Language model intelligent question-answering system-oriented multi-objective optimization method

Abstract

The invention discloses a multi-objective optimization method for a language model intelligent question-answering system, which enables a language model to simultaneously maximize user experience and minimize interaction times under incomplete information so as to carry out online learning and decision-making tasks. The language model intelligent question answering system is composed of a plurality of language generation models and a decision model, the language generation models are responsible for generating a plurality of alternative answers, and the decision model is responsible for selecting two answers as feedback. Firstly, an initial strategy is designed to explore the performance of the language generation models, and the language generation models are simply sorted by using user feedback. For the design of a decision model, a decision method adaptive to dynamic sorting is used, and with the assistance of a pair-by-pair elimination mode, language generation models which do not meet user requirements are gradually eliminated. Compared with an existing question and answer model optimization technology, the question and answer model optimization method and system not only can ensure that the language generation model which most meets the user requirements is utilized as much as possible to provide the answers to maximize the user experience, but also can additionally ensure that the number of interaction times required for searching the language generation model is minimized.

Description

Technical Field

[0001] The invention relates to a multi-objective optimization method for a language model intelligent question answering system, and belongs to the technical field of online learning optimization of language models. Background Art

[0002] With the rapid development of artificial intelligence and natural language processing technologies, language model intelligent question answering systems are playing an increasingly important role in daily life and work. From personal assistants to customer service robots, to information retrieval tools in academic research, language model intelligent question answering systems are being widely used in various fields. However, in practical applications, language models usually only use the accuracy of answers as the only indicator. Although this single-objective optimization method can provide accurate answers to a certain extent, it ignores the user's needs for other performance indicators in practical applications, such as reducing the number of interactions, increasing the speed and relevance of answers, etc. This one-sided pursuit of accuracy may lead to a decline in user experience, especially in application scenarios that require efficient communication.

[0003] However, the user's goal is not only to obtain high-quality answers, but also to minimize the number of interactions required in some scenarios. Existing systems often fail to balance the two, resulting in users facing lengthy interaction processes in actual use, which not only increases time costs, but also may delay user decision-making due to redundant information or irrelevant answers, reducing the practicality of the system and user experience. For example, in an online medical consultation system, users not only want to get accurate answers, but also want to quickly obtain the information they need so that they can make timely judgments. Summary of the invention

[0004] Purpose of the invention: Existing language model question-answering systems usually only consider the accuracy of answers to improve the user's interactive experience, which limits the practicality of existing language model question-answering systems to a great extent. In view of the problems and shortcomings in the prior art, the present invention considers using online learning to design an online decision-making method for a language model, which can simultaneously ensure the maximization of the user's interactive experience and the minimization of the number of interactions.

[0005] Technical solution: A multi-objective optimization method for a language model intelligent question-answering system, wherein the language model intelligent question-answering system includes multiple language generation models and a decision model. The language generation model is responsible for generating alternative answers, and the decision model is responsible for selecting answers and returning them to the user. The multi-objective optimization refers to maximizing the user interaction experience and minimizing the number of interactions.

[0006] First, the language generation model and decision model are initialized. When the user asks a question, the language generation model generates several alternative answers based on the question. This process is determined by the pre-trained language generation model. Next, the decision model makes a decision and selects two answers to return to the user based on the internal state of the current decision model. This process needs to deal with exploring unknown language generation model information or using known language generation model information. Then, the user selects the better of the two answers. Based on this feedback, the decision model updates its internal state. This process needs to deal with potential inaccurate user feedback; finally, the language generation model updates its set of alternative answers. This process needs to deal with the unbalanced information of each language generation model.

[0007] A multi-objective optimization method for language model intelligent question answering system, which needs to deal with exploring unknown language generation model information or utilizing known language generation model information, potential inaccurate user feedback, and unbalanced information of each language generation model.

[0008] Initialize the language generation model and decision model:

[0009] Set the number N of language generation models.

[0010] Initialize an ensemble of language generation models but

[0011] Initialize the weight vector S of N language generation models = 0 N .

[0012] Initialize the qualification vector Q of N language generation models to 1 N .

[0013] Initialize the historical feedback matrix W = 0 N×N .

[0014] Initialize the round number r=1.

[0015] The specific steps for the decision model to make a decision and explore unknown language generation model information or use known language generation model information are as follows:

[0016] Repeat the following steps until

[0017] Select the first language generation model that meets the conditions: the sequence number of the language generation model is i, and the decision model selection weight is S i >0 and qualification Q i = 1, if there is no model that meets the conditions, then select Select one at random.

[0018] Calculate the set except i Historical information of the language generation model in where k=1,...,i-1,i+1,...,N,W m，k Represents the total number of times that user feedback model m is better than model k in all historical feedback.

