Method and system for training dialogue intervention large model for children with autism

Abstract

The invention discloses a method and a system for training a dialogue intervention large model for children with autism, and the system comprises a data processing module which is used for converting original dialogue data from theme dialogue intervention records of children with autism and doctors into a high-quality multi-round dialogue data set; the model training module is used for finely adjusting an existing large language model on the basis of the high-quality multi-round dialogue data set so as to learn dialogue styles and intervention strategies in the high-quality multi-round dialogue data set, and an intervention large model which can follow an ABA application behavior analysis principle to perform theme dialogues with autistic children is obtained; and the model evaluation module is used for comprehensively evaluating the intervention effect of the intervention large model from the sentence dimension and the dialogue dimension so as to feed back the model fine adjustment process. According to the method, the large model following the ABA application behavior analysis principle can be trained to perform dialogue intervention with the autism children, the intervention effect is effectively improved, and accessible, efficient and low-cost social communication rehabilitation support is provided for the autism children.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical artificial intelligence, and particularly relates to a method and system for training an autism child dialogue intervention large model. BACKGROUND

[0002] Autism Spectrum Disorder (ASD) is one of the most common heterogeneous neurodevelopmental disorders in children, characterized by social interaction disorders and repetitive or stereotyped behavior patterns. In order to alleviate the symptoms of social disorders, clinicians usually follow the principles of Applied Behavior Analysis (ABA) to conduct systematic social dialogue intervention on autism children to improve their language expression, emotion recognition and social adaptation ability. However, with the increasing prevalence of autism worldwide, although early diagnosis and continuous intervention have been proven to significantly improve core symptoms, due to the long intervention period and high cost of professional manpower, families and society are facing heavy economic burden. In addition, uneven distribution of medical resources further exacerbates the inequality of intervention services, and a large number of children fail to obtain timely and effective rehabilitation support, resulting in social ability loss, self-care difficulties and other problems. Therefore, exploring new intervention methods that are scalable, low-cost and efficient has become an important direction of current research.

[0003] In recent years, with the rapid development of artificial intelligence technology, especially the breakthrough of Large Language Models (LLMs) in natural language processing, researchers have begun to explore its application potential in autism intervention. Current researches mainly fall into two categories: one is to use the capabilities of large language models to enhance existing intervention processes, such as providing professional advice to parents or providing clues to help children communicate. However, this method only indirectly participates in the intervention process as an auxiliary tool and does not directly interact with the child, and cannot replace the core dialogue intervention link. Another class of research attempts to use large models as "virtual doctors" to simulate clinical doctors for dialogue intervention through prompt engineering. However, the existing methods still have significant gaps in dialogue style, strategy adaptability and personalized feedback compared to real clinical intervention, lack of deep integration and dynamic adjustment ability of ABA principles, and are difficult to achieve precise and coherent rehabilitation guidance.

[0004] In this context, the prior art attempts to build ASD-related models from different perspectives. For example, patent application CN115862839A obtains a training data set based on the metal composition of each detection object's excrement sample and each detection object's autism diagnosis result, thereby iteratively training a preset machine learning model using the training data set to obtain an autism risk assessment model. The advantage of this method is that it simplifies the screening process and reduces the dependence on professional doctors, but its focus is limited to risk assessment at the physiological indicator level and does not involve behavior intervention or social conversation training, which cannot be directly used for rehabilitation treatment. For another example, patent application CN117476183A realizes rehabilitation effect AI evaluation based on AdaBoost ensemble learning, and by integrating multi-dimensional data such as behavior, physiology, and psychology, an evaluation model that can predict rehabilitation progress is constructed. Although this system has certain advancement in data fusion and model optimization, its function is still limited to "evaluation" rather than "intervention", and it cannot realize real-time interaction and dialogue training with children, nor can it simulate the strategic and flexible approach of clinical doctors in ABA intervention.

