Task type multi-round dialogue system enhancement method based on general intention

Abstract

The invention discloses a task type multi-round dialogue system enhancement method based on a general intention. The method comprises the following steps: loading a general intention recognition model of a task type multi-round dialogue; obtaining current user input and a corresponding man-machine conversation history; inputting the current user input and the corresponding man-machine conversation history into a general intention recognition model, and outputting a corresponding general intention list by the model; extracting the intention and task elements of the dialogue from the general intention enhanced task dialogue large model input by the current user and the corresponding man-machine dialogue history, and generating a corresponding reply; and repeating the steps until the dialogue task is completed, and ending the dialogue. According to the method, the capability of a large language model thinking chain is utilized, a general intention is introduced into a task type multi-round dialogue system in a large language model era to serve as an auxiliary means, the intention recognition capability in the task type multi-round dialogue system is effectively improved, and the success rate of the whole dialogue of the system is effectively increased.

Description

Technical Field

[0001] The present invention relates to the field of task-based multi-turn conversations, and particularly to a method for enhancing a task-based multi-turn conversation system based on general intents. Background Art

[0002] Multi-turn conversation systems have become one of the core technologies in the field of human-computer interaction. As an important branch, the task-based multi-turn conversation system (TOD) aims to complete specific domain tasks (such as ticket booking, product query, schedule arrangement, etc.) through multi-turn interactions with users, and it demonstrates significant application value in scenarios such as intelligent customer service and vertical domain assistants. Although current large language models already possess quite powerful multi-turn conversation capabilities, existing technologies still face many challenges in actual implementation.

[0003] Research shows that current models are highly dependent on domain-specific data. Especially in vertical domains such as healthcare and finance, due to the scarcity of labeled data and the high cost of acquisition, model training is insufficient, making it difficult to cover long-tail scenarios. For example, the patent "Task-oriented 1+N-based Multi-turn Conversation Method and System" with the application number CN202211317437.9 can only give responses using slots in a preset domain through the source service corresponding to the intent, and can only respond in a chatting manner in non-preset domains.

[0004] In the prior art, there have been studies attempting to capture general semantic features in conversations through pre-trained language models, but their ability to abstract structured behavior patterns is still insufficient. Most solutions only implicitly model the conversation process through end-to-end learning, lacking decoupled analysis of explicit behavior logic, resulting in unstable performance of the model when facing domain migration. For example, in the patent "A Task Multi-turn Conversation Interaction Method Based on a Combination of Large Models" with the application number CN202410476972.1, the labeled data used to train the intent recognition large model all come from labeled data in specific business scenarios.

[0005] In addition, although existing datasets (such as MultiWOZ, Schema-Guided Dialogue) provide multi-domain annotations, the cross-domain transferability of behavior patterns has not been fully explored. Therefore, there is an urgent need for a technical solution that can effectively identify general conversation behaviors, reduce the need for domain data, and improve the accuracy of multi-turn interactions to promote the large-scale application of task-based conversation systems. Summary of the Invention

[0006] Aiming at the above-mentioned defects of the prior art, the purpose of the present invention is to provide a method for enhancing a task-based multi-turn conversation system based on general intents, aiming to solve the problems of poor domain transferability, large demand for domain data, and low accuracy of intent understanding in the prior art in the field of task-based multi-turn conversation systems.

[0007] The present invention adopts the following technical solutions to solve the technical problems:

[0008] In a first aspect, a method for enhancing a task-based multi-turn dialogue system based on general intents includes the following steps:

[0009] Step S1, load the general intent recognition model for task-based multi-turn dialogue;

[0010] Step S2, obtain the current user input and the corresponding human-machine dialogue history;

[0011] Step S3, input the current user input and the corresponding human-machine dialogue history into the general intent recognition model, and the model outputs the corresponding list of general intents;

[0012] Step S4, input the current user input and the corresponding human-machine dialogue history into the task dialogue large model for general intent enhancement, extract the intent and task elements of the dialogue, and generate corresponding responses;

[0013] Step S5, repeat steps S2 to S4 until the dialogue task is completed and the dialogue ends.

[0014] Further, in step S1, the training process of the general intent recognition model for task-based multi-turn dialogue includes:

[0015] Randomly extract a part of the data from several task-based multi-turn dialogue datasets and perform sentence-level annotation based on the formulated general intents;

[0016] Construct the annotation results into a dataset according to a prompt template structure;

[0017] Fine-tune the model based on this dataset to strengthen the adaptation to the language environment in the task-based multi-turn dialogue system.

