Conversation state tracking method and device, terminal equipment and medium

Abstract

The invention is suitable for the technical field of dialogue systems, and provides a dialogue state tracking method and device, terminal equipment and a medium, and the method comprises the steps: obtaining a current dialogue data set in a dialogue system, the current dialogue data set at least comprising current dialogue data and a to-be-predicted slot of the current dialogue data; the current dialogue data set is input into the target student model, a predicted slot value of the slot to be predicted and current reasoning data corresponding to the predicted slot value are obtained, and the current reasoning data are used for representing the reasoning process that the value of the slot to be predicted is the predicted slot value; and maintaining the conversation state of the conversation system based on the predicted slot value and the current reasoning data. According to the method, the predicted slot value of the to-be-predicted slot and the current reasoning data corresponding to the predicted slot value are output according to the target student model, so that the dialogue state of the dialogue system is maintained, the dialogue system can dynamically integrate and apply knowledge from various fields, disastrous forgetting is effectively relieved, and the dialogue efficiency is improved. And the adaptability and learning efficiency of the system are enhanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of dialogue systems, and particularly relates to a dialogue state tracking method and device, a terminal device and a medium. BACKGROUND

[0002] Dialog state tracking (DST) is crucial for maintaining context in dialogue systems, especially in applications involving complex and continuous interactions such as customer service, virtual assistance, and interactive robots. DST must accurately track and update the state of the dialogue.

[0003] However, current DST methods mainly rely on static or incremental learning methods, which cannot dynamically adapt to knowledge from various fields and easily lose previously acquired knowledge, resulting in low learning and application efficiency. SUMMARY

[0004] The embodiments of the application provide a dialogue state tracking method, device, terminal device and medium, which can dynamically integrate and apply knowledge from various fields, effectively alleviate catastrophic forgetting, and enhance the adaptability and learning efficiency of the system.

[0005] In a first aspect, the embodiments of the application provide a dialogue state tracking method, comprising:

[0006] obtaining a current dialogue data set in a dialogue system, the current dialogue data set at least comprising current dialogue data and a to-be-predicted slot of the current dialogue data;

[0007] inputting the current dialogue data set into a target student model to obtain a predicted slot value of the to-be-predicted slot and current reasoning data corresponding to the predicted slot value, wherein the current reasoning data is used to represent a reasoning process in which the value of the to-be-predicted slot is the predicted slot value;

[0008] maintaining a dialogue state of the dialogue system based on the predicted slot value and the current reasoning data.

[0009] In a second aspect, the embodiments of the application provide a dialogue state tracking device, comprising:

[0010] an obtaining module configured to obtain a current dialogue data set in a dialogue system, the current dialogue data set at least comprising current dialogue data and a to-be-predicted slot of the current dialogue data;

[0011] an inputting module configured to input the current dialogue data set into a target student model to obtain a predicted slot value of the to-be-predicted slot and current reasoning data corresponding to the predicted slot value, wherein the current reasoning data is used to represent a reasoning process in which the value of the to-be-predicted slot is the predicted slot value;

[0012] The maintenance module is used to maintain the dialogue state of the dialogue system based on the predicted slot values ​​and the current inference data.

[0013] Thirdly, embodiments of this application provide a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in any of the first aspects.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as described in any of the first aspects.

[0015] Fifthly, embodiments of this application provide a computer program product that, when run on a terminal device, causes the terminal device to execute any of the methods described in the first aspect above.

[0016] This application provides a dialogue state tracking method, apparatus, terminal device, and medium. The method includes: acquiring a current dialogue dataset in a dialogue system, the current dialogue dataset including at least current dialogue data and slots to be predicted for the current dialogue data; inputting the current dialogue dataset into a target student model to obtain the predicted slot values ​​for the slots to be predicted and the current inference data corresponding to the predicted slot values, wherein the current inference data is used to characterize the inference process of the value of the slot to be predicted being the predicted slot value; and maintaining the dialogue state of the dialogue system based on the predicted slot values ​​and the current inference data. By using the above technical solution, the dialogue state of the dialogue system is maintained by obtaining the predicted slot values ​​for the slots to be predicted and the current inference data corresponding to the predicted slot values ​​based on the output of the target student model. This allows the dialogue system to dynamically integrate and apply knowledge from various domains, effectively mitigating catastrophic forgetting and enhancing the system's adaptability and learning efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a dialogue state tracking method provided by existing technology;

[0019] Figure 2 This is a performance diagram of a dialogue state tracking method provided by existing technology;

[0020] Figure 3This is a schematic diagram illustrating the classification of error slots in a dialogue state tracking method provided by existing technology;

[0021] Figure 4 This is a schematic flowchart of a dialogue state tracking method provided in an embodiment of this application;

[0022] Figure 5 This is a flowchart illustrating a dialogue state tracking method provided in another embodiment of this application;

[0023] Figure 6 This is a schematic diagram of a process for generating perturbation inference data according to an embodiment of this application;

[0024] Figure 7 This is a schematic diagram illustrating a semantic comparison reasoning selection provided in an embodiment of this application;

[0025] Figure 8 This is an overall schematic diagram of a dialogue state tracking method provided in another embodiment of this application;

[0026] Figure 9 This is a structural block diagram of a dialogue state tracking device provided in one embodiment of this application;

[0027] Figure 10 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0028] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0029] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0030] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0031] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0032] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0033] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0034] It should be noted that the information collection process (such as the facial image collection process, the dialogue content collection process, etc.) / feature extraction process involved in this application is carried out with the user's knowledge and permission. That is, the information collection process / feature extraction process complies with the requirements of laws and regulations and does not constitute an act that harms the public interest.

[0035] The dialogue state tracking method provided in this application can be applied to terminal devices such as mobile phones, tablets, wearable devices, in-vehicle devices, augmented reality (AR) / virtual reality (VR) devices, laptops, ultra-mobile personal computers (UMPCs), netbooks, and personal digital assistants (PDAs). This application does not impose any restrictions on the specific type of terminal device.

[0036] It can be argued that an ideal dialogue system needs to possess the ability to continuously acquire and adapt to new tasks while retaining prior knowledge. Dialogue State Tracking (DST) typically involves learning new services and dealing with catastrophic forgetting, as well as the critical capability loss known as the "slot selection dilemma," which is essential in dialogue systems.

[0037] However, traditional research on dialogue systems has primarily focused on domain-specific offline systems, lacking adaptability. With recent advancements in Large Language Models (LLMs), LLM-based systems significantly outperform previous methods. However, the sheer size of LLMs makes training models from scratch extremely time-consuming and challenging. Therefore, for dialogue systems, effective Continual Learning (CL) is crucial for acquiring new skills while retaining prior task knowledge.

