarity threshold, determining that the intermediate answer submitted by any of the agents in the first round of feedback process converges; When the second content similarity is not higher than the fourth similarity threshold, it is determined that the intermediate answers submitted by any of the agents in the first round of feedback process have not converged.

10. The method according to any one of claims 1 to 9, wherein: Determining a first answer for responding to the first question from the intermediate answers submitted by the multiple agents in each round of the feedback process includes: Counting the number of occurrences of each of the intermediate answers; The first answer is determined from the intermediate answers according to the number of occurrences of the intermediate answers.

11. The method according to any one of claims 1 to 9, wherein: Determining a first answer for responding to the first question from the intermediate answers submitted by the multiple agents in each round of the feedback process includes: Evaluate each of the intermediate answers to obtain scores of each of the intermediate answers under multiple evaluation dimensions; Determine a first score for each of the intermediate answers according to the scores of each of the intermediate answers under multiple evaluation dimensions; The first answer is determined from among the intermediate answers based on the first score of each of the intermediate answers.

12. The method according to claim 1, wherein: The number of rounds N of the feedback process is set, and the first round of feedback process in the N rounds of feedback process includes: Classifying the multiple initial answers to obtain initial answers under multiple classification dimensions; Scheduling the multiple agents to refer to the initial answers under the multiple classification dimensions, adjusting the initial answers fed back by the corresponding agents, and obtaining the intermediate answers submitted by the multiple agents in the first round of feedback process; The multiple agents are scheduled to compare and evaluate the initial answers under the multiple classification dimensions, and obtain evaluation scores of the initial answers under the multiple classification dimensions submitted by the multiple agents in the first round of feedback process.

13. The method according to claim 12, wherein: The non-first-round feedback process in the N-round feedback process includes: Classifying the intermediate answers submitted by the multiple agents in the previous round of feedback process to obtain intermediate answers under multiple classification dimensions; Scheduling the multiple agents to refer to the intermediate answers under the multiple classification dimensions, adjusting the intermediate answers submitted by the corresponding agents in the previous round of feedback process, and obtaining the intermediate answers submitted by the multiple agents in the current round of feedback process; The multiple agents are scheduled to compare and evaluate the intermediate answers under the multiple classification dimensions, and evaluation scores of the intermediate answers under the multiple classification dimensions submitted by the multiple agents in this round of feedback process are obtained.

14. The method according to claim 12 or 13, wherein: Determining a first answer for responding to the first question from the intermediate answers submitted by the multiple agents in each round of the feedback process includes: Determining second scores of the plurality of agents based on the evaluation scores of the intermediate answers submitted by the plurality of agents in the N-round feedback process and the evaluation scores of the initial answers; determining a first agent from among the plurality of agents based on the second scores of the plurality of agents; The intermediate answer submitted by the first agent in the last round of feedback process is used as the first answer.

15. The method according to claim 14, wherein: The determining of second scores of the plurality of agents based on the evaluation scores of the intermediate answers submitted by the plurality of agents in the N-round feedback process and the evaluation scores of the initial answers comprises: For any one of the multiple agents, determine a first sub-score of the any one of the agents according to an evaluation score of an intermediate answer submitted by the any one of the agents in the N rounds of feedback process and an evaluation score of an initial answer; Determining a second sub-score of any of the agents according to a degree of change in the content of the intermediate answers submitted by any of the agents in the N rounds of feedback process; A second score of any one of the agents is determined based on the first sub-score and the second sub-score.

16. The method according to claim 15, wherein: The evaluation score includes scores under multiple evaluation dimensions, and for any one of the multiple agents, determining a first sub-score of the any one of the multiple agents according to the evaluation score of the intermediate answer submitted by the any one of the agents in the N-round feedback process and the evaluation score of the initial answer, includes: For any one of the multiple agents, weighted sum of scores of the intermediate answers and initial answers submitted by the any one of the agents in the N rounds of feedback process under the same evaluation dimension to obtain a weighted score of the any one of the agents under the same evaluation dimension; The weighted scores of any of the agents under each of the evaluation dimensions are weightedly summed to obtain the first sub-score of any of the agents.

17. The method according to claim 1, wherein: The target processing requirement includes at least one of the following: Processing timeliness requirements; wherein, the processing timeliness requirements include that the problem processing time limit is not less than the time limit threshold; Processing accuracy requirements; wherein, the processing accuracy requirements include the answer accuracy being higher than an accuracy threshold.

18. The method according to claim 17, wherein: The method further comprises: Acquire a second problem that does not match the target processing requirement; determining a second agent from among the plurality of agents; dispatching the second agent to process the second question and obtain a second answer; Respond to the second question based on the second answer.

