Large model training method and device, interaction method and device, equipment and medium

Abstract

The invention provides a large model training method, an artificial intelligence-based interaction method and device, an intelligent agent, electronic equipment, a storage medium and a program product, and relates to the field of artificial intelligence, in particular to the technical fields of large models, model training, man-machine interaction and the like. The specific implementation scheme is as follows: acquiring a historical interaction dialogue with a target object; under the condition that it is determined that the historical interaction dialogue meets the preset processing condition, sample generation processing is conducted on the historical interaction dialogue, and a training sample is obtained; and training the large model by using the training sample to obtain the large model comprising the model knowledge matched with the target object.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence, and particularly to technical fields such as large models, model training, human-computer interaction, etc. Specifically, it relates to a training method for large models, an interaction method based on artificial intelligence, a device, an intelligent agent, an electronic device, a storage medium, and a program product. Background Art

[0002] Large models can understand and generate natural language and are capable of handling various tasks, such as text generation, knowledge answering, reasoning calculation, reading comprehension, etc. However, the model knowledge learned by large models themselves is generally general knowledge. How to enable large models to have specific memory capabilities has become a research focus. Summary of the Invention

[0003] The present disclosure provides a training method for large models, an interaction method based on artificial intelligence, a device, an intelligent agent, an electronic device, a storage medium, and a program product.

[0004] According to one aspect of the present disclosure, there is provided a training method for a large model, including: obtaining a historical interaction dialogue with a target object; in the case where it is determined that the historical interaction dialogue meets a predetermined processing condition, performing sample generation processing on the historical interaction dialogue to obtain training samples; and using the training samples to train a large model to obtain a large model including model knowledge matching the target object.

[0005] According to another aspect of the present disclosure, there is provided an interaction method based on artificial intelligence, including: in response to receiving an interaction request from a target object, determining a target large model matching the target object from a plurality of candidate large models, where the interaction request carries the interaction content of the target object; and inputting the interaction content into the target large model to obtain a feedback content for the interaction content; where the target large model includes model knowledge matching the target object, and the model knowledge is obtained by using the method described above.

[0006] According to one aspect of the present disclosure, there is provided a training device for a large model, including: an obtaining module for obtaining a historical interaction dialogue with a target object; a sample generation module for performing sample generation processing on the historical interaction dialogue to obtain training samples in the case where it is determined that the historical interaction dialogue meets a predetermined processing condition; and a training module for using the training samples to train a large model to obtain a large model including model knowledge matching the target object.

[0007] According to one aspect of the present disclosure, there is provided an artificial intelligence-based interaction device, including: a response module configured to, in response to receiving an interaction request of a target object, determine a target large model that matches the target object from a plurality of candidate large models, where the interaction request carries interaction content of the target object; and a feedback module configured to input the interaction content into the target large model to obtain feedback content for the interaction content; where the target large model includes model knowledge that matches the target object, and the model knowledge is obtained by using the device as described above.

[0008] According to another aspect of the present disclosure, there is provided an artificial intelligence-based agent, including: an input module configured to receive input information; a processing module configured to execute the method as described above based on the input information received by the input module; and an output module configured to output the output information obtained by the processing module.

[0009] According to another aspect of the present disclosure, there is provided an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; where 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 method as described above.

[0010] According to another aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer instructions, where the computer instructions are used to cause the computer to execute the method as described above.

[0011] According to another aspect of the present disclosure, there is provided a computer program product, including a computer program that, when executed by a processor, implements the method as described above.

[0012] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it used to limit the scope of the present disclosure. Other features of the present disclosure will become easily understandable through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The drawings are used to better understand the solution and do not constitute a limitation to the present disclosure. Among them:

[0014] Figure 1 Schematically shows an exemplary system architecture to which an artificial intelligence-based interaction method and device can be applied according to an embodiment of the present disclosure;

[0015] Figure 2 Schematically shows a flowchart of a method for training a large model according to an embodiment of the present disclosure;

[0016] Figure 3Schematically shows a flowchart of generating training samples for multi-modal question-answering content according to an embodiment of the present disclosure;

[0017] Figure 4 Schematically shows a flowchart of a training method for a large model according to another embodiment of the present disclosure;

[0018] Figure 5 Schematically shows a flowchart of an interaction method based on artificial intelligence according to an embodiment of the present disclosure;

[0019] Figure 6 Schematically shows a flowchart of an interaction method based on artificial intelligence according to another embodiment of the present disclosure;

[0020] Figure 7 Schematically shows a block diagram of a training device for a large model according to an embodiment of the present disclosure;

[0021] Figure 8 Schematically shows a block diagram of an interaction device based on artificial intelligence according to an embodiment of the present disclosure;

[0022] Figure 9 Schematically shows a structural block diagram of an intelligent agent of artificial intelligence according to an embodiment of the present disclosure; and

[0023] Figure 10 Schematically shows a block diagram of an electronic device suitable for implementing the training method of a large model and the interaction method based on artificial intelligence according to an embodiment of the present disclosure. Detailed Embodiments

[0024] The following describes exemplary embodiments of the present disclosure with reference to the accompanying drawings. Various details of the embodiments of the present disclosure are included to facilitate understanding, and they should be considered merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, descriptions of well-known functions and structures are omitted below for clarity and conciseness.

[0025] Understanding, generation, logic, and memory are four important capabilities of a large model. The memory capability is an essential aspect reflecting the intelligence of the large model. The large model itself actually does not have the memory capability, but some memory mechanisms can be used to endow the large model with this memory capability.