[0038] Dialogue state tracking is the core of task-oriented dialogue systems, dynamically updating triples (domain, slot, value) to manage user intent. To extend the DST model, existing research has focused on persistent DST tasks. As a specific task of CL (Consciousness-Based Learning), persistent DST strives to address catastrophic forgetting, but the sequential learning of new tasks hinders the retention of previous tasks. This problem is exacerbated in persistent DST because the cross-domain data distribution changes significantly during CL.

[0039] For example, the learning domains for continuous dialogue state tracking can be, in order, hotel DST, restaurant DST, service DST, and taxi DST. Figure 1 This is a flowchart illustrating a dialogue state tracking method provided by existing technology, such as... Figure 1 As shown, this is a real-world example of the "slot selection dilemma" when transitioning from the hotel domain to the restaurant domain. The figure illustrates three dialogues between the system and the user in the restaurant domain (dialogues in rounds 3, 6, and 7, respectively), which involve three date values. Ultimately, the model incorrectly predicts the reservation date as the time in round 7 (i.e., 10:30 AM) instead of the correct time in round 6 (i.e., 10:45 AM).

[0040] Recent research on continuous DST (Discretionary Searching) focuses on addressing the forgetting problem through methods such as memory replay, regularization, and reconstruction simulation as a question-and-answer task. However, these methods face challenges such as reliance on past data and high computational demands during testing, hindering their real-time application.

[0041] Furthermore, recent LLMs have demonstrated strong DST performance, but face obstacles in practical deployment such as offline computational load and online data privacy issues. Enabling LLMs for CLs requires significant resources, prompting exploration of smaller models. A comprehensive analysis of small DST models reveals a severe loss of capability due to changes in the CL domain, specifically, the model's inability to output the correct slot value when faced with a similar set of candidates (e.g., the "slot selection dilemma"). For example, in... Figure 1 In the example, when tracking the <Service-Appointment Date> slot, the model incorrectly selected the most recent mentioned time, indicating its inability to grasp the nuances of context and its tendency to extract values ​​directly rather than reason logically. This phenomenon arises because, with the evolution of Closed Logic (CL), models need to understand relevant knowledge of emerging domains and selectively forget knowledge that is irrelevant or weakly related to the domain, such as slot selection.

[0042] Figure 2 This is a performance diagram of a dialogue state tracking method provided by existing technology, such as... Figure 2 As shown, by using two different backbone models, T5-small and LLaMA-7B, for dialogue state tracking, it can be observed that the performance of sustained DST (i.e., the average JGA score) decreases as the number of dialogue rounds increases.

[0043] Furthermore, after analyzing 874 mispredicted slot samples in long dialogues (e.g., dialogue rounds > 10), it was found that in 94.5% of cases, the model typically returned a value semantically relevant to the requested slot from the dialogue. For example, a location-related slot usually resulted in selecting any mentioned location. Upon closer examination, it was discovered that the model failed to select the correct value due to a severe loss of ability, termed the "slot value selection dilemma." This problem arises when the model struggles to infer the correct value among multiple ambiguous or semantically similar values ​​in a dialogue.

[0044] By examining 826 errors related to this dilemma, it was observed that in 45% of cases, the model favored the most recently mentioned value; in 30% of cases, the model stuck to the previously chosen value, ignoring modifications made in the dialogue; and the remaining 25% involved random selection from the available values. Furthermore, Figure 3 This is a schematic diagram illustrating the classification of faulty slots in a dialogue state tracking method provided by existing technology, such as... Figure 3 As shown, the error rates of the 49 slots were categorized, and it was found that the slots related to "date", "from", and "to" were the most prone to errors because the time or location in the dialogue often presents multiple values, leading to challenges in selection.

[0045] In summary, current DST methods primarily rely on static or incremental learning approaches, which cannot effectively handle the complexity of real-world applications. Many systems struggle to dynamically adapt to new domains or integrate new information without compromising previously learned knowledge. Furthermore, traditional dialogue state tracking methods often fail to transfer insights across different tasks and domains, resulting in inefficient learning and application.

[0046] Therefore, the main challenges addressed by current RoS methods include:

[0047] (1) Catastrophic forgetting: In continuous learning scenarios, when new tasks or data are introduced, dialogue systems often lose previously acquired knowledge.

[0048] (2) Slot selection dilemma: As systems learn new services, they face difficulties in accurately selecting values ​​from multiple potential options, which is crucial for maintaining a coherent and context-appropriate conversation.

[0049] Based on this, this application provides a dialogue state tracking method that proposes a Reason-of-Select (RoS) distillation framework for continuous dialogue state tracking. By introducing domain-independent meta-reasoning, the system can dynamically integrate and apply knowledge from various domains, effectively mitigating fuzzy slot selection and catastrophic forgetting caused by cross-domain data distribution changes, thereby enhancing its adaptability and learning efficiency. By utilizing meta-knowledge from different domains, a more robust and flexible learning process is ensured. Furthermore, to alleviate the illusion of output from the teacher model in a reasoning framework without manual annotation, a semantic contrastive reasoning selection strategy is proposed, which can obtain credible reasoning reasons from multiple candidates, refining the slot selection and reasoning process, ensuring correct reasoning transitions, and enhancing the reliability of the DST.

[0050] Figure 4 This is a flowchart illustrating a dialogue state tracking method according to an embodiment of this application. It is intended as an example and not a limitation, and the method can be applied to terminal devices.

[0051] S101. Obtain the current dialogue dataset in the dialogue system. The current dialogue dataset includes at least the current dialogue data and the slots to be predicted for the current dialogue data.

[0052] The current dialogue dataset can refer to the dataset corresponding to the current dialogue system. For example, the current dialogue dataset can at least include the current dialogue data and the slots to be predicted within that data. The current dialogue data is the dialogue data collected by the current dialogue system, such as the data of the interaction between the dialogue system and the user. The slots to be predicted can be understood as the dialogue states in the current dialogue data that require slot value prediction. For example, the slots to be predicted can be one or more of a predefined set of dialogue states, such as the restaurant name or the number of diners. Then, based on the target student model, the predicted slot value corresponding to the slot to be predicted needs to be extracted from the current dialogue data, such as the specific restaurant name. Furthermore, the current dialogue dataset can also include other types of data, which can be configured according to the actual situation.

[0053] S102. Input the current dialogue dataset into the target student model to obtain the predicted slot value of the slot to be predicted and the current inference data corresponding to the predicted slot value. The current inference data is used to characterize the inference process of the value of the slot to be predicted being the predicted slot value.

[0054] The target student model can be used to predict slot values ​​and output the current inference data of the predicted slot values. The current inference data can be used to represent the inference process of the value of the slot to be predicted being the predicted slot value. The specific method of obtaining the target student model is not limited. For example, it can be learned and trained from the inference information generated by the teacher model using transfer learning. The specific training process is not limited here. Alternatively, the target student model can also be obtained directly through other methods, etc.