[0005] In summary, the prior art is still in the initial exploration stage of applying large models to autism intervention, and has not yet achieved reliable simulation of the entire process of clinical doctor dialogue intervention. Therefore, there is an urgent need for a dialogue intervention large model training method that can deeply integrate ABA intervention principles, have multi-round dialogue capability, and achieve personalized strategy adjustment. SUMMARY

[0006] In view of the above, the purpose of the present application is to provide a method and system for training a dialogue intervention large model for autistic children, which is based on the collected intervention dialogue recordings between clinical autistic children and doctors to train a large model, so that it can follow the ABA application behavior intervention strategy for dialogue intervention, improve the intervention effect, and make up for the lack of direct intervention capability in the prior art, to provide accessible, efficient, and low-cost social communication rehabilitation support for autistic children.

[0007] To achieve the above-mentioned purpose of the application, the technical solutions provided by the present application are as follows: In a first aspect, the system for training a dialogue intervention large model for autistic children provided by an embodiment of the present application comprises: The data processing module is configured to convert the original dialogue data from the thematic dialogue intervention recordings between autistic children and doctors into a high-quality multi-round dialogue data set; The model training module is configured to fine-tune an existing large language model based on the high-quality multi-round dialogue data set to learn the dialogue style and intervention strategy in the high-quality multi-round dialogue data set, and obtain an intervention large model that can follow the ABA application behavior analysis principles and conduct thematic dialogue with autistic children; The model evaluation module is configured to comprehensively evaluate the intervention effect of the intervention large model from the sentence dimension and the dialogue dimension, so as to feed back the model fine-tuning process.

[0008] Preferably, the data processing module is further configured to filter and correct the original dialogue data to obtain a high-quality multi-turn dialogue data set available for training, including discarding the dialogue in which the autistic children cannot normally carry out semantic dialogue, only repeat the doctor's speech or emit onomatopoeic words, rely on visual aids and cannot be transformed into pure text form, and ensuring that all dialogues conform to the dialogue form in which the doctor starts and alternates with the child.

[0009] Preferably, the data processing module is further configured to identify the child non-response state in the original dialogue data, and supplement the child non-response state to the high-quality multi-turn dialogue data set as dialogue content.

[0010] Preferably, the data processing module is further configured to expand the multi-turn dialogue data set by data augmentation, including: based on the existing high-quality multi-turn dialogue in the high-quality multi-turn dialogue data set and writing expansion prompt words, automatically generating synthetic dialogue with rich dialogue topics and scenes by a large language model, and adding the synthetic dialogue to the high-quality multi-turn dialogue data set.

[0011] Preferably, the model training module uses system prompt words combined with ABA application behavior analysis principles to guide the output of the large language model, wherein the ABA application behavior analysis principles include the requirements of the discrete trial teaching method and the natural environment teaching method.

[0012] Preferably, the model training module adopts a single sentence prediction training method based on historical dialogue, that is, given the dialogue history, the intervention large model predicts the next sentence of the doctor's response, and is trained and optimized through the cross-entropy loss function between the predicted response and the real response.

[0013] Preferably, in the sentence dimension evaluation, the model evaluation module predicts the next sentence of the doctor's response based on the given dialogue context in the test set by the intervention large model, and automatically calculates the indicators with the real response to evaluate the performance of the intervention large model in the sentence dimension.

[0014] Preferably, in the dialogue dimension evaluation, the model evaluation module adopts a role-playing strategy, uses another large language model to play an autistic child, and makes multi-turn dialogue with the intervention large model, and finally scores the multi-turn dialogue content by an expert to evaluate the intervention effect of the intervention large model in the dialogue dimension.

[0015] Preferably, in the dialogue dimension evaluation, the large language model playing the autistic child generates responses according to predefined dialogue intentions, including no response, correct response, wrong response, and repetition, and randomly selects one of them as a reply intention to simulate the social barriers existing in autistic children each time a response is made.

[0016] In a second aspect, the embodiments of the present application also provide a method for training an autistic child dialogue intervention large model, which is implemented by using the above-mentioned system for training an autistic child dialogue intervention large model, and includes the following steps: The data processing module is used to convert the original dialogue data from the topic dialogue intervention recording of the autistic child and the doctor into a high-quality multi-round dialogue data set; The model training module is used to fine-tune an existing large language model based on the high-quality multi-round dialogue data set to learn the dialogue style and intervention strategy in the high-quality multi-round dialogue data set, so as to obtain an intervention large model that can follow the ABA applied behavior analysis principle to have a topic dialogue with the autistic child; The model evaluation module is used to comprehensively evaluate the intervention effect of the intervention large model from the sentence dimension and the dialogue dimension to feed back the model fine-tuning process.