[0018] Further, the prompt structure for identifying the general intent of the user's statement consists of the following parts: {The statement defining the role of the large language model as the dialogue intent classifier}\n{Specific classification task description}\n{Precautions}\n{General intent system}\n{Example dialogue and intent classification results}\n{Input historical dialogue}\n{User dialogue to be classified}\n{Prefix sentence guiding the model to classify}.

[0019] Further, in step S2, the general intents in the list are divided into the following categories: inquiry intent, reply intent, modification intent, request intent, notification intent, greeting intent, complaint intent, gratitude intent, apology intent, farewell intent, and other intents.

[0020] Furthermore, the response intents are divided into 9 secondary intents, namely response notification, response affirmation, response negation, response ignorance, response awareness, response acceptance, response rejection, response completion, and response incompletion.

[0021] Furthermore, in step S4, the training process of the general intent enhanced task dialogue large model includes:

[0022] Obtain the original task dialogue large model; this large model is fine-tuned with a task-based dialogue dataset or only pre-trained without fine-tuning;

[0023] Prepare the fine-tuning data for general intent enhancement; if the original task dialogue large model has been fine-tuned, first obtain the data used for its fine-tuning; if the original large model has not been fine-tuned, use the same dataset as in the above steps; if it is the dataset originally used for its fine-tuning, it is necessary to first annotate the general intent for it; the annotation is performed using the general intent recognition model fine-tuned in the above steps for model-based automatic annotation.

[0024] Furthermore, construct the fine-tuning data into a specific prompt template form, which includes: adding a description of the general intent to the task description section of the instruction template used; adding a description of the general intent of each round of user dialogue to the human-machine dialogue history section of the instruction template used; adding a description of the general intent of the current round of user dialogue to the current dialogue section of the instruction template used; and finally performing model fine-tuning.

[0025] In a second aspect, the present invention also provides a terminal device, which includes a memory, a processor, and a program of the method for enhancing a task-based multi-turn dialogue system based on general intent stored in the memory and executable on the processor. When the processor executes the program of the method for enhancing a task-based multi-turn dialogue system based on general intent, it implements the steps of the method for enhancing a task-based multi-turn dialogue system based on general intent as described above.

[0026] In a third aspect, the present invention also provides a computer-readable storage medium, on which a program of the method for enhancing a task-based multi-turn dialogue system based on general intent is stored. When the program of the method for enhancing a task-based multi-turn dialogue system based on general intent is executed by a processor, it implements the steps of the method for enhancing a task-based multi-turn dialogue system based on general intent as described above.

[0027] Beneficial effects:

[0028] Based on the natural language understanding ability of large language models, the present invention fully explores the general behavior patterns in task-based dialogues, systematically identifies and models such general behaviors, theoretically reducing the dependence on domain-specific data, enhancing the robustness of dialogue strategies, and alleviating the problem of insufficient data. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flowchart of an enhanced method for a task-based multi-turn dialogue system based on general intents according to the present invention;

[0030] Figure 2 is an overview diagram of an intent system including 24 general intents organized in a hierarchical manner according to the present invention;

[0031] Figure 3 are instruction examples of the general intent recognition large model and the task dialogue large model for general intent enhancement in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0033] In the detailed description of the present invention, an enhanced method for a task-based multi-turn dialogue system based on general intents is proposed to solve the problem of poor intent recognition ability caused by the lack of domain data in traditional task-based multi-turn dialogue systems. The method aims to improve the robustness of the task-based multi-turn dialogue system by exploring the general intents in dialogue behaviors.

[0034] To facilitate the understanding of the entire content of this disclosure by readers, the following terms appearing in this disclosure are now explained. It should be noted that the term explanations in this article are only for assisting readers' understanding and do not constitute a limitation to the technical solutions of this disclosure;

[0035] Large language model: refers to a deep learning model trained with a large amount of text data, which can generate natural language text or understand the meaning of language text. Usually, it refers to a language model with scale parameters reaching billions or more;

[0036] Task-based multi-turn dialogue: is a human-computer interaction system specifically designed to help users complete specific tasks. Different from open-domain dialogue systems, TOD systems focus on solving specific problems or performing specific tasks, such as booking a restaurant, querying the weather, purchasing an air ticket, providing navigation guidance, etc.