19. A problem handling device, characterized in that: The device comprises: An acquisition module, used for acquiring a first question matching a target processing requirement; A scheduling module, used for scheduling multiple agents in the distributed system to process the first problem and obtain initial answers fed back by the multiple agents; An execution module, configured to execute at least one round of feedback process based on the plurality of initial answers, so as to obtain intermediate answers submitted by the plurality of agents in each round of the feedback process; A determination module is used to determine a first answer for responding to the first question from the intermediate answers submitted by the multiple agents in each round of the feedback process.

20. The device according to claim 19, wherein The scheduling module is used to: generating a scheduling request based on the first question and context information associated with the first question; Sending the scheduling request to multiple agents in the distributed system, so that the multiple agents process the first question based on the context information to obtain an initial answer; Receive initial answers fed back by the multiple agents.

21. The device according to claim 19, wherein The execution module executes the first round of feedback process in the at least one round of feedback process, specifically: extracting first key response information from the plurality of initial answers; Scheduling the multiple agents to refer to the first key reply information, adjusting the initial answers fed back by the corresponding agents, and obtaining the intermediate answers submitted by the multiple agents in the first round of feedback process; Determining whether the intermediate answers submitted by the multiple agents in the first round of feedback process are all converged; If both converge, the feedback process ends.

22. The device according to claim 21, wherein The execution module executes the non-first-round feedback process in the at least one round of feedback process, specifically: Extracting second key reply information from the intermediate answers submitted by the multiple agents in the previous round of feedback process; Scheduling the multiple agents to refer to the second key reply information, adjusting the intermediate answers submitted by the corresponding agents in the previous round of feedback process, and obtaining the intermediate answers submitted by the multiple agents in the current round of feedback process; Determine whether the intermediate answers submitted by the multiple agents in the current round of feedback process are all converged; If both converge, the feedback process ends.

23. The device according to claim 21, wherein The execution module is used to: For any one of the multiple agents, determining a first content similarity between an intermediate answer submitted by the any one of the agents in the first round of feedback process and an initial answer; In a case where the first content similarity is higher than a first similarity threshold, determining that the intermediate answer submitted by any of the agents in the first round of feedback process converges; When the first content similarity is not higher than the first similarity threshold, it is determined that the intermediate answer submitted by any of the agents in the first round of feedback process has not converged.

24. The device according to claim 21, wherein The execution module is used to: For any of the multiple agents, a semantic classification model is used to classify the text pair to obtain a classification result; wherein the classification result is used to indicate whether the text pair is semantically equivalent, and the text pair includes an intermediate answer and an initial answer submitted by any of the agents in the first round of feedback process; In the case where the classification result indicates that the text pair is semantically equivalent, determining that the intermediate answers submitted by any of the agents in the first round of feedback process converge; When the classification result indicates that the text pairs are not semantically equivalent, it is determined that the intermediate answers submitted by any of the agents in the first round of feedback process have not converged.

25. The device according to claim 21, wherein The execution module is used to: For any one of the multiple agents, annotate an initial answer fed back by the any one of the multiple agents to obtain a first annotation result; wherein the first annotation result is used to indicate the semantic roles in the initial answer and the relationship between the semantic roles; Annotating the intermediate answer submitted by any of the agents in the first round of feedback process to obtain a second annotation result; When a first semantic similarity between the first annotation result and the second annotation result is higher than a second similarity threshold, determining that an intermediate answer submitted by any agent in the first round of feedback process converges; When the first semantic similarity is not higher than the second similarity threshold, it is determined that the intermediate answers submitted by any of the agents in the first round of feedback process have not converged.

26. The device according to claim 21, wherein The execution module is used to: For any one of the multiple agents, a first directed graph corresponding to an initial answer fed back by the any one of the agents is obtained; wherein the nodes in the first directed graph are used to indicate semantic roles in the initial answer, and the edges between the nodes are used to indicate dependency relationships and / or a sequence between the semantic roles; Obtaining a second directed graph corresponding to an intermediate answer submitted by any of the agents in the first round of feedback process; When the second semantic similarity between the first directed graph and the second directed graph is higher than a third similarity threshold, determining that the intermediate answer submitted by any of the agents in the first round of feedback process converges; When the second semantic similarity is not higher than the third similarity threshold, it is determined that the intermediate answers submitted by any of the agents in the first round of feedback process have not converged.