[0026] For example, during the process of a user's conversation with the large model, the historical conversation content generated by continuously splicing the previous context is used for output processing of the current round of conversation, so that the feedback content output by the large model is combined with the user's relevant knowledge. Thus, the specific memory capability of the large model for this user is reflected.

[0027] However, by splicing historical conversation content to enable the large model to have memory information related to the user, it will cause the large model to process a large amount of historical conversation content, which will in turn increase the data processing volume of the large model, resulting in problems such as increased computing costs and reduced feedback efficiency.

[0028] In view of this, the present disclosure provides a training method for a large model, including: obtaining historical interaction conversations with a target object. When it is determined that the historical interaction conversation meets a predetermined processing condition, sample generation processing is performed on the historical interaction conversation to obtain training samples. The large model is trained using the training samples to obtain a large model including model knowledge matching the target object.

[0029] Using this training method for the large model, by training the large model using the historical interaction conversations of the target object, the large model can learn relevant knowledge related to the target object and obtain model knowledge integrated into the large model parameters. Thus, while improving the specific memory ability of the large model for knowledge related to the target object, problems such as increased computing costs and feedback delays are avoided.

[0030] Figure 1 Schematically shows an exemplary system architecture to which an interaction method based on artificial intelligence, a training method for a large model, and an apparatus according to embodiments of the present disclosure can be applied.

[0031] It should be noted that Figure 1 The shown is only an example of a system architecture to which embodiments of the present disclosure can be applied, to help those skilled in the art understand the technical content of the present disclosure, but it does not mean that the embodiments of the present disclosure cannot be used in other devices, systems, environments or scenarios. For example, in another embodiment, an exemplary system architecture to which a content processing method and apparatus can be applied may include a terminal device, but the terminal device can implement the content processing method and apparatus provided by the embodiments of the present disclosure without interacting with a server.

[0032] As Figure 1 shown, the system architecture 100 according to this embodiment may include terminal devices 101, 102, 103, a network 104, and a server 105. The network 104 is used to provide a medium for a communication link between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0033] Users can use the terminal devices 101, 102, 103 to interact with the server 105 through the network 104 to receive or send messages, etc. Various communication client applications may be installed on the terminal devices 101, 102, 103, such as knowledge reading applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (only for example).

[0034] The terminal devices 101, 102, and 103 can be various electronic devices with a display screen and supporting web browsing, including but not limited to smartphones, tablets, laptop computers, desktop computers, and so on.

[0035] The server 105 can be a server that provides various services, such as a background management server (for example only) that supports the content browsed by users using the terminal devices 101, 102, and 103. The background management server can analyze and process data such as user requests received, and feedback the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0036] It should be noted that the artificial intelligence-based interaction method and the large model training method provided by the embodiments of the present disclosure can generally be executed by the terminal devices 101, 102, or 103. Correspondingly, the artificial intelligence-based interaction device and the large model training device provided by the embodiments of the present disclosure can also be set in the terminal devices 101, 102, or 103.

[0037] Alternatively, the artificial intelligence-based interaction method and the large model training method provided by the embodiments of the present disclosure can generally also be executed by the server 105. Correspondingly, the artificial intelligence-based interaction method and the large model training method provided by the embodiments of the present disclosure can generally be set in the server 105. The artificial intelligence-based interaction method and the large model training method provided by the embodiments of the present disclosure can also be executed by a server or a server cluster different from the server 105 and capable of communicating with the terminal devices 101, 102, 103 and / or the server 105. Correspondingly, the artificial intelligence-based interaction device and the large model training device provided by the embodiments of the present disclosure can also be set in a server or a server cluster different from the server 105 and capable of communicating with the terminal devices 101, 102, 103 and / or the server 105.

[0038] It should be understood that Figure 1 the numbers of the terminal devices, the network, and the server in are merely illustrative. According to the implementation requirements, any number of terminal devices, networks, and servers can be provided.

[0039] It should be noted that the sequence numbers of the operations in the following methods are only used as representations of the operations for description, and should not be regarded as indicating the execution order of the respective operations. Unless explicitly stated, the method does not need to be executed exactly in the order shown.

[0040] Figure 2 Schematically shows a flowchart of a large model training method according to an embodiment of the present disclosure.

[0041] AsFigure 2 As shown, the method includes operations S210 to S230.

[0042] In operation S210, obtain the historical interaction dialogue with the target object.

[0043] In operation S220, when it is determined that the historical interaction dialogue meets the predetermined processing conditions, perform sample generation processing on the historical interaction dialogue to obtain training samples.

[0044] In operation S230, use the training samples to train the large model to obtain a large model including model knowledge matching the target object.

[0045] In the technical solution of the present disclosure, the collection, storage, use, processing, transmission, provision, disclosure, and application of the user's personal information involved all comply with the provisions of relevant laws and regulations, take necessary confidentiality measures, and do not violate public order and good customs.

[0046] In the technical solution of the present disclosure, before obtaining or collecting the user's personal information, the user's authorization or consent is obtained.

[0047] The target object can refer to a user, an object used for human-computer interaction with the large model.

[0048] The historical interaction dialogue can be an interaction dialogue generated by the target object interacting with the large model during a historical period, but is not limited to this. It can also be an interaction dialogue generated by the target object interacting with other large models, as long as it is an interaction dialogue related to the target object. The content of the historical interaction dialogue is not limited. For example, it can be one or more of text, voice, video, or image.

[0049] The predetermined processing conditions can include conditions for determining whether to perform sample generation processing on the historical interaction dialogue. For example, it can be a pre-generated sample rule. If the historical interaction dialogue conforms to the sample rule, it does not meet the predetermined processing conditions. If the historical interaction dialogue does not conform to the sample rule, it meets the predetermined processing conditions. When the historical interaction dialogue meets the predetermined processing conditions, sample generation processing needs to be performed on the historical interaction dialogue. When the historical interaction dialogue does not meet the predetermined processing conditions, the historical interaction dialogue can be directly used as a training sample.

