A graph neural network driven zero-shot dialog state tracking method and system
By using a graph neural network-driven approach, the dialogue content and state are encoded, the relationship between dialogue rounds and slot nodes is modeled, and information is integrated using an adaptive attention mechanism. This solves the problem of insufficient adaptability of existing models in complex dialogue scenarios and cross-domain applications, and achieves more efficient zero-shot dialogue state tracking.
Patent Information
- Application Number
- CN202510594842.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-09
- Publication Date
- 2026-11-10
AI Technical Summary
Existing zero-shot dialogue state tracking models struggle to effectively capture implicit relationships when dealing with complex dialogue scenarios, and their cross-domain adaptability and generalization performance are insufficient. In particular, they are unable to maintain efficient and accurate dialogue state tracking when domain-specific labeled data is lacking.
A graph neural network-driven approach is adopted to model the relationship between dialogue rounds and slot nodes by encoding dialogue content and state, extract cross-domain contextual information, and integrate the outputs of each component through an adaptive attention mechanism to predict the slot state.
It enhances the model's cross-domain adaptability and ability to handle complex dialogue scenarios, improves the model's generalization ability and robustness, eliminates the need for domain-specific labeled data, and can better track cross-turn information, thereby improving the accuracy and flexibility of dialogue state tracking.
Smart Images

Figure CN123068534A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of natural language processing technology, specifically to a graph neural network-driven zero-shot dialogue state tracking method and system. Background Technology
[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.
[0003] With the popularization of deep neural networks, the rapid development of natural language processing, and the growing demand for artificial intelligence, the field of dialogue systems has attracted increasing interest from researchers. Task-oriented dialogue systems support users in completing various tasks, such as train booking, hotel reservations, and attraction recommendations, and have achieved significant breakthroughs. Therefore, improving the performance of dialogue state tracking has become an urgent task to be addressed.
[0004] In computer science, natural language processing (NLP) is considered a core area. This interdisciplinary field integrates linguistics, computer science, and mathematics, and its main research focus is on how to process and understand natural language, the language humans frequently use in daily life. The main application areas of NLP technology cover a wide range of fields, including sentiment analysis, automatic summarization, text classification, machine translation, speech recognition, and dialogue systems.
[0005] Dialogue systems, as one of the most popular technologies in natural language processing (NLP), involve multiple other NLP techniques and allow computers to communicate with humans. Their applications are wide-ranging, covering key areas such as customer service, efficient information retrieval, and personalized assistants. Task-oriented dialogue systems, as one type, focus on precisely and efficiently assisting users in completing specific tasks, such as restaurant reservations and weather inquiries, through a series of carefully designed dialogue rounds, greatly improving user experience and efficiency.
[0006] As a crucial component of task-oriented dialogue systems, dialogue state tracking has naturally become a hot research topic. As one of the dialogue management components of task-oriented dialogue systems, dialogue state tracking functions to fill in and correct data from dialogue utterances in natural language understanding. Essentially, it generates the current dialogue state and then transmits it to the policy learning module in dialogue management. Accurate dialogue state tracking allows the system to gain a deeper understanding of user intent and provide appropriate responses and suggestions.
[0007] To achieve this goal, task-oriented dialogue systems need to demonstrate superior intent recognition and dialogue state management capabilities. The system must be able to accurately interpret the actual needs and intentions behind the user's utterances and dynamically track and update contextual information during the dialogue to ensure that every interaction revolves around the user's actual needs, thereby providing personalized and coherent responses and achieving efficient task completion. This process not only demonstrates the depth and breadth of the technology but also marks another important milestone in how artificial intelligence technology serves human life.
[0008] Early Dialogue State Tracking (DST) methods primarily relied on rule-based and statistical techniques. These methods proved effective in simple dialogue scenarios but showed limitations when faced with complex dialogues. Due to the diversity and complexity of dialogues, traditional methods often struggled to capture rich information and heavily depended on domain-specific knowledge. While rule-based DST techniques are widely used in practical dialogue systems, they also face challenges such as poor adaptability to new domains and limited generalization ability. Statistical techniques face challenges such as handling ambiguity and managing noise interference.
[0009] Despite significant progress and diverse innovative methods in zero-shot dialogue state tracking (ZSDST) in recent years, current models still face considerable challenges in handling complex dialogue scenarios. Specifically, complex dialogues are characterized by multi-turn interactions where slot information often spreads across different turns in an implicit and dispersed manner. Capturing these implicit relationships places higher demands on models. Unfortunately, existing models are still insufficient in parsing such complex dialogue structures and accurately extracting and integrating dispersed information, making it difficult to fully understand and track deep state changes in the dialogue.
