A high-functioning autism screening method based on clinician-child dialogue

Abstract

The application relates to a high-function autism screening method based on a clinician-child dialogue, and relates to the technical field of natural language processing and medical auxiliary diagnosis. The application aims to solve the problem that existing screening relies on expensive equipment or manual features and is difficult to capture subtle abnormalities in children's language. The implementation steps are as follows: obtaining a doctor-child question and answer dialogue text; adopting a sliding window to combine multiple continuous question and answer pairs into an input sequence; using a pre-trained BERT model to extract word-level deep semantic representation; dynamically weighting the contribution of each word by an adaptive word aggregation module to obtain sequence-level aggregated features; performing paragraph-level classification; averaging the prediction results of all paragraphs of the same child to obtain individual screening results. The application realizes low-cost and high-efficiency early screening of HFA by analyzing the semantic, syntactic and pragmatic abnormalities in the dialogue language, and highlights fine-grained language clues with diagnostic value.

Description

Technical Field

[0001] This invention belongs to the field of computer natural language processing, specifically involving a method for detecting high-functioning autism based on clinical doctor-child dialogue language. That is, it achieves early screening of high-functioning autistic children by extracting subtle language differences between high-functioning autistic children and normal children. Background Technology

[0002] High-functioning autism (HFA) is a subtype of autism spectrum disorder (ASD). ASD is a neurodevelopmental disorder characterized by deficits in social communication and interaction. Unlike ASD, individuals with HFA typically exhibit normal or near-normal intelligence and appear to have intact language skills. However, their language use often lacks metaphoricality, leans towards concreteness, and lacks pragmatic flexibility, hindering effective social communication. Therefore, individuals with HFA frequently face persistent social communication impairments, higher-order cognitive deficits, and exacerbated social frustration from adolescence to adulthood due to the mismatch between external expectations and sociopragmatic abilities. The long-term needs related to social functioning and behavioral regulation further increase the burden on families and society, thus necessitating early identification and structured intervention.

[0003] HFA (Hypersocial Autism Scale) often leads to missed or misdiagnosed cases due to its insidious manifestations, limiting the effectiveness of traditional assessment methods, and therefore screening remains challenging. Current clinical practice mainly relies on structured social cognitive and executive function assessments, using tools such as the Diagnostic Observation Scale for Autism (Second Edition), the Diagnostic Interview for Autism (Revised Edition), and the Social Responsiveness Scale. To improve diagnostic accuracy, Yaneva et al. proposed a machine learning recognition method using eye-tracking data in their article "Detecting high-functioning autism in adults using eyetracking and machine learning," IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 28, no. 6, pp. 1254–1261, 2020. Riccioni et al. and Deschrijver et al. explored functional magnetic resonance imaging (fMRI) and electroencephalography (EEG) in their articles "Highintellectual potential and high-functioning autism: Clinical and neurophysiological features in a pediatric sample," Brain Sciences, vol. 11, no. 12, p. 1607, 2021 and "Automatic detection of pragmatic deficits in children with autism," in Workshop on Child, Computer and Interaction, 2012, p. 1, respectively. Data-driven auxiliary diagnostic methods (EGD) can provide objective early indicators. However, these technologies require expensive equipment, specialized environments, and trained personnel, limiting their applicability for large-scale screening. Therefore, developing low-cost, scalable computational methods for HFA screening remains an urgent and important goal.

[0004] Given that atypicality associated with HFA is most prominent in higher language behaviors, language has recently become a compelling method for HFA screening. Meir et al., in their article “Syntactic abilities and verbal memory in monolingual and bilingual children with high functioning autism (hFA),” First Language, vol. 40, no. 4, pp. 341–366, 2020, demonstrated that HFA children can be identified by syntactic processing features in sentence repetition and verbal memory tasks. Naman et al., in their article “Narrative production abilities of children with autism,” Autism & Developmental Language Impairments, vol. 10, pp. 23969415251321824, 2025, also indicated that the HFA group has difficulties in constructing complex narrative structures and integrating syntax in production and retelling, making them distinguishable from typically developing (TD) peers. In terms of computation, Rubio-Mart's... In their article "Enhancing ASD detection accuracy: A combined approach of machine learning and deep learning models with natural language processing," *Health Information Science and Systems*, vol. 12, no. 1, p. 20, 2024, n et al. developed a language-based ASD classification model by extracting textual and semantic features from social media posts and applying machine learning techniques. However, these models were designed for the general ASD population and were not optimized for HFA, whose language anomalies are more subtle and often closer to typical development. Furthermore, existing methods typically rely on discrete, context-independent features such as word frequency and syntactic complexity, evaluated on small, single-scenario corpora, lacking a systematic pipeline consistent with real-world clinical conditions. Therefore, they may fail to capture the fine-grained language patterns specific to HFA.

