Psychological assessment method and system based on semantic analysis

Abstract

The invention relates to the technical field of computer software, in particular to a psychological assessment method and system based on semantic analysis. A psychological assessment system based on semantic analysis comprises a user answering module, a question text generation module and a psychological assessment module. According to the psychological assessment method and system, the psychological assessment agent performs multi-round dialogues with the user, the psychological assessment of the user is realized in combination with the dialogue generation model and the psychological assessment model, and the answer of the user and the assessment result of each round are combined in the dialogue generation model, so that the question generated by the psychological assessment agent is more personalized, and the user experience is improved. The psychological change when the user answers is combined in the psychological assessment model, so that the psychological assessment result is higher in accuracy on the basis of personalized information fusion, and the psychological assessment is more flexible.

Description

Technical Field

[0001] The present invention relates to the technical field of computer software, and particularly relates to a psychological assessment method and system based on semantic analysis. Background Art

[0002] As an essential and integral part of the field of mental health, the results of psychological assessment directly affect the accurate evaluation of an individual's mental state, the early identification of potential psychological problems, and the effective formulation of targeted intervention measures, thus concerning the quality of life of the individual and the sound development of social functions. For a long time, traditional psychological assessment methods, especially surveys based on fixed questionnaires and standardized scale tests, have always dominated and been widely used. The core feature of these methods lies in the pre-designed and strictly followed set of structured and standardized questionnaire systems, which contain a series of carefully selected and validated fixed questions. The test takers need to make choices and answers within the preset and limited option ranges according to their subjective feelings and objective situations. For example, in the screening and assessment of common psychological problems such as depression, anxiety disorder, and personality disorder, the widely used Self-Rating Depression Scale (SDS), Self-Rating Anxiety Scale (SAS), Eysenck Personality Questionnaire (EPQ), etc. all adopt this highly standardized fixed questionnaire form without exception.

[0003] However, although traditional fixed-questionnaire-based psychological assessment methods have certain advantages in terms of standardization, operability, and large-scale administration, their inherent limitations are becoming increasingly prominent. Especially in the current context of growing personalized needs, these limitations seriously restrict the accuracy and effectiveness of psychological assessment. Its most fundamental defect is that in order to pursue the universality and comparability of assessment, existing methods often sacrifice the full consideration of individual differences. Summary of the Invention

[0004] The present invention conducts multiple rounds of conversations between a psychological assessment agent and a user, and combines a dialogue generation model and a psychological assessment model to achieve the psychological assessment of the user. By combining the user's answers and the assessment results of each round in the dialogue generation model, the questions generated by the psychological assessment agent become more personalized. By combining the psychological changes of the user when answering in the psychological assessment model, the accuracy of the psychological assessment results is higher on the basis of personalized information integration, and the psychological assessment is also more flexible.

[0005] The present invention provides a psychological assessment method based on semantic analysis, including: Set the psychological assessment agent to output question texts through a dialogue generation model. Construct a question-and-answer text sequence data based on all the historical question texts and all the historical answer texts of the user in the question-and-answer order. Conduct multiple rounds of conversations between the psychological assessment agent and the user. Process the question-and-answer text sequence data through the psychological assessment model to generate psychological assessment results, and use the stabilized psychological assessment results as the final psychological assessment results; Set a psychological assessment result guidance layer and a question-and-answer direction guidance layer in the dialogue generation model. Among them, the psychological assessment result guidance layer is used to reconstruct the question-and-answer text sequence data based on the psychological assessment results output in the previous round; the question-and-answer direction guidance layer is used to query the psychological knowledge graph based on the question-and-answer text sequence data to construct a question-and-answer direction guidance feature, and the question-and-answer direction guidance feature is used to construct a corresponding query vector when the decoder in the dialogue generation model executes the self-attention mechanism; Set a question-and-answer direction guidance analysis layer in the psychological assessment model. The question-and-answer direction guidance analysis layer is used to analyze the question-and-answer direction guidance features of all previous rounds to construct a question-and-answer direction feature, and the question-and-answer direction feature is used to construct a corresponding query vector when the decoder in the psychological assessment model executes the self-attention mechanism.