27. The device according to claim 21, wherein The execution module is used to: For any one of the multiple agents, extract a first key element from an initial answer fed back by the any one of the agents; Extracting a second key element from an intermediate answer submitted by any of the intelligent agents in the first round of feedback process; In a case where the second content similarity between the first key element and the second key element is higher than a fourth similarity threshold, determining that the intermediate answer submitted by any of the agents in the first round of feedback process converges; When the second content similarity is not higher than the fourth similarity threshold, it is determined that the intermediate answers submitted by any of the agents in the first round of feedback process have not converged.

28. The device according to any one of claims 19 to 27, wherein: The determining module is used to: Counting the number of occurrences of each of the intermediate answers; The first answer is determined from the intermediate answers according to the number of occurrences of the intermediate answers.

29. The device according to any one of claims 19 to 27, wherein: The determining module is used to: Evaluate each of the intermediate answers to obtain scores of each of the intermediate answers under multiple evaluation dimensions; Determine a first score for each of the intermediate answers according to the scores of each of the intermediate answers under multiple evaluation dimensions; The first answer is determined from among the intermediate answers based on the first score of each of the intermediate answers.

30. The device according to claim 19, wherein The number of rounds N of the feedback process is set, and the execution module executes the first round of the feedback process in the N rounds of feedback process, specifically: Classifying the multiple initial answers to obtain initial answers under multiple classification dimensions; Scheduling the multiple agents to refer to the initial answers under the multiple classification dimensions, adjusting the initial answers fed back by the corresponding agents, and obtaining the intermediate answers submitted by the multiple agents in the first round of feedback process; The multiple agents are scheduled to compare and evaluate the initial answers under the multiple classification dimensions, and obtain evaluation scores of the initial answers under the multiple classification dimensions submitted by the multiple agents in the first round of feedback process.

31. The device according to claim 30, wherein The execution module executes the non-first-round feedback process in the N-round feedback process, specifically: Classifying the intermediate answers submitted by the multiple agents in the previous round of feedback process to obtain intermediate answers under multiple classification dimensions; Scheduling the multiple agents to refer to the intermediate answers under the multiple classification dimensions, adjusting the intermediate answers submitted by the corresponding agents in the previous round of feedback process, and obtaining the intermediate answers submitted by the multiple agents in the current round of feedback process; The multiple agents are scheduled to compare and evaluate the intermediate answers under the multiple classification dimensions, and evaluation scores of the intermediate answers under the multiple classification dimensions submitted by the multiple agents in this round of feedback process are obtained.

32. The device according to claim 30 or 31, wherein: The determining module is used to: Determining second scores of the plurality of agents based on the evaluation scores of the intermediate answers submitted by the plurality of agents in the N-round feedback process and the evaluation scores of the initial answers; determining a first agent from among the plurality of agents based on the second scores of the plurality of agents; The intermediate answer submitted by the first agent in the last round of feedback process is used as the first answer.

33. The device according to claim 32, wherein: The determining module is used to: For any one of the multiple agents, determine a first sub-score of the any one of the agents according to an evaluation score of an intermediate answer submitted by the any one of the agents in the N rounds of feedback process and an evaluation score of an initial answer; Determining a second sub-score of any of the agents according to a degree of change in the content of the intermediate answers submitted by any of the agents in the N rounds of feedback process; A second score of any one of the agents is determined based on the first sub-score and the second sub-score.

34. The device according to claim 33, wherein The evaluation score includes scores under multiple evaluation dimensions, and the determination module is used to: For any one of the multiple agents, weighted sum of scores of the intermediate answers and initial answers submitted by the any one of the agents in the N rounds of feedback process under the same evaluation dimension to obtain a weighted score of the any one of the agents under the same evaluation dimension; The weighted scores of any of the agents under each of the evaluation dimensions are weightedly summed to obtain the first sub-score of any of the agents.

35. The apparatus of claim 19, wherein: The target processing requirement includes at least one of the following: Processing timeliness requirements; wherein, the processing timeliness requirements include that the problem processing time limit is not less than the time limit threshold; Processing accuracy requirements; wherein, the processing accuracy requirements include the answer accuracy being higher than an accuracy threshold.

36. The device according to claim 35, wherein The acquisition module is further used to acquire a second problem that does not match the target processing requirement; The determination module is further used to determine a second agent from the multiple agents; The scheduling module is further used to schedule the second agent to process the second question and obtain a second answer; The device further comprises: A reply module is used to reply to the second question based on the second answer.

37. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the problem handling method according to any one of claims 1 to 18.

38. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to enable the computer to execute the problem handling method according to any one of claims 1-18.

39. A computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the problem handling method according to any one of claims 1 to 18.

Images