A language model-based personalized cognitive training method and system, and a storage medium

Abstract

The application relates to the technical field of cognitive health and artificial intelligence, and discloses a personalized cognitive training method and system based on a language model and a storage medium. The method comprises the following steps: obtaining user language interaction data; inputting the language interaction data into a pre-trained language model to determine language features associated with the cognitive ability of the user; determining a comprehensive cognitive evaluation result according to the language features; generating training content for the user to interact according to the evaluation result; obtaining response data of the user to the training content, taking the response data as new language interaction data to re-determine the language features and update the comprehensive cognitive evaluation result. The pre-trained language model is used to analyze the language features associated with the cognitive ability in the user language interaction data, so that the evaluation result can comprehensively reflect the cognitive state of the user from multiple language dimensions. Meanwhile, the training content is dynamically adapted according to the evaluation result, and the evaluation result is continuously updated through the response data backflow, so that the cognitive evaluation has the ability of continuous dynamic updating.

Description

Technical Field

[0001] This application relates to the fields of cognitive health and artificial intelligence technology, and in particular to a personalized cognitive training method, system and storage medium based on a language model. Background Technology

[0002] Cognitive ability assessment and training are important components of cognitive health management, involving the monitoring and intervention of multiple cognitive dimensions such as memory, language, attention, and executive function. With the deepening of population aging and the increasing awareness of health management, how to continuously and objectively monitor cognitive status and provide effective cognitive interventions accordingly has become a key technical issue of concern in related fields.

[0003] In current technologies, cognitive assessment typically relies on standardized neuropsychological scales, clinical interviews, or computerized cognitive tasks. These assessment methods generally require subjects to actively cooperate under the guidance of professionals at specific times and locations. Because the assessment process demands a high degree of cooperation from subjects, it is difficult to conduct frequently and over a long period in everyday life. Furthermore, the assessment results may be affected by factors such as the subject's emotional state, fatigue level, testing environment, and practice effects, thus their objectivity and continuity still need improvement.

[0004] In cognitive training, existing solutions mostly employ pre-set question banks, fixed training tasks, or gamified exercises. Training content and difficulty are typically pre-set manually or adjusted according to fixed rules, making it difficult to adaptively update in response to changes in the user's cognitive state. Furthermore, in existing technologies, the evaluation and training processes are often independent, making it difficult to provide timely feedback on evaluation results to the training phase, and hindering the continuous quantification and verification of training effectiveness. This results in a lack of dynamic adjustment capabilities in intervention programs.

[0005] Therefore, how to achieve continuous monitoring and personalized intervention of cognitive state without significantly increasing the user's operational burden and professional personnel's reliance has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the problems in existing technologies, such as the difficulty in conducting cognitive assessments continuously in daily scenarios, insufficient objectivity of assessments, difficulty in adaptively adjusting training content according to user status, and the disconnect between assessment and training, this application provides a personalized cognitive training method, system, and storage medium based on a language model.

[0007] According to a first aspect of the embodiments of this application, this application provides a personalized cognitive training method based on a language model, which adopts the following technical solution: A personalized cognitive training method based on a language model includes the following steps: Obtain user language interaction data; The language interaction data is input into a pre-trained language model, and the language features associated with the user's cognitive ability in the language interaction data are determined by the pre-trained language model; wherein, the language features include language fluency feature values, vocabulary extraction difficulty feature values, and semantic coherence feature values; Based on the language characteristics, determine the user's comprehensive cognitive assessment result; Based on the comprehensive cognitive assessment results, training content is generated for the user to interact with; The system acquires the user's response data to the training content, uses the response data as new language interaction data, redetermines the language features, and updates the comprehensive cognitive assessment result.

[0008] Optionally, acquiring user language interaction data includes: Collect voice data and / or instant messaging text data generated by users in non-task-oriented natural scenarios; The speech data is subjected to speech recognition to obtain speech recognition text data and speech timestamp information; The instant messaging text data and the speech recognition text data are subjected to de-identification processing to obtain de-identified text data; wherein, The de-identified text data includes the de-identified instant messaging text data and the de-identified speech recognition text data; The instant messaging text data after de-identification is segmented based on sender switching and message time interval, pause events are determined based on message time interval, writing features are quantified and statistically analyzed, and emoticons, internet slang and abbreviations are filtered and standardized and mapped. The speech recognition text data after de-identification is segmented based on the speech timestamp information and pause events exceeding a preset duration are detected to calculate pause density and / or filler word ratio.

[0009] Optionally, the language fluency feature value is determined in the following way: Based on the de-identified text data, the number of effective words and the effective dialogue duration are counted, and the original speech rate is calculated. For the speech recognition text data, the original speech rate is corrected by the pause density and the proportion of filler words to obtain the corrected speech rate; For the instant messaging text data, pause events are determined based on message time intervals and the instant messaging pause density is statistically analyzed. A writing feature correction factor is obtained based on the quantitative statistics of the writing features. The original speech rate is corrected using the instant messaging pause density and the writing feature correction factor to obtain the instant messaging corrected speech rate. The speech rate correction and instant messaging rate correction are normalized relative to the average reference speech rate to obtain the language fluency feature value. The vocabulary extraction difficulty feature value is determined in the following way: Detect vocabulary extraction difficulties in the de-identified text data, wherein the vocabulary extraction difficulties include at least one of named entity replacement, semantically related word shift, and extraction interruption; The frequency of events indicating difficulty in word extraction within the effective word count of a statistical unit; The occurrence frequency is normalized relative to a reference benchmark to obtain the word extraction difficulty feature value; The semantic coherence feature values ​​are determined in the following way: The de-identified text data is segmented by sentence. The pre-trained language model is used to encode each pair of adjacent sentences into a semantic vector; Calculate the cosine similarity of the semantic vectors of each pair of adjacent statements; The average similarity is obtained by averaging the cosine similarities of all adjacent sentence pairs. Detect unexpected topic switching events between adjacent statements and calculate topic switching penalties; The semantic coherence feature value is obtained by subtracting the topic switching penalty from the average similarity.

[0010] Optionally, determining the user's comprehensive cognitive assessment result based on the language features includes: The language features are normalized. The normalized language features of each dimension are weighted and summed according to the preset weight coefficients to obtain the comprehensive cognitive assessment result. For feature dimensions where higher values ​​indicate lower cognitive ability, their inverse values ​​are used in the weighted summation to unify the direction of each dimension so that higher scores indicate better cognitive ability. The preset weighting coefficients are adjusted according to the age, language background, or cognitive ability distribution of the target user group, and the comprehensive cognitive assessment results are mapped to discrete ability levels.