[0006] As a preferred aspect, the dialogue generation model specifically includes the following: In the psychological assessment result guidance layer, obtain the psychological assessment results output by the psychological assessment model in the previous round, and append the psychological assessment results output by the psychological assessment model in the previous round at the end of each line of the question-and-answer text sequence data to construct the reconstructed question-and-answer text sequence data, and input the reconstructed question-and-answer text sequence data into the encoder of the subsequent dialogue generation model; In the question-and-answer direction guidance layer, query in the psychological knowledge graph based on the words corresponding to each line in the question-and-answer text sequence data to obtain all associated psychological knowledge graph triples, splice all the obtained psychological knowledge graph triples from top to bottom to construct a psychological knowledge graph feature, construct a corresponding query vector based on the psychological knowledge graph feature, construct corresponding value vectors and key vectors based on the reconstructed question-and-answer text sequence data, execute the self-attention mechanism, and construct a question-and-answer direction guidance feature; In the decoder of the dialogue generation model, construct corresponding value vectors and key vectors based on the output of the last encoder in the dialogue generation model, construct a corresponding query vector based on the question-and-answer direction guidance feature, and execute the self-attention mechanism.

[0007] As a preferred aspect, the psychological assessment model specifically includes the following: In the Q&A direction guidance analysis layer, the Q&A direction guidance features output by the dialogue generation model in all previous rounds are arranged in chronological order to construct a time series set of Q&A direction guidance features. The Q&A direction analysis network includes a ConvLSTM layer and a fully connected layer. The ConvLSTM layer is used to perform time series analysis on the time series set of Q&A direction guidance features through the Q&A direction analysis network, and the fully connected layer is used to perform a fully connected operation on the output of the ConvLSTM layer to construct Q&A direction features. In the decoder of the psychological assessment model, a corresponding value vector and key vector are constructed based on the output of the last encoder in the psychological assessment model, a corresponding query vector is constructed based on the Q&A direction features, and the self-attention mechanism is executed.

[0008] As a preferred aspect, training the dialogue generation model specifically includes the following steps: Obtain a number of dialogue generation training samples, where the dialogue generation training samples include the Q&A records in the continuous psychological assessment process. Combine all the dialogue generation training samples to form a dialogue generation training set, and perform unsupervised training on the dialogue generation model through the dialogue generation training set. During the unsupervised training process, randomly mask the question text of the dialogue generation training samples and use the masked question text as the training target.

[0009] As a preferred aspect, training the psychological assessment model specifically includes the following steps: Obtain a number of psychological assessment training samples, where the psychological assessment training samples include the Q&A records and the corresponding round of Q&A direction guidance features in the psychological assessment process executed by the trained dialogue generation model. Label the psychological assessment training samples through the psychological assessment results, combine all the labeled psychological assessment training samples to form a psychological assessment training set, and perform supervised training on the psychological assessment model through the psychological assessment training set. During the training process, use the labeled psychological assessment results as the training target.

[0010] As a preferred aspect, determining that the output psychological assessment results tend to be stable specifically includes the following steps: Taking the currently output psychological assessment results as the end, obtain the previous N - 1 output psychological assessment results to form a psychological assessment result set. Perform clustering analysis on the psychological assessment result set to determine the central psychological assessment result, calculate the sum of the squares of the similarities between all the psychological assessment results in the psychological assessment result set and the central psychological assessment result, denoted as the stability value. Determine whether the stability value is higher than the stability threshold, which is set artificially. If the stability value is higher than the stability threshold, it is considered that the output psychological assessment results tend to be stable; otherwise, the output psychological assessment results tend to be stable.

[0011] The present invention also provides a psychological assessment system based on semantic analysis, including: A user response module for obtaining response texts of a user to question texts of a psychological assessment intelligent agent; A question text generation module for constructing question-and-answer text sequence data from all historical question texts of the psychological assessment intelligent agent and all historical response texts of the user in the order of question and answer, sending the question-and-answer text sequence data into a dialogue generation model for processing, and outputting the question text for the next round output by the psychological assessment intelligent agent; A psychological assessment module for sending the question-and-answer text sequence data into a psychological assessment model for processing. The psychological assessment model is established based on the Transformer model, generates psychological assessment results corresponding to each round, and when the output psychological assessment results tend to be stable, outputs the psychological assessment results finally output by the psychological assessment model as the final psychological assessment results; A psychological assessment result guidance layer and a question-and-answer direction guidance layer are set in the dialogue generation model. The psychological assessment result guidance layer is used to reconstruct the question-and-answer text sequence data based on the psychological assessment results output in the previous round; the question-and-answer direction guidance layer is used to query a psychological knowledge graph based on the question-and-answer text sequence data to construct a question-and-answer direction guidance feature, and the question-and-answer direction guidance feature is used to construct a corresponding query vector when the decoder in the dialogue generation model executes the self-attention mechanism; A question-and-answer direction guidance analysis layer is set in the psychological assessment model. The question-and-answer direction guidance analysis layer is used to analyze the question-and-answer direction guidance features of all previous rounds to construct a question-and-answer direction feature, and the question-and-answer direction feature is used to construct a corresponding query vector when the decoder in the psychological assessment model executes the self-attention mechanism.