[0055] Specifically, after obtaining the current dialogue dataset in the dialogue system, the obtained current dialogue dataset can be input into the target student model, and the predicted slot value of the slot to be predicted and the current inference data corresponding to the predicted slot value can be output, so as to maintain the dialogue state of the dialogue system based on the obtained data.

[0056] S103. Based on the predicted slot values ​​and current inference data, maintain the dialogue state of the dialogue system.

[0057] This embodiment provides a dialogue state tracking method. It acquires the current dialogue dataset in a dialogue system, which includes at least current dialogue data and slots to be predicted for that data. The current dialogue dataset is input into a target student model to obtain the predicted slot values ​​and the corresponding current inference data. The current inference data characterizes the inference process of determining the value of the slot to be predicted as the predicted slot value. Based on the predicted slot values ​​and the current inference data, the dialogue state of the dialogue system is maintained. Using this method, by obtaining the predicted slot values ​​and corresponding current inference data from the target student model output, the dialogue state of the dialogue system is maintained. This allows the dialogue system to dynamically integrate and apply knowledge from various domains, effectively mitigating catastrophic forgetting and enhancing the system's adaptability and learning efficiency.

[0058] Figure 5 This is a flowchart illustrating a dialogue state tracking method according to another embodiment of this application. This embodiment further optimizes the process before inputting the current dialogue dataset into the target student model to obtain the predicted slot value of the slot to be predicted and the corresponding current inference data. The optimization is as follows: A sample dialogue dataset is constructed based on the sample dialogue data. The sample dialogue dataset includes at least sample slots of the sample dialogue data, the sample slot values ​​corresponding to the sample slots, and inference prompt data. Sample inference data corresponding to the sample dialogue data is generated based on the sample dialogue dataset and the teacher model. The initial student model is trained using the sample inference dataset containing the sample inference data to obtain the target student model. Figure 5 As shown, the method includes:

[0059] S201. Construct a sample dialogue dataset based on the sample dialogue data. The sample dialogue dataset shall include at least the sample slots of the sample dialogue data, the sample slot values ​​corresponding to the sample slots, and the inference hint data.

[0060] Sample dialogue data can be dialogue data used to train the target student model, such as historical dialogue data from a dialogue system or dialogue data collected from other dialogue systems. The sample dialogue dataset can be considered input data for the teacher model, used to generate sample inference data corresponding to the sample dialogue data. The content of the sample dialogue dataset is not limited; it can include sample slots, corresponding slot values, and inference hints. Sample slots can be slots identified or selected from the sample dialogue data, and slot values ​​are the corresponding slot values. Inference hints can be used to prompt the teacher model to generate sample inference data corresponding to the sample dialogue data; these hints can be pre-set hints by relevant personnel.

[0061] For example, the following is a specific example of a sample dialogue dataset provided in this embodiment, which can be used to generate inference data from a teacher model.

[0062] {

[0063] "Instruction": Simply return a concise reasoning process;

[0064] "input": Performs dialogue state tracking tasks;

[0065] Consider the dialogue content: "[The previous dialogue was omitted...]"

[0066] [User]: I would like to make an appointment at 10:45 AM.

[0067] [SYSTEM]: When is the appointment time?

[0068] [User]: The appointment is for March 11th.

[0069] [System]: Schedule an appointment with Rig Christie at 10:45 AM on March 11th.

[0070] [The rest of the dialogue is omitted...]

[0071] The slot value for the slot <services-appointment_date> is "10:45 AM";

[0072] Some information about this slot: services means finding the right therapist for you and making it easy to book, and appointment_date means the appointment date.

[0073] Inference Hints: Can you tell me why and your reasoning process? In the context of dialogue state tracking, there are usually multiple possible values ​​associated with a requested slot. By carefully analyzing the dialogue context, user and system intent, and considering any confirmation or rejection information, briefly explain how to choose the most appropriate slot value for the requested slot.

[0074] }

[0075] S202. Based on the sample dialogue dataset and the teacher model, generate sample inference data corresponding to the sample dialogue data.

[0076] Among them, sample inference data can be understood as the process of inferring the value of the sample slot as the sample slot value, that is, how to identify the value of the sample slot as the sample slot value from the sample dialogue data.

[0077] In this embodiment, sample inference data corresponding to the sample dialogue data can be generated based on the sample dialogue dataset and teacher model constructed in the previous steps. The specific process of generating sample inference data can be, for example, inputting the sample dialogue dataset into the teacher model and using the directly output data as the sample inference data corresponding to the sample dialogue data. Alternatively, a series of processing steps can be performed on the output data to generate the sample inference data corresponding to the sample dialogue data. For example, further filtering or verification operations can be performed on the data output by the teacher model to obtain the sample inference data corresponding to the sample dialogue data. The specific filtering or verification operations are not limited here, as long as the sample inference data corresponding to the sample dialogue data can be generated.

[0078] In some embodiments, sample inference data corresponding to the sample dialogue data is generated based on the sample dialogue dataset and the teacher model, including:

[0079] Input the sample dialogue dataset into the teacher model to generate multiple candidate inference data corresponding to the sample dialogue data;

[0080] For each candidate inference data, inference parameters are calculated based on the sample dialogue data. The inference parameters are used to characterize the semantic similarity between each candidate inference data and the sample dialogue data.

[0081] Based on the inference parameters of each candidate inference data, determine the sample inference data corresponding to the sample dialogue data.

[0082] Candidate inference data can be understood as multiple inference data initially generated by the teacher model, which can be filtered to select the final sample inference data through subsequent calculations. Inference parameters can be used to characterize the semantic similarity between each candidate inference data and the sample dialogue data;

[0083] Specifically, the process of generating sample inference data can be as follows: First, input the sample dialogue dataset into the teacher model to generate multiple candidate inference data corresponding to the sample dialogue data. Then, by calculating the inference parameters of each candidate inference data, the final sample inference data corresponding to the sample dialogue data is determined. The method for calculating the inference parameters is not limited. For example, the semantic distance between each candidate inference data and the sample dialogue data can be directly calculated, and the calculated semantic distance can be used as the inference parameter corresponding to the candidate inference data. Alternatively, for each candidate inference data, in addition to calculating the semantic distance between the candidate inference data and the sample dialogue data, the semantic distance between the candidate inference data and other inference data can be further calculated. The inference parameters corresponding to the sample dialogue data are then calculated comprehensively based on the calculated semantic distances. Other inference data can be, for example, candidate inference data other than the current candidate inference data, or interfering inference data, etc.

[0084] Furthermore, after obtaining the inference parameters of each candidate inference data, the sample inference data corresponding to the sample dialogue data can be determined by a preset selection strategy. Specifically, this can be determined by relevant personnel. For example, the minimum inference parameter among the inference parameters of each candidate inference data can be selected, and the candidate inference data corresponding to the selected minimum inference parameter can be used as the sample inference data corresponding to the sample dialogue data. That is, the minimum inference parameter among each inference parameter can be selected to determine the final sample inference data. Alternatively, the final sample inference data can be selected by comparing the size of each inference parameter with the preset parameter. This embodiment does not limit this.