[0011] Optionally, generating training content for user interaction based on the comprehensive cognitive assessment results includes: Obtain user profile information, which includes at least one of the user's age group, geographic background, and points of interest; The training difficulty level is determined based on the comprehensive cognitive assessment results. Based on the comprehensive cognitive assessment results, the training difficulty level, and the user profile information, prompt words are constructed and generated. The generated prompt words are input into a pre-trained language model, which then generates training content corresponding to the training difficulty level in real time. The training content includes nostalgic-themed training materials, which are generated based on the user's historical background, age group, or regional background. The training difficulty level includes multiple progressively increasing levels, with lower levels corresponding to basic vocabulary and simple questions and answers, and higher levels corresponding to complex logical reasoning and abstract topic discussions.

[0012] Optionally, the step of acquiring the user's response data to the training content, using the response data as new language interaction data, redetermining the language features, and updating the comprehensive cognitive assessment result includes: The training content is embedded into the user's interaction process in the form of natural dialogue; Collect user response data to the training content; The response data is then used as new language interaction data to re-perform feature extraction and evaluation; Among them, the delay from obtaining response data to updating the comprehensive cognitive assessment results shall not exceed a preset delay threshold; The method further includes at least one of the following boundary condition processing methods: When the voice data contains multiple speakers, the user's voice segments are extracted using voiceprint recognition and speaker separation technology. When the recognition confidence level of the speech recognition is lower than the preset recognition threshold, the corresponding segment data is downweighted or discarded. When the effective dialogue duration or number of effective dialogues based on the de-identified text data is lower than a preset threshold, it is marked as insufficient data and the comprehensive cognitive assessment result is not updated for the time being. If a user has no valid conversation records for a preset number of consecutive days, the most recent comprehensive cognitive assessment result will be maintained and the user will be marked as silent.

[0013] Optional features include tiered privacy protection deployment: In the first deployment mode, speech recognition, de-identification processing and language feature extraction are completed locally on the user terminal, and only the de-identified language features are uploaded to the cloud. In the second deployment mode, after the user terminal completes speech recognition and de-identification processing, the de-identified text data is encrypted and transmitted to the cloud. The cloud-pre-trained language model performs language feature extraction, comprehensive cognitive evaluation, and training content generation. After the cloud processing is completed, the de-identified text data is deleted.

[0014] Optionally, the pre-trained language model is a large language model based on the Transformer architecture; The pre-trained language model receives the language interaction data through a structured prompt template. The structured prompt template instructs the pre-trained language model to output language feature scores and observation records, and includes chained reasoning instructions to instruct the pre-trained language model to first analyze the language feature performance sentence by sentence, and then output scores for each dimension.

[0015] According to a second aspect of the embodiments of this application, this application provides a personalized cognitive training system based on a language model, which adopts the following technical solution: A personalized cognitive training system based on a language model includes: The data acquisition module is used to acquire user language interaction data; The feature determination module is used to input the language interaction data into a pre-trained language model and determine the language features in the language interaction data that are related to the user's cognitive ability through the pre-trained language model. An assessment module is used to determine the user's comprehensive cognitive assessment result based on the language features. The training content generation module is used to generate training content for user interaction based on the comprehensive cognitive assessment results. The update module is used to acquire the user's response data to the training content, use the response data as new language interaction data, redetermine the language features, and update the comprehensive cognitive assessment result.

[0016] According to a third aspect of the embodiments of this application, this application provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the language model-based personalized cognitive training method described in any of the first aspects of the embodiments of this application.

[0017] In summary, this application provides a personalized cognitive training method, system, and storage medium based on a language model. The personalized cognitive training method provided in this application acquires user language interaction data and inputs it into a pre-trained language model. The pre-trained language model is then used to determine language features associated with the user's cognitive abilities, thereby achieving automated semantic analysis of the user's cognitive language abilities. Based on this, a comprehensive cognitive assessment result is generated according to the determined language features, enabling the cognitive assessment result to comprehensively reflect the user's cognitive state from multiple language dimensions. Furthermore, training content for user interaction is generated based on the comprehensive cognitive assessment result, ensuring that the difficulty and type of the training content match the user's current cognitive assessment result and adaptively adjusts as the user's cognitive state changes. Finally, by acquiring the user's response data to the training content, the response data is used as new language interaction data to redetermine language features and update the comprehensive cognitive assessment result, allowing the cognitive assessment result to be continuously updated as training interaction continues.

[0018] Other features and advantages disclosed in this application will be described in detail in the following detailed description section. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the present application and form part of the specification. They are used together with the following detailed description to explain the present application, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a personalized cognitive training method based on a language model, according to an exemplary embodiment.

[0020] Figure 2 This is a flowchart illustrating a method for acquiring user language interaction data according to an exemplary embodiment.

[0021] Figure 3 This is a flowchart illustrating a method for determining language fluency feature values ​​according to an exemplary embodiment.

[0022] Figure 4 This is a flowchart illustrating a method for determining the difficulty feature value of word extraction according to an exemplary embodiment.

[0023] Figure 5 This is a flowchart illustrating a method for determining semantic coherence feature values ​​according to an exemplary embodiment.

[0024] Figure 6 This is a flowchart illustrating a method for determining the results of a comprehensive cognitive assessment according to an exemplary embodiment.

[0025] Figure 7 This is a flowchart illustrating a method for generating training content according to an exemplary embodiment.

[0026] Figure 8 This is a flowchart illustrating a method for updating comprehensive cognitive assessment results according to an exemplary embodiment.

[0027] Figure 9 This is a structural diagram illustrating a personalized cognitive training system based on a language model, according to an exemplary embodiment. Detailed Implementation

[0028] The specific embodiments disclosed in this application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this application.

[0029] In this application, language interaction data refers to language materials generated by users during natural language use that reflect their language abilities. Unlike standardized tasks required by structured cognitive tests (such as describing pictures, reading designated text, answering pre-set questions, etc.), the language interaction data in this application originates from users' non-task-oriented natural language use scenarios, including but not limited to daily family conversations, telephone or video calls, instant messaging text chat, and conversations in community social activities. The characteristics of this type of data are: topics are freely initiated and changed by the participants, and the purpose of speaking is daily social interaction (greetings, chatting, narration, sharing), rather than completing a designated language task. Language interaction data can be represented as voice data, instant messaging text data, or a combination of both.

[0030] In this application, language features include fluency features, vocabulary extraction difficulty features, and semantic coherence features. Fluency features characterize a user's ability to produce coherent language per unit of time, reflecting language ability from a quantitative perspective; vocabulary extraction difficulty features characterize the frequency with which a user experiences difficulty extracting words or making naming errors in conversation, reflecting language ability from a qualitative perspective; and semantic coherence features characterize the semantic cohesion between sentences in a user's conversation, reflecting language ability from a structural perspective. These three dimensions complement each other, providing a more comprehensive reflection of a user's overall language state.