[0012] The present invention has the following advantages: The present invention conducts multiple rounds of conversations between a psychological assessment intelligent agent and a user, and combines a dialogue generation model and a psychological assessment model to implement psychological assessment of the user. By combining the user's answers and the assessment results of each round in the dialogue generation model, the questions generated by the psychological assessment intelligent agent are made more personalized. By combining the psychological changes of the user when answering in the psychological assessment model, the psychological assessment results are more accurate on the basis of personalized information fusion, and the psychological assessment is also made more flexible. Brief Description of the Drawings

[0013] Figure 1 It is a schematic structural diagram of a psychological assessment system based on semantic analysis adopted in an embodiment of the present invention. Detailed Embodiments

[0014] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0015] Example 1, a psychological assessment method based on semantic analysis, includes: Step S1: Start the psychological assessment agent, obtain the response text of the user to the initial question text of the psychological assessment agent, construct the Q&A text sequence data based on the initial question text of the psychological assessment agent and the user's response text, send the Q&A text sequence data into the dialogue generation model for processing. The dialogue generation model is established based on the Transformer model and outputs the question text to be output by the psychological assessment agent in the next round. It should be noted that during the process of performing the psychological assessment, the user will have multiple rounds of Q&A with the psychological assessment agent. Generally, the initial questions of the psychological assessment agent are set in advance, such as "Which aspect of psychological problems do you want to consult?" or "Have you felt unhappy recently?" Since the model can only process numerical data, it is necessary to convert the Q&A records of the user and the agent into numerical data. Generally, the Q&A text sequence data constructed from the initial question text of the psychological assessment agent and the user's response text is arranged in a text sequence according to the Q&A order, and usually, identity information is added in front of the corresponding text, such as "[CLS], [USER], you, good, [SEP], [AGENT], you, good,!, very, glad, to, chat, with, you, [SEP], [USER], I, recently, feel, a, little, anxious, [SEP], [AGENT], can, you, specifically, talk, about, your, anxious, feeling,?, [SEP], [USER], mainly, due, to, work, pressure, big, can't, sleep, well, at, night, [SEP]". Then it is converted into a numerical form through word embedding. This numerical conversion step is also implemented in the word embedding layer of the Transformer model; Step S2: Obtain the response text of the user to the question text of the psychological assessment agent, construct the Q&A text sequence data according to the Q&A order from all the historical question texts of the psychological assessment agent and all the historical response texts of the user, send the Q&A text sequence data into the dialogue generation model for processing, and output the question text to be output by the psychological assessment agent in the next round; at the same time, send the Q&A text sequence data into the psychological assessment model for processing. The psychological assessment model is established based on the Transformer model and generates the corresponding psychological assessment results for each round. For each round of Q&A, a psychological assessment of the user is performed based on the historical Q&A records, and each psychological assessment result will be used as prior knowledge to guide the question generation of the psychological assessment agent. Considering the actual situation of the user during the psychological assessment process, the personalized information of the user is integrated, and the psychological assessment results here are generally psychological state categories, such as normal, mild anxiety, moderate anxiety, severe anxiety, no depression, mild depression, moderate depression, and severe depression, etc., and are numerically represented through one-hot encoding; Step S3: Repeat Step S2 until the output psychological assessment results tend to be stable. Output the psychological assessment results output by the psychological assessment model in the last round as the final psychological assessment results. The psychological assessment results tending to be stable means that the psychological assessment results output in the recent several rounds are the same or semantically similar, which is regarded as the completion of the psychological assessment of the user. A psychological assessment result guidance layer and a Q&A direction guidance layer are set in the dialogue generation model. The psychological assessment result guidance layer is used to reconstruct the Q&A text sequence data based on the psychological assessment results output in the previous round, so that the generation of the question text pays more attention to the psychological assessment results corresponding to the user's answers. The Q&A direction guidance layer is used to query the psychological knowledge graph based on the Q&A text sequence data to construct Q&A direction guidance features. The Q&A direction guidance features are used to construct corresponding query vectors when the decoder in the dialogue generation model executes the self-attention mechanism. The Q&A direction guidance features can determine the psychological assessment direction corresponding to the Q&A records between the psychological assessment agent and the user based on the psychological knowledge graph, such as anxiety or depression, so that the questions generated by the psychological assessment agent are more targeted. A Q&A direction guidance analysis layer is set in the psychological assessment model. The Q&A direction guidance analysis layer is used to analyze the Q&A direction guidance features of all previous rounds to construct Q&A direction features. The Q&A direction features are used to construct corresponding query vectors when the decoder in the psychological assessment model executes the self-attention mechanism, so as to pay attention to the user's psychological changes during the psychological assessment process and obtain more accurate psychological assessment results. This application conducts multiple rounds of conversations between the psychological assessment agent and the user, and combines the dialogue generation model and the psychological assessment model to achieve the psychological assessment of the user. By combining the user's answers and the assessment results of each round in the dialogue generation model, the questions generated by the psychological assessment agent are more personalized. By combining the user's psychological changes when answering in the psychological assessment model, the psychological assessment results have higher accuracy on the basis of personalized information fusion, and the psychological assessment is also more flexible.