[0050] Sample generation processing can refer to information processing, such as information expansion, rewriting, etc., but is not limited to this. It can also be processing such as noise reduction and deletion of redundant data, or information format conversion.

[0051] For example, the historical interaction dialogue A includes: "What is my job?" output by the target object and "You are a programmer" output by the large model. This historical interaction dialogue A conforms to the sample rules and does not meet the predetermined processing conditions, so the historical interaction dialogue A can be directly used as a training sample.

[0052] Also for example, the historical interaction dialogue B includes: "I found a new job today and I'm very happy" output by the target object; "Really? Congratulations. What job did you find?" output by the large model; "Programmer" output by the target object; and "Wish you smooth development" output by the large model subsequently. This historical interaction dialogue B does not conform to the sample rules and meets the predetermined processing conditions. Sample generation processing is performed on this historical interaction dialogue B to generate the historical interaction dialogue A as a training sample.

[0053] Using sample generation processing can make the training samples conform to the sample format and training requirements for large model training, and improve the training effect of the large model.

[0054] Optionally, the large model can include one or a combination of more of a large language model (LLM), a large vision model (LVM), or a multimodal large model (MLM). As long as it is a large model with general understanding and information generation capabilities, for example, it can perform conventional human-computer interaction, which will not be elaborated here.

[0055] Training the large model with the training samples obtained from interacting with the target object can enable the large model to learn model knowledge related to the target object. For example, adjusting the parameters in the large model and taking the adjusted parameters as model knowledge. It can simulate the ability of the human neural network to store memory information, so that the model knowledge of the trained large model has memory information related to the target object, achieving the effect of personalized memory. Thus, when applying the trained large model to the scenario of interacting with the target object, it can use the model knowledge to have an effective dialogue with the target object that meets the personalized needs of the target object, while improving the communication ability of the large model for specific objects and solving problems, and avoiding problems such as increased computational costs and feedback delays.

[0056] The above has made an overall description of the training of the large model. The following will specifically describe how to obtain historical interaction dialogues.

[0057] The candidate historical interaction dialogue generated by interacting with the target object can be directly used as the historical interaction dialogue. However, it is not limited to this. The candidate historical interaction dialogue can also be screened to obtain the historical interaction dialogue.

[0058] According to an embodiment of the present disclosure, when performing as Figure 2Before the operation S210 shown, the training method of the large model may further include an operation of obtaining candidate historical interaction dialogues for interacting with the target object. In the case where the semantic information representation of the candidate historical interaction dialogue represents positive sentiment information, based on the candidate historical interaction dialogue, a historical interaction dialogue is obtained.

[0059] Positive sentiment information can represent information with a positive attitude. For example, when the target object outputs content such as "You are amazing" or "So humorous", it indicates that the target object expresses positive sentiment information. Opposite to positive sentiment information is negative sentiment information. For example, when the target object outputs content such as "I'm so disappointed in you" or "That makes me so sad", it reflects that the target object expresses negative sentiment information.

[0060] Content representing positive sentiment information can be extracted from the candidate historical interaction dialogue to determine whether the candidate historical interaction dialogue is used as a historical interaction dialogue. For example, a dictionary is established in advance, and the dictionary includes reference words representing positive sentiment information. The dictionary is used to match keywords in the candidate historical interaction dialogue, such as entity words, verbs, or adjectives. If the matching result indicates that there are reference words in the dictionary in the candidate historical interaction dialogue, it is determined that the candidate historical interaction dialogue can be used as a historical interaction dialogue. If the matching result indicates that there are no reference words in the dictionary in the candidate historical interaction dialogue, it is determined that the candidate historical interaction dialogue can be deleted.

[0061] The semantic information of the candidate historical interaction dialogue can also be extracted. In the case where the semantic information represents positive sentiment information, the candidate historical interaction dialogue is used as a historical interaction dialogue. For example, a semantic recognition model is used to extract the semantic information of the candidate historical interaction dialogue. The model structure of the semantic recognition model is not limited. For example, it can include a convolutional neural network, an encoder-decoder, etc. The semantic information is classified to determine the classification result. A classifier, such as a support vector machine (SVM) or a classification model, can be used to classify the semantic information. In the case where the classification result indicates that the semantic information represents positive sentiment information, the candidate historical interaction dialogue is used as a historical interaction dialogue.

[0062] Preferably, the above two means can be used simultaneously. In the case where the matching result indicates that there are reference words in the dictionary in the candidate historical interaction dialogue and the classification result indicates that the semantic information represents positive sentiment information, the candidate historical interaction dialogue is used as a historical interaction dialogue. In the case where the matching result indicates that there are no reference words in the dictionary in the candidate historical interaction dialogue or the classification result does not indicate that the semantic information represents positive sentiment information, the candidate historical interaction dialogue is deleted.

[0063] Exemplarily, the semantic information represents candidate historical interactive dialogues of positive emotional information, which generally includes emotional memory or logical memory. Emotional memory can be understood as the memory of having specific emotional fluctuations towards certain things. Logical memory, also known as procedural memory, can be understood as the memory of processing tasks, such as the memory of doing exercises.

[0064] Filtering out such candidate historical interactive dialogues as historical interactive dialogues for training the large model can increase the amount of knowledge related to emotional memory or logical memory in the model's knowledge, improve the large model's ability to simulate the human brain's storage of emotional memory and logical memory, and thus improve the training effect. In addition, screening candidate historical interactive dialogues by combining the two means of dictionary matching and semantic analysis can improve the effectiveness of historical interactive dialogues, and thus improve the training efficiency while avoiding the waste of training resources caused by ineffective training.