[0031] The three language feature dimensions mentioned above are related to users' cognitive abilities in the following ways: The fluency features, reflecting the rate and coherence of language production, are related to a user's attentional maintenance ability, information processing speed, and executive function. The more effective words a user can coherently produce per unit of time, and the fewer pauses and filler words, the stronger their ability to allocate attentional resources to maintain task orientation, suppress irrelevant interference, and quickly extract and assemble language units during language production. Conversely, decreased fluency may reflect reduced efficiency in attentional resource allocation or slower information processing speed. Therefore, fluency features reflect a user's cognitive state from the perspective of cognitive processing speed.

[0032] The frequency of vocabulary retrieval difficulties reflected by the vocabulary retrieval difficulty feature value is related to the user's semantic memory storage and retrieval capabilities. Phenomena such as named entity substitution, semantically related word shifting, and retrieval interruption reflect the difficulties users encounter when retrieving target words from their semantic memory. Semantic memory is an important component of long-term memory, and changes in its retrieval efficiency can reflect the state of memory function. Therefore, the vocabulary retrieval difficulty feature value reflects the user's cognitive state from the perspective of memory retrieval accuracy.

[0033] Semantic coherence features, reflecting the degree of semantic connection between sentences, are related to a user's planning and organizing abilities and working memory capacity within their executive functions. Maintaining topic consistency and reducing unexpected topic shifts in conversation requires working memory to retain and update contextual information between sentences, as well as to perform topic planning and monitoring. Decreased semantic coherence may reflect insufficient working memory capacity or a weakening of planning and organizing abilities within executive functions. Therefore, semantic coherence features reflect a user's cognitive state from the perspective of discourse organization ability.

[0034] The three dimensions mentioned above provide quantitative indicators from three cognitive function aspects: processing speed, memory retrieval, and discourse organization. These are also the cognitive function aspects that cognitive training can influence. By continuously extracting language features from these three dimensions in daily conversations, the system can identify which cognitive function aspect the user has a relative deficiency in, and use this identification result as the basis for generating training content, making the training content suitable for the user's current state in terms of difficulty and type.

[0035] The establishment of the above-mentioned relationship enables this application to reflect the user's cognitive state from multiple cognitive function aspects simply by extracting language features from daily conversations, without requiring the user to complete a structured cognitive test task. Based on this, training content adapted to the user's current cognitive state can be generated, realizing the integration of cognitive assessment and cognitive training in daily conversation scenarios.

[0036] It should be noted that the cognitive assessment involved in this application differs from disease diagnosis in clinical medicine. The linguistic features extracted from everyday conversation data in this application are quantitative indicators at the linguistic level, used to characterize the user's overall language state. The assessment results are used to adapt the difficulty of training content and do not constitute a diagnosis of any disease or a judgment of health status. The training content generated in this application is conversational information material not intended for treatment, presented in a natural dialogue format, and the user's response to the training content is part of everyday social conversation.

[0037] like Figure 1 As shown in the embodiments of this application, the personalized cognitive training method based on a language model includes the following steps: Step S101: Obtain user language interaction data.

[0038] In this step, language interaction data refers to language materials generated by users during natural language use that reflect their language proficiency. Unlike standardized tasks required by structured cognitive tests, the language interaction data in this application originates from users' non-task-oriented natural language use scenarios. Language interaction data can be represented as voice data, instant messaging text data, or a combination of both.

[0039] like Figure 2 As shown, step S101 specifically includes steps S201 to S205.

[0040] Step S201: Collect voice data and / or instant messaging text data generated by the user in a non-task-oriented natural scenario.

[0041] Voice data is collected through microphone arrays built into user terminal devices, including smart speakers, smartphones, tablets, and wearable devices. The microphone arrays utilize beamforming technology to enhance the voice signal from the direction of the target speaker and suppress ambient noise. Instant messaging text data is extracted from instant messaging applications, such as WeChat and SMS, with user authorization. The data collection process can be continuous or paused and resumed by the user via voice commands or the application itself. Non-task-oriented natural scenarios refer to real-life conversations in the user's daily life, including everyday family conversations, phone or video calls, instant messaging text chats, and conversations in community social activities, rather than task-induced language production aimed at assessment, such as clinical interviews, cognitive tests, structured interviews, reading tasks, or picture description tasks.

[0042] Step S202: Perform speech recognition on the speech data to obtain speech recognition text data and speech timestamp information.

[0043] In step S202, an Automatic Speech Recognition (ASR) engine is used to transcribe the speech signal into text, while retaining the timestamp information of the speech data during the transcription process. The timestamp information records the start and end times of each word or sentence in the speech stream, serving as the foundational data for subsequent calculations of time-related features such as pause density and speech rate. In this step, the ASR engine can be deployed locally on the user's terminal, thus avoiding the need to upload the raw speech data to the cloud. Speech recognition text data refers to the text data output by the ASR engine.

[0044] Step S203: Perform de-identification processing on the instant messaging text data and the speech recognition text data to obtain the de-identified text data.

[0045] De-identification refers to the process of replacing information in text data that can be traced back to a specific user's identity with preset placeholder labels. Specifically, it involves replacing names with...<PERSON_N> Tags, replace place names with<LOCATION_N> Tags that replace sensitive information such as phone numbers and ID card numbers with <sensitive>Tags. The purpose of de-identification is to remove personally identifiable information from text without affecting language feature analysis, allowing subsequent processing to be performed without revealing the user's identity. De-identified text data includes de-identified instant messaging text data and de-identified speech recognition text data.

[0046] Step S204: For the de-identified instant messaging text data, segment it based on sender switching and message time interval, determine pause events based on message time interval, quantify and statistically analyze writing features, and filter and standardize the mapping of emoticons, internet slang and abbreviations.

[0047] Instant messaging text data and voice data differ significantly in their data format. Voice data possesses continuity and rich temporal information, while instant messaging text data is presented in discrete message units, lacking paralinguistic information such as pauses and speech rate found in voice signals. Therefore, this step takes the following processing measures based on the unique attributes of instant messaging text data: Segmentation: Using sender switching as the primary segmentation signal, multiple consecutive messages sent by the same user are merged into a single conversation segment, while consecutive messages from different users are divided into independent segments. Simultaneously, the time interval between messages serves as an auxiliary segmentation boundary: when the time interval between two messages exceeds a preset multiple of the user's historical average response time, the two messages are divided into different conversation segments.

