Multi-round dialogue method, device, equipment and storage medium based on DST

By identifying user intentions, filling semantic slots and updating conversation history, the problem of incomplete information transmission in multiple rounds of conversations is solved, and the fluency and user experience of conversations are improved.

CN118627631BActive Publication Date: 2024-11-15ZHONGDIAN DATA IND CO LTD
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411110394.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-14
Publication Date
2024-11-15
Estimated Expiration
2044-08-14

AI Technical Summary

Technical Problem

During multiple rounds of dialogue, it is difficult for the model to maintain accurate tracking and effective utilization of the entire dialogue history, which affects the integrity of information transmission and poor fluency.

Method used

By obtaining the conversation information input by the user equipment, intent recognition and filling the preset semantic slot, it is determined whether the state transfer condition is met. If it is met, the historical context information is obtained and the generation model is generated, the response is generated and the context information is updated until the dialogue task is completed.

Benefits of technology

Effectively convey historical dialogue information, enhance the consistency and fluency of dialogue, ensure information integrity, and improve user experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118627631B_ABST
    Figure CN118627631B_ABST
Patent Text Reader

Abstract

The present application discloses a multi-round dialogue method, device, equipment and storage medium based on DST, which relates to the technical field of natural language processing. The multi-round dialogue method based on DST includes: obtaining first dialogue information input by a user device, performing intent recognition on the first dialogue information, and determining intent information; filling at least one preset semantic slot according to the intent information; if the preset state transition condition is satisfied according to the filling state of the preset semantic slot, then obtaining the historical context information of the dialogue, inputting the slot value and the historical context information of the preset semantic slot into the preset generation model, generating the second dialogue information, sending the second dialogue information to the user device, and updating the historical context information according to the first dialogue information and the second dialogue information; judging whether the dialogue task is completed according to the filling state of all preset semantic slots and the updated historical context information; if the dialogue task is completed, then ending the dialogue. The present application can improve the fluency of multi-round dialogues.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of natural language processing technology, and in particular to a multi-round dialogue method, device, equipment and storage medium based on DST. Background Art

[0002] In recent years, with the rapid development of deep learning technology, multi-round dialogue systems based on large models have made breakthrough progress in the field of natural language processing. Through deep neural networks, especially large-scale pre-trained models, these systems can generate natural and fluent responses in a wide range of dialogue scenarios, significantly improving the quality of human-computer interaction.

[0003] However, large models have certain limitations in their context understanding capabilities when processing long conversations. Due to limitations in memory or computing resources, the model may gradually forget early context information during the conversation, resulting in a lack of coherence in subsequent communication. Especially in the case of long texts or long conversations, information loss is prone to occur, making it difficult for the model to accurately track and effectively utilize the entire conversation history, affecting the integrity of information transmission during multiple rounds of conversations, resulting in poor fluency in multiple rounds of conversations.

[0004] Therefore, the inability to transmit historical conversation information during multi-round conversations leads to poor fluency in multi-round conversations, which has become a problem that needs to be solved urgently.

[0005] The above contents are only used to assist in understanding the technical solution of the present application and do not constitute an admission that the above contents are prior art. Summary of the invention

[0006] The main purpose of this application is to provide a multi-round dialogue method, device, equipment and storage medium based on DST, aiming to solve the technical problem that historical dialogue information cannot be transmitted during the multi-round dialogue process, resulting in poor fluency of the multi-round dialogue.

[0007] To achieve the above objectives, the present application proposes a multi-round dialogue method based on DST, and the multi-round dialogue method based on DST includes:

[0008] Acquire first dialogue information input by a user device, perform intent recognition on the first dialogue information, and determine intent information;

[0009] Fill at least one preset semantic slot according to the intent information, and determine whether a preset state transition condition is met according to the filling state of the preset semantic slot;

[0010] If the preset state transition condition is met, historical context information of the conversation is obtained, the slot value of the preset semantic slot and the historical context information are input into a preset generation model, second conversation information is generated, the second conversation information is sent to the user device, and the historical context information is updated according to the first conversation information and the second conversation information;

[0011] Determining whether the dialogue task is completed according to the filling status of all preset semantic slots and the updated historical context information;

[0012] If the dialogue task is completed, the dialogue ends.

[0013] In one embodiment, the step of filling at least one preset semantic slot according to the intent information includes:

[0014] Determine the intent type present in the intent information and the intent data corresponding to the intent type;

[0015] Get all preset semantic slots;

[0016] Determining, according to the intent type, from the preset semantic slots, a preset semantic slot that matches the intent type;

[0017] Fill the intent data corresponding to the intent type into the preset semantic slot corresponding to the intent type.

[0018] In one embodiment, the step of judging whether a preset state transition condition is satisfied according to the filling state of the preset semantic slot includes:

[0019] Determine the first semantic slot in all preset semantic slots whose filling status is that it has been filled;

[0020] Using the preset semantic slot whose filling state is unfilled as the second semantic slot;

[0021] Determining whether a next target semantic slot can be determined from the second semantic slot according to the first semantic slot;

[0022] If the next target semantic slot can be determined from the second semantic slot according to the first semantic slot, it is determined that the preset state transition condition is met;

[0023] If the next target semantic slot cannot be determined from the second semantic slot according to the first semantic slot, it is determined that the preset state condition is not satisfied.

[0024] In one embodiment, the step of determining whether the next target semantic slot can be determined from the second semantic slot according to the first semantic slot includes:

[0025] Inputting the first semantic slot into a preset state transition model, calculating a state transition value of each of the second semantic slots, wherein the state transition value is a quantitative indicator of the possibility of transitioning from the first semantic slot to each of the second semantic slots;

[0026] If there is a second semantic slot whose state transition value is greater than a preset state transition value, determining that a next target semantic slot can be determined from the second semantic slot according to the first semantic slot;

[0027] If there is no second semantic slot whose state transition value is greater than the preset state transition value, it is determined that the next target semantic slot cannot be determined from the second semantic slot according to the first semantic slot.

[0028] In one embodiment, the step of inputting the slot value of the preset semantic slot and the historical context information into a preset generation model to generate the second dialogue information further includes:

[0029] The second semantic slot with the highest state transition value among the second semantic slots is used as the next target semantic slot;

[0030] The slot value of the preset semantic slot and the historical context information are input into a preset generation model to generate second dialogue information related to the next target semantic slot, wherein the second dialogue information is dialogue information that inquires about the intention information corresponding to the next target semantic slot.