[0005] To address these challenges, this invention proposes a computational framework based on doctor-child dialogue language, specifically designed for HFA recognition.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To overcome the shortcomings of the existing technologies, this invention provides a method for detecting high-functioning autism through clinical doctor-child dialogue, which aims to predict whether a child has high-functioning autism based on the diagnostic dialogue between the doctor and the child, thereby achieving early screening for high-functioning autism in children.

[0008] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0009] According to a first aspect of the present invention, a high-functioning autism screening method based on clinician-child dialogue is provided, the method comprising: Obtain the text of multi-turn question-and-answer dialogues between clinicians and children; A sliding window strategy is used to merge multiple consecutive question-answer pairs into a single input sequence to construct multiple input paragraphs; The input sequence is fed into a pre-trained language model to extract the deep contextual semantic representation of each word. The adaptive word aggregation module performs weighted aggregation on the deep contextual semantic representation of the word to obtain a sequence-level aggregated feature representation. The adaptive word aggregation module dynamically allocates the weights of different word elements through a learnable attention mechanism. Based on the sequence-level aggregated feature representation, each input paragraph is classified and predicted to obtain paragraph-level prediction results; The prediction results at multiple paragraph levels for the same child are aggregated to generate individual-level screening results, which are used to indicate whether the child has high-functioning autism.

[0010] In some exemplary embodiments, the use of a sliding window strategy to merge multiple consecutive question-answer pairs into a single input sequence further includes: Continuous The questions in each question-answer pair are concatenated into a question sequence, and the corresponding answers are concatenated into an answer sequence. The question sequence and answer sequence are constructed into the input sequence according to a preset format; When you scroll to the end of the text, the number of remaining question-answer pairs... Less than window length When using an adaptive window adjustment strategy, the settings should be adjusted accordingly. .

[0011] In some exemplary embodiments, the pre-trained language model is a transformer-based bidirectional encoder representation model BERT; the extraction of deep contextual semantic representations for each word is achieved by stacking multiple transformer encoders, each encoder including a multi-head attention mechanism and a feedforward network, used to capture semantic alignment relationships and syntactic patterns in question-and-answer interactions.

[0012] In some exemplary embodiments, the adaptive lexical aggregation module is a two-layer feedforward attention network, which implements weighted aggregation through the following steps: The hidden state of each word in the input sequence The signal is sequentially passed through a first linear transformation and a nonlinear activation function, and then projected onto the intermediate attention space. The representation of the intermediate attention space is converted into a scalar score through a second linear transformation. ; Softmax normalization is applied to the scalar scores of all words to obtain the attention weight for each word. ; Based on the attention weight Hidden states of all lexical units By performing a weighted summation, the sequence-level aggregated feature representation is obtained. .

[0013] In some exemplary embodiments, the step of classifying and predicting each input paragraph to obtain paragraph-level prediction results further includes: Representing the sequence-level aggregation features Input a prediction head, which consists of a linear layer and a classification layer; The linear layer represents the aggregated features. Mapped to a low-dimensional latent space; The classification layer maps the representation of the low-dimensional latent space to category probabilities to determine whether the language behavior pattern corresponding to the paragraph is closer to high-functioning autism spectrum disorder (HFA) or typical developmental dysplasia (TD).

[0014] In some exemplary embodiments, generating individual-level screening results further includes: All corresponding to a child Paragraph level prediction probability of each paragraph The child's overall predicted score is obtained by averaging the results. ; The overall prediction score is compared with a preset decision threshold. The results are compared, and if the result is higher than the threshold, it is considered a positive high-functioning autism diagnosis; otherwise, it is considered a negative diagnosis.