[0048] Pause Event Determination: Since there are no physical pauses in instant messaging text messages as in voice signals, this step uses the time interval between message sending as the basis for determining pause events. A valid pause event is counted when the time interval between two consecutive messages exceeds a preset multiple of the user's historical average response interval. Instant messaging pause density can be expressed as: ,in To determine the effective number of pause events, This represents the number of valid words.

[0049] Quantitative statistics of handwriting features: This involves quantifying handwriting features such as ellipses and recurring punctuation marks in the text. For example, consecutive ellipses may reflect hesitation or thought during expression, while repeated punctuation marks may reflect emotional intensity or emphasis. The handwriting feature correction factor can be expressed as: ,in The number of times the ellipsis appears. To indicate the number of times the punctuation mark appears repeatedly. This represents the total number of messages. The above writing characteristics can be used as an auxiliary indicator for calculating language fluency.

[0050] Non-standard expression filtering and normalization mapping: Instant messaging texts often contain non-standard language expressions such as emojis, internet slang, and abbreviations. Emojis are replaced with placeholders, are not counted in the valid word count, and do not trigger the detection of vocabulary extraction difficulty features. Internet slang and abbreviations are normalized and mapped using a dynamically updated dictionary. For example, internet slang is mapped to standard expressions to ensure the semantic accuracy of language feature extraction.

[0051] Step S205: For the de-identified speech recognition text data, segment it based on the speech timestamp information and detect pause events that exceed the preset duration in order to count the pause density and / or the proportion of filler words.

[0052] Speech recognition text data retains the timestamp information of the original speech data, which can be used for segmentation and pause event detection. Specifically, based on the time interval between adjacent words or sentences in the timestamp information, a pause event is determined when the time interval exceeds a preset duration (e.g., 500 milliseconds). The total number of pause events is then counted. With effective word count The ratio gives the pause density: At the same time, statistical filler words were analyzed. (Such as "um", "that", "this", "it is", etc.) account for the number of effective words. The proportion of filler words is obtained by calculating the ratio of filler words: Pause density and filler word ratio are important input parameters for subsequent calculation of language fluency features.

[0053] Step S102: Input the language interaction data into the pre-trained language model, and use the pre-trained language model to determine the language features in the language interaction data that are related to the user's cognitive ability.

[0054] A pre-trained language model refers to a machine learning model that has been pre-trained on a large-scale text corpus and possesses the ability to deeply understand the semantics of natural language. In this application, the pre-trained language model is a large language model based on the Transformer architecture, such as DeepSeek, Kimi, Minimax, GPT-4, and Claude. The pre-trained language model receives language interaction data through structured prompt templates, which instruct the pre-trained language model to output language feature scores and observation records.

[0055] In one implementation, the structured prompt template includes chained reasoning instructions to instruct the pre-trained language model to first analyze the language features sentence by sentence, and then output scores for each dimension to improve the interpretability and consistency of the scores.

[0056] Language features refer to quantitative indicators that reflect the language dimension of a user's cognitive ability. In this application, language features include language fluency feature values, vocabulary extraction difficulty feature values, and semantic coherence feature values.

[0057] Example 1: All three language feature values ​​are determined through a pre-trained language model.

[0058] In this embodiment, de-identified text data is input into a pre-trained language model. A structured prompt template instructs the model to analyze the text data from three dimensions: fluency, vocabulary retrieval difficulty, and semantic coherence, outputting scores and observation records for each dimension. Specifically, the structured prompt template can include two parts: system prompts and user prompts. System prompts define the model's role and task, such as, "You are a language ability analysis assistant. You will receive a daily conversation record. Please analyze the speaker's language performance from the following three dimensions. Do not provide diagnostic conclusions, only provide objective linguistic descriptions and numerical scores." User prompts provide the dialogue content and output format requirements, such as requiring the model to output analysis results in JSON format, including scores and observation records for the three fields: fluency, vocabulary retrieval, and coherence.

[0059] Example 2: Some language features were determined using statistical algorithms.

[0060] In this embodiment, the language fluency feature value is determined by a statistical algorithm, the vocabulary extraction difficulty feature value is determined by a statistical algorithm, and the semantic coherence feature value is determined by a pre-trained language model.

[0061] Specifically, such as Figure 3 As shown, the language fluency feature values ​​are determined in the following way: Step S301: Calculate the original speech rate by counting the number of effective words and the effective dialogue duration based on the de-identified text data.

[0062] Effective word count refers to the number of words in the de-identified text data after removing non-linguistic content such as emoji placeholders and purely symbolic messages. Effective dialogue duration refers to the actual length of time a user participates in the dialogue. The formula for calculating the original speech rate is: ; in, This is the original speech rate (words per minute). For the number of valid words, Effective conversation duration (minutes).

[0063] Step S302: For the speech recognition text data, the original speech rate is corrected by the pause density and the proportion of the filler words to obtain the corrected speech rate.

[0064] The formula for calculating speech rate correction is: ; in, To correct speech rate, Pause density (the ratio of the number of pauses exceeding a preset duration to the number of valid words). The proportion of filler words (the proportion of filler words to the total number of words). and These are the speech correction coefficients. For example, 0.08 is acceptable. A value of 0.05 is acceptable. A higher pause density indicates more frequent pauses during language production; a higher proportion of filler words indicates difficulties in vocabulary retrieval or sentence organization. Therefore, adjusting the original speech rate based on pause density and filler word proportion can more accurately reflect the user's actual fluency.

[0065] Step S303: For instant messaging text data, determine pause events based on message time intervals and count instant messaging pause density, obtain writing feature correction factor based on writing feature quantification statistics, and correct the original speech rate with instant messaging pause density and writing feature correction factor; thus obtaining the corrected instant messaging speech rate.

[0066] The formula for calculating the corrected speech rate in instant messaging is: ; in, To adjust speech rate for instant messaging, This refers to the pause density in instant messaging (the ratio of the number of pause events determined based on message intervals to the number of valid words). The writing feature correction factor (obtained through quantitative statistics based on writing features such as ellipses and repeated punctuation) is used. and This is a correction factor for instant messaging. For example, 0.08 is acceptable. 0.05 is acceptable.

[0067] Step S304: Normalize the speech rate correction and instant messaging rate correction relative to the average reference speech rate to obtain the speech fluency feature value.

[0068] The formula for normalizing the eigenvalues ​​of fluency is: ; in, For language fluency features, To correct speech rate for voice recordings or instant messaging. This is for reference to the average speaking speed. For example, It can take 120 words per minute. The function is used to limit the calculation result to a range of 0 to 100. The reference average speech rate can be adjusted according to the age and language background of the target user group.

[0069] like Figure 4 As shown, the vocabulary extraction difficulty feature value is determined in the following way: Step S401: Detect difficult events in word extraction from the de-identified text data.

