Multi-modal multi-round dialogue agent construction method and device in IT operation and maintenance service field and medium

Abstract

The invention relates to a multi-modal multi-round dialogue agent construction method in the IT operation and maintenance service field, and the method comprises the steps: extracting multi-modal information from each historical IT operation and maintenance case, so as to form a training data set; performing preliminary training on the joint neural network to establish a multi-round dialogue agent, performing behavior judgment, tracking a dialogue state, and preliminarily determining intention and named entity information; based on the dialogue action, a DQN reinforcement learning model is adopted to perform reinforcement training on a multi-round dialogue agent, intention-based training modeling is performed on dialogue action generation, and a dialogue strategy is clarified; based on a vertical domain question and answer data knowledge base, a multi-round dialogue agent generates a natural language in a rule matching mode, supplemental information is fed back to a user, and then perfect dialogue state information is combined to clarify the intention of the user. The method has the beneficial effects that fuzzy information can be preliminarily provided according to the user, the user is gradually guided to focus specific intentions through dialogues, and a solid foundation is provided for the intelligence of the I-T operation and maintenance service.

Description

Technical Field

[0001] The present invention relates to the field of IT operation and maintenance service technology, and specifically to a method, device and medium for constructing a multimodal multi-round dialogue intelligent agent in the field of IT operation and maintenance services. Background Art

[0002] IT operation and maintenance refers to the work of service providers to maintain, manage and support the information technology systems of enterprises or organizations. It includes monitoring, maintenance, troubleshooting, upgrading and other operations of IT resources such as hardware, software, networks and databases to ensure the stability, security and availability of the system. IT operation and maintenance is vital to the normal operation of enterprises or organizations. It can help enterprises improve efficiency, reduce costs, and increase customer satisfaction, while also protecting the information security of enterprises. In IT operation and maintenance, users often communicate with service providers to solve technical problems encountered in daily life. Due to the uneven professional skills of customer service personnel, there is a problem that customer service personnel cannot understand the true intentions of users based on the pictures and text information provided by users, and there is a problem of misinterpreting the user's intentions, which makes it impossible for them to quickly provide humane answers and targeted solutions. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method, device and medium for constructing a multimodal multi-round dialogue intelligent agent in the field of IT operation and maintenance services, so as to overcome the deficiencies in the above-mentioned prior art.

[0004] The technical solution of the present invention to solve the above technical problems is as follows:

[0005] A method for constructing a multimodal multi-turn dialogue agent in the IT operation and maintenance service field includes the following steps:

[0006] S10. Obtain several historical IT operation and maintenance cases, and extract multimodal information including user requirement text, image data, and intent and entity information corresponding to each type of data from each historical IT operation and maintenance case to form a training data set;

[0007] S20. Preliminary training of the joint neural network based on the training dataset to establish a multi-round dialogue agent, perform behavioral judgment, track the dialogue status, and preliminarily clarify the intent and named entity information;

[0008] S30. Based on dialogue actions, the DQN reinforcement learning model is used to conduct reinforcement training on the multi-round dialogue agent to conduct intention-based training modeling for dialogue action generation and clarify the dialogue strategy.

[0009] S40, a vertical domain question and answer data knowledge base is established, multi-round dialogue intelligent agents are enabled to generate natural language in a rule matching manner based on the vertical domain question and answer data knowledge base, and supplementary information is fed back to the user, and clear user intent is realized by combining perfect dialogue state information.

[0010] Based on the above technical solutions, the application can also be improved as follows.

[0011] Further, the IT operation case is extracted from the enterprise historical work order.

[0012] Further, the joint neural network is preliminarily trained based on the training data set, specifically as follows:

[0013] Text processing: the Transformer model is used to encode the dialogue context in the user demand text to extract text semantic features;

[0014] Image processing: the DenseNet169 model is used to extract image features;

[0015] Feature fusion: the encoded text semantic features and image features are input into the joint neural network to perform joint model training of the semantic features and the image features, so as to identify the named entities in the text and the image, extract the named entity information, and preliminarily determine the intent.