[0017] Compared with the prior art, the present application has at least the following beneficial effects: (1) In the data processing stage of the present application, the dialogue data is automatically transcribed and processed by screening, correction and augmentation, a high-quality, topic-rich autistic intervention dialogue data set is constructed, and the child's non-response state is further considered to be included, which enhances the authenticity and integrity of the data, helps to improve the ability of the model to cope with typical situations such as silence and avoidance in real intervention scenarios, and avoids misinterpretation or interruption of interaction caused by data missing, so as to more effectively implement dialogue intervention.

[0018] (2) In the model training stage of the present application, the system prompt words fused with the ABA principle are used to guide the efficient fine-tuning of the model, so that the model can learn the dialogue style and intervention strategy in the high-quality multi-round dialogue data set, and ensure that the output language style is close to the clinical interventionist to more effectively achieve the intervention target.

[0019] (3) In the model evaluation stage of the present application, through the sentence and dialogue double-dimension evaluation, especially by randomly simulating the social barriers such as child non-response and repetition through role-playing, the automatic, close-to-real comprehensive evaluation of the intervention effect of the model is realized, forming a closed-loop feedback mechanism from training to evaluation, providing reliable basis for continuous iteration and optimization of the model, and guaranteeing the reliability and robustness of the finally deployed model. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort based on these drawings.

[0021] Figure 1 is a framework process schematic diagram of a system for training an autism child dialogue intervention large model provided by the embodiments of the present application; Figure 2 is a subject distribution schematic diagram of clinical data processed by a data processing module provided by the embodiments of the present application; Figure 3 is a system prompt word schematic diagram of a model training module provided by the embodiments of the present application; Figure 4 is a comparison diagram of scoring results of dialogue dimension model evaluation; Figure 5 is an example diagram of multi-round dialogue between a large model trained as a doctor and a large model playing an autism child in the present application; Figure 6 is a flowchart of a method for training an autism child dialogue intervention large model provided by the embodiments of the present application. DETAILED DESCRIPTION

[0022] In order to make the objects, technical solutions and advantages of the present application clearer, the following will further describe the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and do not limit the protection scope of the present application.

[0023] The inventive concept of the present application is that, in order to solve the problem of insufficient model intervention professionalism in the prior art, the embodiments of the present application provide a method and system for training an autism child dialogue intervention large model, a high-quality multi-round dialogue data set containing a "child non-response" state is constructed, a system prompt word fusing ABA principle is used to guide model fine-tuning, so that the model masters professional intervention strategies and dialogue style, and innovatively adopts a role-playing strategy to simulate child social communication disorders for multi-dimensional evaluation, forms a training-evaluation closed loop, and finally realizes the goal of professional and personalized dialogue intervention for autism children.

[0024] As shown in Figure 1 , the embodiments provide a system for training an autism child dialogue intervention large model, which comprises a data processing module, a model training module and a model evaluation module, and a usable autism child dialogue intervention large model is trained according to the system framework process to verify the effectiveness of the present application.

[0025] (1) Data processing module.

[0026] In embodiments, to ensure the authenticity and quality of the data, after obtaining ethical approval, six autism children treatment centers were cooperated, involving a total of 20 experienced clinicians. With the full informed consent of parents and children, audio recordings were collected during the thematic conversation intervention using a recording device. The audio sampling rate was 16000 Hz, and the WAV format was stored.

[0027] Then the original audio recording was transcribed into multi-turn conversation text by using a three-stage processing method: First, the automatic transcription stage: using automatic transcription tools for audio segmentation, speaker recognition, and text transcription. Specifically, Deep-FMSN was used for audio segmentation, and Cam++ was used to determine whether the segmented audio belonged to the child or the doctor. Paraformer was used for audio-to-text transcription, and finally CT-Transformer was used for punctuation prediction.