[0070] Vocabulary extraction difficulty events refer to situations where users encounter difficulties in vocabulary retrieval or make naming errors during language production. These events include at least one of the following: named entity substitution, semantically related word shifting, and extraction interruption.

[0071] Named entity replacement detection can be achieved as follows: Named entity recognition is performed on the de-identified text data to identify entities such as names of people, places, and items. When a generic term (such as "that thing," "that person," or "that place") appears in the text and has a clear referent in the context, it is considered a named entity replacement event. Specifically, a generic term dictionary can be maintained, and generic terms appearing in the text can be matched against the dictionary. When a match is successful and there is a referentable entity in the preceding context of the generic term, it is counted as a named entity replacement event. For example, if a user mentions "I went to the hospital today" and later says "there were many people in that place," where "that place" refers to "the hospital," but the user does not directly say "hospital" but uses a generic term instead, this is considered a named entity replacement event.

[0072] Semantic word shift detection can be achieved by using word vector models (such as Word2Vec, GloVe, or word embeddings based on pre-trained language models) to map content words in the text into semantic vectors; for content words that have a clear referent in the context, calculate the semantic similarity between the content word and the referent; when the semantic similarity is higher than a preset similarity threshold (indicating that the two are semantically related) but the content word is not the correct name of the referent, it is determined as a semantic word shift event.

[0073] Specifically, the detection of semantically related word shifts can include the following steps: First, determine the target entity and its correct name in the text through named entity recognition or context analysis; second, detect whether there are any substitute words in the text that are semantically related to the target entity but have different names; third, use a word vector model to calculate the cosine similarity between the substitute word and the correct name of the target entity; fourth, when the cosine similarity is higher than a preset similarity threshold (such as 0.6) and the substitute word is not equal to the correct name, it is determined as a semantically related word shift event.

[0074] For example, a user mentions "I have a dog" in the first part of the text and then says "I take my cat for a walk every day". The cosine similarity between "cat" and "dog" in the word vector space is high (both are pet categories), but "cat" is not the correct name for "dog". This is determined to be a semantically related word shift event.

[0075] Extraction interruption detection can be achieved as follows: Maintain an extraction interruption marker dictionary containing explicit word retrieval interruption markers such as "what was it called again?", "that thing", "it's that one", "can't remember", etc.; match the de-identified text data with the dictionary, and when a match is successful, it is determined as an extraction interruption event. In addition, consecutive ellipses (such as "...") or long pauses in the text (based on speech timestamp information) can also be detected as auxiliary criteria for extraction interruption determination.

[0076] Step S402: Count the frequency of difficult events in vocabulary extraction within the effective word count of a unit.

[0077] The formula for calculating the frequency of occurrence is: ; in, The frequency of occurrence of events that make word retrieval difficult. This represents the number of times named entity replacement occurred. This represents the number of times semantically related word shifts occur. To extract the number of interrupts, This represents the number of valid words.

[0078] Step S403: Normalize the occurrence frequency relative to the reference baseline to obtain the word extraction difficulty feature value.

[0079] The normalization formula for the feature value of vocabulary extraction difficulty is: ; in, Extracting difficulty feature values ​​for words. For reference, For example, A value of 0.02 is acceptable, meaning that, based on publicly available naming ability benchmark data, normal adults experience difficulty retrieving words that appear less than twice per hundred words in natural conversation.

[0080] like Figure 5 As shown, semantic coherence feature values ​​are determined in the following way: Step S501: Segment the de-identified text data according to sentences.

[0081] Step S502: Encode each pair of adjacent sentences into semantic vectors using a pre-trained language model.

[0082] The pre-trained language model encodes each statement as a fixed-length semantic vector. For example, the semantic vector can have 768 dimensions (based on BERT-base) or more.

[0083] Step S503: Calculate the cosine similarity of the semantic vectors of each pair of adjacent statements.

[0084] The formula for calculating cosine similarity is: ; in, For the first Cosine similarity of adjacent statements For the first The semantic vector of each statement. For the first The semantic vector of each statement. The cosine similarity function is used. This indicates the first statement segment within a dialogue block. The text content of each statement.

[0085] Step S504: Take the average cosine similarity of all adjacent sentence pairs to obtain the average similarity.

[0086] The formula for calculating the average similarity is: ; in, For average similarity, This represents the number of statements within the dialogue segment.

[0087] Step S505: Detect unexpected topic switching events between adjacent statements and calculate topic switching penalty terms.

[0088] The detection of unexpected topic switching events is based on the cosine similarity of adjacent statements calculated in step S503. Implementation. When When the similarity is below a preset threshold, the first... An unexpected topic switch occurs between adjacent statements. The total number of unexpected topic switch events between adjacent statements within a dialogue segment is recorded as follows: The preset similarity threshold is determined as follows: For normal dialogue samples and dialogue samples containing unexpected topic switching in the labeled dataset, the cosine similarity distribution of adjacent sentences is calculated respectively. The similarity value that maximizes the distinction between the two types of samples is taken as the preset similarity threshold. The formula for calculating the preset similarity threshold is: .

[0089] in, Set a preset similarity threshold; This is the set of cosine similarities between adjacent sentences in a normal dialogue sample in the labeled dataset. The set of cosine similarities between adjacent sentences in a labeled dataset containing dialogue samples with unexpected topic switching. and These represent the number of elements in the corresponding set; The cosine similarity of adjacent statements in the set; that is, the similarity calculated in step S503. ; This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. This indicates taking the option that minimizes the objective function value. value.

[0090] For example, The value can be between 0.3 and 0.5, with 0.4 being the preferred value.

[0091] The penalty for switching topics is calculated as follows: ,in Switch the penalty coefficient for the topic. The number of statements in the dialogue segment. This represents the total number of unexpected topic switching events. For example, Version 5.0 is acceptable. The method for determining this is as follows: For dialogue samples in the labeled dataset that contain unexpected topic switching, the semantic coherence score as the objective is to fit... The value of minimizes the mean square error between the semantic coherence feature value and the human score.

[0092] Step S506: Subtract the topic switching penalty from the average similarity to obtain the semantic coherence feature value.

[0093] The formula for calculating semantic coherence feature values ​​is: ; in, For semantic coherence feature values, The total number of unexpected topic switching events. Switch the penalty coefficient for the topic.

[0094] Step S103: Determine the user's comprehensive cognitive assessment result based on language characteristics.

[0095] like Figure 6 As shown, step S103 specifically includes the following steps: Step S601: Normalize the language features.

[0096] The language fluency feature values ​​obtained above Vocabulary extraction difficulty feature value and semantic coherence feature values Normalize each eigenvalue to the 0-100 range. If the above eigenvalues ​​have already been normalized to 0-100 using the formula, this step can be used directly.