[0065] According to an embodiment of the present disclosure, as described above, for the obtained historical interactive dialogues, the rule matching method can be used to determine whether the historical interactive dialogue meets the predetermined processing conditions. For example, the question-and-answer form rule can be set, and the question-and-answer content rule can also be set. In the case where the historical interactive dialogue conforms to the question-and-answer form rule or the question-and-answer content rule, it is determined that the historical interactive dialogue does not meet the predetermined processing conditions, otherwise it is determined that the historical interactive dialogue meets the predetermined processing conditions.

[0066] By setting a normative reference standard as the matching rule, the training samples that conform to the rule can be accurately determined as the training samples that do not require sample generation processing. Thereby, while improving the effectiveness and normativity of the training samples, the judgment difficulty is reduced.

[0067] The classifier can also be used to classify the historical interactive dialogues to determine whether the historical interactive dialogue meets the predetermined processing conditions. Furthermore, it is determined whether sample generation processing needs to be performed on the historical interactive dialogue.

[0068] For example, after performing the operation S210 as shown in Figure 2 , the training method of the large model can further include: performing corpus recognition on the historical interactive dialogue to obtain a recognition result. Based on the recognition result, it is determined whether the historical interactive dialogue meets the predetermined processing conditions.

[0069] The classifier such as a support vector machine (SVM) or a classification model can be used to perform corpus recognition on the historical interactive dialogue to obtain a recognition result. For example, the recognition result can include the probability of indicating whether the predetermined processing conditions are met. In the case where the recognition result is greater than the preset threshold, it is determined that the historical interactive dialogue meets the predetermined processing conditions. In the case where the recognition result is less than or equal to the preset threshold, it is determined that the historical interactive dialogue does not meet the predetermined processing conditions.

[0070] Compared with the method of matching rules, determining whether a historical interactive dialogue meets a predetermined condition by using the above method can improve the recognition accuracy while improving the processing efficiency and application scope.

[0071] The following will elaborate on the sample generation process for historical interactive dialogues that meet the predetermined processing conditions.

[0072] According to an embodiment of the present disclosure, for the operation S220 as shown in Figure 2 obtaining training samples through sample generation processing of historical interactive dialogues may include: using a sample generation large model to perform sample generation processing on historical interactive dialogues to obtain training samples.

[0073] Optionally, a sample generation large model may be used to perform sample generation processing on historical interactive dialogues in a chain of thought manner to obtain training samples. However, it is not limited thereto. One or more of deep learning models such as convolutional neural networks, encoder-decoders, or long short-term memory networks may also be used to perform sample generation processing on historical interactive dialogues to obtain training samples.

[0074] The sample generation large model may include one or more combinations of a large language model (LLM), a large vision model (LVM), or a multimodal large model (MLM). As long as it can perform sample generation processing in the way of artificial intelligence generated content (AIGC), for example, a large model for sample generation processing that takes historical interactive dialogues as input data and training samples as output data, will not be elaborated here.

[0075] The chain of thought manner may refer to using a sample generation large model to execute multiple tasks, filling the processing order of the multiple tasks in the form of a chain of thought in the prompt information prompt, so that the sample generation large model executes tasks according to the processing order reflected by the chain of thought and achieves the expected effect.

[0076] For example, the chain of thought may include tasks, task execution orders, and expected results such as "extracting key information, expanding or rewriting the key information to generate labels for training samples. Generating questions based on the labels. Using the questions as sample input data for training samples. Obtaining training samples based on the sample input data and labels".

[0077] Exemplarily, the key information in the historical interactive dialogue may characterize the emotional information or logical knowledge information of the target object.

[0078] Exemplarily, the above-mentioned chain of thought can be generated when the historical interaction dialogue includes text Q&A content. When the historical interaction dialogue includes multimodal Q&A content, the following chain of thought can be used to guide the sample generation large model to execute multiple tasks.

[0079] For example, the chain of thought may include tasks such as "determine the image to be processed and the processed image from the historical interaction dialogue. Generate the label of the training sample based on the processed image. Generate the sample input data of the training sample based on the image to be processed and the image description information in the historical interaction dialogue", the task execution order, and the expected results.

[0080] Using the method of the sample generation large model to obtain the training sample can improve the generation speed of the training sample. By adding the chain of thought to the prompt information, it can reasonably guide the sample generation large model to execute multiple tasks according to the expected execution order and achieve the expected results. Thereby improving the generation effect of the sample generation large model and reducing the probability of the sample generation large model having hallucinations.

[0081] The sample generation large model can be used to perform sample generation processing on the historical interaction dialogue to obtain the training sample. However, it is not limited to this. Other deep learning models and tools can also be directly used to process the historical interaction dialogue to obtain the training sample.

[0082] For example, when the historical interaction dialogue includes text Q&A content, key information in the historical interaction dialogue can be extracted using dictionary matching or an information extraction model. Optionally, the key information represents the emotional information or logical knowledge information of the target object. The key information is rewritten using a rewriting model such as an encoder-decoder to obtain the label of the training sample. Another rewriting model is used to generate a question based on the label as the sample input data of the training sample.

[0083] Processing the historical interaction dialogue in the above manner can transform the historical interaction dialogue into a sample format suitable for model training, thereby improving the training efficiency and training effect of training the large model using this training sample.