[0085] S203. Train the initial student model using the sample reasoning dataset containing sample reasoning data to obtain the target student model.

[0086] A sample inference dataset can be considered as a dataset used to train an initial student model. A sample inference dataset can include sample inference data and other data related to training.

[0087] In some embodiments, before training the initial student model using a sample inference dataset containing sample inference data to obtain the target student model, the method further includes:

[0088] A sample inference dataset is constructed based on sample dialogue data, sample slots, sample slot values, and sample inference data.

[0089] After obtaining the sample inference data corresponding to the sample dialogue data through the above steps, a sample inference dataset for training can be constructed. For example, a fully labeled sample inference dataset can be constructed based on the obtained sample dialogue data, sample slots, sample slot values ​​and sample inference data to train the initial student model.

[0090] For example, this embodiment can continue training a target student model within a self-rationalization framework based on the annotated dataset {X, S, R, V}. The target student model emphasizes prediction and explanation capabilities. The target student model trained in this embodiment differs from previous post-rationalization models, such as those where inference data is formulated after prediction, or those that employ multi-task formation and creation of basic principles as auxiliary tasks. Here, the sample dialogue data is X, the sample inference data can be R, and S and V are the sample slots and slot values, respectively.

[0091] Specifically, the student model is conditional on generating a sequence that combines sample inference data, sample slots, and corresponding slot values ​​in response to a given dialogue context and slot request. This task is performed by fine-tuning a pre-trained language model. This embodiment also employs a standard language modeling loss for model training, which can be referred to as the fact-reasoning loss. J is the total number of sample slots.

[0092] S204. Obtain the current dialogue dataset in the dialogue system. The current dialogue dataset includes at least the current dialogue data and the slots to be predicted for the current dialogue data.

[0093] S205. Input the current dialogue dataset into the target student model to obtain the predicted slot value of the slot to be predicted and the current inference data corresponding to the predicted slot value. The current inference data is used to characterize the inference process of the value of the slot to be predicted being the predicted slot value.

[0094] S206. Based on the predicted slot values ​​and current inference data, maintain the dialogue state of the dialogue system.

[0095] In a specific implementation, a model f: X×T→Y can be trained in a continuous DST based on a series of dialogue domains T1,…,TK. This model is based on the input X and the task T. k ∈T to predict target Y. In a specific task T k In this context, T interactive dialogues between the system and the user can be represented as X. T ={(A1,U1),(A2,U2)...,(A T U T )}, where A represents the system response and U represents the user input; a predefined set of slots S = {S1,...,S} is provided. J}, where J is a specific task T k The total number of slots. The goal of continuous DST can be based on the dialogue context X. t To predict dialogue state B t Dialogue state B t It can be represented as a set of (slot, slot value) pairs, that is, as in It is slot S J The slot value at time t. Essentially, the Dialogue State Tracker (DST) problem can be defined as training a dialogue state tracker model f: in This represents a simple text connection.

[0096] Therefore, this embodiment can utilize the LLM (Large Language Model) framework for extracting the reasons for selection, and through a "multi-valued parsing" strategy and semantic contrastive reasoning, create a correct selection reasoning process. This rich knowledge is then transferred to the student model for training. This process can begin with dialogue-centric cue data to derive the reasoning process R, which includes dialogue content X and target slot S. j and its value V j And the "multivalued parsing" hint P R Traditional prompts, such as "Tell me why," are different. j >Is V j Hints like "Let's think through this example step by step" tend to highlight only the position of the correct answer in the dialogue to generate reasoning. For example, the input and output of this stage can be represented as f teacher :

[0097] The "Multi-valued parsing" prompt P in this embodiment R It facilitates a more refined reasoning process and enhances the teacher model's ability to move from simple location recognition to in-depth selection reasoning. This methodological advancement from basic recognition to complex process reasoning represents a substantial leap in the model's cognitive capabilities.

[0098] The following is a specific example of a current dialogue dataset provided in this embodiment, which can be used to generate prediction slot values ​​and the current inference data corresponding to the prediction slot values.

[0099] {

[0100] "Instruction": Given the following dialogue, track the state of the dialogue and predict the value of the slot.<alarm_1-new_alarm_name> .

[0101] "input": [User]: I want to check my alarm clock.

[0102] [SYSTEM]: You currently have two alarms set, one of which is set at 6:30 AM and is called Wake Up.

[0103] [User]: Okay, very good.

[0104] [SYSTEM]: Are you interested in adding another alarm clock?

[0105] [User]: Actually, I do want to add another alarm clock, and I want to call the alarm clock "grocery store".

[0106] The slot <alarm_1-new_alarm_name> represents the name used for the new alarm, so the value of slot <alarm_1-new_alarm_name> is ? ​

[0107] "Output": General Store

[0108] }

[0109] This embodiment provides a dialogue state tracking method. The teacher model generates sample inference data corresponding to the sample dialogue data based on the inference prompt data. This enables the initial student model to learn from the sample inference dataset containing the sample inference data, enriching the small model's ability to analyze and interpret complex dialogue scenarios. At the same time, due to the domain independence of the target learning model, the data distribution is aligned across different domains, effectively reducing catastrophic forgetting.

[0110] In some embodiments, inference parameters for each candidate inference data are calculated based on sample dialogue data, including:

[0111] Calculate the first semantic distance between each candidate inference data and the sample dialogue data;

[0112] Calculate the second semantic distance between each candidate inference data and each perturbation inference data in the perturbation inference dataset;

[0113] Based on the first semantic distance and each of the second semantic distances, the inference parameters for each candidate inference data are determined.

[0114] In this embodiment, the perturbation inference dataset can be a collection containing one or more perturbation inference datasets. The perturbation inference dataset can be distinguished from the candidate inference dataset and is used to better generate sample inference datasets. The source of the perturbation inference dataset is not limited. For example, it can be generated in advance based on sample dialogue data and teacher models, or it can be obtained from other terminal devices.

[0115] The first semantic distance can refer to the semantic distance between each candidate inference data and the sample dialogue data, while the second semantic distance can refer to the semantic distance between the candidate inference data and the perturbed inference data.

[0116] In some embodiments, before calculating the second semantic distance between each candidate inference data and each perturbation inference data in the perturbation inference dataset, the method further includes:

[0117] Multiple perturbation dialogue datasets are constructed based on sample dialogue data. Each perturbation dialogue dataset includes at least the perturbation slot value corresponding to the sample slot or the perturbation slot value corresponding to the perturbation slot.

[0118] Multiple perturbation dialogue datasets are input into the teacher model to generate perturbation inference datasets corresponding to the sample dialogue data.