[0015] According to a second aspect of the present invention, a high-functioning autism screening system based on clinician-child dialogue is provided, characterized in that it comprises: The data acquisition module is used to acquire the text of multi-turn question-and-answer dialogues between clinicians and children; the text processing module is used to merge multiple consecutive question-and-answer pairs into an input sequence using a sliding window strategy to construct multiple input paragraphs. The feature extraction module is used to input the input sequence into a pre-trained language model and extract the deep contextual semantic representation of each word. An adaptive aggregation module is used to dynamically allocate weights for different lexical units through a learnable attention mechanism, and to perform weighted aggregation on the deep contextual semantic representation of the lexical units to obtain a sequence-level aggregated feature representation. The paragraph classification module is used to classify and predict each input paragraph based on the sequence-level aggregated feature representation to obtain paragraph-level prediction results. The individual classification module is used to aggregate prediction results at multiple paragraph levels for the same child to generate individual-level screening results, which are used to indicate whether the child has high-functioning autism.

[0016] According to a third aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the high-functioning autism screening method based on clinician-child dialogue described in the first aspect above.

[0017] According to a fourth aspect of the present invention, a computer program product is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, it implements the high-functioning autism screening method based on clinician-child dialogue described in the first aspect above.

[0018] According to a fifth aspect of the present invention, an electronic device is provided, comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to implement the high-functioning autism screening method based on clinician-child dialogue as described in the first aspect above by executing the executable instructions.

[0019] The high-functioning autism screening method based on clinician-child dialogue provided by the embodiments of the present invention has the following advantages compared with the prior art: 1) Unlike traditional high-functioning autism screening methods based on scales and manual feature learning, this invention can automatically identify high-functioning autism simply by conducting a simple verbal dialogue, which is low-cost and highly efficient.

[0020] 2) The language-based HFA computational classification model proposed in this invention uses pre-trained BERT to learn high-level context-aware language representations from conversation sequences, and uses an adaptive tagging lexical mechanism to dynamically weight feature contributions, amplifying subtle discrimination cues and making it more helpful in capturing language behavior deviations between HFA and TD children.

[0021] 3) The attention-based design of this invention can quantify the contribution of words and activity levels, support personalized screening and intervention, and lay the foundation for language- and multimodal HFA recognition methods.

[0022] This invention not only has important applications in the field of high-functioning autism identification, but can also be extended to other language-based disease screening and identity recognition tasks. The model proposed in this invention also helps to analyze the language behavior of children with high-functioning autism and provides ideas for exploring subtle language differences in children with the disease.

[0023] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0025] Figure 1 This is a flowchart illustrating the overall implementation method of the present invention; Figure 2 This is a schematic diagram of the doctor-child question-and-answer text flow embedded in the sliding window in this invention; Figure 3 A framework diagram for extracting text features from doctor-child question-and-answer pairs in this invention; Figure 4 This is a framework diagram of the adaptive word aggregation module in this invention; Figure 5 This is a framework diagram of the question-answering pair level prediction module in this invention; Figure 6 This is a flowchart of the individual-level prediction process in this invention. Detailed Implementation

[0026] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0027] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0028] To address the shortcomings and deficiencies of existing technologies, this example implementation provides a high-functioning autism screening method based on a doctor-child dialogue. For a diagnostic dialogue between a doctor and a child, a text encoder based on a pre-trained bidirectional encoder BERT is used to model the deep contextual semantic representation of the question-and-answer dialogue, thereby capturing fine-grained language patterns beyond handcrafted features. To further improve discriminative performance, an adaptive lexical aggregation module is introduced to dynamically weight lexical-level contributions, emphasizing behavior-related linguistic cues. Finally, based on the extracted language features reflecting the child's lexical grammar and Theory of Mind reasoning abilities, the method further classifies the input question-and-answer pairs and predicts whether the child has HFA. (Reference) Figure 1 As shown, the specific steps may include: Step 1: Embed the doctor-child question-and-answer text in the sliding window. A sliding window approach is used to combine the dialogue text between the child and the doctor, so that it can be subsequently input into a text encoder for text feature extraction. Specifically, the text will be... Consecutive question-answer pairs are merged into a single input sequence, enabling the model to simultaneously encode semantic delivery, turn-taking, and topic continuation relationships between multiple question-answer pairs. Within each window, consecutive questions are concatenated to form a new question sequence, and their corresponding answers are also concatenated to form a new question-answer sequence. An adaptive window adjustment strategy is used to set the window size to ensure that all question-answer pairs are included in model training.