[0016] Further, the contrast loss function used in the training of the joint neural network is the InfoNCE loss function:

[0017]

[0018] wherein f i is the embedding vector of the i-th image, g j is the embedding vector of the j-th text, sim(x, y) is a function for calculating the cosine similarity between two vectors, the cosine similarity is used to measure the similarity between the text and the image, the named entities in the image and the text are corresponded, and τ is a temperature parameter for controlling the distribution of the similarity score.

[0019] Further, the dialogue actions include: asking, supplementing, and stating.

[0020] Further, the DQN reinforcement learning model is used to reinforce the training of the multi-round dialogue intelligent agent, which is divided into two training, specifically: first, reinforcement training is performed for the coarse intent, and then reinforcement learning training is performed for the fine-grained intent of each coarse intent.

[0021] Further, a coarse classification controller classification, namely DC, is first performed to determine the approximate classification of the intent, and the controller strategy selects an action according to the current multi-modal dialogue state (S):

[0022]

[0023] Among them, π c is the controller strategy, represents the nonlinear action in the action space, that is, the k-th action under the c controller;

[0024] For each action ac, the multi-round dialogue agent can receive a reward or penalty, which then selects the multi-round dialogue agent's actions and guides the user to supplement information to better improve the intent recognition task:

[0025]

[0026] Then, based on the coarse classification controller, we conduct the refined reinforcement learning training. After that, based on the coarse intent classification under each coarse classification controller, we select the fine classification controller again to conduct the fine-grained intent classification reinforcement learning training, i.e., DP i ;

[0027] The detailed reinforcement learning training conditions are as follows: hardware—hardware missing, hardware—hardware damaged, software—software expired;

[0028] The sub-classification controller will select one (ai) as:

[0029] a i =argmax i Q i (A ij |S,π i )

[0030] Among them, Q i is i th Intention Strategy (π i )’s state-action function value, A ij is i th Intentional Strategy J th action, S is multimodal information, and the size of the action space of each strategy is N+1, where N is the number of intentions of the neural network. For each action The activated department policy receives rewards / penalties from the environment based on its response to the current dialogue state;

[0031] As shown below:

[0032]

[0033] Among them, s, ts, and p are the rewards / penalties for successful judgment, recognition of user intention, and repeated behavior, respectively.

[0034] Furthermore, the vertical field question-answering data knowledge base established is as follows:

[0035] Database structure:

[0036] Based on the enterprise SOP as the basic template, we build a dedicated database that includes IT operation and maintenance entities, solution templates, and multimodal dialogue sample templates. Based on the captured named entity information, we fill it into the multimodal dialogue sample template and output it to the user.

[0037] Reply template:

[0038] Build different response templates based on different actions. Design multiple response text templates for inquiry, supplement, and statement actions. Embed named entities and intents into the templates for personalized responses.

[0039] Dynamic update mechanism:

[0040] The data knowledge base continuously optimizes database content through user feedback and engineer review to ensure that the templates match the latest operation and maintenance scenarios.

[0041] Based on the above technical solution, the present invention also provides a multimodal multi-round dialogue intelligent agent construction device, including: a processor, a memory, and a computer program stored in the memory and runnable on the processor, and the processor implements the steps of the above construction method when executing the computer program.

[0042] Based on the above technical solution, the present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the above construction method are implemented.