[0031] In one embodiment, after the step of determining whether the dialog task is completed according to the filling status of all preset semantic slots and the updated historical context information, the method further includes:

[0032] If the dialogue task is not completed, the third dialogue information input by the user device in response to the second dialogue information is obtained, the third dialogue information is used as the first dialogue information, and the execution step is returned to: performing intent recognition on the first dialogue information to determine intent information.

[0033] In one embodiment, after the step of ending the conversation if the conversation task is completed, the method further includes:

[0034] Sending a preset feedback request to the user equipment, and obtaining feedback information sent by the user equipment in response to the preset feedback request;

[0035] If the feedback information is positive feedback information, the historical context information is used as a positive sample, and the preset generation model is optimized based on the positive sample;

[0036] If the feedback information is negative feedback information, the historical context information is used as a negative sample, and the preset generation model is optimized based on the negative sample.

[0037] In addition, to achieve the above-mentioned purpose, the present application also proposes a multi-round dialogue device based on DST, and the multi-round dialogue device based on DST includes:

[0038] an identification module, configured to obtain first dialogue information input by a user device, perform intent identification on the first dialogue information, and determine intent information;

[0039] A first judgment module is used to fill at least one preset semantic slot according to the intention information, and judge whether a preset state transition condition is met according to the filling state of the preset semantic slot;

[0040] a generation module, configured to obtain historical context information of the conversation if the preset state transition condition is met, input the slot value of the preset semantic slot and the historical context information into a preset generation model, generate second conversation information, send the second conversation information to the user device, and update the historical context information according to the first conversation information and the second conversation information;

[0041] A second judgment module is used to judge whether the dialogue task is completed according to the filling status of all preset semantic slots and the updated historical context information;

[0042] The ending module is used to end the dialogue if the dialogue task is completed.

[0043] In addition, to achieve the above-mentioned objectives, the present application also proposes a multi-round dialogue device based on DST, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the multi-round dialogue method based on DST as described above.

[0044] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by the processor, the steps of the multi-round dialogue method based on DST as described above are implemented.

[0045] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the multi-round dialogue method based on DST as described above.

[0046] The present application obtains first dialogue information input by a user device, identifies the intent of the first dialogue information, and determines the intent information; fills at least one preset semantic slot according to the intent information, and determines whether a preset state transition condition is satisfied according to the filling state of the preset semantic slot; if the preset state transition condition is satisfied, obtains historical context information of the dialogue, inputs the slot value of the preset semantic slot and the historical context information into a preset generation model, generates second dialogue information, sends the second dialogue information to the user device, and updates the historical context information according to the first dialogue information and the second dialogue information; determines whether the dialogue task is completed according to the filling state of all preset semantic slots and the historical context information; if the dialogue task is completed, ends the dialogue. This application first obtains and parses the user's input, fills the semantic slots, and guides the advancement of the dialogue process by evaluating the state transition conditions. If the currently filled slot value meets specific conditions, it indicates that the dialogue can enter the next stage, thereby allowing the system to access historical context information, combine the historical context information and the slot value of the preset semantic slot, and generate the corresponding information of the system through the preset generation model. Subsequently, the response information is sent back to the user device, and the system updates the historical context information internally to reflect the latest dialogue status; throughout the process, it is also necessary to detect the filling status of all preset semantic slots and historical context information to determine whether the dialogue task has been completed. When it is determined that the dialogue task has been completed, the dialogue will officially end. In this application, through the state transfer mechanism, the historical state information can be saved as context and provided to the model as a prompt for the dialogue scenario, so that the model can respond based on the historical context information, thereby enhancing the memory of historical dialogues and generating more fluent replies. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0049] Figure 1 A flowchart of a multi-round dialogue method based on DST provided in Example 1 of the present application;

[0050] Figure 2 A flowchart of the second embodiment of the multi-round dialogue method based on DST of the present application is provided;

[0051] Figure 3 A flowchart of the third embodiment of the multi-round dialogue method based on DST of the present application is provided;

[0052] Figure 4 This is a schematic diagram of the module structure of a multi-round dialogue device based on DST in an embodiment of the present application;

[0053] Figure 5 Schematic diagram of the device structure of the hardware operating environment involved in the DST-based multi-round dialogue method in the embodiment of the present application.

[0054] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0055] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0056] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0057] The present application embodiment provides a multi-round dialogue method based on DST. Figure 1 , Figure 1 This is a flowchart of the first embodiment of the multi-round dialogue method based on DST of the present application.

[0058] In this embodiment, the multi-round dialogue method based on DST includes steps S10 to S50:

[0059] Step S10, obtaining first dialogue information input by the user device, performing intent recognition on the first dialogue information, and determining intent information;

[0060] It should be noted that the execution entity and user device can be a smart phone, computer, smart speaker or any other terminal device that can interact with the user; the first conversation information can be data in various forms, such as text information entered by the user through the keyboard, voice commands issued by the user through the microphone, which are then converted into processable text form, or in some application scenarios, the user may upload images or videos as part of the conversation, and the system converts them into descriptive text; the received raw input may need to be preprocessed, such as speech recognition, image recognition or natural language processing, to convert it into a form that the system can understand and further process.

[0061] It should be noted that the purpose of intent recognition is to understand the essential purpose of the user's request, such as checking the weather, booking a restaurant, or playing music. Through the dialogue state tracking model (such as LSTM neural network model, statistical model, etc.), the key information in the dialogue is extracted to identify the potential intention needs. For example, natural language processing (NLP) technology and machine learning models, such as recurrent neural network (RNN), long short-term memory network (LSTM) or more advanced Transformer architecture, can be used to classify text or converted voice text to determine the user's intent category; semantic slots refer to the key information units required to achieve specific functions, such as "date", "number of people" and "time" in restaurant reservations.

[0062] Step S20, filling at least one preset semantic slot according to the intention information, and judging whether a preset state transition condition is satisfied according to the filling state of the preset semantic slot;

[0063] It should be noted that DST (Dialogue State Tracking) is a key part of the task-based dialogue system. Its main purpose is to predict and maintain the current dialogue state based on the dialogue history information, and to obtain the user's true purpose based on the dialogue state. This process involves extracting the user's goals, intentions, and related slot values ​​from multiple rounds of dialogue, and formalizing this information into dialogue states for subsequent dialogue strategy selection and system action generation. The dialogue state is generally represented in the form of slot-value pairs.