[0084] Also, for example, when the historical interaction dialogue includes multimodal Q&A content, an image recognition model is used to determine the image to be processed and the processed image from the historical interaction dialogue. Based on the processed image, the label of the training sample is generated. For example, the processed image is used as the label of the training sample. However, it is not limited to this. Post-processing such as noise reduction, distortion correction, or segmentation can also be performed on the processed image and then used as the label of the training sample. An information extraction model is used to determine the image description information in the historical interaction dialogue. Based on the image to be processed and the image description information in the historical interaction dialogue, the sample input data of the training sample is generated.

[0085] Figure 3 Schematically shows a schematic flow diagram of generating training samples for multi-modal question-answering content according to an embodiment of the present disclosure.

[0086] As Figure 3 shown, the historical interaction dialogue includes multi-modal question-answering content. For example, the target object inputs the processed image 310 including "a small car" and the text content 321 "Do you know what this image is generated based on?". The robot loaded with the large model outputs "I don't know". Then the target object inputs the image to be processed 330 including "a small animal" and the text content 322 "The small car image is generated by using the small animal image for image-to-image processing. You need to study hard and learn this skill to become smarter".

[0087] The processed image 310 can be extracted from the historical interaction dialogue as a label, the image to be processed 330 is extracted, and the image description information 323 "generating a small car image using a small animal image" in the historical interaction dialogue is extracted. Based on the image to be processed 330 and the image description information 323, the sample input data is obtained.

[0088] Processing multi-modal question-answering content in the above manner can improve the richness of training samples, enable the large model to learn the model knowledge of multi-modal content processing, and improve the intelligence and logical memory ability.

[0089] The above has described in detail how to obtain training samples. The following will describe in detail how to train a large model using training samples.

[0090] According to an embodiment of the present disclosure, for the operation S230 as Figure 2 shown, training the large model using training samples may include: inputting the sample input data of the training samples into the large model to obtain an output result. Comparing the output result with the label of the training sample, and adjusting the model parameters of the large model based on the comparison result, thereby making the output result increasingly approach the training sample.

[0091] For example, the output result and the label can be input into a loss function to obtain a loss value. Adjusting the model parameters of the large model based on the loss value until the loss value converges. The type of the loss function is not limited. For example, it can be a cross-entropy loss function.

[0092] Exemplarily, a memory matrix with adjustable parameters can be built into the large model. During the process of training using training samples, the model parameters of the large model except the memory matrix are frozen, and the adjustable parameters in the memory matrix are adjusted based on the output result and the label of the training sample until the output result approaches the label. Thereby, in addition to the model knowledge related to the target object, the memory matrix also maintains the general model knowledge of the large model and improves the training efficiency.

[0093] According to an embodiment of the present disclosure, the above is a description of a single training of a large model, but it is not limited thereto. Newly generated historical interaction dialogues can also be continuously used as training samples to optimize the training of the large model, resulting in a large model with continuously enriched model knowledge.

[0094] For example, for the operation S230 as Figure 2 shown, training the large model with training samples to obtain a large model including model knowledge matching the target object may include: in response to determining that the current moment is within a predetermined time period, training the large model with training samples to obtain a large model including model knowledge.

[0095] The predetermined time period may be, for example, the time period when the target object rests at night, but it is not limited thereto. It may also be the time period when the target object goes to work. As long as it is a time period that does not affect the interaction. The large model can be optimized and trained, such as fine-tuned, according to the predetermined time period and predetermined frequency, thereby improving the efficiency and ability of the large model optimization.

[0096] The large model can be loaded into the robot. When the robot is charging, historical interaction dialogues generated during the day and meeting the predetermined processing conditions with the target object are subjected to sample generation processing to obtain training samples, and the large model is fine-tuned using the training samples to implement a mechanism for automatically organizing memories when humans sleep every day. When the fine-tuning is completed, the dialogue between the robot and the target object will be more biased towards the expression mode of interest of the target object.

[0097] Exemplarily, training the large model during the predetermined time period can control the iteration period of the large model. For example, through corpus alignment and fine-tuning at the day level, meaningful and valuable logical information or emotional information can be summarized from the historical interaction dialogues with the target object, and the knowledge and experience are quickly trained into the neural network or memory matrix of the large model, thereby improving the task processing ability of the large model for specific tasks or specific scenarios.

[0098] Figure 4 Schematically shows a flowchart of a method for training a large model according to another embodiment of the present disclosure.

[0099] As Figure 4As shown, obtain a candidate historical interaction dialogue 420 that interacts with the target object 410. When it is determined that the semantic information of the candidate historical interaction dialogue 420 represents positive emotional information, based on the candidate historical interaction dialogue 420, obtain a historical interaction dialogue 430. When it is determined that the historical interaction dialogue 430 meets the predetermined processing conditions, perform sample generation processing on the historical interaction dialogue 430 to obtain a training sample 440. In response to the current moment being within a predetermined time period, determine the large model M411 of the target object from the large model set M410, and use the training sample 440 to fine-tune the large model M411 of the target object to obtain an optimized large model.

[0100] The above describes the training method of the large model. The following will detail the actual application of the trained large model.

[0101] Figure 5 Schematically shows a flowchart of an interaction method based on artificial intelligence according to an embodiment of the present disclosure.

[0102] As Figure 5 shown, the method includes operations S510 to S520.

[0103] In operation S510, in response to receiving an interaction request from the target object, determine a target large model that matches the target object from multiple candidate large models.

[0104] In operation S520, input the interaction content into the target large model to obtain feedback content for the interaction content.

[0105] The interaction request carries the interaction content of the target object.

[0106] The target large model includes model knowledge that matches the target object, and the model knowledge is obtained by using the large model training method described above.