[0043] The beneficial effects of the present invention are: this method constructs a multimodal multi-round dialogue agent through deep learning and natural language processing technology. When it is applied to IT operation and maintenance services, the multi-round dialogue agent comprehensively analyzes the user's expression in different dialogue rounds, not only identifying the direct intention of the current dialogue, but also inferring the user's potential intention and overall demand trend based on the context. That is, the multimodal multi-round dialogue agent can use the dialogue to gradually guide the user to focus on specific intentions based on the fuzzy information initially provided by the user, providing a solid foundation for IT operation and maintenance service intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Flowchart of the method for building a multimodal multi-turn conversational agent in the field of IT operation and maintenance services. DETAILED DESCRIPTION

[0045] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0046] Example 1

[0047] like Figure 1As shown in FIG, a method for constructing a multimodal multi-round dialogue agent in the IT operation and maintenance service field includes the following steps:

[0048] S10. Obtain several historical IT operation and maintenance cases, and extract multimodal information including user requirement text, image data, and intent and entity information corresponding to each type of data from each historical IT operation and maintenance case to form a training data set;

[0049] S20. Preliminary training of the joint neural network based on the training dataset to establish a multi-round dialogue agent, perform behavioral judgment, track the dialogue status, and preliminarily clarify the intent and named entity information;

[0050] S30. Based on dialogue actions, the DQN reinforcement learning model is used to conduct reinforcement training on the multi-round dialogue agent to conduct intention-based training modeling for dialogue action generation and clarify the dialogue strategy.

[0051] S40. Establish a vertical field question-answering data knowledge base. Based on this vertical field question-answering data knowledge base, allow multi-round dialogue agents to generate natural language in a rule-matching manner, and provide supplementary information to users, and then combine it with the complete dialogue status information to clarify user intentions.

[0052] Example 2

[0053] like Figure 1 As shown, this embodiment is a further optimization based on Example 1, and its details are as follows:

[0054] IT operation and maintenance cases are extracted from the company's historical work orders. The number of IT operation and maintenance cases can be 10, 100, 200, etc. This is just an example description. The specific number can be adjusted according to the accuracy of the training. IT operation and maintenance cases are preferably typical IT operation and maintenance cases. Image data can be system error screenshots, and entity information can be entity labels (BIO sequence annotations).

[0055] Example 3

[0056] like Figure 1 As shown, this embodiment is a further optimization based on embodiment 1 or 2, and its details are as follows:

[0057] The preliminary training of the joint neural network based on the training data set is as follows:

[0058] Text processing: The Transformer model is used to encode the conversation context in the user demand text to extract text semantic features;

[0059] Image processing: DenseNet169 model is used to extract image features;

[0060] Feature fusion: The encoded text semantic features and image features are input into a joint neural network to perform joint model training of semantic features and image features to identify named entities in text and images, extract named entity information, and preliminarily clarify intent.

[0061] During the training process, multiple sets of data matching text and images are input simultaneously.

[0062] Furthermore, the contrast loss function used in joint neural network training is the InfoNCE loss function:

[0063]

[0064] Where: f i is the embedding vector of the i-th image, g j is the embedding vector of the jth text, sim(x, y) is a function that calculates the cosine similarity between two vectors. Cosine similarity is used to measure the similarity between text and image, corresponding to the named entities in the image and text. τ is a temperature parameter used to control the distribution of similarity scores.

[0065] Example 4

[0066] like Figure 1 As shown, this embodiment is a further optimization based on any one of Embodiments 1 to 3, and its details are as follows:

[0067] Dialogue actions can be questions, supplements, or statements.

[0068] Furthermore, the DQN reinforcement learning model is used to conduct reinforcement training for the multi-round dialogue agent. The training is divided into two stages:

[0069] First, reinforcement training is performed on the coarse intent, and then reinforcement learning training of fine-grained intent is performed on each coarse intent.

[0070] Specifically, it can be understood as:

[0071] First, a coarse controller classification (DC) is performed to determine the approximate classification of intent, such as hardware issues and software issues. The controller strategy then selects actions based on the current multimodal dialogue state (S):

[0072]

[0073] Among them, π c is the controller strategy, represents the nonlinear action in the action space, that is, the k-th action under the c controller;

[0074] For each action ac, the multi-round dialogue agent receives a reward or penalty, called an external reward. This reward then selects the multi-round dialogue agent's actions and guides the user to supplement information to better improve the intent recognition task:

[0075]

[0076] Then, based on the coarse classification controller, we conduct the refined reinforcement learning training. After that, based on the coarse intent classification under each coarse classification controller, we select the fine classification controller again to conduct the fine-grained intent classification reinforcement learning training, i.e., DP i .