[0064] Optionally, at the beginning of the conversation, the conversation state is initialized and all preset semantic slots are set to specific initial values. Based on the identified intent information, specific entities or attribute values ​​can be extracted from the user conversation and filled into the preset semantic slots. This action also relies on NLP technology, and a sequence labeling model (such as CRF, BiLSTM+CRF) may be used to identify and extract entities. State transfer is a key technology used to describe and control the conversation process in a multi-turn conversation system. It involves the process of determining the next conversation state based on the current conversation state and user input. The state usually includes information such as user intent, identified entities, and conversation history. State transfer enables the conversation system to understand the user's intent changes, maintain the continuity of the conversation, and make appropriate responses based on the conversation goals. By defining clear transfer rules and using natural language processing technology, state transfer helps the system navigate in complex conversation scenarios and provide accurate and timely answers, thereby improving user experience and conversation efficiency. The design and optimization of state transfer is an important part of building an efficient multi-turn conversation system.

[0065] In addition, it should be noted that in some complex scenarios, intent recognition and semantic slot filling may need to be performed alternately multiple times to ensure that user needs are fully understood. For example, the system may first recognize that the user wants to query the weather, but does not fill the semantic slot corresponding to the query weather. It is necessary to further inquire about the specific location and date and gradually fill the relevant slots. In this step, at least one preset semantic slot can be filled. The filled semantic slot is used to transfer from the current dialogue state to the next state. By checking whether all necessary semantic slots have been effectively filled, that is, whether enough information has been collected to meet the requirements of a specific dialogue state, based on the preset state transition conditions, it is verified whether each required semantic slot has been filled with a valid value; the preset state transition conditions are a set of rules that specify when and how to transfer from the current dialogue state to the next state; according to the above state evaluation of the semantic slot and the result of matching the preset state transition conditions, it is decided whether to continue the current dialogue process, ask for more information, or transfer to the next state.

[0066] Step S30: if the preset state transition condition is met, historical context information of the conversation is obtained, the slot value of the preset semantic slot and the historical context information are input into a preset generation model, second conversation information is generated, the second conversation information is sent to the user device, and the historical context information is updated according to the first conversation information and the second conversation information;

[0067] It should be noted that the historical context information includes previous conversation records, semantic slots, etc., which are used to ensure that the generated conversation information can take into account the information in the previous conversation to make the conversation more natural and coherent; the slot value of the preset semantic slot is the value filled in step S20; the preset generation model generates contextual and targeted second conversation information based on the input historical context and semantic slot value; by updating the historical context information, the dialogue system maintains the latest understanding of the conversation history and provides more accurate context for subsequent conversations.

[0068] In addition, it should be noted that the preset generation model needs to be trained with a large amount of conversation data to improve the accuracy and naturalness of the generated conversation information. In addition, continuous optimization of the model is the key to improving the quality of conversations. By comprehensively considering the historical context and the slot value of the current semantic slot, using the preset generation model to generate high-quality conversation information, and timely updating the system's understanding of the conversation history, the user's conversation experience and satisfaction can be significantly improved.

[0069] After the historical context information is updated, slots can be added, deleted, and updated according to the historical context information, and the slot values ​​of the current state can be stored.

[0070] Step S40, judging whether the dialogue task is completed according to the filling status of all preset semantic slots and the updated historical context information;

[0071] It should be noted that each semantic slot has a filling status, indicating whether the information point has been fully or partially collected. When all key semantic slots are filled correctly, the system believes that sufficient information has been collected to meet user needs. By analyzing historical context information, the system helps understand the overall context of the conversation and determine whether specific ending conditions have occurred or whether the user has expressed satisfaction or the intention to end the conversation. Combining the filling status of the semantic slot and the context information, if all necessary information has been collected and no issues that require further clarification or resolution are detected, the conversation task is considered to be completed.

[0072] In addition, it should be noted that if the user raises new requirements or has unresolved questions, even if the main semantic slot has been filled, the system should consider the dialogue task as unfinished and continue the dialogue to solve additional questions. When the dialogue task is determined to be completed, the system can end the dialogue, give a final confirmation or perform corresponding operations. If the dialogue task is not completed, the system will continue to ask or provide information until the completion condition is met.

[0073] Step S50: If the dialogue task is completed, the dialogue ends.

[0074] It should be noted that before the end of the conversation, the system can solicit feedback from the user to understand the quality of the conversation experience, which will help with subsequent system optimization and user experience improvement. The collected feedback information can be used to train the conversation system model, optimize the conversation process, and improve the user experience. The system can send a friendly closing message or confirmation message to the user to inform the user that the conversation has ended, thank the user for their time, and provide necessary follow-up operation instructions. In addition, the system state can be reset to the initial state to prepare for receiving new conversation requests, including clearing all temporary variables, resetting the semantic slot state, etc., to ensure that each conversation is independent and not affected by the previous conversation.

[0075] In a feasible implementation, the step B10 is also included:

[0076] Step B10, if the dialogue task is not completed, obtain the third dialogue information input by the user device in response to the second dialogue information, use the third dialogue information as the first dialogue information, and return to the execution step: perform intent recognition on the first dialogue information to determine the intent information.

[0077] It should be noted that if it is determined that more semantic slot information needs to be collected, the user provides incomplete or ambiguous information, or there is an unexpected turn in the conversation, then it can be evaluated that the conversation task is not yet completed. Once it is detected that the conversation task is not completed, it is necessary to wait for the response of the user device, that is, the third conversation message. The third conversation message is usually a direct response to the previous round of inquiry (the second conversation message), but it may also contain additional information actively provided by the user to clarify or supplement the previous conversation content.

[0078] After receiving the third dialogue message, it is not regarded as an independent input, but as the starting point of a new round of dialogue, that is, as the first dialogue message. This is to restart the dialogue process internally, taking the user's latest input as the latest state of the dialogue, so that the system can continue to advance the dialogue task based on the latest information. Subsequently, the steps of filling the semantic slots are re-executed, and information is gradually accumulated until all necessary semantic slots are correctly filled, thus completing the dialogue task.

[0079] It is understandable that this implementation is essentially an iterative process. The system will continuously obtain user feedback, use it as the starting point for a new round of dialogue, and try to fill the semantic slot until the dialogue task is finally completed. This mechanism ensures that even in complex dialogue scenarios, the system can gradually collect information through multiple iterations and user interactions, and ultimately achieve the dialogue goal. It not only improves the efficiency and success rate of the dialogue system, but also improves the user experience, making it feel more natural and smooth in interacting with the system.