[0028] Second, the manual cleaning stage: autism children may not respond due to attention loss or misunderstanding of sentence meaning, in which case the intervention doctor will take appropriate measures such as repeating the question or simplifying the inquiry to maintain the conversation and encourage the child to speak. This is also the intervention strategy that the model of the present invention hopes to learn, however, during the audio-to-text transcription process, automatic transcription cannot identify this state. Therefore, the assistance of professional intervention doctors is needed to insert the [child non-response] label in the multi-turn conversation to supplement the child's non-response state. At the same time, errors in automatic transcription are corrected, such as inaccurate audio segmentation, speaker recognition errors, and misspelling. Due to the need for privacy protection, specific names or addresses involved in the conversation will be anonymized.

[0029] Third, the data augmentation stage: to enrich the theme and conversation scenarios of the data set, another large language model (such as GPT-4.1) was used to synthesize multi-turn conversations. Using the above processed real conversations as references, GPT-4.1 was used to imitate their style to generate new conversations of different themes and insert [child non-response] in appropriate places, thereby enriching the intervention conversation data set. In embodiments, 10 major themes were summarized from real clinical conversations and expanded to 27 sub-themes for data synthesis. For each real multi-turn conversation instance, it was used as a reference conversation to generate 27 new conversations of different sub-themes to enrich the data set. A total of about 100 hours of audio data was collected and transcribed into 751 thematic multi-turn conversations.

[0030] To further obtain high-quality data suitable for training large models for autism children intervention, the present invention designed the following data set construction and data selection rules: (i) If the autistic child cannot normally carry out semantic dialogue, can only repeat what the doctor says or make onomatopoeic words, the dialogue is abandoned; (ii) If the topic of the dialogue is a physical object such as a picture book and cannot be transformed into a dialogue that does not rely on visual aids, the dialogue is abandoned, because the large language model does not have multi-modal understanding ability, such dialogue will cause reference confusion, and the model cannot understand whether the child's response is correct or not; (iii) All dialogues should conform to the dialogue alternation form of the doctor and the child, and the dialogue should start with the doctor's part, so that the model can learn the intervention strategy of the current round according to the historical context.

[0031] Through the above strategies, 287 high-quality multi-round dialogues were obtained. Then, high-quality data synthesis was performed using GPT-4.1, and finally 8035 topic multi-round dialogues were obtained as the training data set for enhancing the general dialogue ability of the model. As shown in Figure 2 Fig. 2 is a schematic diagram of the topic distribution of dialogue data processed by the data processing module.

[0032] (2) Model training module.

[0033] In the embodiment, the model is trained using system prompt words following the ABA applied behavior analysis principle. ABA is a structured treatment method commonly used to treat autism, and its internal principle is the intervention strategy that the model wants to learn. The treating doctor will intervene in the autistic child in combination with Discrete Trial Teaching (DTT) and Natural Environment Teaching (NET) methods. DTT includes five basic elements: instruction, response, reinforcement, assistance, and pause. The pause refers to the interval time between each round of dialogue for the child to think and establish a mapping relationship. In the human-computer dialogue intervention process, the pause process naturally exists due to the requirements of model calculation, so the remaining requirements are organized and added to the system prompt words.

[0034] Specifically, as shown in Figure 3 The system prompt words following the ABA applied behavior analysis principle are divided into three parts: The first part is role setting, which includes the setting of the role and target of the large model.

[0035] The second part is the intervention principle, which gives the core elements of ABA application behavior analysis principles, including the points of instruction, response, assistance, and reinforcement. Among them, instruction refers to the part of the doctor's speech that guides the child to understand language and learn social skills, which needs to ensure that these instructions are concise and easy for children to understand. Reinforcement refers to the stimulation provided by the doctor when the child responds to the instruction, the purpose of which is to encourage the continuous occurrence of appropriate behavior, while inappropriate behavior is weakened or eliminated due to the lack of reinforcement. Reinforcement can be physiological, such as favorite food or toys, or social, such as praise. In social conversation intervention, social reinforcement is emphasized, and the child's socialization is reinforced through verbal praise and empathy. Assistance refers to the support and help provided by the doctor when the autistic child encounters response difficulties, which can take the form of physical action, visual guidance, or language, and must be timely and appropriate to prevent frustration caused by failure and dependence on assistance. In the theme conversation intervention, assistance usually takes the form of language, such as restating the question, breaking it down, or hinting at the answer. Since the constructed dataset contains the [child non-response] state, the model can learn from multiple rounds of conversation how the doctor responds to the non-response state, which is crucial for the model to successfully guide the child in social conversation. At the same time, how to handle the child's response also needs to follow the ABA application behavior intervention principles. Response refers to the child's response to the doctor's instruction, and according to the child's response, the doctor will implement different intervention strategies (instruction, assistance, reinforcement) to guide the child to respond correctly. The response strategy for different response types is also given in the system prompt words.