[0097] Step S602: The normalized language features of each dimension are weighted and summed according to the preset weight coefficients to obtain the comprehensive cognitive assessment result.

[0098] The comprehensive cognitive assessment results can be calculated using the following weighted summation formula: ; in, To comprehensively assess cognitive results, For language fluency features, Extracting difficulty feature values ​​for words. For semantic coherence feature values, , and These are preset weighting coefficients. For example, 0.35 is acceptable. 0.35 is acceptable. 0.30 is acceptable. The preset weighting coefficients are adjusted based on the age, language background, or cognitive ability distribution of the target user group. Finally, the comprehensive cognitive assessment results are mapped to discrete ability levels, making it easier to present users with an intuitive ability profile.

[0099] For example, in some implementations of the application, a preset weighting coefficient is used. , and Adjustments can be made based on the age of the target user group. Specifically, a mapping relationship between age groups and preset weight coefficients can be established in advance. The following is an example table of an age-stratified mapping: The aforementioned mapping table can be determined as follows: For each age group, daily conversation data of users in that age group is collected. Following steps S102 to S103, three-dimensional language features for each user are extracted, namely, language fluency feature values, vocabulary extraction difficulty feature values, and semantic coherence feature values. Simultaneously, professionals provide reference evaluation scores for the language abilities of the same group of users under offline conditions. These reference evaluation scores are only used for offline calibration of preset weight coefficients and do not participate in the online operation of the method described in this application. This forms a labeled dataset for that age group.

[0100] Using three-dimensional language features as independent variables and reference assessment scores as dependent variables, multiple linear regression was performed to obtain the contribution coefficients of language fluency, vocabulary extraction difficulty, and semantic coherence features to the reference assessment scores. A larger contribution coefficient indicates a greater predictive contribution of that dimension to the reference assessment score. Each of the three contribution coefficients was divided by the sum of the three values ​​and normalized until the sum of the normalized coefficients equaled 1. These normalized coefficients are then used as the preset weighting coefficients for that age group. , And fill in the mapping table.

[0101] For example, for the 70-79 age group, if regression analysis shows that the contribution coefficient of the vocabulary extraction difficulty feature value is the largest, then the normalized β value should be the largest. For instance, the preset weight coefficient for the vocabulary extraction difficulty feature value is 0.40, and the preset weight coefficients for the language fluency feature value and the semantic coherence feature value are each 0.30. This indicates that for the 70-79 age group, the vocabulary extraction difficulty feature value should have the largest weight in the comprehensive score.

[0102] In other implementations of the application, for user groups with different language backgrounds, such as dialect speakers, the importance of semantic coherence features is relatively increased due to differences in vocabulary and word order between dialects and Standard Mandarin, and the value of γ can be adjusted accordingly. For user groups with different cognitive ability distributions, such as those at higher risk of cognitive decline, the weight β of the vocabulary extraction difficulty feature can be increased accordingly. The adjustments to the above-mentioned preset weight coefficients can all be determined by collecting labeled data of the corresponding groups and using the aforementioned multiple linear regression method. Specifically, for user groups with a dialect language background, the degree of dialect use is added as a covariate in the regression analysis; for groups at higher risk of cognitive decline, the level of cognitive risk is added as a covariate in the regression analysis, thereby obtaining preset weight coefficients for different groups.

[0103] In other implementations of the application, the preset weight coefficients can also be determined through a machine learning model. Specifically, a regression model is constructed that takes user group characteristics as input and preset weight coefficients as output. The training dataset is constructed as follows: labeled data from multiple groups of different user groups are collected, and the aforementioned multiple linear regression is performed on each group to obtain the preset weight coefficients for that group as training labels; a gradient boosting tree or lightweight neural network model is trained using group characteristics as input features and preset weight coefficients as output labels. After training, for a new user group, its group characteristics are input into the trained model to output the corresponding preset weight coefficients. This model can be deployed on a cloud server, and the initial preset weight coefficients are automatically determined based on user profile information when the user uses the model for the first time.

[0104] Step S603: For feature dimensions where higher values ​​indicate lower cognitive ability, use their inverse values ​​to participate in the weighted summation, so that the direction of each dimension is unified so that higher scores indicate better cognitive ability.

[0105] Vocabulary extraction difficulty feature value The higher the value, the more frequently the user experiences difficulty in word retrieval, indicating a lower cognitive ability. Therefore, the weighted summation formula uses... As a contribution value for this dimension, its direction is aligned with other dimensions, with higher scores indicating better cognitive ability.

[0106] In other embodiments of this application, the comprehensive cognitive assessment results can also be calculated using a machine learning regression model. Specifically, a labeled dataset is collected: three-dimensional language features from daily conversations of N subjects are simultaneously collected (…). , , ) and standardized language proficiency assessment scores, using multiple linear regression or gradient boosting tree (GBDT) models to ( , , The input is a standardized score, which is used as the target to fit a scoring function. Alternatively, an end-to-end neural network approach can be adopted: the dialogue text is directly input into a pre-trained language model (such as BERT or RoBERTa), and a fully connected layer is added after the CLS token to output the comprehensive cognitive evaluation result. In this case, the pre-trained language model implicitly assumes the dual role of feature extraction and mapping.

[0107] Step S104: Based on the comprehensive cognitive assessment results, generate training content for user interaction.

[0108] like Figure 7 As shown, step S104 specifically includes the following steps: Step S701: Obtain user profile information, which includes at least one of the user's age group, geographic background, and interests. User profile information can be obtained through user registration information, analysis of historical dialogue content, or data entry by caregivers.

[0109] Step S702: Determine the training difficulty level based on the comprehensive cognitive assessment results.

[0110] The mapping relationship between comprehensive cognitive assessment results and training difficulty levels can be set according to actual needs. For example, comprehensive cognitive assessment results can be divided into multiple increasing levels: 80-100 corresponds to level 5, 60-79 to level 4, 40-59 to level 3, 20-39 to level 2, and 0-19 to level 1. Lower levels correspond to basic vocabulary and simple question-and-answer sessions, while higher levels correspond to complex logical reasoning and abstract topic discussions.

[0111] Step S703: Based on the comprehensive cognitive assessment results, training difficulty level, and user profile information, construct and generate prompt words.

[0112] The generated prompts include the current comprehensive cognitive assessment results, training difficulty level indicators, and user profile information. For example, the generated prompts could be: "Please generate a nostalgic-themed dialogue training content with a difficulty level of 3 for a user born in the 1950s and living in the Jiangnan region. The theme should revolve around Spring Festival customs, presented in a natural chat tone, and avoid repetition." Step S704: Input the generated prompt words into the pre-trained language model, which will then generate training content corresponding to the training difficulty level in real time.