[0107] The large model obtained by using the above large model training method has model knowledge that matches the target object, can supplement the emotional memory and logical memory that existing large models cannot achieve, and trains relevant memory knowledge into the parameters or memory matrix of the neural network, so that while the target large model reflects memory capabilities such as emotional memory and logical memory, it effectively reduces the cost and latency of a single interaction of the large model.

[0108] The above gives an overall description of how to perform human-computer interaction. The following will describe how to screen a target large model that matches the target object.

[0109] Based on the object identification information of the target object and the object model mapping relationship, a target large model that matches the target object can be determined from multiple candidate large models. The object model mapping relationship represents the matching relationship between the object identification information and the candidate large models.

[0110] The object identification information can be information used to uniquely identify the target object. For example, the account number, QR code, or other information of the object identification information.

[0111] An object model mapping relationship used to characterize the relationship between the object and the large model can be established in advance. Based on the object model mapping relationship and the object identification information, a target large model that matches the target object can be determined from multiple candidate large models.

[0112] The object model mapping relationship is used to map the target object to the large model, improving the screening efficiency of the large model, thereby improving the interactive feedback efficiency of the large model.

[0113] The following will elaborate on how to further generate feedback content using the target large model.

[0114] Optionally, the interaction content can be input into the target large model to obtain feedback content for the interaction content. However, it is not limited to this. The previous content associated with the interaction content can also be concatenated with the current round of interaction content and input into the target large model to obtain feedback content for the interaction content.

[0115] The previous content can be the historical interaction previous text within a predetermined time interval from the interaction content. However, it is not limited to this. It can also be the content obtained by performing artificial intelligence-based generation processing on the historical interaction previous text. For example, the summary or key information obtained from the historical interaction previous text can be used as the previous content.

[0116] The previous content and the interaction content can also be filled into the prompt information template to generate prompt information. The prompt information is input into the target large model to obtain feedback content for the interaction content.

[0117] Figure 6 Schematically shows a flowchart of an interaction method based on artificial intelligence according to another embodiment of the present disclosure.

[0118] As Figure 6 shown, based on the object identification information of the target object 610 and the object model mapping relationship, a target large model M611 that matches the target object can be determined from multiple candidate large models M610.

[0119] As Figure 6 shown, the previous content 640 associated with the interaction content 630 is obtained from the historical interaction previous text 620. Based on the previous content 640 and the interaction content 630, prompt information is generated. The prompt information is input into the target large model M611 to obtain feedback content for the interaction content.

[0120] Using the feedback content generation method provided by the embodiments of the present disclosure, by splicing the above content with the interaction content, historical conversation records can be added to the prompt information, thereby enabling the target large model to utilize the short-term memory mechanism during the interaction process. Thus, by combining two memory mechanisms with different model knowledge and prompt information, the memory capabilities such as emotional memory and logical memory of the target large model are further improved.

[0121] Figure 7 A block diagram of a training device for a large model according to an embodiment of the present disclosure is schematically shown.

[0122] As Figure 7 shown, the training device 700 for the large model includes: an acquisition module 710, a sample generation module 720, and a training module 730.

[0123] The acquisition module 710 is configured to acquire historical interaction conversations with the target object.

[0124] The sample generation module 720 is configured to perform sample generation processing on the historical interaction conversation to obtain training samples when it is determined that the historical interaction conversation meets a predetermined processing condition.

[0125] The training module 730 is configured to train the large model using the training samples to obtain a large model including model knowledge matching the target object.

[0126] According to an embodiment of the present disclosure, the training module includes: a training sub-module.

[0127] The training sub-module is configured to train the large model using the training samples to obtain a large model including model knowledge in response to determining that the current moment is within a predetermined time period.

[0128] According to an embodiment of the present disclosure, the sample generation module includes: a sample generation sub-module.

[0129] The sample generation sub-module is configured to perform sample generation processing on the historical interaction conversation using a sample generation large model to obtain training samples.

[0130] According to an embodiment of the present disclosure, the sample generation module includes: an extraction sub-module, a rewriting sub-module, and a question generation sub-module.

[0131] The extraction sub-module is configured to extract key information from the historical interaction conversation when the historical interaction conversation includes text Q&A content, where the key information represents the emotional information or logical knowledge information of the target object.

[0132] The rewriting sub-module is configured to rewrite the key information to obtain the label of the training sample.

[0133] A question generation sub-module, configured to generate questions based on tags as sample input data for training samples.

[0134] According to an embodiment of the present disclosure, the sample generation module includes: an image determination sub-module, a first image generation sub-module, and a second image generation sub-module.

[0135] The image determination sub-module is configured to determine a to-be-processed image and a processed image from the historical interaction dialogue when the historical interaction dialogue includes multimodal Q&A content.

[0136] The first image generation sub-module is configured to generate labels for training samples based on the processed image.

[0137] The second image generation sub-module is configured to generate sample input data for training samples based on the to-be-processed image and the image description information in the historical interaction dialogue.

[0138] According to an embodiment of the present disclosure, the training device for the large model further includes: an identification module and a result determination module.

[0139] The identification module is configured to perform corpus identification on the historical interaction dialogue to obtain an identification result.

[0140] The result determination module is configured to determine whether the historical interaction dialogue meets a predetermined processing condition based on the identification result.

[0141] According to an embodiment of the present disclosure, the training device for the large model further includes: a candidate acquisition module and a screening module.

[0142] The candidate acquisition module is configured to acquire candidate historical interaction dialogues for interacting with the target object.

[0143] The screening module is configured to obtain a historical interaction dialogue based on the candidate historical interaction dialogue when it is determined that the semantic information of the candidate historical interaction dialogue represents positive sentiment information.

[0144] Figure 8 Schematically shows a block diagram of an interaction device based on artificial intelligence according to an embodiment of the present disclosure.