[0080] In a feasible implementation manner, after step S50, steps S60 to S90 are further included:

[0081] Step S60, sending a preset feedback request to the user equipment, and obtaining feedback information sent by the user equipment in response to the preset feedback request;

[0082] It should be noted that after one or more rounds of conversations, the system automatically or according to preset rules sends a preset feedback request to the user device. The preset feedback request may include a series of questions or evaluation criteria, covering all aspects of the conversation, such as the fluency of the conversation process, the accuracy of the system response, the practicality of the information provided, etc., to collect the user's true feelings and opinions, and provide data support for subsequent system optimization.

[0083] After sending the preset feedback request, wait for and receive feedback information from the user's device. This information can include questionnaires filled out by the user, ratings, comments, or other forms of feedback. Feedback information can be quantitative, such as ratings, or qualitative, such as text comments or voice feedback, or a combination of multiple feedback forms. Feedback information is an important basis for system optimization and iteration. By analyzing this information, users' needs, preferences, and pain points can be identified, providing direction for future dialogue design and function upgrades.

[0084] Collect and analyze all received feedback information. Based on the results of feedback analysis, the system can adjust the dialogue strategy, update model parameters or improve UI design to improve the efficiency, accuracy and user satisfaction of the dialogue, and incorporate user feedback into the closed loop of system optimization, forming a continuous improvement cycle. This not only improves the performance of the model in the system, but also enhances user trust and satisfaction with the system.

[0085] Step S70: if the feedback information is positive feedback information, the historical context information is used as a positive sample, and the preset generation model is optimized based on the positive sample;

[0086] Step S80: If the feedback information is negative feedback information, the historical context information is used as a negative sample, and the preset generation model is optimized based on the negative sample.

[0087] It should be noted that when receiving positive feedback information from the user, the system will mark the "historical context information" related to this conversation as a positive sample. Positive samples are used as positive examples in the model optimization process to enhance the model's ability to make correct decisions in similar scenarios. Based on these positive samples, the preset generation model is retrained or fine-tuned. The weight parameters of the model are updated to increase the probability of generating similar positive sample outputs. In this way, the model gradually learns to more accurately predict and generate user-satisfactory responses when faced with the same or similar situations.

[0088] On the contrary, if the user provides negative feedback information, the relevant historical context information is marked as a negative sample. Negative samples reveal the deficiencies of the model in certain aspects and help identify error patterns or deviations so as to avoid similar mistakes in the future. Based on negative samples, the preset generation model is adjusted or trained in a targeted manner, and the model parameters are adjusted accordingly to reduce the possibility of similar errors in the future, thereby improving the overall performance of the model and user satisfaction.

[0089] In this embodiment, by sending a preset feedback request to the user device and obtaining feedback information sent by the user device in response to the preset feedback request, the user's real feedback is collected to understand the performance of the dialogue system in actual use and the user's satisfaction with the system function, performance and interaction design; if the feedback information is positive feedback information, the historical context information is used as a positive sample, and the preset generation model is optimized based on the positive sample; if the feedback information is negative feedback information, the historical context information is used as a negative sample, and the preset generation model is optimized based on the negative sample, and the user feedback is used to continuously optimize the preset generation model to improve the dialogue quality and user experience. In this embodiment, by effectively collecting and using user feedback, not only can errors be corrected in a timely manner, but also continuous learning and evolution can be achieved, serving users in a smarter and more efficient way, and realizing sustainable optimization and development of the dialogue system.

[0090] In this embodiment, first dialogue information input by a user device is obtained, intent recognition is performed on the first dialogue information, and intent information is determined; at least one preset semantic slot is filled according to the intent information, and whether a preset state transition condition is satisfied is determined according to the filling state of the preset semantic slot; if the preset state transition condition is satisfied, historical context information of the dialogue is obtained, the slot value of the preset semantic slot and the historical context information are input into a preset generation model, second dialogue information is generated, the second dialogue information is sent to the user device, and the historical context information is updated according to the first dialogue information and the second dialogue information; whether the dialogue task is completed is determined according to the filling state of all preset semantic slots and the historical context information; if the dialogue task is completed, the dialogue is ended. This application first obtains and parses the user's input, fills the semantic slots, and guides the advancement of the dialogue process by evaluating the state transition conditions. If the currently filled slot value meets specific conditions, it indicates that the dialogue can enter the next stage, thereby allowing the system to access historical context information, combine the historical context information and the slot value of the preset semantic slot, and generate the corresponding information of the system through the preset generation model. Subsequently, the response information is sent back to the user device, and the system updates the historical context information internally to reflect the latest dialogue status; throughout the process, it is also necessary to detect the filling status of all preset semantic slots and historical context information to determine whether the dialogue task has been completed. When it is determined that the dialogue task has been completed, the dialogue will officially end. In this application, through the state transfer mechanism, the historical state information can be saved as context and provided to the model as a prompt for the dialogue scenario, so that the model can respond based on the historical context information, thereby enhancing the memory of historical dialogues and generating more fluent replies.

[0091] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can be referred to the above introduction, and will not be repeated in the following. Figure 2, step S10 may include steps A10 to A40:

[0092] Step A10, determining the intent type existing in the intent information and the intent data corresponding to the intent type;

[0093] It should be noted that the intent type refers to the specific task or type of inquiry that the user wants to complete through the dialogue system. For example, the user may want to query the weather forecast, book a restaurant, purchase goods, etc., which usually relies on natural language understanding (NLU) technology, especially deep learning models, such as recurrent neural networks (RNN), long short-term memory networks (LSTM), Transformer models, etc., which can identify specific intent types from the text entered by the user; intent data refers to the entity information or parameters required to complete a specific intent, such as "place" and "date" when querying the weather, or "departure place", "destination", "departure date" when booking air tickets, etc., involving entity recognition (ER) and slot filling. Entity recognition is used to identify key entities in the text, while slot filling is to fill these entities into the corresponding intent slots to form structured information.

[0094] In addition, it should be noted that even if the intent type is correctly identified, the system cannot complete the user request if sufficient intent data is not extracted. For example, if the user wants to query the weather in Beijing tomorrow, but the system fails to extract the two key entities "Beijing" and "tomorrow", the query cannot be performed.