[0019] By probability A language generation model is obtained by sampling, where k=1, ..., i-1, i+1, ..., N. The serial number of the language generation model is j, and its corresponding answer is taken as the second answer.

[0020] The specific steps for the decision model to update its internal state and respond to potentially inaccurate user feedback are:

[0021] Repeat the following steps until the weight S of a language generation model among language generation model i and language generation model j is -r:

[0022] User asks a question.

[0023] The answers generated by the two language generation models numbered i and j are returned to the user.

[0024] If the user feedback model i's answer is better than model j, then W i，j =W i，j +1, S i =S i +1, S j =S j -1; otherwise W j，i =W j，i +1, S i =S i -1, S j =S j +1.

[0025] The language generation model updates its candidate answer set. The specific steps to deal with the imbalanced information of each language generation model are as follows:

[0026] If S i =-r,Q i =0; if S j =-r,Q j =0

[0027] For language generation model i and model j, let the maximum acceptable failure probability be δ. renew like Update Update

[0028] If the qualification vector Q of all language generation models is 0N ,but Language Generation Model

[0029] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the multi-objective optimization method for the language model intelligent question-answering system as described above is implemented.

[0030] A computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of the multi-objective optimization method for a language model intelligent question-answering system as described above.

[0031] Beneficial effects: Compared with the prior art, the multi-objective optimization method for the language model intelligent question-answering system provided by the present invention has the following advantages: the present invention not only focuses on the accuracy of the answer in design, but also optimizes the number of interactions, achieving an effective balance between accuracy and user interaction burden. Compared with the prior art, the present invention makes the language model intelligent question-answering system more flexible in practical applications by introducing a dynamic learning mechanism, enhancing anti-noise performance, and balancing information processing, which significantly improves the practicality of the system and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flowchart of the language model intelligent question answering system in an embodiment of the present invention;

[0033] Figure 2 is a flow chart of a method for exploring unknown language generation model information or utilizing known language generation model information in an embodiment of the present invention;

[0034] Figure 3 is a flow chart of a method for dealing with potentially inaccurate user feedback in an embodiment of the present invention;

[0035] Figure 4 It is a flow chart of a method for dealing with imbalanced information of each language generation model in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0037] Take the intelligent medical question-and-answer system as an example. Existing intelligent medical question-and-answer systems usually only consider the accuracy of answers to improve the user's interactive experience, which limits the practicality of existing intelligent medical question-and-answer systems to a great extent. For example, in an intelligent medical question-and-answer system, users not only want to get accurate answers, but also want to quickly obtain the required information so that they can make timely judgments. However, existing systems can often only optimize a single goal, resulting in a poor user experience.

[0038] The multi-objective optimization technology for the language model intelligent question answering system disclosed in the present invention can effectively address the above problems, thereby greatly expanding the practicality of the intelligent question answering system.

[0039] A multi-objective optimization method for a language model intelligent question answering system. The language model intelligent question answering system includes multiple language generation models and a decision model. The language generation model is responsible for generating alternative answers, and the decision model is responsible for selecting answers and returning them to the user. Multi-objective optimization refers to maximizing the user interaction experience and minimizing the number of interactions.

[0040] First, the language generation model and decision model are initialized. When the user asks a question in the intelligent medical question-answering system, the language generation model first generates several alternative answers based on the knowledge base and past experience. This process is determined by the pre-trained language generation model. Then, the decision model makes a decision and selects two answers that may be more helpful to the user based on the internal state of the current decision model and returns them to the user. This process needs to deal with exploring unknown language generation model information or using known language generation model information. Then, the user chooses the one that feels better among the two answers based on the answer given by the selection. Based on this feedback, the decision model updates its internal state. This process needs to deal with potential inaccurate user feedback; finally, the language generation model updates its set of alternative answers based on the preferences corresponding to the user feedback to answer the user's subsequent medical consultation questions. This process needs to deal with the unbalanced information of each language generation model.

[0041] Multi-objective optimization technology for language model intelligent question answering system, which needs to deal with exploring unknown language generation model information or utilizing known language generation model information, potential inaccurate user feedback, and unbalanced information of each language generation model.

[0042] Initialize the language generation model and decision model:

[0043] Set the number N of language generation models.

[0044] Initialize an ensemble of language generation models but

[0045] Initialize the weight vector S of N language generation models = 0N .

[0046] Initialize the qualification vector Q of N language generation models to 1 N .

[0047] Initialize the historical feedback matrix W = 0 N×N .

[0048] Initialize the round number r=1.