[0077] Specifically, the detailed reinforcement learning training conditions are as follows: hardware—hardware missing, hardware—hardware damaged, software—software expired;

[0078] The sub-classification controller will select one (ai) as:

[0079] a i =argmax i Q i (A ij |S,π i )

[0080] Among them, Q i is i th Intention Strategy (π i )’s state-action function value, A ij is i th Intentional Strategy J th action, S is multimodal information, and the size of the action space of each strategy is N+1, where N is the number of intentions of the neural network. For each action The activated department policy receives rewards / penalties from the environment based on its response to the current dialogue state;

[0081] As shown below:

[0082]

[0083] Among them, s, ts, and p are the rewards / penalties for successful judgment, recognition of user intention, and repeated behavior, respectively.

[0084] The multi-round dialogue agent estimates the expected cumulative rewards that can be obtained by starting from a given state and following a specific strategy based on reinforcement learning. By continuously updating the value function, the multi-round dialogue agent learns to distinguish which states and actions are more likely to lead to high rewards, thereby making more accurate intention recognition. Through hierarchical reinforcement learning strategies, it makes real-time intention judgments on user needs, achieves accurate decision-making, and reduces the need for manual intervention.

[0085] Example 5

[0086] like Figure 1 As shown, this embodiment is a further optimization based on any one of embodiments 1 to 4, and its details are as follows:

[0087] The vertical field question-answering data knowledge base established is as follows:

[0088] Database structure:

[0089] Using the enterprise SOP as the basic template, we build a dedicated database containing IT operation and maintenance entities (such as "server model" and "error code"), solution templates, and multimodal dialogue sample templates. Based on the captured named entity information, we fill it into the multimodal dialogue sample templates and output it to the user.

[0090] Reply template:

[0091] Build different response templates based on different actions. Design multiple response text templates for inquiry, supplement, and statement actions. Embed named entities and intents into the templates for personalized responses.

[0092] Dynamic update mechanism:

[0093] The data knowledge base continuously optimizes database content through user feedback and engineer review to ensure that the templates match the latest operation and maintenance scenarios.

[0094] The vertical field question-answering data knowledge base provides structured knowledge support for multi-round dialogue agents, improving the accuracy of solutions and user experience.

[0095] Example 6

[0096] A multimodal multi-round dialogue agent construction device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the construction method described in any one of embodiments 1 to 5 are implemented.

[0097] Example 7

[0098] A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the construction method described in any one of embodiments 1 to 5 are implemented.

[0099] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for constructing a multimodal multi-round dialogue agent in the IT operation and maintenance service field, characterized by: The steps include: S10. Obtain several historical IT operation and maintenance cases, and extract multimodal information including user requirement text, image data, and intent and entity information corresponding to each type of data from each historical IT operation and maintenance case to form a training data set; S20. Preliminary training of the joint neural network based on the training dataset to establish a multi-round dialogue agent, perform behavioral judgment, track the dialogue status, and preliminarily clarify the intent and named entity information; S30. Based on dialogue actions, the DQN reinforcement learning model is used to conduct reinforcement training on the multi-round dialogue agent to conduct intention-based training modeling for dialogue action generation and clarify the dialogue strategy. S40. Establish a vertical field question-answering data knowledge base. Based on this vertical field question-answering data knowledge base, allow multi-round dialogue agents to generate natural language in a rule-matching manner, and provide supplementary information to users, and then combine it with the complete dialogue status information to clarify user intentions.

2. The method for constructing a multimodal multi-round dialogue agent in the IT operation and maintenance service field according to claim 1 is characterized in that: IT operation and maintenance cases are extracted from the company's historical work orders.