[0036] The third part is the start of the conversation, giving the theme {topic_content} for the intervention large model to start the conversation. Then the existing Chinese open source large language model such as Qwen2.5-7b-Instruct, Llama3-Chinese-8b-Instruct, Yi-1.5-9B-Chat, InternLM3-8b-Instruct, Baichuan2-7b-Chat, GLM4-9B-Chat is fine-tuned with LoRA method, by giving the conversation history to the intervention large model, let it predict the next round of conversation should output sentence, and through the cross entropy loss function to optimize the sentence dimension output, to learn the doctor's conversation style and intervention strategy in the dataset. In the embodiment, the ms-swift framework is used to train on 8 cards GTX 4090 GPUs, for hyperparameters, set epoch to 5, seed to 42, learning rate to 1e-4, LoRA rank to 8, and LoRA alpha to 32.

[0037] (3) Model evaluation module.

[0038] In the embodiments, the performance of the intervention large model in the topic dialogue intervention task for autistic children is comprehensively measured from two aspects of sentence dimension and dialogue dimension. 100 multi-turn dialogues are selected from the dataset as the test set, and the rest of the data is used as the training set. In order to ensure the richness of the test set, 10 dialogues are randomly selected from each of the 10 topics to form the test set.

[0039] On the sentence dimension, the dialogues between the doctor and the child in the test set from the first to the t-1th turn are selected, and then the intervention large model is used to predict the response of the doctor in the tth turn (predicted sentence) and compare it with the real response of the doctor (reference sentence). Common evaluation indicators are used to evaluate the difference between the predicted sentence and the reference sentence, including BLEU (BLEU-4), GLEU, ROUGE (R-1., R-2., R-L.), METEOR (MET.) and BERTScore (BS.) indicators, so as to measure whether the dialogue style and intervention strategy of the intervention large model on the sentence dimension are similar to those of the professional clinical interventionist. The evaluation results are shown in Table 1.

[0040] Table 1

[0041] The models ending with SFT in the table are the models obtained after training according to the present application. It can be observed that after training, the indicators of the model are significantly improved. All the fine-tuned models are superior to the existing models in most indicators, including GPT-4.1, Deepseek-V3, Gemini2.0-flash, etc. This indicates the effectiveness of the training framework proposed in the present application, and the dialogue style and intervention strategy of the model are very similar to those of the real clinical dialogue intervention. For Qwen2.5-7b-Instruct, compared with before training, the BLEU-4 indicator is improved by 5.96%, the GLEU is improved by 9.21%, the Rouge-1 is improved by 9.3%, the Rouge-2 is improved by 7.03%, the Rouge-L is improved by 11.07%, the METEOR is improved by 2.62%, and the BertScore is improved by 4.64%.

[0042] On the dialogue dimension, the performance of the intervention large model in the topic multi-turn dialogue intervention is mainly evaluated. After discussing with the autism intervention doctor, the scoring method is used for scoring. The evaluation is based on 10 dimensions in 3 aspects to score the intervention large model playing the role of the doctor in the multi-turn dialogue. The scoring criteria are shown in Table 2. Each dimension is evaluated according to the scoring criteria of 0 to 4. The higher the score, the better the quality of the doctor's output, and the closer to the level of the real autism child interventionist. Three experienced clinical intervention doctors specializing in autism are invited to score. The comparison of the scoring results is shown in Table 3. Figure 4 ​