[0119] In specific implementations, neural language models (LMs), such as teacher models, often exhibit hallucinations, generating text data with weak correlations to the input dataset, especially exacerbated by unsupervised annotation. To address this, this embodiment introduces an improved semantic contrastive reasoning selection method, which is crucial for accurately aligning reasoning with corresponding answers.

[0120] Specifically, this embodiment can create a series of candidate inference data and select the most semantically consistent candidate through semantic comparison inference selection technology. For example, this embodiment can introduce policy input perturbation to simulate and correct inference errors, thereby improving the output of the teacher model.

[0121] Figure 6 This is a schematic diagram of a process for generating perturbation inference data according to an embodiment of this application, such as... Figure 6 As shown, this embodiment can introduce two forms of perturbation: value-level and slot-level. When the perturbation slot value subtly modifies the truth value, the slot-level perturbation will completely replace the slot value pair. For example, when the normal slot request is (scheduled date, 10:45 AM), the value-level perturbation can be (scheduled date, 9:15 AM), which is the perturbation slot value corresponding to the sample slot; the slot-level perturbation can be (therapist name, blank gary), which is the perturbation slot value corresponding to the perturbation slot. Thus, the dialogue context content and the value-level or slot-level perturbation are respectively input into the teacher model to obtain two disordered inference data, namely perturbation inference data.

[0122] The perturbation data, from logical to non-logical, demonstrates that the model is particularly effective in inducing a range of inference responses, thus allowing the model to be used with N perturbation data points (PR1,...,PR2). N This enriches the negative sample library of this embodiment. Positive samples can be represented as DC, centered on the dialogue content (i.e., sample dialogue data); the candidate inference data initially generated by the teacher model can be G different inferences (R1,...,R...). G ).

[0123] Figure 7 This is a schematic diagram of a semantic comparison reasoning selection provided in an embodiment of this application, such as... Figure 7 As shown, by utilizing a fixed, pre-trained context encoder and sentence converter, this embodiment can convert these candidate inference data R, perturbation inference data PR, and sample dialogue data DC into semantic representations within the semantic space ε, such as converting DC, R1, ..., R... G ,PR1,...,PR N After passing through the context encoder, a semantic vector can be obtained, which can then be projected into the semantic space on the right (the figure illustrates this with N=3 and G=4 as an example).

[0124] The key to this method lies in the decision-making process based on the semantic proximity of positive samples and the semantic deviation of negative samples, selecting the most suitable prediction from the generated G candidates. This selection process can be expressed as an optimization problem: minDistance(R1,DC) and

[0125] To further quantify the relationship between each candidate inference data and the sample set (i.e., the positive and negative sample libraries), this embodiment introduces a complex scoring mechanism, namely, calculating R for each candidate inference data. i Corresponding inference parameters Where d is the distance function and τ is the temperature scalar.

[0126] By calculating the inference parameters for each candidate inference data, this embodiment can identify the optimal choice for training the student model—the one with the lowest score—based on the consistency between the student model and the actual dialogue context, as well as its difference from the disturbed content. This is the final sample inference data. The method of this embodiment can be applied to black-box LLMs like ChatGPT.

[0127] Figure 8 This is an overall schematic diagram of a dialogue state tracking method provided in another embodiment of this application, as shown below. Figure 8 As shown, the teacher model, through a "multi-valued parsing" strategy and a semantic contrastive reasoning selection method, prompts the large LM to generate correct inference data given the contextual dialogue and the values ​​of the requested time period in the training set. Specifically, a prompt template (i.e., a sample dialogue dataset) can be constructed based on dialogue-related prompts, solution-guided prompts, and multi-valued parsing prompts. In practice, the content of the prompt template can be filled in, and the filled prompt template is input into the large LM model. Sample inference data is generated through semantic contrastive reasoning selection. Then, the student model can train the small LM model based on the labeled sample inference dataset {X, S, R, V} to perform subsequent dialogue state tracking based on the trained target student model. That is, the target student model can output the true slot value and the corresponding inference process based on the input data. Here, X represents the sample dialogue data, S and V represent the sample slots and their corresponding sample slot values, respectively, and R represents the sample inference data.

[0128] For example, the specific content of the sample dialogue dataset can be:

[0129] <Dialogue Context> = [SYSTEM]: What time would you like? [User]: I'd like to schedule an appointment at 10:45 AM... [System]: Would you like to change your call to Rig Christie to 10:30 AM on March 5th? [User]: No, that's not suitable for me.

[0130] <Slot Name> = Service - 4 - Appointment Date, <Actual Data> = 10:45 AM, <Slot Description> = Service means finding a suitable therapist for you, and you can easily make an appointment. appointment_date represents the appointment date.

[0131] <Multiple Value Resolution Hint> = Can you tell me why and your reasoning process? In the context of dialogue state tracking, there are often multiple possible values ​​associated with a requested slot. Please briefly explain how you could choose the most appropriate value for the requested slot by carefully analyzing the dialogue context, user and system intent, and considering any confirmation or rejection information.

[0132] The sample inference data output by the teacher model could include: In a given dialogue, a user initially requests an appointment at 9:15 AM. However, the user subsequently changes the time to 10:45 AM. The system then suggests an alternative at 10:30 AM on March 5th, but the user declines. Based on this information, the requested time slot…<services-appointment_date> The most suitable value for the time slot is "10:45 AM". Although the system suggested an alternative time, the user rejected it, indicating that they still prefer the original time of 10:45 AM. Therefore, considering the user's preference and their rejection of the alternative time, "10:45 AM" is the most suitable time slot for the requested period.<services-appointment_date> The most suitable value for "".

[0133] Extensive empirical verification shows that, compared with the prior art, the method provided in this embodiment has made substantial progress by addressing the serious defects in the current DST method, significantly improving the performance and generalization ability of the DST system, and setting a new standard for a powerful and adaptable dialogue system that can continuously learn and develop.

[0134] For example, this embodiment uses a pattern-guided dialogue dataset for experiments. The dataset can include 44 service and slot descriptions across 19 domains. This embodiment focuses on a single service dialog box, randomly selecting 15 tasks from the 44 services. The training sample size and time period vary for each service. To ensure robustness, this embodiment experiments with five tasks in a randomized order. The performance of the DST is evaluated using the widely adopted Joint Objective Accuracy (JGA) metric, which requires accurate prediction of all slot values. After training on task Ti, this embodiment will... j,i This is represented as JGA on the test set for task Ti. CL performance can be evaluated using the following three metrics.

[0135] (i) This can represent the average JGA for all tasks after the final task TK training.

[0136] (ii) Forward transfer Generalization can be evaluated by measuring zero-sample performance.

[0137] (iii) Reverse transfer Resistance to forgetting can be quantified by assessing the impact of new learning on previous tasks.