[0049] The decision model makes a decision. The specific steps for exploring unknown language generation model information or using known language generation model information are as follows:

[0050] Repeat the following steps until the collection

[0051] 1. Select the first language generation model that meets the conditions: the sequence number of the language generation model is i, and the decision model selection weight is S i >0 and qualification Q i = 1, if there is no model that meets the conditions, then select Select one at random.

[0052] 2. Calculate the set other than i Historical information of the language generation model in Among them, k=1,…,i-1,i+1,…,N.

[0053] 3. Based on probability A model is obtained by sampling, where k = 1,…,i-1,i+1,…,N. The serial number of the language generation model is j, and its corresponding answer is taken as the second answer.

[0054] The specific steps for the decision model to update its internal state and respond to potentially inaccurate user feedback are:

[0055] Repeat the following steps until the weight S of a language generation model in model i and model j is -r:

[0056] 1. The user asks a question.

[0057] 2. Return the answers generated by the two language generation models numbered i and j to the user.

[0058] 3. If the user feedback model i's answer is better than model j, then W i,j =W i,j +1, S i =S i +1,S j =S j -1; otherwise

[0059] Wj,i =W j,i +1, S i =S i -1,S j =S j +1

[0060] 4. The language generation model updates its candidate answer set. The specific steps to deal with the imbalanced information of each language generation model are:

[0061] 5. If S i =-r,Q i =0; if S j =-r,Q j =0

[0062] 6. For model i and model j, if renew like Update Update

[0063] If the qualification vector Q of all language generation models is 0 N ,but

[0064] Next, the specific implementation of the present invention is described in detail by taking the intelligent medical question-answering system as an example.

[0065] Figure 1 This is the workflow diagram of the intelligent medical question-answering system proposed by the present invention. The system consists of multiple language generation models and a decision model. The language generation model is responsible for generating alternative answers, and the decision model is responsible for selecting answers and returning them to the user. Step 1 initializes N alternative language generation models, which have different medical knowledge and expertise. Step 2 determines whether there is only one alternative language model left. If so, the process ends, which means that the system has determined the model that best suits the current medical scenario. Step 3 Each language model generates several alternative answers based on the user's medical questions, such as diagnostic suggestions or treatment plans. Step 4 The decision model selects two medical suggestions that best meet the current user's needs from all alternative answers based on the internal state of the current decision model and returns them to the user. Step 5 The user selects a more satisfactory one based on the quality of the two returned medical suggestions and feeds it back to the system. Step 6 Based on the user's feedback, the decision model updates its internal state. Step 7 The decision model determines whether the alternative answer set needs to be updated. If not, it returns to step 3 to answer the user's next question. If updated, it enters step 8. Step 8 updates the alternative model based on user feedback and internal state, so that subsequent medical suggestions are more in line with user needs.

[0066] Figure 2This is a workflow diagram of the decision model proposed in the present invention making a decision in the intelligent medical question-answering system and selecting a model that meets the conditions from all language generation models. For example, at a certain moment, after a user asks a medical question to the intelligent medical system, several language generation models in the system have generated multiple alternative medical suggestions. At this time, the decision model should select two best suggestions from these alternative suggestions as output and return them to the user. The first selected generation model is numbered i, and the decision model in step 41 establishes the judgment condition as weight S i >0 and qualification Q i =1, which means that the model has sufficient credibility and qualifications. Step 42 determines whether there is a language generation model that meets the condition. If so, go to step 43, and step 43 selects the medical advice generated by the model that meets the condition as the first medical advice; if there is no model that meets the condition in step 42, go to step 44, and step 44 selects the medical advice generated by the model that meets the condition from the set. After the first answer is selected, step 45 calculates the set except i. Historical information of the language generation model in Where k = 1, ..., i-1, i+1, ..., N. The purpose of this step is to evaluate the relative merits of other models in providing medical advice in history. Step 46 The second model j is sampled, where k=1, ..., i-1, i+1, ..., N, and the answer generated by the model is selected as the second medical advice. Step 47 finally returns the selected two medical advices to the user.

[0067] Figure 3 This is a workflow diagram of the decision model proposed in the present invention updating its internal state in the intelligent medical question-and-answer system. Take the case where a user at a certain moment receives two medical suggestions returned by the intelligent medical system and selects one of the suggestions that better meets his needs. At this time, the decision model should update its internal state based on user feedback. The serial numbers of the models corresponding to the two selected answers are recorded as i and j. Step 61 determines whether there is a language generation model with a weight S of -r in model i and model j, and ends if so. Step 62 jumps to step 63 and step 65 respectively according to user feedback that i is better than j or j is better than i. If i is better than j, step 63 updates the historical feedback matrix W i，j =W i，j +1, step 64 updates the weight vector S i =S i +1, S j =S j -1; if j is better than i, then step 65 updates the historical feedback matrix W j，i =W j，i +1, step 66 updates the weight vector Si =S i -1, S j =S j +1.