[0043] Table 2

[0044] To simulate real intervention, the present application designs a role-playing strategy for GPT-4o to play the role of an autistic child and simulate a dialogue with the intervention large model. In order to simulate the social barrier state that may exist in autistic children, the child's response is divided into four states according to the ABA principle: no response, repetition, correct response and error response. A random intention is extracted from these intentions to generate the child's response. When the intention is no response, return [child no response] directly. When the intention is repetition, in order to simulate the fact that children will subconsciously repeat the last few words of the doctor in a real scenario, use the jieba package to tokenize the current doctor's instructions and exclude stop words, and use the last word as the child's response. For intentions representing correct response or error response, use GPT-4o combined with different system prompts to simulate the child's response. Under this strategy, the autistic intervention large model can cope with the chaotic expression of autistic children, while still effectively using ABA intervention strategies and maintaining topic dialogue. In addition, in order to simulate the child's short-term attention, the current doctor's output is spliced with the last round of doctor-child dialogue content as the dialogue history for GPT-4o to generate a response. By providing a topic and a dialogue turn, the dialogue between the doctor model and the autistic child model can be simulated, thereby testing the intervention large model's ability to intervene at the dialogue level.

[0045] In an embodiment, the role-playing strategy described above is used to generate multiple rounds of dialogue required for dialogue dimension evaluation. The topics in the test set are used, and the number of rounds is specified as 10 rounds to generate multiple rounds of dialogue. At the same time, the real multi-round dialogue of the test set is also scored by professional doctors as a scoring benchmark to evaluate its authority. The results show that the various capabilities of the model trained using the framework proposed by the present application are highly consistent with those of clinical intervention doctors. In terms of professionalism, all four indicators point to the ABA application analysis principle, and the trained model has significantly improved in all four indicators. This shows that after fine-tuning on the dataset, the model has effectively learned the intervention strategies of the doctor and can appropriately apply these strategies in different rounds. In terms of language, the dialogue style of the trained model is closest to that of the clinical intervention doctor, with an improvement of more than 1.6 points in language style. The improvement in these indicators is consistent with the fine-tuning goal of the present application: to make the model learn the clinical intervention dialogue strategy and improve its dialogue style, indicating the feasibility and effectiveness of the present application.

[0046] Finally, as Figure 5As shown, the dialogue style and intervention strategy are close to the professional doctor level. In the dialogue intervention with the theme of color, after training by the proposed autism child intervention large model training framework, the dialogue style and intervention strategy of the model are very close to the real doctor, the dialogue is simple and easy to understand, and the model can skillfully apply the ABA application behavior analysis principle to respond appropriately according to the different responses of the child, thereby improving the intervention effect.

[0047] Based on the same inventive concept, as Figure 6 As shown, the present embodiment also provides a method for training an autism child dialogue intervention large model, including the following steps: S1, using a data processing module to convert the original dialogue data from the theme dialogue intervention recording of the autism child and the doctor into a high-quality multi-round dialogue data set.

[0048] S2, using a model training module to fine-tune an existing large language model based on the high-quality multi-round dialogue data set to learn the dialogue style and intervention strategy in the high-quality multi-round dialogue data set, and obtain an intervention large model that can follow the ABA application behavior analysis principle to conduct theme dialogue with the autism child.

[0049] S3, using a model evaluation module to comprehensively evaluate the intervention effect of the intervention large model from the sentence dimension and the dialogue dimension, to feedback the model fine-tuning process.

[0050] It should be noted that the method for training an autism child dialogue intervention large model provided by the above embodiment belongs to the same inventive concept as the system for training an autism child dialogue intervention large model, and the specific implementation process is detailed in the system for training an autism child dialogue intervention large model. Embodiments, which will not be repeated here.

[0051] The specific embodiments described above have detailed the technical solutions and beneficial effects of the present application. It should be understood that the above description is only the most preferred embodiment of the present application and is not intended to limit the present application. Any modifications, supplements, and equivalent replacements made within the principle range of the present application shall be included in the protection scope of the present application.