[0095] In this step, the system first receives instructions or requests from the user in natural language, performs lexical and grammatical analysis on the user input, converts it into a form that can be processed by the computer, uses the trained NLU model to identify the user's true intention type from the parsed input, further analyzes the input, identifies and extracts the entity information and parameters required to complete the intention, checks whether the extracted data is complete and reasonable, makes corrections if necessary, or asks the user to provide more information, and passes the identified intent type and extracted intent data to the system's back-end module in a structured form to perform corresponding operations or generate appropriate responses, ensuring that the system can accurately understand the user's intentions and obtain all necessary information required to complete the task.

[0096] Step A20, obtaining all preset semantic slots;

[0097] It should be noted that semantic slots are structural units used to store and organize key information in user requests in the dialogue system. Each slot usually represents a specific type of information, such as time, location, quantity, preference, etc., which is essential to completing the user's request. Semantic slots help the system to clarify what information is needed to meet the user's specific needs. They also serve as a bridge for interaction between the system and the user, guiding the user to gradually provide the missing information.

[0098] Preset semantic slots mean that these slots are predefined based on the expectations of possible user intentions and scenarios. For example, in a restaurant reservation system, possible preset semantic slots include "number of people", "date", "time" and "special requirements". During the system development phase, based on domain analysis and user needs, the designer defines a series of semantic slots, each of which corresponds to a type of information. When the system starts, all preset semantic slots are loaded into the memory, ready for subsequent dialogue processing. During operation, the system may dynamically adjust the use of slots according to the actual dialogue situation, such as adding new slots or changing the status of certain slots.

[0099] Step A30, determining a preset semantic slot matching the intent type from the preset semantic slots according to the intent type;

[0100] It should be noted that each intent type is usually associated with a specific set of semantic slots. For example, if the user's intention is to query the weather, the system will pay attention to slots such as "place" and "time". The system will filter out slots that match the current intent type from the preset list of semantic slots, and prepare to collect relevant information in subsequent conversations. During the system design phase, it is necessary to create a mapping table between intent types and semantic slots to clarify which slots correspond to each intent type. It is usually built based on domain knowledge and historical user behavior data. During the conversation, the system may need to dynamically adjust the slot selection based on the user's real-time feedback. For example, if the user changes the intent, the system needs to be able to adapt quickly and re-match the appropriate slot.

[0101] Through sophisticated intent type recognition and slot matching, the system can accurately locate the type of information that needs to be collected based on the user's specific needs, thereby improving conversation efficiency and user experience, enabling the conversation system to respond to users more intelligently and flexibly, and promoting the smooth progress of conversations.

[0102] Step A40, filling the intent data corresponding to the intent type into the preset semantic slot corresponding to the intent type.

[0103] This step fills the intent data extracted from the user input, such as location and time, into the corresponding preset semantic slots. In some cases, the user may not provide all the necessary information in the first input. At this time, it is necessary to dynamically adjust the filling strategy to guide the user to fill in the missing data by asking questions or prompting. For user input that cannot be recognized or understood, an error handling mechanism should be provided, such as requesting user clarification, providing default values, or skipping the current slot to ensure the fluency of the conversation.

[0104] In this embodiment, the intent type existing in the intent information and the intent data corresponding to the intent type are determined, so as to extract key information from the user input; all preset semantic slots are obtained to provide a target and framework for subsequent filling of intent information; the preset semantic slots matching the intent type are determined from the preset semantic slots according to the intent type, and which slots need to be filled by comparing the correlation between the intent type and the preset semantic slots; the intent data corresponding to the intent type is filled into the preset semantic slot corresponding to the intent type. This implementation parses the user input, identifies the intent type and its specific content contained therein, and selects the preset semantic slots corresponding thereto according to the type of user intent, and fills the extracted intent data into the corresponding preset semantic slots according to the principle of matching the intent type. By filling the preset semantic slots, user needs can be better understood, clear guidance can be provided for the subsequent dialogue process, and user needs can be responded to more intelligently and efficiently, driving the dialogue process toward the goal.

[0105] Based on the second embodiment of the present application, in the third embodiment of the present application, the same or similar contents as those in the above-mentioned second embodiment can refer to the above introduction, and will not be repeated in the following. Figure 3 , step S20 may include steps B10 to B50:

[0106] Step B10, determining the first semantic slot in all preset semantic slots whose filling status is that it has been filled;

[0107] It should be noted that at different stages of the conversation, the filling status of the preset semantic slots reflects the progress of the conversation and the degree of satisfaction of user needs. By monitoring the status of these slots, it is determined whether the conversation is ready for the next step or whether more information needs to be collected. Traverse all preset semantic slots and check whether each slot has been filled. If a slot contains complete and clear information, it is considered to be the first semantic slot that has been filled. The first semantic slot represents part of the goal that has been achieved in the conversation, which helps the system determine whether state transfer can be made or whether specific information needs to be continued to be collected. Each preset semantic slot should have a status mark to indicate its filling status. This may include states such as "unfilled", "partially filled", and "filled".

[0108] When determining whether a slot is the first semantic slot, it is necessary to check whether the filled information is complete and logical. For example, if the "time" slot only contains "morning" without a specific number of hours, it may not be considered "filled" yet, and it is necessary to return to obtain new first conversation information to obtain the slot value corresponding to the complete "time" slot.

[0109] Step B20, using the preset semantic slot whose filling state is unfilled as the second semantic slot;

[0110] It should be noted that the system will check all preset semantic slots and mark those slots that have not been filled as "second semantic slots". These slots represent the information that may need to be collected or asked in the next step. The determination of the second semantic slot helps the system clarify the next dialogue goal, that is, what information to collect to move the dialogue forward and ensure the coherence and efficiency of the dialogue. As the dialogue progresses, the set of second semantic slots may change. For example, when an originally unfilled slot is provided with information by the user, it is no longer a second semantic slot, and the remaining unfilled slots need to be re-evaluated.

[0111] Step B30, determining whether the next target semantic slot can be determined from the second semantic slot according to the first semantic slot;

[0112] It should be noted that by analyzing the information contained in the first semantic slot, it is evaluated whether the information is sufficient to deduce which information in the second semantic slot should be collected next. For example, if the user has provided the "restaurant location" and "number of diners", the system may be able to determine that the next step should be to ask "dining time" based on this information. This judgment is based on the preset dialogue process and logical rules, that is, the dependencies and order between the semantic slots preset by the system. If the information in the first semantic slot can meet the preset logical conditions, then the next target semantic slot can be determined.