[0037] Reference Figure 1 , the present invention discloses a method for enhancing a task-based multi-turn dialogue system based on general intents, including the following steps:

[0038] Step S1, load the general intent recognition model for task-based multi-turn dialogue;

[0039] Step S2, obtain the current user input and the corresponding human-machine dialogue history;

[0040] Step S3, input the current user input and the corresponding human-machine dialogue history into the general intent recognition model, and the model outputs the corresponding list of general intents; the general intents in the list can be classified into the following categories: inquiry intent, reply intent, modification intent, request intent, notification intent, greeting intent, complaint intent, gratitude intent, apology intent, farewell intent, and other intents.

[0041] Step S4, input the current user input and the corresponding human-machine dialogue history into the task dialogue large model for general intent enhancement, extract the intent and task elements of the dialogue, and generate the corresponding reply;

[0042] Step S5, repeat steps S2 to S4 until the dialogue task is completed and the dialogue ends.

[0043] For further optimizing the technical solution, in step S1, the training process of the general intent recognition model for task-based multi-turn dialogue includes:

[0044] Randomly extract a part of the data from several task-based multi-turn dialogue datasets and perform sentence-level annotation based on the formulated general intents;

[0045] Construct the annotation results into a dataset according to a prompt template structure;

[0046] Based on this dataset, perform model fine-tuning to strengthen the adaptation to the language environment in the task-based multi-turn dialogue system.

[0047] The prompt structure for identifying the general intent of the user's statement consists of the following parts: {The role definition statement for setting the large language model as the dialogue intent classifier}\n{Specific classification task description}\n{Precautions}\n{General intent system}\n{Example dialogue and intent classification results}\n{Input historical dialogue}\n{User dialogue to be classified}\n{Prefix sentence to guide the model for classification}\n. Among them, {General intent system} includes the names and meanings of all the above intents. In addition, several examples of each intent can be added after the meaning, or there can be no examples. {Example dialogue and intent classification results} can include a single example or multiple examples, or there can be no examples. The example content includes the complete dialogue between the user side and single-round or multi-round, the prefix sentence to guide the model for classification, and the classification result.

[0048] Further optimize the technical solution. The general intention system consists of 24 intentions organized hierarchically as shown in Figure 2 the following, including:

[0049] I. Inquiry intention. The meaning of this intention is that the speaker asks the other party about a certain concept, event, or entity. According to the direction of the inquiry, this intention can be divided into 4 secondary intentions, namely ask_info (inquire about information), ask_if (inquire whether), ask_compare (inquire for comparison), and ask_reason (inquire about the reason). Specifically, ask_info means that the speaker asks about a specific value; ask_if means that the speaker hopes to get a yes or no answer, and asking about time, place, method, etc. also belongs to the category of this intention; ask_compare means that the speaker gives several options and hopes that the other party can choose a better answer from them; ask_reason means that the speaker hopes to get the reason for a certain result.

[0050] II. Reply intention. The meaning of this intention is that the speaker replies to a certain question or request from the other party. According to the reply method, this intention can be divided into 9 secondary intentions, namely reply_inform

[0051] (reply to inform), reply_affirm (reply affirmatively), reply_deny (reply negatively), reply_dont_know

[0052] (reply not knowing), reply_ack (reply acknowledging), reply_accpet (reply accepting), reply_refuse (reply refusing), reply_finish (reply finishing), and reply_not_finish (reply not finishing). Specifically, reply_inform means that the speaker replies with a specific value to the question asked by the other party in the previous few rounds of conversation, corresponding to ask_info; reply_affirm and reply_deny mean that the speaker gives an affirmative or negative answer to the question of ask_if asked by the other party in the previous few rounds; reply_ack means that the speaker only indicates that they understand the other party's words, such as descriptions like "good", "mmm"; reply_accpet and reply_refuse mean that the speaker accepts or refuses the request intention of the other party; reply_finish and reply_not_finish mean that the speaker replies to the other party that they have completed or not completed the request intention of the other party.

[0053] III. Modification intention. The meaning of this intention is that the speaker hopes to modify the previously given information or request. According to the modification method, it can be divided into modify_cover, modify_complement, and modification cancellation. Modify_cover means that the speaker hopes to completely cover the value of a certain slot originally given, such as a change in the number of people; modify_complement means that the speaker hopes to supplement new values to the original value of the slot, such as an addition to the requirements; modification cancellation means that the speaker hopes to invalidate the value of a certain slot previously given.