Claims

1. A system for training a large model for autism children dialogue intervention, characterized in that, The application comprises the following steps: A data processing module, a model training module, and a model evaluation module; The data processing module is used to convert the original dialogue data from the thematic dialogue intervention recording between autistic children and doctors into a high-quality multi-turn dialogue dataset; The model training module is used to fine-tune an existing large language model based on the high-quality multi-turn dialogue dataset to learn the dialogue style and intervention strategies in the high-quality multi-turn dialogue dataset, thereby obtaining an intervention large model that can follow the ABA application behavior analysis principles and conduct thematic dialogue with autistic children; The model evaluation module is used to comprehensively evaluate the intervention effect of the intervention large model from the sentence dimension and the dialogue dimension to feedback the model fine-tuning process.

2. The system for training autism children dialogue intervention large model according to claim 1, wherein, The data processing module is also used to filter and correct the original dialogue data to obtain a high-quality multi-turn dialogue dataset for training, including discarding the dialogue that the autistic children cannot normally conduct semantic dialogue, only repeat the doctor's speech, or emit onomatopoeic words, relying on visual aids and cannot be transformed into pure text form, and ensuring that all dialogues conform to the form of the doctor starting and alternating with the child.

3. The system for training an autism child dialogue intervention large model according to claim 1, wherein, The data processing module is also used to identify the child's non-response state in the original dialogue data and supplement the child's non-response state as the dialogue content to the high-quality multi-turn dialogue dataset.

4. The system for training an autism child dialogue intervention large model according to claim 1, wherein, The data processing module is also used to expand the multi-turn dialogue dataset through data augmentation, including generating synthetic dialogue with rich dialogue topics and scenes through a large language model based on the existing high-quality multi-turn dialogue in the high-quality multi-turn dialogue dataset and writing extension prompt words, and adding the synthetic dialogue to the high-quality multi-turn dialogue dataset.

5. The system for training an autism child dialogue intervention large model according to claim 1, wherein, The model training module uses system prompt words combined with ABA application behavior analysis principles to guide the output of the large language model, wherein the ABA application behavior analysis principles include the requirements of the discrete trial teaching method and the natural environment teaching method.

6. The system for training an autism child dialogue intervention large model according to claim 1, wherein, The model training module adopts a single sentence prediction training method based on historical dialogue, that is, given the dialogue history, the intervention large model predicts the next sentence of the doctor's response, and the training optimization is performed through the cross-entropy loss function between the predicted response and the real response.

7. The system for training an autism children dialogue intervention large model according to claim 1, wherein, In the sentence dimension evaluation of the model evaluation module, the intervention large model is used to predict the next sentence of the doctor's response based on the given dialogue context in the test set, and the real response is calculated through automatic indicators to evaluate the performance of the intervention large model in the sentence dimension.

8. The system for training an autism child dialogue intervention large model according to claim 1, wherein, In the dialogue dimension evaluation of the model evaluation module, a role-playing strategy is adopted, another large language model is used to play the autistic child, and the large language model is used to conduct multi-turn dialogue with the intervention large model, and finally the multi-turn dialogue content is scored by experts to evaluate the intervention effect of the intervention large model in the dialogue dimension.

9. The system for training autism children dialogue intervention large model according to claim 8, characterized in that, In the dialogue dimension evaluation, the large language model playing the autistic child generates responses according to predefined dialogue intentions, including non-response, correct response, error response, and repetition, and randomly selects one of them as a reply intention to simulate the social barriers existing in autistic children.

10. A method for training a large model for autism children dialogue intervention, implemented by the system for training a large model for autism children dialogue intervention according to any one of claims 1-9. The application comprises the following steps: The data processing module is used to convert the original dialogue data from the thematic dialogue intervention recording between autistic children and doctors into a high-quality multi-turn dialogue dataset; The model training module is used for fine-tuning an existing large language model based on a high-quality multi-turn dialogue data set to learn dialogue styles and intervention strategies in the high-quality multi-turn dialogue data set, so as to obtain an intervention large model capable of conducting theme dialogue with autistic children in compliance with ABA application behavior analysis principles; The model evaluation module is used for comprehensively evaluating intervention effects of the intervention large model from a sentence dimension and a dialogue dimension, so as to feed back a model fine-tuning process.

Images