[0138] Specifically, the model in this embodiment can be evaluated based on the existing continuous DST baseline, and the model can be continuously fine-tuned based on new task data; |M| instances of each task Ti are stored in memory Mi for joint training with new tasks; memory is maintained through Elastic Weight Consolidation (EWC) and used to compute the regularized Fisher information matrix; the pre-trained model is frozen through Adapter-based Continual Learning (AdapterCL), and the remaining adapters are trained independently for each task; the backbone model is frozen through Continuous Prompt Tuning (CPT), and soft prompts are continuously trained through forward and backward knowledge transfer; and DST is reformatted into a question-answering task using enhanced retrieval in context learning through Dialogue State Tracking as Example Guided Question Answering (DST-EGQA).

[0139] Furthermore, this embodiment may also include a method of using memory playback and multi-task learning as performance limits.

[0140] Specifically, this embodiment selects ChatGPT (using the gpt-3.5-turbo API) and LLaMA-2-70B as teacher models, and uses the generated inference data to fine-tune smaller student models, including T5-small, T5-base, FlanT5-XL, and LLaMA-7B. The teacher model's temperature is set to 0.7, generating five candidate inferences (G=5) and three slot perturbations and three slot perturbations for negative inference (N=6); the parameter τ is set to 0.8, using Euclidean distance as the metric; and the memory size |M| for each task is maintained at 50. Table 1 lists the results of various methods.

[0141] Table 1

[0142]

[0143]

[0144] Table 1 shows the CL metric results and dependencies on other continuous learning techniques, displaying the mean and standard deviation for each method. This example compares models trained sequentially on 15 tasks in the SGD dataset and summarizes the results for five domains. The last four rows provide multi-task processing results, which is an upper limit. All rows using memory for data replay have M=50.

[0145] Firstly, it can be observed that the RoS method in this embodiment effectively improves CL performance by injecting inference knowledge. Specifically, the RoS method in this embodiment significantly outperforms the standard fine-tuning on T5-small, increasing the average JGA from 44.1% to 59.0%, and also showing improvements in FWT and BWT. This highlights the value of integrating inference into continuous DST. Compared to the previous best-performing DST-EGQA, this embodiment achieves new state-of-the-art performance across all metrics, significantly improving the average.

[0146] Without additional memory, JGA improved from 55.5% to 59.0%, demonstrating the effectiveness of the method in reducing historical task forgetting. Furthermore, when memory was available, RoS jumped from 59.0% to 72.1%, even surpassing CPT's multitasking performance. Among various student models, the LLaMA-7B RoS, fine-tuned based on ChatGPT inference, stood out, significantly improving the average JGA from 59.0% to 68.7% and the FWT from 25.5% to 51.9%.

[0147] Secondly, it can be observed that the RoS method in this embodiment has strong generalization ability; that is, the method in this embodiment significantly improves the FWT metric, demonstrating strong generalization and zero-shot learning capabilities. It bridges the distribution gap across domains by creating a systematic inference chain, which is particularly useful when dealing with semantically similar but differently described time periods (such as...).<services-appointment_date> and<hotels-check_in_date> The similarities are quite obvious. Unlike traditional models, the model in this embodiment uses reasoning to identify similarities, enhancing its adaptability to unseen slots.

[0148] To further evaluate this generalization ability, this embodiment also introduces a new 16th task using the MultiWOZ 2.4 dataset. Table 2 shows the zero-shot performance of the 16th cross-dataset task using the MultiWOZ 2.4 dataset.

[0149] Table 2

[0150]

[0151] As shown in Table 2, the zero-shot performance of this task follows the sequential training of the first 15 tasks of SGD. Compared to a backbone with similar parameter sizes, RoS (T5-small) outperforms the T5-base CPT baseline, with a 3.7% increase in average JGA across domains, from 10.92% to 14.62%.

[0152] This improvement was particularly noticeable in the tourism sector, where JGA rose from 10.05% to 15.74%. Compared to various common backbones, the improvement in domain-agnostic reasoning increased by an average of 6%, further confirming the robust generalization ability of RoS.

[0153] To evaluate the illusion rate of the teacher model, this embodiment also uses SelfCheckGPT and evaluates it using BERTScore. BERTScore provides scores ranging from [0.0, 1.0], where higher values ​​indicate a greater likelihood of non-factual content, meaning an increased probability of illusion. The evaluation in this embodiment includes the average BERTScore of the five candidate inferences initially generated by the teacher model, the BERTScore after applying the semantic contrastive inference selection method, and the illusion rate of the extracted students. Table 3 shows the illusion rate of the inferences generated by the teacher model after applying the contrastive selection method in this embodiment.

[0154] Table 3

[0155]

[0156] As shown in Table 3, the BERTS scores for the initial five inference candidates were 19.31% and 21.7% for both the ChatGPT and LLaMA-2-70B teacher models, respectively. After applying the semantic contrastive inference selection method, the BERTS scores decreased to 14.2% and 15.9%, respectively, confirming the effectiveness of the proposed method in enhancing the realism of inference. There was no significant difference in the BERTS scores for the student model when different teacher models were used for refinement. Furthermore, this embodiment observed that the quality of inference generated by the student model improved with increasing model size.

[0157] In addition to the above description, the dialogue state tracking method of this embodiment also provides potential alternatives to the aforementioned key functions: for example, techniques such as reinforcement learning of the decision process or using generative adversarial networks (GANs) to simulate and learn various reasoning scenarios can enrich the RoS method and be used for reasoning enhancement methods in RoS; semantic contrastive reasoning selection involves separating the variable factors in the dialogue data, which can more accurately manipulate and compare semantic features to gain a clearer understanding of causal relationships in the dialogue, thereby potentially improving the accuracy of reasoning; multi-value resolution strategies can apply probabilistic graphical models, such as Bayesian networks, to infer the most likely dialogue state based on observed data, effectively managing multiple values. These models provide robust statistical methods for uncertainty management in complex interactions, which can improve state resolution.

[0158] Furthermore, this embodiment can enhance the capabilities of the model in the following aspects.

[0159] Firstly, by providing more sophisticated meta-reasoning algorithms, increasingly complex dialogues can be handled more efficiently. This includes the integration of neuro-symbolic reasoning methods, which combine deep learning with symbolic reasoning to improve interpretability and flexibility. This allows the system to understand and process user intent and context more deeply, thereby improving decision-making accuracy and user satisfaction.

[0160] Secondly, by enhancing the domain bootstrapping process through transfer learning techniques, the system can adapt to new domains more quickly and effectively, and leverage domain-specific data more efficiently. This increased adaptability will make the system more versatile and deployable across different industries and service sectors without requiring extensive retraining.

[0161] By integrating a real-time learning mechanism, the system can continuously update its knowledge base as it processes new dialogue instances. This real-time update ensures that the system stays up-to-date with the latest information and user preferences, thereby improving responsiveness and relevance.