[0113] The training content includes nostalgic-themed training materials, generated based on the user's historical period, age group, or geographical location. Examples of these materials include historical event dialogues generated based on the user's historical period, cultural topics related to the era generated based on the user's age group, and local memories generated based on the user's geographical location. Because the training content is generated in real-time by a pre-trained language model, each generated content differs in its specific expressions, case selections, and topic angles, thus avoiding content repetition and rigidity. The training difficulty levels include multiple increasing levels, with lower levels corresponding to basic vocabulary and simple question-and-answer sessions, and higher levels corresponding to complex logical reasoning and abstract topic discussions.

[0114] Regarding content generation control strategies, explicit constraints can be embedded in the prompts, such as "only involving everyday topics" and "avoiding politically or religiously sensitive content," supplemented by an output filtering mechanism to conduct security checks on the generated text. More refined control methods can draw on constraint decoding technology, imposing hard constraints during the word-by-word generation process of the pre-trained language model to ensure that the output always adheres to the preset structural framework and topic boundaries.

[0115] Step S105: Obtain the user's response data to the training content, use the response data as new language interaction data, redetermine language features, and update the comprehensive cognitive assessment results.

[0116] like Figure 8 As shown, step S105 specifically includes the following steps: Step S801: Embed the training content into the user's interaction process in the form of natural dialogue.

[0117] The training content is not presented in an explicit form as "exercises" or "tasks," but rather embedded in the user's daily interactions in a natural conversational tone and rhythm. For example, the system may bring up a topic related to the user's background in a casual manner, guiding the user to recall and narrate.

[0118] Step S802: Collect user response data to the training content.

[0119] The language data generated by the user during the response process is collected by the system as new language interaction data. The collection method is the same as in step S101.

[0120] Step S803: Re-perform feature extraction and evaluation using the response data as new language interaction data.

[0121] The delay from obtaining response data to updating the comprehensive cognitive assessment results shall not exceed a preset delay threshold. For example, this preset delay threshold can be 30 minutes to ensure timely feedback and avoid the training content becoming out of touch with the user's current state due to excessive delay.

[0122] The method also includes at least one of the following boundary condition processing: When the voice data contains multiple speakers, the user's voice segments are extracted using voiceprint recognition and speaker separation technologies. Voiceprint recognition technology creates an independent voiceprint template for each registered user, and the speaker separation algorithm segments the mixed voice stream into exclusive voice segments for each user.

[0123] When the confidence level of speech recognition is lower than a preset confidence level threshold, the corresponding segment data is downweighted or discarded. For example, the preset confidence level threshold can be 0.70.

[0124] When the effective dialogue duration or number of effective dialogues based on de-identified text data falls below a preset threshold, it is marked as insufficient data and the comprehensive cognitive assessment results are not updated temporarily. For example, if the effective dialogue duration is less than 2 minutes or the number of effective dialogues is less than 10 sentences, it can be marked as insufficient data.

[0125] If a user has no valid conversation records for a preset number of days, the most recent comprehensive cognitive assessment result will be maintained and the user will be marked as silent. For example, the preset number of days can be 7 days.

[0126] For example, the personalized cognitive training method based on language models provided in this application also includes a privacy-protected tiered deployment: In the first deployment mode, speech recognition, de-identification processing, and language feature extraction are completed locally on the user terminal, with only the de-identified language features uploaded to the cloud. Specifically, after the speech is transcribed into text locally by ASR and de-identified locally, the pre-trained language model on the terminal performs three-dimensional feature extraction, uploading only the de-identified language features. Uploaded to the cloud for scoring calculation and content generation. Original instant messaging text data, voice data, and any personally identifiable information remain on the device.

[0127] In the second deployment mode, after speech recognition and de-identification processing are completed on the user terminal, the de-identified text data is encrypted and transmitted to the cloud. A pre-trained language model in the cloud performs language feature extraction, comprehensive cognitive assessment, and training content generation. After cloud processing, the de-identified text data is deleted. The cloud-based pre-trained language model (such as GPT-4 or Claude) performs 3D feature extraction, calculates the comprehensive cognitive assessment results, and generates training content. The generated comprehensive cognitive assessment results are returned to the terminal via an encrypted channel. The de-identified text data is deleted immediately after cloud processing and is not persistently stored.

[0128] The above options can be flexibly selected based on the actual deployment scenario. The first-level option is preferred by default, and the second-level option is only enabled when the user explicitly authorizes it and higher-quality analysis is required.

[0129] For example, in an embodiment of this application, the pre-trained language model is a large language model based on the Transformer architecture. The pre-trained language model receives language interaction data through a structured prompt template. The structured prompt template instructs the pre-trained language model to output language feature scores and observation records, and includes chained reasoning instructions to instruct the pre-trained language model to first analyze the language feature performance sentence by sentence, and then output scores for each dimension.

[0130] like Figure 9 As shown, this application also provides a personalized cognitive training system 900 based on a language model, used to execute any of the methods described in the above embodiments. The system includes: Data acquisition module 901 is used to acquire user language interaction data; The feature determination module 902 is used to input language interaction data into a pre-trained language model and determine the language features in the language interaction data that are related to the user's cognitive ability through the pre-trained language model. Assessment module 903 is used to determine the user's comprehensive cognitive assessment result based on language characteristics; The training content generation module 904 is used to generate training content for user interaction based on the comprehensive cognitive assessment results. The update module 905 is used to acquire the user's response data to the training content, use the response data as new language interaction data, redetermine language features, and update the comprehensive cognitive assessment results.

[0131] This application also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the language model-based personalized cognitive training method described in any of the above method embodiments. The storage medium can be any form of computer-readable storage medium, such as a read-only memory, random access memory, flash memory, hard disk, or optical disk. When the computer program stored in the storage medium is executed by a processor, the processor performs the steps in the above method embodiments.

[0132] The preferred embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this application, various simple modifications can be made to the technical solution of this application, and these simple modifications all fall within the protection scope of this application.

[0133] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this application will not describe the various possible combinations separately.

[0134] Furthermore, various different implementations of this application can be combined in any way, as long as they do not violate the spirit of this application, they should also be regarded as the content disclosed in this application.< / sensitive>

Claims

1. A method of personalized cognitive training based on a language model, characterized in that, Includes the following steps: Obtain user language interaction data; The language interaction data is input into a pre-trained language model, and the language features associated with the user's cognitive ability in the language interaction data are determined by the pre-trained language model; wherein, the language features include language fluency feature values, vocabulary extraction difficulty feature values, and semantic coherence feature values; Based on the language characteristics, determine the user's comprehensive cognitive assessment result; Based on the comprehensive cognitive assessment results, training content is generated for the user to interact with; The user's response data to the training content is obtained, and the response data is used as new language interaction data to redetermine the language features and update the comprehensive cognitive assessment result.