[0029] Step 2: Extract features from the doctor-child dialogue text. This paper utilizes a BERT-pretrained text encoder to extract features from the embedded doctor-child question-and-answer text, capturing relevant semantic and grammatical abnormalities in children with language impairments. Under the proposed sliding window question-and-answer text embedding strategy, each input is embedded... As input to the text encoder, deep feature extraction is performed by stacking multiple encoder layers sequentially. Each encoder layer consists of a multi-head attention (MHSA) and a feed-forward network (FFN), with residual connections and layer normalization mechanisms added after each sub-layer to improve training stability, gradient propagation efficiency, and the model's expressive power.

[0030] Step 3, Adaptive Lexical Aggregation The text encoding module outputs a series of contextualized lexical-level representations, which contain deep semantic, syntactic, and cross-sentence dependency information from multiple Transformer layers. In standard Transformer-based classification model designs, the hidden states of the lexical units (CLS) in the last layer's output features are typically used directly as the aggregated representation of the entire input sequence for downstream classifiers. However, although the hidden states of the lexical units (CLS) can theoretically integrate information from the entire sequence, their representation often leans towards global semantic summarization, lacking fine-grained sensitivity to local key cues. In clinical language analysis tasks, this unidirectional, coarse-grained aggregation approach may obscure sparse but highly diagnostic features in the sequence, such as specific emotion words, language deviations, pragmatic violations, referential errors, or logical breaks.

[0031] Therefore, this invention proposes an adaptive lexical aggregation module. It explicitly models the relative importance of each lexical unit and adaptively integrates diagnostically relevant fine-grained language signals in a data-driven manner. Unlike traditional average pooling, max pooling, or fixed-structure [CLS] representations, By employing a learnable attention mechanism and dynamically assigning weights to different lexical units, the model can identify and amplify clinically valuable language segments while suppressing surface noise or task-irrelevant content, such as small talk, hesitation markers, or filler words. This mechanism significantly improves the model's ability to identify anomalous language patterns.

[0032] Step 4, Text prediction at the paragraph level After adaptive lexical aggregation module After processing, the model obtains compact and discriminative feature representations for each text pair of paragraphs. These features, which integrate deep semantic patterns, pragmatic dependencies, and local language anomalies from the question-and-answer sequence, serve as key inputs for subsequent classification tasks. To further map these high-dimensional features to a separable category space, a prediction head was subsequently designed. It is used to perform text-to-paragraph level classification inference, and realizes the prediction of the category of input text pairs.

[0033] Step 5, Individual-level prediction Building upon paragraph-level predictions based on question-and-answer pairs, it is also necessary to design individual-level predictions applicable to clinical diagnosis. This allows for a comprehensive assessment of the participant's overall language behavior throughout the entire interview session, leading to more stable and robust final classification results. Specifically, all paragraph-level prediction results (0 or 1) for the same child participant are summed and averaged to obtain the child's overall prediction score. A set threshold is then used to determine the child's final category.

[0034] The steps in this exemplary embodiment will now be described in more detail with reference to the accompanying drawings and embodiments.

[0035] Step 1: Embed the doctor-child question-and-answer text in the sliding window. Reference Figure 2 The doctor-child question-and-answer text is embedded using a sliding window. Consecutive question-answer pairs are merged into a single input sequence, enabling the model to simultaneously encode semantic transfer, turn-taking, and topic continuation relationships between multiple question-answer pairs. Within each window, consecutive questions are concatenated to form... The corresponding answer sequence is .in For sliding window index, This represents the window length. It also represents the number of question-answer pairs remaining when the window slides to the end of the sequence. Less than When using an adaptive window adjustment strategy, the settings should be adjusted accordingly. This is to ensure that all question-answer pairs are included in the model training.

[0036] The multi-question-answer pair input structure based on sliding windows can be uniformly constructed as follows: . Subsequently, the WordPiece tokenizer was used to tokenize the above sequence, mapping it to a word embedding sequence. : . To enable the model to distinguish the source of questions and answers, corresponding paragraph embeddings are added. and with position embedding The summation results in the combined embedding vector of the input Transformer: . Step 2: Extract text features from doctor-child question-and-answer pairs Reference Figure 3 For the first Encoder layer Record the output of the previous layer as hidden. The initial state is The MHSA module allows each lexical unit in the input sequence to simultaneously focus on the question segment. With the answer paragraph All lexical units are used to explicitly simulate the semantic alignment relationships in question-and-answer interactions.