[0145] As Figure 8 shown, the interaction device 800 based on artificial intelligence includes: a response module 810 and a feedback module 820.

[0146] The response module 810 is configured to determine a target large model that matches the target object from multiple candidate large models in response to receiving an interaction request from the target object. The interaction request carries the interaction content of the target object.

[0147] A feedback module 820 is configured to input interaction content into a target large model to obtain feedback content for the interaction content. The target large model includes model knowledge that matches a target object, and the model knowledge is obtained by a training device of the large model.

[0148] According to an embodiment of the present disclosure, the feedback module includes: an acquisition sub-module, a prompt generation sub-module, and a feedback sub-module.

[0149] The acquisition sub-module is configured to acquire previous content associated with the interaction content.

[0150] The prompt generation sub-module is configured to generate a prompt message based on the previous content and the interaction content.

[0151] The feedback sub-module is configured to input the prompt message into the target large model to obtain feedback content for the interaction content.

[0152] According to an embodiment of the present disclosure, the response module includes: a response sub-module.

[0153] The response sub-module is configured to determine a target large model that matches the target object from a plurality of candidate large models based on the object identification information of the target object and the object model mapping relationship, where the object model mapping relationship represents the matching relationship between the object identification information and the candidate large models.

[0154] Figure 9 A schematic block diagram of an intelligent agent of artificial intelligence according to an embodiment of the present disclosure is shown.

[0155] In an embodiment of the present disclosure, as Figure 9 shown, the AI intelligent agent 900 may include an input module 910, a processing module 920, and an output module 930.

[0156] The input module 910 is configured to receive input information.

[0157] The processing module 920 is configured to execute an interaction method based on a large language model provided by an embodiment of the present disclosure or execute a training method of a large model provided by an embodiment of the present disclosure to obtain output information based on the input information received by the input module.

[0158] The output module 930 is configured to output the output information obtained by the processing module.

[0159] According to an embodiment of the present disclosure, the input module 910 is responsible for receiving or perceiving information such as queries, requests, instructions, signals, or data from the external world (e.g., users or the external environment), and converting it into a format that the AI agent 900 can understand and process. The input module 910 is the primary link for the AI agent 900 to interact with the external world, enabling the AI agent 900 to efficiently and accurately obtain necessary "sensory" information from the external world and respond to this information.

[0160] In an example, the input module 910 can input the historical interaction dialogues with the target object described above, etc.

[0161] In an example, the processing module 920 is the core support for the AI agent 900's ability to handle complex tasks. The processing module 920 can execute the interaction method based on the large language model and the training method of the large language model described above.

[0162] In an example, the performance of the processing module 920 can be closely related to the large model on which the AI agent 900 is based. To fully utilize the capabilities of the large model, the internal structure of the processing module 920 can be designed to be highly configurable and extensible to handle various different types of tasks and requirements in real scenarios.

[0163] In an example, after the AI agent 900 obtains the historical interaction dialogue, when the processing module 920 determines that the historical interaction dialogue meets the predetermined processing conditions, the processing module 920 can perform sample generation processing on the historical interaction dialogue to obtain training samples; and use the training samples to train the large model to obtain a large model including model knowledge matching the target object, and transfer the large model to the output module 930.

[0164] It can be understood that although the large language model has excellent language understanding and generation capabilities, like humans, the tasks it can solve without any tools are very limited. When the AI agent 900 is given the ability to call tools, it can achieve tasks such as performing mathematical operations with the help of a calculator, performing data analysis with the help of Python, and obtaining weather forecasts with the help of a search engine.

[0165] In an example, the output module 930 can output the feedback content or the trained large model described above.

[0166] The AI agent 900 according to the embodiment of the present disclosure can simply and effectively improve the degree of intelligence and enhance flexibility and versatility.

[0167] 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.

[0168] According to an embodiment of the present disclosure, an electronic device includes: 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 to enable the at least one processor to execute the method as described above.

[0169] According to an embodiment of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the method as described above.

[0170] According to an embodiment of the present disclosure, a computer program product includes a computer program, and the computer program implements the method as described above when executed by a processor.

[0171] Figure 10 FIG. shows a schematic block diagram of an exemplary electronic device 1000 that can be used to implement the embodiments of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as, a laptop computer, a desktop computer, a workbench, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as, a personal digital processor, a cellular phone, a smart phone, a wearable device, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely exemplary and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0172] As Figure 10 shown, the device 1000 includes a computing unit 1001, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1002 or a computer program loaded from a storage unit 1008 into a random access memory (RAM) 1003. In the RAM 1003, various programs and data required for the operation of the device 1000 can also be stored. The computing unit 1001, the ROM 1002, and the RAM 1003 are connected to each other via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0173] A plurality of components in the device 1000 are connected to the input / output (I / O) interface 1005, including: an input unit 1006, such as a keyboard, a mouse, etc.; an output unit 1007, such as various types of displays, speakers, etc.; a storage unit 1008, such as a magnetic disk, an optical disc, etc.; and a communication unit 1009, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1009 allows the device 1000 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0174] The computing unit 1001 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1001 executes the various methods and processes described above, such as the training method of a large model and the AI-based interaction method. For example, in some embodiments, the training method of a large model and the AI-based interaction method can be implemented as a computer software program that is tangibly included in a machine-readable medium, such as the storage unit 1008. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device 1000 via the ROM 1002 and / or the communication unit 1009. When the computer program is loaded into the RAM 1003 and executed by the computing unit 1001, one or more steps of the training method of the large model and the AI-based interaction method described above can be executed. Alternatively, in other embodiments, the computing unit 1001 can be configured to execute the training method of the large model and the AI-based interaction method by any other suitable means (e.g., by means of firmware).