[0113] It is understandable that a clear set of logical rules is needed to guide how to infer the next target semantic slot from the first semantic slot. These rules can be based on domain knowledge, historical conversation data, or machine learning models. The state transition conditions are usually based on the design of the conversation process. For example, the system will only ask about the "time" after the "location" and "number of people" are determined. This logic ensures the naturalness and coherence of the conversation. In some cases, the user's answer may change the expected flow of the conversation. For example, the user may provide the "time" information in advance. At this time, the system needs to be able to dynamically adjust the state transition logic to adapt to the user's actual answer. By continuously optimizing the logical rules and dynamic decision-making mechanism, it is ensured that each step of the conversation can be reasonably advanced based on the existing information, avoiding invalid or repeated inquiries, while also ensuring the coherence and efficiency of the conversation. By accurately judging the next target semantic slot, the conversation process can be guided more intelligently, improving user satisfaction and the success rate of the conversation.

[0114] In a feasible implementation, step B30 may include steps B31 to B33:

[0115] Step B31, inputting the first semantic slot into a preset state transition model, and calculating the state transition value of each second semantic slot, wherein the state transition value is a quantitative index of the possibility of transitioning from the first semantic slot to each second semantic slot;

[0116] It should be noted that the preset state transition model is a model used in the dialogue system to predict the possibility of transitioning from the current dialogue state to the next state. The first semantic slot is used as input, and the state transition value of each second semantic slot is calculated through the preset state transition model. The state transition value is a quantitative indicator that represents the possibility of transitioning from the current information state to a specific second semantic slot state. The state transition value can be a value between 0 and 1. The larger the value, the greater the possibility of transitioning from the first semantic slot to a specific second semantic slot. The preset state transition model needs to be trained with a large amount of dialogue data to learn the correlation and transition rules between different types of information.

[0117] Step B32, if there is a second semantic slot whose state transition value is greater than the preset state transition value, determining that a next target semantic slot can be determined from the second semantic slot according to the first semantic slot;

[0118] Step B33: If there is no second semantic slot whose state transition value is greater than the preset state transition value, it is determined that the next target semantic slot cannot be determined from the second semantic slot according to the first semantic slot.

[0119] It should be noted that the preset state transition value is a threshold value set by the system and is used to determine whether the state transition value is high enough to determine whether the next target semantic slot can be determined from the first semantic slot.

[0120] Check the state transition values ​​of all second semantic slots. If there is at least one second semantic slot with a state transition value greater than the preset state transition value, then it is considered that the next target semantic slot can be determined based on the first semantic slot. In other words, the possibility of transferring from the current information state to a specific second semantic slot is large enough to be used as the next action of the conversation.

[0121] If after checking the state transition values ​​of all second semantic slots, it is found that none of the state transition values ​​exceeds the preset state transition value, then it is determined that the next target semantic slot cannot be determined from the first semantic slot, that is, the current information state is not sufficient to support the next action of the conversation, and other strategies need to be adopted, such as sending follow-up questions or ambiguous clarification words to collect necessary information or adjust the conversation process. When it is determined that the next target semantic slot cannot be determined from the first semantic slot, it is necessary to re-ask the user or provide more detailed guidance to help the user supplement the missing information in order to reach the preset state transition value.

[0122] In this embodiment, by inputting the first semantic slot into the preset state transition model, the state transition value of each second semantic slot is calculated, wherein the state transition value is a quantitative index of the possibility of transferring from the first semantic slot to each second semantic slot, and a higher state transition value means that the probability of transferring from the current information state to a specific second semantic slot is greater; if there is a second semantic slot with a state transition value greater than the preset state transition value, it is determined that the next target semantic slot can be determined from the second semantic slot according to the first semantic slot; if there is no second semantic slot with a state transition value greater than the preset state transition value, it is determined that the next target semantic slot cannot be determined from the second semantic slot according to the first semantic slot. This embodiment provides a more scientific and quantitative method to determine whether the next target semantic slot can be determined by introducing the concepts of the preset state transition model and the state transition value, which not only improves the intelligent decision-making ability of the dialogue system, but also ensures the coherence and efficiency of the dialogue process, and can more intelligently respond to various dialogue scenarios and determine the next target semantic slot in a timely manner, thereby providing a more fluent, efficient and user-friendly dialogue experience.

[0123] In a feasible implementation, step S30 may include steps C10 to C20:

[0124] Step C10, taking the second semantic slot with the highest state transition value among the second semantic slots as the next target semantic slot;

[0125] It should be noted that the semantic slot with the highest state transition value indicates that this slot is the most critical for driving the conversation toward the expected goal. By prioritizing inquiries to the semantic slot with the highest state transition value, the system can effectively avoid unnecessary inquiries, thereby optimizing resource usage.

[0126] It is understandable that during a multi-round conversation, the states of multiple semantic slots are tracked, which represent information points to be filled in the conversation. At a certain moment, it is necessary to decide the next semantic slot to be asked or focused on, that is, the next target semantic slot. This decision is based on the state transition value of each second semantic slot, which reflects the possibility of transition from the current conversation state to the next potential state. The second semantic slot with the highest state transition value is generally the one that needs the most attention in the next step, because it is most likely to drive the conversation forward and meet the purpose of the conversation.

[0127] Step C20, inputting the slot value and historical context information of the preset semantic slot into the preset generation model to generate second dialogue information related to the next target semantic slot, wherein the second dialogue information is dialogue information for inquiring the intention information corresponding to the next target semantic slot.

[0128] It should be noted that the preset generation model receives the slot value and historical context information of the preset semantic slot as input, and after a series of complex calculations, outputs the second dialogue information related to the next target semantic slot. After the next target semantic slot is determined, the slot value of the currently known preset semantic slot and the entire historical context information are input into the preset generation model. Based on this information, combined with the understanding of the user's intention and the consideration of the dialogue goal, the model generates a dialogue message that asks for the intent information corresponding to the next target semantic slot, that is, the second dialogue message. For example, if the next target semantic slot is "meal time", a second dialogue message similar to "When do you plan to eat?" may be generated.

[0129] Once the next target semantic slot is determined, the second dialogue information generation phase begins. The “second dialogue information” here refers specifically to the dialogue content that collects or confirms the intent information of the next target semantic slot. In order to make the generated dialogue information both relevant and natural, the slot value of the preset semantic slot and the historical context of the entire dialogue are considered. This context-based dialogue generation method helps maintain the coherence of the dialogue and improves user satisfaction.