[0037] For each attention head The input sequence is generated through linear projection. , , matrix:

[0038]

[0039]

[0040] in, , This indicates the number of attention heads. The multi-head mechanism allows the model to capture multi-granularity language relations in parallel across different subspaces, including word-level associations, syntactic patterns, and cross-sentence semantic connections.

[0041] Based on these projections, each attention head The scaled dot product attention is represented as:

[0042] Subsequently, the outputs of all attention heads are concatenated and linearly projected to form the final output of MHSA:

[0043] in, This is the learnable projection matrix. Further, residual connections are added and layer normalization is performed to obtain the intermediate representation:

[0044] FFN is applied independently to each word, and GELU activation is used to enhance the model's nonlinear representation capability of high-order language features as follows:

[0045] After another residual path and layer normalization, we obtain the first... Layer encoder output:

[0046] Finally, the output hidden state can be decomposed into:

[0047] in, This represents the global semantic aggregation representation of the entire sliding window question-and-answer pair. and This provides fine-grained word-level semantic encoding within the question-and-answer segment.

[0048] Step 3, Adaptive Lexical Aggregation Reference Figure 4 Adaptive word aggregation module The implementation is a two-layer feedforward attention network. The first layer projects the lexical representations into an intermediate attention space through a linear transformation combined with the Tanh activation function, enabling the model to capture non-linear feature interactions, including intra-word semantic enhancement, syntactic relevance enhancement, and cross-lexical discriminative comparison. The second layer further generates a scalar attention score for each lexical and normalizes it using Softmax to form adaptive attention weights, thereby obtaining a probability distribution at the lexical level.

[0049] Given an input sequence ,in For batch size, The total number of lexical units, To hide the state dimension, the aggregation process is defined as follows:

[0050]

[0051]

[0052] in, Indicates the first The hidden state of each word element It is its normalized attention score. These are the Softmax weights, and This represents the final sequence-level aggregate representation. Projection matrix and and bias terms , This creates a bottleneck structure that can suppress redundant dimensions and highlight semantically significant dimensions, thereby enhancing the compactness, stability, and discriminativeness of the output features.

[0053] Step 4, Question-Answer Pair Level Prediction Reference Figure 5 , First, apply linear projection to... Mapping to a low-dimensional latent space:

[0054] This linear layer reduces the representation dimensionality while retaining key discriminative factors, enabling the model to more effectively model fine-grained linguistic features related to HFA. Using the LeakyReLU activation function alleviates the death gradient problem of ReLU, allowing the model to maintain stable gradient updates when dealing with noisy dialogue corpora and sparse features. Furthermore, adding dropout after linear projection has a regularization effect, effectively suppressing overfitting, especially in scenarios with small question-answer pair datasets or clinical corpora with significant cross-individual variability.

[0055] Subsequently, the prediction head projects the latent representation to class logits through a fully connected layer, and calculates the class probability distribution using Softmax:

[0056] in, and The learnable parameters of the prediction head, and The model represents the first Each question-and-answer pair is assigned a posterior probability that a paragraph belongs to different categories (HFA or TD). The final predicted label is given by the category corresponding to the highest posterior probability.

[0057] Step 5, Individual-level prediction Reference Figure 6 For each subject Assuming it contains For each question-and-answer pair paragraph, aggregate the predicted probabilities of all question-and-answer pair paragraphs at the question-and-answer pair level:

[0058] in, Indicates the subject The The predicted probability vectors for each question-and-answer pair segment. Through this simple and effective averaging strategy, the model can integrate linguistic signals from multiple dialogue segments, thereby significantly reducing the impact of random noise or contextual bias in individual question-and-answer pairs on the final classification.

[0059] In obtaining Then, by applying decision thresholds This process generates final individual-level classification labels for the subjects. This process reflects the model's overall performance in real-world diagnostic situations, enabling the framework to perform both local language anomaly detection and generate overall diagnostic judgments that meet practical needs.

[0060] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0061] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0062] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is defined only by the appended claims.