[0010] Furthermore, these models often perform poorly when attempted to be applied across different domains. This exposes a significant weakness in their generalization ability and transfer performance: they are unable to flexibly adapt to new domain environments and struggle to maintain efficient and accurate dialogue state tracking in the absence of any domain-specific labeled data.
[0011] Therefore, how to enhance the implicit relationship capture ability of the ZSDST model and improve its cross-domain adaptability and generalization performance remains a key issue that needs to be addressed in current research. Summary of the Invention
[0012] To address the shortcomings of existing technologies, this disclosure proposes a graph neural network-driven zero-shot dialogue state tracking method and system. By introducing graph neural networks and leveraging their unique advantages in processing graph-structured data, and by explicitly modeling the relationship network between dialogue rounds and slot pairs, it is possible to deeply explore the complex relationships in multi-turn dialogues.
[0013] To achieve the above objectives, the present disclosure adopts the following technical solution:
[0014] The first aspect of this disclosure provides a zero-shot dialogue state tracking method, comprising the following steps:
[0015] S101: Encode the dialogue content and status;
[0016] S102: Model the relationship between dialogue rounds and slot nodes;
[0017] S103: Extract cross-domain contextual information;
[0018] S104: Integrate the outputs of each component to predict the slot status.
[0019] A second aspect of this disclosure provides a zero-sample dialogue state tracking system, comprising the following steps:
[0020] The encoding module is configured to encode the dialogue content and state.
[0021] The modeling module is configured to model the relationships between dialogue rounds and slot nodes.
[0022] The extraction module is configured to extract cross-domain context information.
[0023] The prediction module is configured to integrate the outputs of various components to predict the slot status.
[0024] A third aspect of this disclosure provides a medium having a program stored thereon that, when executed by a processor, implements the steps of a graph neural network-driven zero-shot dialogue state tracking method as described in the first aspect of this disclosure.
[0025] This fourth aspect of the disclosure provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements steps in a graph neural network-driven zero-sample dialogue state tracking method.
[0026] Compared with the prior art, the beneficial effects of this disclosure are:
[0027] (1) This disclosure enhances cross-domain adaptability and the ability to handle complex dialogue scenarios. By introducing graph neural networks and adaptive attention mechanisms, it effectively captures implicit relationships in dialogues, improves the model's transferability across different domains, and reduces dependence on domain-specific labeled data. In addition, it can better handle multi-round, information-dispersed dialogues, accurately model the relationship between dialogue rounds and slot nodes, thereby effectively tracking cross-round information and better coping with complex dialogue scenarios.
[0028] (2) This disclosure has stronger generalization ability and does not require domain-specific labeled data. Compared with traditional rule-based and statistical techniques, it exhibits higher accuracy and efficiency in different dialogue tasks and domains, has better generalization ability and robustness, supports zero-shot dialogue state tracking, does not require domain-specific labeled data, and provides greater flexibility and adaptability.
[0029] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0030] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative examples and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention.
[0031] Figure 1 This is an overall flowchart of the method in the first embodiment.
[0032] Figure 2 This is a graph neural network module diagram of the method in the first embodiment.
[0033] Figure 3 This is a block diagram of the self-attention mechanism of the method in the first embodiment. Detailed Implementation
[0034] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.
[0035] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this disclosure. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.
[0036] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover a non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or apparatus.
[0037] Where there is no conflict, the embodiments and features described herein can be combined with each other.
[0038] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.
[0039] Example 1
[0040] This embodiment provides a zero-shot dialogue state tracking method, including:
[0041] S101: Encode the dialogue content and status;
[0042] S102: Model the relationship between dialogue rounds and slot nodes;
[0043] S103: Extract cross-domain context information;
[0044] S104: Integrate the outputs of each component to predict the slot status.
[0045] Furthermore, the zero-shot dialogue state tracking model includes: a pre-trained language model T5-small, a graph convolutional neural network (GCN), and a self-attention mechanism connected in sequence.