[0130] In this implementation, the second semantic slot with the highest state transition value among the second semantic slots is used as the next target semantic slot; the slot value and historical context information of the preset semantic slot are input into the preset generation model to generate the second dialogue information related to the next target semantic slot, wherein the second dialogue information is dialogue information for inquiring the intention information corresponding to the next target semantic slot. By accurately selecting the next target semantic slot and generating dialogue information that closely fits the context, the dialogue system can significantly improve the interaction quality and user satisfaction, and provide users with a more fluent, efficient and humane dialogue service.

[0131] Step B40, if the next target semantic slot can be determined from the second semantic slot according to the first semantic slot, it is determined that the preset state transition condition is met;

[0132] Step B50: If the next target semantic slot cannot be determined from the second semantic slot according to the first semantic slot, it is determined that the preset state condition is not satisfied.

[0133] It should be noted that when key information is missing in the first semantic slot, the system may not be able to determine what information should be collected next. For example, if the user only provides the "restaurant type" but does not provide the "dining time" or "number of diners", the system may not be able to decide which information to ask next, so measures need to be taken to supplement the missing information, which may include re-asking the user or providing more detailed guidance to help the user provide the required information.

[0134] In this embodiment, the first semantic slots whose filling status is filled in all preset semantic slots are determined; the preset semantic slots whose filling status is unfilled are used as the second semantic slots, wherein the second semantic slots represent the information that the system may need to obtain from the user in the next step, and are the key to the continuation of the dialogue process; it is determined whether the next target semantic slot can be determined from the second semantic slot according to the first semantic slot; if the next target semantic slot can be determined from the second semantic slot according to the first semantic slot, it is determined that the preset state transition condition is met; if the next target semantic slot cannot be determined from the second semantic slot according to the first semantic slot, it is determined that the preset state condition is not met. If the next target semantic slot can be determined, it means that the current dialogue state allows state transition, that is, the preset state transition condition is met; on the contrary, if the next target semantic slot cannot be determined, it means that the current dialogue state does not allow state transition, that is, the preset state condition is not met. In this embodiment, judging whether the preset state transition condition is met according to the filling status of the preset semantic slots is the key to ensuring the continuity and intelligence of the dialogue process. By dynamically adjusting the dialogue strategy and optimizing the state transition logic, it can respond to user needs more intelligently and improve the fluency of interaction.

[0135] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the multi-round dialogue method based on DST of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0136] This application also provides a multi-round dialogue device based on DST, please refer to Figure 4 , the DST-based multi-round dialogue device includes:

[0137] The identification module 10 is used to obtain first dialogue information input by the user device, perform intent identification on the first dialogue information, and determine intent information;

[0138] A first judgment module 20 is used to fill at least one preset semantic slot according to the intention information, and judge whether a preset state transition condition is met according to the filling state of the preset semantic slot;

[0139] A generation module 30 is configured to obtain historical context information of the conversation if the preset state transition condition is met, input the slot value of the preset semantic slot and the historical context information into a preset generation model, generate second conversation information, send the second conversation information to the user device, and update the historical context information according to the first conversation information and the second conversation information;

[0140] A second judgment module 40 is used to judge whether the dialogue task is completed according to the filling status of all preset semantic slots and the updated historical context information;

[0141] The ending module 50 is used to end the dialogue if the dialogue task is completed.

[0142] The DST-based multi-round dialogue device provided by the present application adopts the DST-based multi-round dialogue method in the above embodiment, which can solve the technical problem that the historical dialogue information cannot be transmitted during the multi-round dialogue process, resulting in poor fluency of the multi-round dialogue. Compared with the prior art, the beneficial effects of the DST-based multi-round dialogue device provided by the present application are the same as the beneficial effects of the DST-based multi-round dialogue method provided by the above embodiment, and other technical features in the DST-based multi-round dialogue device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0143] The present application provides a multi-round dialogue device based on DST, and the multi-round dialogue device based on DST 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 so that the at least one processor can execute the multi-round dialogue method based on DST in the above-mentioned embodiment 1.

[0144] Reference below Figure 5, which shows a schematic diagram of the structure of a DST-based multi-round dialogue device suitable for implementing the embodiment of the present application. The DST-based multi-round dialogue device in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 5 The DST-based multi-round dialogue device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0145] like Figure 5 As shown, the multi-round dialogue device based on DST may include a processing device 1001 (such as a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. Various programs and data required for the operation of the multi-round dialogue device based on DST are also stored in RAM1004. The processing device 1001, ROM1002 and RAM1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 1003 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 1009. The communication device 1009 can allow the DST-based multi-turn dialogue device to communicate with other devices wirelessly or by wire to exchange data. Although the DST-based multi-turn dialogue device with various systems is shown in the figure, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems may be implemented or have alternatively.

[0146] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0147] The DST-based multi-round dialogue device provided by the present application adopts the DST-based multi-round dialogue method in the above embodiment, which can solve the technical problem that the historical dialogue information cannot be transmitted during the multi-round dialogue process, resulting in poor fluency of the multi-round dialogue. Compared with the prior art, the beneficial effects of the DST-based multi-round dialogue device provided by the present application are the same as the beneficial effects of the DST-based multi-round dialogue method provided by the above embodiment, and the other technical features of the DST-based multi-round dialogue device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0148] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0149] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0150] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, wherein the computer-readable program instructions are used to execute the DST-based multi-round dialogue method in the above-mentioned embodiment.

[0151] The readable storage medium provided by the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned multi-round dialogue method based on DST, and can solve the technical problem that historical dialogue information cannot be transmitted during the multi-round dialogue process, resulting in poor fluency of the multi-round dialogue. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by the present application are the same as the beneficial effects of the multi-round dialogue method based on DST provided in the above-mentioned embodiment, and will not be repeated here.

[0152] The present application also provides a computer program product, including a computer program, which implements the steps of the DST-based multi-round dialogue method as described above when executed by a processor.

[0153] The computer program product provided by the present application can solve the technical problem that historical conversation information cannot be transmitted during a multi-round conversation, resulting in poor fluency in the multi-round conversation. Compared with the prior art, the beneficial effects of the computer program product provided by the present application are the same as the beneficial effects of the multi-round conversation method based on DST provided in the above embodiment, and will not be repeated here.