Claims

1. A high-functioning autism screening method based on clinician-child dialogue, characterized in that, The method includes: Obtain the text of multi-turn question-and-answer dialogues between clinicians and children; A sliding window strategy is used to merge multiple consecutive question-answer pairs into a single input sequence to construct multiple input paragraphs; The input sequence is fed into a pre-trained language model to extract the deep contextual semantic representation of each word. The adaptive word aggregation module performs weighted aggregation on the deep contextual semantic representation of the word to obtain a sequence-level aggregated feature representation. The adaptive word aggregation module dynamically allocates the weights of different word elements through a learnable attention mechanism. Based on the sequence-level aggregated feature representation, each input paragraph is classified and predicted to obtain paragraph-level prediction results; The prediction results at multiple paragraph levels for the same child are aggregated to generate individual-level screening results, which are used to indicate whether the child has high-functioning autism.

2. The method according to claim 1, characterized in that, The method of employing a sliding window strategy to merge multiple consecutive question-answer pairs into a single input sequence further includes: Continuous The questions in each question-answer pair are concatenated into a question sequence, and the corresponding answers are concatenated into an answer sequence. The question sequence and answer sequence are constructed into the input sequence according to a preset format; When you scroll to the end of the text, the number of remaining question-answer pairs... Less than window length When using an adaptive window adjustment strategy, the settings should be adjusted accordingly. .

3. The method according to claim 1, characterized in that, The pre-trained language model is the BERT bidirectional encoder representation model based on transformers; the extraction of deep contextual semantic representations for each word is achieved by stacking multiple transformer encoders, each encoder including a multi-head attention mechanism and a feedforward network, used to capture semantic alignment relationships and syntactic patterns in question-and-answer interactions.

4. The method according to claim 1, characterized in that, The adaptive lexical aggregation module is a two-layer feedforward attention network, which implements weighted aggregation through the following steps: The hidden state of each word in the input sequence The signal is sequentially passed through a first linear transformation and a nonlinear activation function, and then projected onto the intermediate attention space. The representation of the intermediate attention space is converted into a scalar score through a second linear transformation. ; Softmax normalization is applied to the scalar scores of all words to obtain the attention weight for each word. ; Based on the attention weight Hidden states of all lexical units By performing a weighted summation, the sequence-level aggregated feature representation is obtained. .

5. The method according to claim 1, characterized in that, The step of classifying and predicting each input paragraph to obtain paragraph-level prediction results further includes: Representing the sequence-level aggregation features Input a prediction head, which consists of a linear layer and a classification layer; The linear layer represents the aggregated features. Mapped to a low-dimensional latent space; The classification layer maps the representation of the low-dimensional latent space to category probabilities to determine whether the language behavior pattern corresponding to the paragraph is closer to high-functioning autism spectrum disorder (HFA) or typical developmental dysplasia (TD).

6. The method according to claim 1, characterized in that, The generation of individual-level screening results further includes: All corresponding to a child Paragraph level prediction probability of each paragraph The child's overall predicted score is obtained by averaging the results. ; The overall prediction score is compared with a preset decision threshold. The results are compared, and if the result is higher than the threshold, it is considered a positive high-functioning autism diagnosis; otherwise, it is considered a negative diagnosis.

7. A high-functioning autism screening system based on clinician-child dialogue, characterized in that, include: The data acquisition module is used to acquire the text of multi-turn question-and-answer dialogues between clinicians and children; The text processing module is used to merge multiple consecutive question-answer pairs into a single input sequence using a sliding window strategy to construct multiple input paragraphs. The feature extraction module is used to input the input sequence into a pre-trained language model and extract the deep contextual semantic representation of each word. An adaptive aggregation module is used to dynamically allocate weights for different lexical units through a learnable attention mechanism, and to perform weighted aggregation on the deep contextual semantic representation of the lexical units to obtain a sequence-level aggregated feature representation. The paragraph classification module is used to classify and predict each input paragraph based on the sequence-level aggregated feature representation to obtain paragraph-level prediction results. The individual classification module is used to aggregate prediction results at multiple paragraph levels for the same child to generate individual-level screening results, which are used to indicate whether the child has high-functioning autism.

8. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the high-functioning autism screening method based on clinician-child dialogue as described in any one of claims 1 to 6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the high-functioning autism screening method based on clinician-child dialogue as described in any one of claims 1 to 6.

10. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the high-functioning autism screening method based on clinician-child dialogue as described in any one of claims 1 to 6 by executing the executable instructions.

Images