[0068] Figure 4 is a flowchart of the intelligent medical question-answering system proposed by the present invention updating its candidate answer set. Step 81 determines which model of model i or j will be temporarily suppressed as a candidate after the internal state is updated. For example, if S i = -r, then step 82 sets the qualification Q of model i i =0, indicating that the model will not participate in answer generation in the next question; if S j = -r, then step 83 sets the qualification Q of model j j = 0. Step 84 determines whether the candidate model needs to be updated. The specific determination is based on two conditions: Condition 1: If condition 1 is met, step 85 updates the selected language model to Condition 2: If condition 2 is met, step 86 updates the selected language model to Step 87 determines whether the qualification vectors of all language generation models are 0, that is, condition 3: Q = 0 N If condition 3 is met, step 88 will initialize the qualifications of all remaining candidate language models, i.e. Ensure that the smart medical system can continue to provide effective medical advice to users.

[0069] Obviously, those skilled in the art should understand that the various steps of the multi-objective optimization method of the intelligent medical question-answering system of the above-mentioned embodiment of the present invention can be implemented by a general-purpose computing device, which can be concentrated on a single computing device or distributed on a network composed of multiple computing devices. Optionally, they can be implemented with program codes executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present invention are not limited to any specific combination of hardware and software.

Claims

1. A multi-objective optimization method for a language model intelligent question answering system, characterized in that: The language model intelligent question answering system includes multiple language generation models and a decision model; the language generation model is responsible for generating alternative answers, and the decision model is responsible for selecting answers and returning them to the user; the multi-objective optimization refers to maximizing the user interaction experience and minimizing the number of interactions; First, the language generation model and the decision model are initialized. When the user asks a question, the language generation model generates several alternative answers based on the question. This process is determined by the pre-trained language generation model. Next, the decision model makes a decision and selects two answers to return to the user based on the internal state of the current decision model. This process needs to deal with exploring unknown language generation model information or using known language generation model information. Then, the user selects the better of the two answers. Based on this feedback, the decision model updates its internal state. This process needs to deal with potential inaccurate user feedback. Finally, the language generation model updates its set of alternative answers. This process needs to deal with the unbalanced information of each language generation model.

2. The multi-objective optimization method for the language model intelligent question answering system according to claim 1, characterized in that: The initialization language generation model and decision model include: Set the number of language generation models N; Initialize an ensemble of language generation models but Initialize the weight vector S of N language generation models = 0 N ; Initialize the qualification vector Q of N language generation models to 1 N ; Initialize the historical feedback matrix W = 0 N×N ; Initialize the round number r=1.

3. The multi-objective optimization method for the language model intelligent question answering system according to claim 1, characterized in that: The specific steps for the decision model to make a decision and explore unknown language generation model information or use known language generation model information are as follows: Repeat the following steps until Select the first language generation model that meets the conditions: the sequence number of the language generation model is i, and the decision model selection weight is S i >0 and qualification Q i = 1, if there is no model that meets the conditions, then select Choose one at random; Calculate the set except i Historical information of the language generation model in where k=1,…,i-1,i+1,…,N,W m,k Represents the total number of times that user feedback model m is better than model k in all historical feedback; By probability A language generation model is obtained by sampling, where k = 1,…,i-1,i+1,…,N. The serial number of the language generation model is j, and its corresponding answer is taken as the second answer.

4. The multi-objective optimization method for the language model intelligent question answering system according to claim 1, characterized in that: The specific steps for the decision model to update its internal state and respond to potentially inaccurate user feedback are: Repeat the following steps until the weight S of a language generation model in language generation model i and language generation model j is -r: Users raise questions; Return the answers generated by the two language generation models numbered i and j to the user; If the user feedback model i's answer is better than model j, then W i,j =W i,j +1, S i =S i +1,S j =S j -1; otherwise W j,i =W j,i +1, S i =S i -1,S j =S j +1.

5. The multi-objective optimization method for the language model intelligent question answering system according to claim 1, characterized in that: The language generation model updates its candidate answer set. The specific steps to deal with the imbalanced information of each language generation model are as follows: If S i =-r,Q i =0; if S j =-r,Q j =0 For language generation model i and model j, if renew like Update Update If the qualification vector Q of all language generation models is 0 N ,but Language Generation Model 6. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the multi-objective optimization method for the language model intelligent question answering system as described in any one of claims 1 to 5 is implemented.

7. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the multi-objective optimization method for a language model intelligent question answering system as described in any one of claims 1 to 5 are implemented.

Images