[0054] IV. Request intention. The meaning of this intention is that the speaker (usually the user) requests a certain service from the other party, which is generally bound to a specific business scenario. As a sign that the user hopes to start a certain task, this intention can be utilized by the task dialogue system.

[0055] V. Informing intention. The meaning of this intention is that the speaker is informing the other party of some information. Different from the reply to inform, this intention means that the speaker actively informs the other party of information.

[0056] VI. In addition, there are a total of 6 special intentions including greeting, apology, complaint, gratitude, goodbye, and others.

[0057] Further optimize the technical solution. In step S4, the training process of the task dialogue large model with enhanced general intention includes:

[0058] Obtain the original task dialogue large model; this large model can be fine-tuned with a task-based dialogue dataset, or it can be only pre-trained without fine-tuning;

[0059] Prepare the fine-tuning data for enhanced general intention; the preparation process of this fine-tuning data for enhanced general intention is as follows: if the original task dialogue large model has been fine-tuned, first obtain the data used for its fine-tuning; if the original large model has not been fine-tuned, use the same dataset as in step S1; if it is the dataset it was originally fine-tuned on, it is necessary to first annotate it with general intention; the annotation is performed using the general intention recognition model fine-tuned in step S1 for model-based automatic annotation.

[0060] Construct the fine-tuning data into a specific prompt template form, which includes: adding a description of the general intention to the task description section of the instruction template used; adding a description of the general intention of each round of user dialogue to the human-machine dialogue history section of the instruction template used; adding a description of the general intention of the current round of user dialogue to the current dialogue section of the instruction template used; and finally performing model fine-tuning.

[0061] Second aspect, the present invention also provides a terminal device, which includes a memory, a processor, and a program of the method for enhancing a task-based multi-turn dialogue system based on general intent stored in the memory and executable on the processor. When the processor executes the program of the method for enhancing a task-based multi-turn dialogue system based on general intent, the steps of the method for enhancing a task-based multi-turn dialogue system based on general intent as described above are implemented.

[0062] Third aspect, the present invention also provides a computer-readable storage medium, on which a program of the method for enhancing a task-based multi-turn dialogue system based on general intent is stored. When the program of the method for enhancing a task-based multi-turn dialogue system based on general intent is executed by a processor, the steps of the method for enhancing a task-based multi-turn dialogue system based on general intent as described above are implemented.

[0063] The present invention utilizes the ability of the thought chain of large language models, introduces general intent as an auxiliary means in the task-based multi-turn dialogue system in the era of large language models, and effectively improves the intent recognition ability in the task-based multi-turn dialogue system and the success rate of the overall system dialogue.

[0064] Embodiment

[0065] The present embodiment provides a method for enhancing a task-based multi-turn dialogue system based on general intent, including the following steps:

[0066] Step S1, load the general intent recognition model of the task-based multi-turn dialogue;

[0067] In the embodiment of the present invention, training the general intent recognition model includes the following steps: randomly extract data with a dialogue sample number of n from multiple task-based multi-turn dialogue data sets {D1, D2,..., D3} to form a new data set D new And manually perform sentence-level annotation based on the formulated general intent.

[0068] A specific implementation process of the annotation is to use three annotators to respectively perform full-scale annotation on D new When annotating, the general intent of each sentence of both the user and the system is annotated in the way of self-dividing sentences and taking the sentence level as the unit. Finally, the Kappa coefficient is used to measure the consistency of the three annotation results.

[0069] On the new data set obtained by manual annotation, different large language models are used for sentence-level intent classification based on the same general intent system as a verification means, and the classification performance of the large language model on this intent system is given according to the classification results. Weigh and select the large model with the most appropriate performance and cost for the subsequent steps.

[0070] Step S2, obtain the current user input and the corresponding human-machine conversation history;

[0071] In the embodiments of the present invention, a dialogue window on the web page is developed, which supports operations such as users conversing with the system and viewing the conversation history. The method obtains the current input information of the user from the web page backend and obtains the human-machine conversation history from the data structure in the backend.

[0072] Step S3, input the current user input and the corresponding human-machine conversation history into the general intention recognition model, and the model outputs the corresponding general intention list;

[0073] In the embodiments of the present invention, the previously labeled dataset D new is further divided into a training set and a test set. The training set is assembled into prompt words according to specific templates and then used for model fine-tuning of the large language model for general intention recognition.