3. According to the method for constructing a multimodal multi-round dialogue agent in the IT operation and maintenance service field in claim 1, the preliminary training of the joint neural network based on the training data set is specifically as follows: Text processing: The Transformer model is used to encode the conversation context in the user demand text to extract text semantic features; Image processing: DenseNet169 model is used to extract image features; Feature fusion: The encoded text semantic features and image features are input into a joint neural network to perform joint model training of semantic features and image features to identify named entities in text and images, extract named entity information, and preliminarily clarify intent.

4. The method for constructing a multimodal multi-round dialogue agent in the IT operation and maintenance service field according to claim 3 is characterized in that: The contrast loss function used in joint neural network training is the InfoNCE loss function: Where: f i is the embedding vector of the i-th image, g j is the embedding vector of the jth text, sim(x, y) is a function that calculates the cosine similarity between two vectors. Cosine similarity is used to measure the similarity between text and image, corresponding to the named entities in the image and text. τ is a temperature parameter used to control the distribution of similarity scores.

5. The method for constructing a multimodal multi-round dialogue agent in the IT operation and maintenance service field according to claim 1 is characterized in that: Dialogue actions include: asking, supplementing, and stating.

6. The method for constructing a multimodal multi-round dialogue agent in the IT operation and maintenance service field according to claim 1 or 5, characterized in that: The DQN reinforcement learning model is used to conduct reinforcement training for multi-round dialogue agents, which is divided into two training phases: first, reinforcement training is performed on coarse intents, and then reinforcement learning training is performed on fine-grained intents for each coarse intent.

7. The method for constructing a multimodal multi-round dialogue agent in the IT operation and maintenance service field according to claim 6 is characterized in that: First, a coarse controller classification, DC, is performed to determine the approximate classification of the intent, and the controller strategy selects actions based on the current multimodal dialogue state (S): Among them, π c is the controller strategy, represents the nonlinear action in the action space, that is, the k-th action under the c controller; For each action ac, the multi-round dialogue agent can receive a reward or penalty, which then selects the multi-round dialogue agent's actions and guides the user to supplement information to better improve the intent recognition task: Then, based on the coarse classification controller, we conduct the refined reinforcement learning training. After that, based on the coarse intent classification under each coarse classification controller, we select the fine classification controller again to conduct the fine-grained intent classification reinforcement learning training, i.e., DP i ; The detailed reinforcement learning training conditions are as follows: hardware—hardware missing, hardware—hardware damaged, software—software expired; The sub-classification controller will select one (ai) as: a i =argmax i Q i (A ij |S,π i ) Among them, Q i is i th Intention Strategy (π i )’s state-action function value, A ij is i th Intentional Strategy J th action, S is multimodal information, and the size of the action space of each strategy is N+1, where N is the number of intentions of the neural network. For each action The activated department policy receives rewards / penalties from the environment based on its response to the current dialogue state; As shown below: Among them, s, ts, and p are the rewards / penalties for successful judgment, recognition of user intention, and repeated behavior, respectively.

8. The method for constructing a multimodal multi-round dialogue agent in the IT operation and maintenance service field according to claim 1 is characterized in that: The vertical field question-answering data knowledge base established is as follows: Database structure: Based on the enterprise SOP as the basic template, we build a dedicated database that includes IT operation and maintenance entities, solution templates, and multimodal dialogue sample templates. Based on the captured named entity information, we fill it into the multimodal dialogue sample template and output it to the user. Reply template: Build different response templates based on different actions. Design multiple response text templates for inquiry, supplement, and statement actions. Embed named entities and intents into the templates for personalized responses. Dynamic update mechanism: The data knowledge base continuously optimizes database content through user feedback and engineer review to ensure that the templates match the latest operation and maintenance scenarios.

9. Multimodal multi-round dialogue agent construction equipment, including: A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the construction method according to any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the construction method according to any one of claims 1 to 8 are implemented.

Images