[0046] Furthermore, the training steps for the trained zero-shot dialogue state tracking model include:
[0047] Construct a training set, which includes the dialogue content of all rounds and the dialogue state of previous rounds;
[0048] Encoding is performed using a pre-trained language model T5-small based on the training set;
[0049] GCN is constructed based on the dialogue content and state encoding as vector representations;
[0050] The output of GCN is then input into the self-attention mechanism model to accurately capture the dynamic importance distribution of the current dialogue context;
[0051] The context representations generated by the GCN and adaptive attention mechanism modules are input into the integrated reasoning module, and then integrated to predict the dialogue state, resulting in a trained zero-shot dialogue state tracking model.
[0052] Furthermore, based on the training set, encoding is performed using the pre-trained language model T5-small; specifically including:
[0053] The dialogue content and state are encoded into vector representations using the pre-trained language model T5-small.
[0054] Furthermore, based on the dialogue content and state encoding as vector representations, a GCN is constructed, specifically including:
[0055] This method utilizes dynamic graphs to capture the complex relationships between dialogue turns and slot pairs. It comprises two nodes: a dialogue turn node and a slot pair node. These are connected by four different edge types to capture the direct relationships between dialogue context and slot pairs, the interrelationships between different slots, the latest context information for each slot pair, and shared information between slots within the same domain. Node initialization is based on the output of the dialogue encoding module. Through message passing and node updates using a graph convolutional neural network, the global embedding representation of the nodes is learned progressively, providing strong support for dialogue state tracking tasks.
[0056] This application constructs a dynamic Generative Communication Network (GCN) to model the relationship between dialogue rounds and slot nodes, thereby comprehensively capturing contextual information and inter-slot dependencies and improving the accuracy of zero-shot dialogue state tracking. For example, in a dialogue round containing slot pairs such as "book a hotel" and "provide breakfast," modeling only the direct relationship between individual slots might overlook the shared contextual information between "hotel" and "breakfast." However, through the dynamic mechanism of the GCN, the inter-slot dependencies and their association with the context can be captured simultaneously, avoiding information omissions and significantly improving adaptability and prediction accuracy in actual dialogue state prediction for unknown domains.
[0057] Further, S103: Extract cross-domain context information. The output of GCN is input into a self-attention mechanism to accurately capture the dynamic importance distribution of the current dialogue context; specifically including:
[0058] By embedding the dialogue rounds and slot-value pairs of nodes output by GCN into vectors, and using domain-specific attention heads and dynamic weight adjustments, the dynamic importance of the dialogue context is accurately captured. Combined with the multi-head self-attention mechanism, the model's generalization ability in different domains and the accuracy of dialogue state tracking are effectively improved.
[0059] The context representations generated by GCN and the self-attention mechanism are input into the integrated reasoning module, and then integrated to predict the dialogue state, resulting in a trained zero-shot dialogue state tracking model.
[0060] Furthermore, the dialogue content of all rounds and the dialogue state of previous rounds, the dialogue content of all rounds refers to the dialogue process of the current round and all rounds before it, and the dialogue state of previous rounds refers to all dialogue states before the currently predicted round, but does not include the dialogue state of the current round.
[0061] The training phase of this invention comprises four parts. First, the dialogue encoding module encodes the dialogue content and state into vector representations. Next, the graph convolutional neural network construction module explicitly models the relationship between dialogue rounds and slot nodes, capturing implicit information in the dialogue. Subsequently, the self-attention mechanism module flexibly allocates attention weights, accurately extracts cross-domain contextual information, and improves the model's generalization ability. Finally, the integrated reasoning and prediction module integrates the outputs of each component to efficiently predict the slot state.
[0062] The first step involves encoding using a pre-trained language model, T5-small. The model's input includes the dialogue content and previous dialogue states. The dialogue content is represented as I = {(A1, R1), ..., (A...}. t R t )}, where A t R represents the user's utterance in round t. t This represents the system's utterance in round t. The previous dialogue states are represented as S = {S1, ..., S2}. t-1},in k represents the dialogue state in round t. i Indicates the slot name. This represents the value of the i-th slot in round t.
[0063] like Figure 1 As shown, the dialogue content and dialogue state are encoded.
[0064] E t =T5 small ([A t ,R t ])
[0065]
[0066] in, This represents the encoded representation of the i-th slot in round t. This indicates that the sequences are connected in order.
[0067] The second step, in the graph convolutional neural network construction module, involves building a dynamic graph to capture the complex relationships between dialogue turns and slot pairs. This graph includes two types of nodes and four types of edges, such as... Figure 2 As shown, we utilize this graph structure to model the dialogue state tracking task using GCN. GCN is an efficient variant of graph neural networks, particularly well-suited for processing graph data with complex structures.