[0074] The structural composition of the prompt words assembled by the template is as follows: "{Role definition statement for setting the large language model as a dialogue intention classifier}\n{Specific classification task description}\n{Precautions}\n{General intention system}\n{Example dialogue and intention classification results}\n{Input historical dialogue}\n{User dialogue to be classified}\n{Prefix sentence to guide the model for classification}\n".

[0075] Among them, the {Specific classification task description} guides the model to classify the user's utterance by selecting an intention from the given general intentions through the method of chain of thought. The approach is to first divide the user's utterance into multiple sentences, then consider the general intention of each sentence separately, and when selecting a category, classify layer by layer according to the hierarchical organization form of the general intention.

[0076] The {General intention system} includes the names and meanings of all the above intentions. In addition, several examples of each intention can be added after the meaning, or there can be no examples.

[0077] The {Example dialogue and intention classification results} can include a single example or multiple examples, or there can be no examples. The example content includes the complete dialogue between the user side and single-round or multi-round, the prefix sentence to guide the model for classification, and the classification result.

[0078] In some specific embodiments, according to the structural composition of the prompt words assembled by the above template, the specific content of the template is determined as follows (already translated into Chinese):

[0079] "You are an expert in analyzing dialogue behaviors. You will be provided with a list of

General Intentions

Multi-turn Dialogue History

Multi-turn Dialogue History

General Intentions

Examples

[0080]

Precautions

[0081] - Some intentions are history-based, such as those prefixed with "reply_" or "modify_". Please analyze in combination with the dialogue history.

[0082] - You can select more than one intention.

[0083] - When you are unable to classify, please select the "other" intention.

[0084] - Do not create or rename intentions on your own.

[0085]

General Intentions

[0086] List of general intentions, including descriptions and intention examples

[0087]

Example Dialogues and Intention Classification Results

[0088] Example 1: {Example Dialogue 1}

[0089] The speaker's general intention is: {Classification Result}

[0090] Example 2: {Example Dialogue 2}

[0091] The speaker's general intention is: {Classification Result}

[0092]

Multi-turn Dialogue History

[0093] Historical dialogue

[0094]

User Dialogue to be Classified

[0095] User dialogue to be classified

[0096] The speaker's general intention is: ".

[0097] In the embodiments of the present invention, the instruction format for fine-tuning the large model for general intention recognition is unified with the instruction format for actual intention recognition. The output intention results are dynamically maintained in the program for use in subsequent rounds of dialogue."

[0098] Step S4: Input the current user input and the corresponding human-machine dialogue history into the task dialogue large model with enhanced general intent, extract the intent and task elements of the dialogue, and generate corresponding responses.

[0099] In the embodiments of the present invention, the original task dialogue large model without intent enhancement adopts two methods that achieve the best performance on the MultiWOZ dataset, both of which use large language models to implement dialogue policy generation.

[0100] Prepare the fine-tuning data for enhanced general intent. The preparation process of the fine-tuning data for enhanced general intent is as follows: if the original task dialogue large model has been fine-tuned, first obtain the data used for its fine-tuning. If the original large model has not been fine-tuned, use the same dataset as that used for training the general intent recognition large model. If it is the dataset on which it was originally fine-tuned, it is necessary to first perform general intent annotation on it. The annotation is performed using the general intent recognition large model for model-based automatic annotation.

[0101] Construct the above data into a specific prompt template form. This form includes adding an explanation of the general intent to the task description section of the instruction template it uses. An example is as follows: "For each turn in the dialogue, you will be provided with the general intent of the user's utterance in that turn. These general intents represent the speaker's higher-level communication purposes and can be used to better understand the speaker's needs. Carefully analyze the dialogue and the intent, extract the task elements and give corresponding responses."

[0102] Add a description of the general intent of each turn of the user dialogue to the human-machine dialogue history section of the instruction template used. An example is as follows: "<Think> Let's think step by step. The general intent of the user is {example general intent}, so the user may be indicating that his {slot} is {slot value} (corresponding to the informing intent) / requesting us to {task} (corresponding to the requesting intent) / answering our previous question about {slot}, and his answer is {slot value} (corresponding to the responding intent)", etc., as Figure 3 shown;

[0103] Add a description of the general intent of the current turn of the user dialogue to the current dialogue section of the instruction template used; perform model fine-tuning.