2. The method of claim 1, wherein, The acquisition of user language interaction data includes: Collect voice data and / or instant messaging text data generated by users in non-task-oriented natural scenarios; The speech data is subjected to speech recognition to obtain speech recognition text data and speech timestamp information; The instant messaging text data and the speech recognition text data are subjected to de-identification processing to obtain de-identified text data; wherein, The de-identified text data includes the de-identified instant messaging text data and the de-identified speech recognition text data; The instant messaging text data after de-identification is segmented based on sender switching and message time interval, pause events are determined based on message time interval, writing features are quantified and statistically analyzed, and emoticons, internet slang and abbreviations are filtered and standardized and mapped. The speech recognition text data after de-identification is segmented based on the speech timestamp information and pause events exceeding a preset duration are detected to calculate pause density and / or filler word ratio.

3. The method of claim 2, wherein, The language fluency feature value is determined in the following way: Based on the de-identified text data, the number of effective words and the effective dialogue duration are counted, and the original speech rate is calculated. For the speech recognition text data, the original speech rate is corrected by the pause density and the proportion of filler words to obtain the corrected speech rate; For the instant messaging text data, pause events are determined based on message time intervals and the instant messaging pause density is statistically analyzed. A writing feature correction factor is obtained based on the quantitative statistics of the writing features. The original speech rate is corrected using the instant messaging pause density and the writing feature correction factor. Get instant messaging speed correction; The speech rate correction and instant messaging rate correction are normalized relative to the average reference speech rate to obtain the language fluency feature value. The vocabulary extraction difficulty feature value is determined in the following way: Detect vocabulary extraction difficulties in the de-identified text data, wherein the vocabulary extraction difficulties include at least one of named entity replacement, semantically related word shift, and extraction interruption; The frequency of events indicating difficulty in word extraction within the effective word count of a statistical unit; The occurrence frequency is normalized relative to a reference benchmark to obtain the word extraction difficulty feature value; The semantic coherence feature values ​​are determined in the following way: The de-identified text data is segmented by sentence. The pre-trained language model is used to encode each pair of adjacent sentences into a semantic vector; Calculate the cosine similarity of the semantic vectors of each pair of adjacent statements; The average similarity is obtained by averaging the cosine similarities of all adjacent sentence pairs. Detect unexpected topic switching events between adjacent statements and calculate topic switching penalties; The semantic coherence feature value is obtained by subtracting the topic switching penalty from the average similarity.

4. The method according to claim 1, characterized in that, The step of determining the user's comprehensive cognitive assessment result based on the language features includes: The language features are normalized. The normalized language features of each dimension are weighted and summed according to the preset weight coefficients to obtain the comprehensive cognitive assessment result. For feature dimensions where higher values ​​indicate lower cognitive ability, their inverse values ​​are used in the weighted summation to unify the direction of each dimension so that higher scores indicate better cognitive ability. The preset weighting coefficients are adjusted according to the age, language background, or cognitive ability distribution of the target user group, and the comprehensive cognitive assessment results are mapped to discrete ability levels.

5. The method according to claim 1, characterized in that, The step of generating training content for user interaction based on the comprehensive cognitive assessment results includes: Obtain user profile information, which includes at least one of the user's age group, geographic background, and points of interest; The training difficulty level is determined based on the comprehensive cognitive assessment results. Based on the comprehensive cognitive assessment results, the training difficulty level, and the user profile information, prompt words are constructed and generated. The generated prompt words are input into a pre-trained language model, which then generates training content corresponding to the training difficulty level in real time. The training content includes nostalgic-themed training materials, which are generated based on the user's historical background, age group, or regional background. The training difficulty level includes multiple progressively increasing levels, with lower levels corresponding to basic vocabulary and simple questions and answers, and higher levels corresponding to complex logical reasoning and abstract topic discussions.

6. The method according to claim 3, characterized in that, The step of acquiring the user's response data to the training content, using the response data as new language interaction data, redefining the language features, and updating the comprehensive cognitive assessment result includes: The training content is embedded into the user's interaction process in the form of natural dialogue; Collect user response data to the training content; The response data is then used as new language interaction data to re-perform feature extraction and evaluation; Among them, the delay from obtaining response data to updating the comprehensive cognitive assessment results shall not exceed a preset delay threshold; The method further includes at least one of the following boundary condition processing methods: When the voice data contains multiple speakers, the user's voice segments are extracted using voiceprint recognition and speaker separation technology. When the recognition confidence level of the speech recognition is lower than the preset recognition threshold, the corresponding segment data is downweighted or discarded. When the effective dialogue duration or number of effective dialogues based on the de-identified text data is lower than a preset threshold, it is marked as insufficient data and the comprehensive cognitive assessment result is not updated for the time being. If a user has no valid conversation records for a preset number of consecutive days, the most recent comprehensive cognitive assessment result will be maintained and the user will be marked as silent.

7. The method according to claim 2, characterized in that, This also includes the deployment of privacy protection tiers: In the first deployment mode, speech recognition, de-identification processing and language feature extraction are completed locally on the user terminal, and only the de-identified language features are uploaded to the cloud. In the second deployment mode, after the user terminal completes speech recognition and de-identification processing, the de-identified text data is encrypted and transmitted to the cloud. The cloud-pre-trained language model performs language feature extraction, comprehensive cognitive evaluation, and training content generation. After the cloud processing is completed, the de-identified text data is deleted.

8. The method according to claim 1, characterized in that, The pre-trained language model is a large language model based on the Transformer architecture; The pre-trained language model receives the language interaction data through a structured prompt template. The structured prompt template instructs the pre-trained language model to output language feature scores and observation records, and includes chained reasoning instructions to instruct the pre-trained language model to first analyze the language feature performance sentence by sentence, and then output scores for each dimension.

9. A personalized cognitive training system based on a language model, characterized in that, include: The data acquisition module is used to acquire user language interaction data; The feature determination module is used to input the language interaction data into a pre-trained language model and determine the language features in the language interaction data that are related to the user's cognitive ability through the pre-trained language model. An assessment module is used to determine the user's comprehensive cognitive assessment result based on the language features. The training content generation module is used to generate training content for user interaction based on the comprehensive cognitive assessment results. The update module is used to acquire the user's response data to the training content, use the response data as new language interaction data, redetermine the language features, and update the comprehensive cognitive assessment result.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the personalized cognitive training method based on a language model as described in any one of claims 1 to 8.

Images