[0175] Various embodiments of the systems and techniques described above in this document can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGA), application-specific integrated circuits (ASIC), application-specific standard products (ASSP), systems-on-a-chip (SOC), complex programmable logic devices (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special or general programmable processor, and can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit the data and instructions to the storage system, the at least one input device, and the at least one output device.

[0176] The program code for implementing the methods of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general purpose computer, a special purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program codes cause the functions / operations specified in the flowchart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0177] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0178] In order 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 a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds 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).

[0179] The systems and techniques described herein can be implemented in a computing system including backend components (e.g., as a data server), or a computing system including middleware components (e.g., an application server), or a computing system including frontend components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described herein), or a computing system including any combination of such backend components, middleware components, or frontend components. The components of the system can be interconnected to each other by digital data communication in any form or medium (e.g., a communication network). Examples of communication networks include: local area network (LAN), wide area network (WAN), and the Internet.

[0180] A computer system can include a client and a server. The client and the server are generally remote from each other and typically interact through a communication network. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, or a server of a distributed system, or a server incorporating a blockchain.

[0181] It should be understood that various forms of the processes shown above can be used, with steps reordered, added, or deleted. For example, the steps recited in this disclosure can be executed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. There is no limitation herein in this regard.

[0182] The above specific embodiments do not constitute a limitation on the protection scope of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the protection scope of this disclosure.

Claims

1. A large model training method, comprising: Obtain historical interactive dialogues with the target object; When it is determined that the historical interactive dialogue meets a predetermined processing condition, performing sample generation processing on the historical interactive dialogue to obtain a training sample; as well as The large model is trained using the training samples to obtain a large model including model knowledge matching the target object.

2. The method according to claim 1, wherein: The step of training the large model using the training samples to obtain the large model including model knowledge matching the target object includes: In response to determining that the current moment is in a predetermined time period, the large model is trained using the training samples to obtain the large model including the model knowledge.

3. The method according to claim 1 or 2, wherein: The performing sample generation processing on the historical interactive dialogue to obtain a training sample includes: The sample generation large model is used to perform sample generation processing on the historical interactive dialogue to obtain the training sample.

4. The method according to claim 1 or 2, wherein: The performing sample generation processing on the historical interactive dialogue to obtain a training sample includes: In the case where the historical interactive dialogue includes text question-and-answer content, extracting key information from the historical interactive dialogue, wherein the key information represents the emotional information or logical knowledge information of the target object; Rewriting the key information to obtain a label of the training sample; and Based on the label generation problem, sample input data is used as the training sample.

5. The method according to claim 1 or 2, wherein: The performing sample generation processing on the historical interactive dialogue to obtain a training sample includes: In a case where the historical interactive dialogue includes multimodal question-and-answer content, determining the image to be processed and the processed image from the historical interactive dialogue; Based on the processed image, generating a label of the training sample; and Based on the image to be processed and the image description information in the historical interactive dialogue, sample input data of the training sample is generated.

6. The method according to any one of claims 1 to 6, further comprising: Performing corpus recognition on the historical interactive dialogue to obtain a recognition result; as well as Based on the recognition result, it is determined whether the historical interactive dialogue satisfies the predetermined processing condition.

7. The method according to any one of claims 1 to 6, further comprising: Acquire candidate historical interaction dialogues for interacting with the target object; as well as In a case where it is determined that the semantic information of the candidate historical interaction dialogue represents positive sentiment information, the historical interaction dialogue is obtained based on the candidate historical interaction dialogue.

8. An interactive method based on artificial intelligence, comprising: In response to receiving an interaction request of a target object, determining a target macromodel matching the target object from a plurality of candidate macromodels, wherein the interaction request carries interaction content of the target object; and Inputting the interactive content into the target macro model to obtain feedback content for the interactive content; The target large model includes model knowledge matching the target object, and the model knowledge is obtained by using the method as described in any one of claims 1 to 7.

9. The method according to claim 8, wherein: The step of inputting the interactive content into the target macro model to obtain feedback content for the interactive content includes: Acquire the above content associated with the interactive content; Generate prompt information based on the above content and the interactive content; and The prompt information is input into the target macro model to obtain feedback content for the interactive content.

10. The method according to claim 8 or 9, wherein: The step of determining a target large model matching the target object from a plurality of candidate large models comprises: Based on the object identification information and the object model mapping relationship of the target object, a target large model matching the target object is determined from the multiple candidate large models, wherein the object model mapping relationship represents the matching relationship between the object identification information and the candidate large models.

11. A large model training device, comprising: An acquisition module is used to acquire historical interactive dialogues with a target object; A sample generation module, configured to perform sample generation processing on the historical interactive dialogue to obtain a training sample when it is determined that the historical interactive dialogue meets a predetermined processing condition; as well as The training module is used to train the large model using the training samples to obtain the large model including model knowledge matching the target object.

12. An interactive device based on artificial intelligence, comprising: a response module, configured to determine a target macromodel matching the target object from a plurality of candidate macromodels in response to receiving an interaction request of the target object, wherein the interaction request carries interaction content of the target object; and A feedback module, used for inputting the interaction content into the target macro model to obtain feedback content for the interaction content; Wherein, the target large model includes model knowledge matching the target object, and the model knowledge is obtained using the device as described in claim 11.

13. An artificial intelligence-based intelligent agent, comprising: An input module, used for receiving input information; A processing module, configured to execute the method according to any one of claims 1 to 10 based on the input information received by the input module; An output module is used to output the output information obtained by the processing module.

14. 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 method according to any one of claims 1 to 10.

15. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-10.

16. A computer program product comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 10.

Images