[0104] Step S5: Repeat steps S2 to S4 until the dialogue task is completed and then end the dialogue.

[0105] In the description of this specification, the descriptions referring to the terms "one embodiment", "example", "specific example", etc. mean that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. However, such modifications or replacements 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 the present invention.

Claims

1. An enhancement method for a task-based multi-turn dialogue system based on general intents, characterized in that It includes the following steps: Step S1, load the general intent recognition model for task-based multi-turn conversations; Step S2, obtain the current user input and the corresponding human-machine conversation history; Step S3, input the current user input and the corresponding human-machine conversation history into the general intent recognition model, and the model outputs the corresponding list of general intents; Step S4, input the current user input and the corresponding human-machine conversation history into the task dialogue large model with enhanced general intent, extract the intent and task elements of the conversation, and generate corresponding responses; Step S5, repeat steps S2 to S4 until the conversation task is completed and the conversation ends.

2. The method for enhancing a task-based multi-turn dialogue system based on general intention according to claim 1, wherein In step S1, the training process of the general intent recognition model for task-based multi-turn conversations includes: Randomly extract a part of the data from several task-based multi-turn conversation datasets and perform sentence-level annotation based on the proposed general intent; Construct the annotation results into a dataset according to a prompt template structure; Based on this dataset, perform model fine-tuning to strengthen the adaptation to the language environment under the task-based multi-turn conversation system.

3. An enhancement method for a task-based multi-turn dialogue system based on general intention according to claim 2, characterized in that, The prompt word structure for identifying the general intent of user statements consists of the following parts: {Set the large language model as the role definition statement of the dialogue intent classifier}\n{Specific classification task description}\n{Precautions}\n{General intent system}\n{Example dialogue and intent classification results}\n{Input historical dialogue}\n{User dialogue to be classified}\n{Prefix sentence to guide the model for classification}\n.

4. An enhancement method for a task-based multi-turn dialogue system based on general intention according to claim 1, characterized in that, In step S2, the general intents in the list are divided into the following categories: inquiry intent, reply intent, modification intent, request intent, notification intent, greeting intent, complaint intent, gratitude intent, apology intent, farewell intent, and other intents.

5. An enhancement method for a task-based multi-turn dialogue system based on general intent according to claim 4, characterized in that, The reply intent is divided into 9 secondary intents, namely reply notification, reply affirmation, reply negation, reply ignorance, reply awareness, reply acceptance, reply rejection, reply completion, and reply incompletion.

6. An enhancement method for a task-based multi-turn dialogue system based on general intention according to claim 2, characterized in that, In step S4, the training process of the task dialogue large model with enhanced general intent includes: Obtain the original task dialogue large model; this large model is fine-tuned with task-based dialogue datasets or only pre-trained without fine-tuning; Prepare the fine-tuning data for enhanced general intent; if the original task dialogue large model has been fine-tuned, first obtain the data used for its fine-tuning; if the original large model has not been fine-tuned, use the same dataset as in claim 2; if it is the dataset it was originally fine-tuned with, it is necessary to first perform general intent annotation on it; the annotation is performed using the general intent recognition model fine-tuned in claim 2 for model-based automatic annotation.

7. The method for enhancing a task-based multi-turn dialogue system based on general intent according to claim 6, wherein Construct the fine-tuning data into a specific prompt template form, which includes: adding a description of the general intent to the task description section of the instruction template it uses; adding a description of the general intent of each round of user dialogue to the human-machine conversation history section of the instruction template it uses; adding a description of the general intent of the current round of user dialogue to the current dialogue section of the instruction template it uses; and finally perform model fine-tuning.

8. A terminal device, characterized in that, The terminal device includes a memory, a processor, and a program of the method for enhancing a task-based multi-round dialogue system based on general intents, which is stored in the memory and can run on the processor. When the processor executes the program of the method for enhancing a task-based multi-round dialogue system based on general intents, the steps of the method for enhancing a task-based multi-round dialogue system based on general intents as described in any one of claims 1-7 are implemented.

9. A computer-readable storage medium, characterized in that, A program of the method for enhancing a task-based multi-round dialogue system based on general intents is stored on the computer-readable storage medium. When the program of the method for enhancing a task-based multi-round dialogue system based on general intents is executed by the processor, the steps of the method for enhancing a task-based multi-round dialogue system based on general intents as described in any one of claims 1-7 are implemented.

Images