[0154] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A multi-round dialogue method based on DST, characterized in that: The DST-based multi-round dialogue method includes: Acquire first dialogue information input by a user device, perform intent recognition on the first dialogue information, and determine intent information; Fill at least one preset semantic slot according to the intent information, and determine whether a preset state transition condition is met according to the filling state of the preset semantic slot; If the preset state transition condition is met, historical context information of the conversation is obtained, the slot value of the preset semantic slot and the historical context information are input into a preset generation model, second conversation information is generated, the second conversation information is sent to the user device, and the historical context information is updated according to the first conversation information and the second conversation information; Determining whether the dialogue task is completed according to the filling status of all preset semantic slots and the updated historical context information; If the dialogue task is completed, the dialogue ends; The step of inputting the slot value of the preset semantic slot and the historical context information into a preset generation model to generate the second dialogue information further includes: The second semantic slot with the highest state transition value among the second semantic slots is used as the next target semantic slot, wherein the first semantic slot is a semantic slot whose filling state is completed among all preset semantic slots, the second semantic slot is a preset semantic slot whose filling state is unfilled, and the state transition value is a quantitative indicator of the possibility of transitioning from the first semantic slot to each of the second semantic slots; Inputting the slot value of the preset semantic slot and the historical context information into a preset generation model to generate second dialogue information related to the next target semantic slot, wherein the second dialogue information is dialogue information for inquiring about the intention information corresponding to the next target semantic slot; The step of judging whether the dialogue task is completed according to the filling status of all preset semantic slots and the updated historical context information further includes: If it is determined based on the filling status of the preset semantic slots and the context information that all necessary information has been collected and no issues requiring further clarification or resolution are detected, it is determined that the dialogue task has been completed.

2. The multi-round dialogue method based on DST as claimed in claim 1, characterized in that: The step of filling at least one preset semantic slot according to the intent information comprises: Determine the intent type present in the intent information and the intent data corresponding to the intent type; Get all preset semantic slots; Determining, according to the intent type, from the preset semantic slots, a preset semantic slot that matches the intent type; Fill the intent data corresponding to the intent type into the preset semantic slot corresponding to the intent type.

3. The multi-round dialogue method based on DST as claimed in claim 2, characterized in that: The step of judging whether a preset state transition condition is satisfied according to the filling state of the preset semantic slot comprises: Determine the first semantic slot in all preset semantic slots whose filling status is that it has been filled; Using the preset semantic slot whose filling state is unfilled as the second semantic slot; Determining whether a next target semantic slot can be determined from the second semantic slot according to the first semantic slot; If the next target semantic slot can be determined from the second semantic slot according to the first semantic slot, it is determined that the preset state transition condition is met; If the next target semantic slot cannot be determined from the second semantic slot according to the first semantic slot, it is determined that the preset state condition is not satisfied.

4. The multi-round dialogue method based on DST as claimed in claim 3, characterized in that: The step of determining whether the next target semantic slot can be determined from the second semantic slot according to the first semantic slot comprises: Inputting the first semantic slot into a preset state transition model, and calculating the state transition value of each of the second semantic slots; If there is a second semantic slot whose state transition value is greater than a preset state transition value, determining that a next target semantic slot can be determined from the second semantic slot according to the first semantic slot; If there is no second semantic slot whose state transition value is greater than the preset state transition value, it is determined that the next target semantic slot cannot be determined from the second semantic slot according to the first semantic slot.

5. The multi-round dialogue method based on DST as claimed in claim 1, characterized in that: After the step of judging whether the dialogue task is completed according to the filling status of all preset semantic slots and the updated historical context information, the method further includes: If the dialogue task is not completed, the third dialogue information input by the user device in response to the second dialogue information is obtained, the third dialogue information is used as the first dialogue information, and the execution step is returned to: performing intent recognition on the first dialogue information to determine intent information.

6. The multi-round dialogue method based on DST as claimed in claim 1, characterized in that: After the step of ending the dialogue if the dialogue task is completed, the method further includes: Sending a preset feedback request to the user equipment, and obtaining feedback information sent by the user equipment in response to the preset feedback request; If the feedback information is positive feedback information, the historical context information is used as a positive sample, and the preset generation model is optimized based on the positive sample; If the feedback information is negative feedback information, the historical context information is used as a negative sample, and the preset generation model is optimized based on the negative sample.

7. A multi-round dialogue device based on DST, characterized in that: The DST-based multi-round dialogue device includes: an identification module, configured to obtain first dialogue information input by a user device, perform intent identification on the first dialogue information, and determine intent information; A first judgment module is used to fill at least one preset semantic slot according to the intention information, and judge whether a preset state transition condition is met according to the filling state of the preset semantic slot; a generation module, configured to obtain historical context information of the conversation if the preset state transition condition is met, input the slot value of the preset semantic slot and the historical context information into a preset generation model, generate second conversation information, send the second conversation information to the user device, and update the historical context information according to the first conversation information and the second conversation information; A second judgment module is used to judge whether the dialogue task is completed according to the filling status of all preset semantic slots and the updated historical context information; An ending module, used for ending the dialogue if the dialogue task is completed; Wherein, the generating module further includes: The second semantic slot with the highest state transition value among the second semantic slots is used as the next target semantic slot, wherein the first semantic slot is a semantic slot whose filling state is completed among all preset semantic slots, the second semantic slot is a preset semantic slot whose filling state is unfilled, and the state transition value is a quantitative indicator of the possibility of transitioning from the first semantic slot to each of the second semantic slots; Inputting the slot value of the preset semantic slot and the historical context information into a preset generation model to generate second dialogue information related to the next target semantic slot, wherein the second dialogue information is dialogue information for inquiring about the intention information corresponding to the next target semantic slot; Wherein, the second judgment module further includes: If it is determined based on the filling status of the preset semantic slots and the context information that all necessary information has been collected and no issues requiring further clarification or resolution are detected, it is determined that the dialogue task has been completed.

8. A multi-round dialogue device based on DST, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the DST-based multi-round dialogue method according to any one of claims 1 to 6.

9. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the multi-round dialogue method based on DST as described in any one of claims 1 to 6 are implemented.

Citation Information

Patent Citations

  • Multi-round conversation interaction processing method, device, electronic equipment and storage medium

    CN109616108A

  • Task-driven system based on natural language

    CN113515616A

  • Multi-round dialogue method and system supporting multi-person real-time group chat

    CN116680374A

  • Task-based dialogue method based on vector retrieval and large language model

    CN117829295A

  • Multi-round interaction semantic understanding method and apparatus, and computer storage medium

    WO2020125457A1