[0068] In GCN, the initial embedding vector for each node is generated from the output of the dialogue encoding module. Dialogue content and slot pair information are encoded into the node's initial feature representation.
[0069] In GCN, message passing is one of the core operations. Each node updates its embedding representation progressively by exchanging information with its neighbors. For each node i, message passing and aggregation are performed based on the embeddings of its neighbors and the types of edges. The formula is as follows:
[0070]
[0071] in, This represents the embedding representation of node i at layer l+1. r It is the weight matrix of edge type r. Let i be the set of neighboring nodes of node i, where the edge type is r. σ is the average coefficient of the neighbor node embeddings. σ is the non-linear activation function.
[0072] After message passing and aggregation, the node's embedding representation is updated and fed into the next layer of the GCN. First, for each node in each layer, the new embedding information obtained through message passing is applied to the node's update process. Then, the node's embedding vector is updated in each layer through linear transformations and non-linear activation functions, progressively capturing the complex relationships between nodes in the dialogue. The formula for node update is as follows:
[0073]
[0074] This process transfers local information from nodes to the global level, gradually learning global embedding representations of nodes through multiple layers of GCN. These representations capture semantic information and relationships between slot pairs in the dialogue, providing strong support for subsequent dialogue state tracking tasks. Through these steps, GCN can effectively encode dialogue turn nodes and slot pair nodes, generating node embedding vectors, thereby clarifying the complex relationships between nodes and slot pairs in the dialogue. This provides a solid foundation for the model to handle implicit relationships and cross-domain transfer tasks.
[0075] The third step involves inputting the self-attention mechanism into the embedding vectors of the dialogue turn nodes and slot-value pair nodes output by the GCN. The dialogue turn node embeddings represent the semantic information of each turn in the dialogue. The slot-value pair node embeddings represent the semantic information of each slot and its corresponding value. These node embedding vectors are obtained through the message passing and aggregation mechanisms in the GCN, providing a deep relationship between dialogue turns and slot-value pairs. The structure of the self-attention mechanism is as follows: Figure 3 As shown.
[0076] To effectively capture diverse information from the dialogue context, the model employs a multi-head self-attention mechanism. Each attention head independently computes its query, key, and value vectors, generating different context representations. Finally, the outputs of all attention heads are concatenated, and a linear transformation is applied to obtain the final context representation, as shown in the following formula:
[0077]
[0078] Where Q is the query vector, K is the key vector, and V is the value vector. Attention weights are applied to the weighted context representation obtained from the value vector. W o It is a linear transformation weight matrix, and H is the number of attention heads.
[0079] In the self-attention mechanism, the node embedding vectors generated by the GCN module serve as input for calculating attention weights. These embedding vectors include semantic representations of dialogue turn nodes and slot-pair nodes. The GCN module provides information about the relationships between nodes, enabling the self-attention mechanism to adjust attention weights for different domains based on the dynamic characteristics of the dialogue context, thereby improving the accuracy and domain adaptability of dialogue state tracking.
[0080] The fourth step is to integrate the outputs of each component in the comprehensive reasoning and prediction module. This module is responsible for applying the knowledge learned during the training phase to new domains and for predicting the dialogue state in unseen domains.
[0081] The node embedding vectors processed by the self-attention mechanism module contain comprehensive information about both the dialogue turn nodes and slot-pair nodes. Specifically, these vectors undergo weighted processing by a multi-head self-attention mechanism to reflect important contextual information in the dialogue. These contextual representation vectors are then fused into a unified dialogue context representation through linear transformations and aggregation operations, defined as follows:
[0082]
[0083] The synthesized context representation is processed through a fully connected layer and an activation function to generate a prediction of the dialogue state. The definition is as follows:
[0084] St =softmax(W P ·h context +b P )
[0085] Among them, S t This is the predicted state of the dialogue in round t, W. P and b P These are the weight matrix and bias term of the prediction layer, respectively.
[0086] During the training phase, the model learns general features for dialogue state tracking through multi-domain data. When faced with new, unseen domains, the model needs to adapt to the new dialogue features through domain-specific adjustments. This includes dynamically adjusting domain-specific vocabulary and slots to better adapt to the new domain's dialogue scenarios.
[0087] Example 2
[0088] This embodiment provides a zero-sample dialogue state tracking system, including:
[0089] The encoding module is configured to encode the dialogue content and state.
[0090] The modeling module is configured to model the relationships between dialogue rounds and slot nodes.
[0091] The extraction module is configured to extract cross-domain context information.