[0162] Then, by using differential privacy technology and secure multi-party computation, stronger data privacy and security measures can be implemented to protect the sensitive information processed by the system, making the dialogue system trustworthy for more sensitive applications, such as healthcare or financial services.

[0163] Finally, more intuitive and user-friendly interfaces can be developed for system administrators and end users, which may include more sophisticated natural language understanding components to handle different language expressions, thereby making the technology more accessible and usable, and potentially increasing its adoption and effectiveness.

[0164] As can be seen from the above description, the goal of the Choice of Cause (RoS) distillation method is to improve the effectiveness and adaptability of the Continuous Dialogue State Tracking (DST) system by introducing advanced meta-reasoning functions that go beyond the limitations of traditional learning. It aims to address specific challenges in dialogue systems, improve the operational efficiency of dialogue systems, and expand the applicability of dialogue systems.

[0165] The RoS method further enhances the student model in two novel ways. First, it endows the student model with the ability to perform reasoning (i.e., improves the reasoning ability of the small model, which can be called meta-reasoning). For example, it proposes a "multi-valued analysis" strategy corresponding to DST, prompting LLM to generate a "choice chain" to identify and articulate the optimal slot value selection from various options. This principle is extracted into a smaller student model to enhance its meta-reasoning ability, making it easier to deploy to end systems. Second, by utilizing higher-level abstractions during the reasoning process, it enhances the model's ability to integrate and reason about new information from various domains, possessing cross-domain knowledge transfer capabilities. This helps it adapt faster with less data requirements without requiring extensive retraining.

[0166] Meta-reasoning, guided by the view that domain-specific dialogues represent only a portion of underlying meta-knowledge, can be seen as a form of multi-domain expansion in the CL process. It employs an enhanced multi-domain perspective, incorporating fragments of meta-knowledge from domain-specific dialogues during continuous learning. This domain-guided strategy facilitates the transfer and expansion of meta-knowledge, surpassing traditional single-view reasoning. The domain-guided process enhances the model's ability to dissect complex dialogues from multiple possible values. Its domain-agnostic nature, through the teacher model's reasoning across different tasks and domains rather than merely imparting specific information or tasks, effectively reduces forgetting during continuous learning.

[0167] Secondly, to enhance the correct reasoning of the teacher model and reduce illusions, pattern-guided prompts are integrated, improving source-target relevance and semantic coherence. Based on semantic contrastive reasoning selection, the distillation process is optimized by comparing causes from the teacher model, reducing errors in existing reasoning methods. This ensures that only the most relevant and accurate information is retained and strengthened in the student model, leading to more precise updates and reduced noise during model training. This improves the student model's reasoning ability, accuracy, and robustness, promoting more reliable and effective knowledge transfer. This is applicable even in black-box model scenarios. Furthermore, improvements to the multi-valued resolution strategy ensure the accuracy of dialogue interaction in multi-valued scenarios, guaranteeing the clarity and accuracy of state tracking.

[0168] In summary, the dialogue state tracking method provided in this embodiment, by leveraging the meta-reasoning capabilities and semantic contrastive reasoning selection of RoS and optimizing the knowledge distillation process, significantly improves the accuracy and robustness of the DST model, and more effectively achieves cross-domain learning and adaptation.

[0169] At the same time, RoS minimizes the need for large datasets and continuous retraining by effectively managing knowledge transfer and retention, and enabling zero-shot functionality, which reduces computational overhead and the resources required to maintain and update DST systems. This not only simplifies the deployment process but also makes the technology more cost-effective to maintain and scale.

[0170] Furthermore, it addresses existing technical problems: for example, this embodiment directly solves the catastrophic forgetting problem prevalent in DST continuous learning scenarios by integrating task-specific knowledge protection and intelligent parameter update mechanisms. Traditional DST systems, without specific training, are typically unable to adapt to new domains. The RoS method in this embodiment enhances domain adaptability, enabling the system to perform effectively across various domains and tasks without prior domain-specific training.

[0171] Therefore, the dialogue state tracking method provided in this embodiment not only improves performance, but also expands the applicability of the DST system in practical applications, making it more practical and effective in dynamic environments.

[0172] Corresponding to the dialogue state tracking method in the above embodiments, Figure 9 This is a structural block diagram of a dialogue state tracking device provided in one embodiment of this application. For ease of explanation, only the parts related to the embodiment of this application are shown.

[0173] Reference Figure 9 The device includes:

[0174] The acquisition module 301 is used to acquire the current dialogue dataset in the dialogue system. The current dialogue dataset includes at least the current dialogue data and the slots to be predicted for the current dialogue data.

[0175] The input module 302 is used to input the current dialogue dataset into the target student model to obtain the predicted slot value of the slot to be predicted and the current inference data corresponding to the predicted slot value. The current inference data is used to characterize the inference process in which the value of the slot to be predicted is the predicted slot value.

[0176] Maintenance module 303 is used to maintain the dialogue state of the dialogue system based on the predicted slot value and the current inference data.

[0177] This embodiment provides a dialogue state tracking device. An acquisition module obtains the current dialogue dataset from the dialogue system. The current dialogue dataset includes at least the current dialogue data and the slots to be predicted for the current dialogue data. An input module inputs the current dialogue dataset into a target student model to obtain the predicted slot values ​​and the corresponding current inference data. The current inference data represents the inference process of determining the value of the slot to be predicted. A maintenance module maintains the dialogue state of the dialogue system based on the predicted slot values ​​and the current inference data. Using this device, by obtaining the predicted slot values ​​and corresponding current inference data from the target student model output, the dialogue state of the dialogue system is maintained. This allows the dialogue system to dynamically integrate and apply knowledge from various domains, effectively mitigating catastrophic forgetting and enhancing the system's adaptability and learning efficiency.

[0178] Optionally, the dialogue state tracking device provided in this embodiment further includes:

[0179] The first construction module is used to construct a sample dialogue dataset based on the sample dialogue data before inputting the current dialogue dataset into the target student model to obtain the predicted slot value of the slot to be predicted and the current inference data corresponding to the predicted slot value. The sample dialogue dataset includes at least the sample slot of the sample dialogue data, the sample slot value corresponding to the sample slot, and the inference hint data.

[0180] The generation module is used to generate sample inference data corresponding to the sample dialogue data based on the sample dialogue dataset and the teacher model.

[0181] The training module is used to train the initial student model using a sample reasoning dataset containing sample reasoning data to obtain the target student model.

[0182] Optionally, the generation module includes:

[0183] The generation unit is used to input the sample dialogue dataset into the teacher model and generate multiple candidate inference data corresponding to the sample dialogue data.

[0184] The computational unit is used to calculate the inference parameters of each candidate inference data based on the sample dialogue data. The inference parameters are used to characterize the semantic similarity between each candidate inference data and the sample dialogue data.