[0092] The prediction module is configured to integrate the outputs of various components to predict the slot status.
[0093] It should be noted that the encoding module, modeling module, extraction module, and prediction module described above correspond to steps S101 to S104 in Embodiment 1. The examples and application scenarios implemented by these modules and their corresponding steps are the same, but they are not limited to the content disclosed in Embodiment 1. It should also be noted that these modules, as part of the system, can be executed in a computer system, such as a set of computer-executable instructions.
[0094] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0095] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.
[0096] Example 3
[0097] This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform the method described in Embodiment 1.
[0098] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.
[0099] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.
[0100] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.
[0101] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.
[0102] Those skilled in the art will recognize that the units and algorithm steps described in connection with the various examples of this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.
[0103] Example 4
[0104] This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.
[0105] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.
Claims
1. A graph neural network-driven zero-shot dialogue state tracking method, characterized in that, include: S101: Encode the dialogue content and status; S102: Model the relationship between dialogue rounds and slot nodes; S103: Extract cross-domain context information; S104: Integrate the outputs of each component to predict the slot status.
2. The zero-sample dialogue state tracking method as described in claim 1, characterized in that, The zero-shot dialogue state tracking model includes: a pre-trained language model T5-small, a GCN, and a self-attention mechanism connected in sequence.
3. The zero-shot dialogue state tracking method as described in claim 1, characterized in that, The training steps for the trained zero-shot dialogue state tracking model include: Construct a training set, which includes the dialogue content of all rounds and the dialogue state of previous rounds; Encoding is performed using a pre-trained language model T5-small based on the training set; GCN is constructed based on the dialogue content and state encoding as vector representations; The output of GCN is then input into the self-attention mechanism model to accurately capture the dynamic importance distribution of the current dialogue context; The context representations generated by the GCN and adaptive attention mechanism modules are input into the integrated reasoning module, and then integrated to predict the dialogue state, resulting in a trained zero-shot dialogue state tracking model.
4. The zero-sample dialogue state tracking method as described in claim 1, characterized in that, The dialogue content of all rounds and the dialogue state of previous rounds are described. The dialogue content of all rounds refers to the dialogue process of the current round and all rounds before it. The dialogue state of previous rounds refers to all dialogue states before the current predicted round, but does not include the dialogue state of the current round.
5. The zero-sample dialogue state tracking method as described in claim 1, characterized in that, Based on the training set, the dialogue content and state are encoded into vector representations using a pre-trained language model T5-small.
6. The zero-shot dialogue state tracking method as described in claim 1, characterized in that, The construction of GCN based on dialogue content and state encoding as vector representations specifically includes: In the GCN process, a dynamic graph is used to capture the complex relationships between dialogue rounds and slot pairs. This graph consists of two nodes: a dialogue round node and a slot pair node. These are connected by four different edge types to capture the direct relationships between the dialogue context and slot pairs, the interrelationships between different slots, the latest context information for each slot pair, and the shared information between slots within the same domain. Node initialization is based on the output of the dialogue encoding module. Through message passing and node updates in GCN, the global embedding representation of the nodes is gradually learned, providing strong support for the dialogue state tracking task.
7. The zero-shot dialogue state tracking method as described in claim 1, characterized in that, The output of GCN is then input into a self-attention mechanism model to accurately capture the dynamic importance distribution of the current dialogue context, specifically including: By embedding the dialogue rounds and slot-value pairs of nodes output by GCN into vectors, and using domain-specific attention heads and dynamic weight adjustments, the dynamic importance of the dialogue context is accurately captured. Combined with a multi-head self-attention mechanism, the model's generalization ability in different domains and the accuracy of dialogue state tracking are effectively improved. The context representations generated by GCN and the self-attention mechanism module are input into the integrated reasoning module, and integrated to predict the dialogue state, resulting in a trained zero-shot dialogue state tracking model.
8. A zero-shot dialogue state tracking method system, characterized by: include: The encoding module is configured to encode the dialogue content and state. The modeling module is configured to model the relationships between dialogue rounds and slot nodes. The extraction module is configured to extract cross-domain context information. The prediction module is configured to integrate the outputs of various components to predict the slot status.
9. An electronic device, characterized in that it comprises: Memory is used to store computer-readable instructions in a non-transitory manner. Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in any one of claims 1-7.
10. A storage medium, characterized in that, The computer-readable instructions are stored non-transitory, wherein when the non-transitory computer-readable instructions are executed by a computer, the instructions of the method according to any one of claims 1-7 are executed.