[0185] The determining unit is used to determine the sample inference data corresponding to the sample dialogue data based on the inference parameters of each candidate inference data.

[0186] Optionally, the computing unit includes:

[0187] The first computational subunit is used to calculate the first semantic distance between each candidate inference data and the sample dialogue data;

[0188] The second computational subunit is used to calculate the second semantic distance between each candidate inference data and each perturbation inference data in the perturbation inference dataset, respectively.

[0189] The sub-unit is determined to determine the inference parameters for each candidate inference data based on the first semantic distance and each second semantic distance.

[0190] Optionally, the computing unit may also include:

[0191] Construct sub-units to build multiple perturbation dialogue datasets based on sample dialogue data before calculating the second semantic distance between each candidate inference data and each perturbation inference data in the perturbation inference dataset. The perturbation dialogue datasets include at least the perturbation slot value corresponding to the sample slot or the perturbation slot value corresponding to the perturbation slot.

[0192] The generation sub-unit is used to input multiple perturbation dialogue datasets into the teacher model before calculating the second semantic distance between each candidate inference data and each perturbation inference data in the perturbation inference dataset, thereby generating the perturbation inference dataset corresponding to the sample dialogue data.

[0193] Optionally, the determination unit is specifically used for:

[0194] Select the smallest inference parameter among the inference parameters of each candidate inference data, and use the candidate inference data corresponding to the smallest inference parameter as the sample inference data corresponding to the sample dialogue data.

[0195] Optionally, the dialogue state tracking device provided in this embodiment further includes:

[0196] The second building module is used to construct a sample inference dataset based on sample dialogue data, sample slots, sample slot values, and sample inference data before training the initial student model using a sample inference dataset containing sample inference data to obtain the target student model.

[0197] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0198] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0199] This application also provides a terminal device. Figure 10 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application, as shown below. Figure 10 As shown, the terminal device includes: at least one processor 401, a memory 402, an input device 403, an output device 404, and a computer program stored in the memory 402 and executable on at least one processor 401. When the processor 401 executes the computer program, it implements the steps in any of the above-described method embodiments.

[0200] Input device 403 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the terminal device. Output device 404 may include display devices such as a display screen.

[0201] This application also provides a computer-readable storage medium storing a computer program, which, when executed by processor 401, implements the steps in the above-described method embodiments.

[0202] This application provides a computer program product that, when run on a mobile terminal, enables the mobile terminal to implement the steps described in the above-described method embodiments.

[0203] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by the processor 401, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a device / terminal equipment, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0204] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0205] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein 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 application.

[0206] In the embodiments provided in this application, it should be understood that the disclosed apparatus / terminal devices and methods can be implemented in other ways. For example, the apparatus / terminal device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0207] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0208] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A dialog state tracking method, characterized by, The method comprises the following steps: obtaining a current dialogue data set in a dialogue system, the current dialogue data set comprising at least current dialogue data and a to-be-predicted slot of the current dialogue data; inputting the current dialogue data set into a target student model to obtain a predicted slot value of the to-be-predicted slot and current reasoning data corresponding to the predicted slot value, wherein the current reasoning data is used to represent a reasoning process of the to-be-predicted slot taking the predicted slot value as the value; maintaining a dialogue state of the dialogue system based on the predicted slot value and the current reasoning data.

2. The dialog state tracking method of claim 1, wherein, Before the step of inputting the current dialogue data set into the target student model to obtain the predicted slot value of the to-be-predicted slot and the current reasoning data corresponding to the predicted slot value, the method further comprises the following steps: constructing a sample dialogue data set based on sample dialogue data, the sample dialogue data set comprising at least a sample slot of sample dialogue data, a sample slot value corresponding to the sample slot, and reasoning prompt data; generating sample reasoning data corresponding to the sample dialogue data according to the sample dialogue data set and a teacher model; training an initial student model by using a sample reasoning data set containing the sample reasoning data to obtain the target student model.

3. The dialog state tracking method of claim 2, wherein, The step of generating sample reasoning data corresponding to the sample dialogue data according to the sample dialogue data set and the teacher model comprises the following steps: inputting the sample dialogue data set into the teacher model to generate a plurality of candidate reasoning data corresponding to the sample dialogue data; for each candidate reasoning data, calculating a reasoning parameter of each candidate reasoning data based on the sample dialogue data, wherein the reasoning parameter is used to represent a semantic similarity between each candidate reasoning data and the sample dialogue data; determining the sample reasoning data corresponding to the sample dialogue data according to the reasoning parameter of each candidate reasoning data.

4. The dialog state tracking method of claim 3, wherein, The step of calculating the reasoning parameter of each candidate reasoning data based on the sample dialogue data comprises the following steps: calculating a first semantic distance between each candidate reasoning data and the sample dialogue data; respectively calculating a second semantic distance between each candidate reasoning data and each perturbation reasoning data in a perturbation reasoning data set; determining the reasoning parameter of each candidate reasoning data according to the first semantic distance and the second semantic distance.

5. The dialog state tracking method of claim 4, wherein, Before the step of respectively calculating the second semantic distance between each candidate reasoning data and each perturbation reasoning data in the perturbation reasoning data set, the method further comprises the following steps: constructing a plurality of perturbation dialogue data sets based on sample dialogue data, wherein each perturbation dialogue data set comprises at least a perturbation slot value corresponding to the sample slot or a perturbation slot value corresponding to a perturbation slot; inputting the plurality of perturbation dialogue data sets into the teacher model to generate a perturbation reasoning data set corresponding to the sample dialogue data.

6. The dialog state tracking method of claim 3, wherein, The step of determining the sample reasoning data corresponding to the sample dialogue data according to the reasoning parameter of each candidate reasoning data comprises the following steps: selecting a minimum reasoning parameter in the reasoning parameter of each candidate reasoning data, and taking the candidate reasoning data corresponding to the minimum reasoning parameter as the sample reasoning data corresponding to the sample dialogue data.

7. The dialog state tracking method of claim 2, wherein, Before the initial student model is trained by using the sample inference data set containing the sample inference data to obtain a target student model, the method further comprises: According to the sample dialogue data, the sample slot, the sample slot value and the sample inference data, the sample inference data set is constructed.

8. A dialog state tracking apparatus characterized by comprising: Comprise: An acquisition module is configured to acquire a current dialogue data set in a dialogue system, the current dialogue data set comprising at least current dialogue data and a to-be-predicted slot of the current dialogue data; An input module is configured to input the current dialogue data set into a target student model to obtain a predicted slot value of the to-be-predicted slot and current inference data corresponding to the predicted slot value, wherein the current inference data is used to represent an inference process in which the value of the to-be-predicted slot is the predicted slot value; A maintenance module is configured to maintain a dialogue state of the dialogue system based on the predicted slot value and the current inference data.

9. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the method of any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the method of any one of claims 1-7.

Images