[0016] Both the dialogue generation model and the psychological assessment model are established based on the Transformer model. Generally, the Transformer model includes an encoding layer and a decoding layer. The encoding layer includes a word embedding layer and six encoders. The word embedding layer can construct the Q&A text sequence data, and the Q&A text sequence data is stored in the form of a two-dimensional matrix. Each row in the Q&A text sequence data is the word vector corresponding to a word in the text sequence. Here, word embedding generally uses Word2Vec. The encoder generally performs the self-attention mechanism and the feed-forward neural propagation. The decoding layer generally includes a word encoding layer and six decoders. The word encoding layer in the decoder generally encodes the input data of the model to construct the corresponding query vector for the self-attention mechanism in the decoder. However, in the dialogue generation model and the psychological assessment model of this application, this word encoding layer is deleted, and instead, the Q&A direction guiding feature and the Q&A direction feature are used to construct the corresponding query vector. The decoder generally performs the self-attention mechanism and the feed-forward neural propagation. The self-attention mechanism generally constructs the corresponding value vector V, key vector K, and query vector Q based on the input features, and then through the formula: H = softmax(QK T / D 05 ) to implement the self-attention mechanism. H is the feature output by the self-attention mechanism, T is the matrix transpose operation, and D is the dimension size of the corresponding key vector K. The following explains the dialogue generation model and the psychological assessment model based on the basis of the Transformer model: For the dialogue generation model, in the psychological assessment result guiding layer, the psychological assessment result output by the psychological assessment model in the previous round is obtained, and the psychological assessment result output by the psychological assessment model in the previous round is added to the end of each row of the Q&A text sequence data, so that the subsequent encoder pays more attention to the psychological assessment result corresponding to the user's answer when performing the self-attention mechanism, and constructs the reconstructed Q&A text sequence data. The reconstructed Q&A text sequence data is input into the encoder of the subsequent dialogue generation model. In the Q&A direction guiding layer, based on the words corresponding to each row in the Q&A text sequence data, queries are made in the psychological knowledge graph to obtain all associated psychological knowledge graph triples. All the obtained psychological knowledge graph triples are concatenated from top to bottom to construct the psychological knowledge graph feature. Based on the psychological knowledge graph feature, the corresponding query vector is constructed. Based on the reconstructed Q&A text sequence data, the corresponding value vector and key vector are constructed, and the self-attention mechanism is executed to construct the Q&A direction guiding feature. In the decoder of the dialogue generation model, based on the output of the last encoder in the dialogue generation model, the corresponding value vector and key vector are constructed. Based on the Q&A direction guiding feature, the corresponding query vector is constructed, and the self-attention mechanism is executed. By obtaining the corresponding triples of the psychological knowledge graph from the user's answer in the psychological knowledge graph, and obtaining the corresponding psychological assessment direction of the user's answer from the psychological knowledge graph, the process of psychological assessment for the user can be made more personalized. Moreover, the psychological knowledge graph can adopt existing and publicly available data. The psychological knowledge graph consists of several triples of the psychological knowledge graph, specifically in the form of entity - entity relationship - entity; For the psychological assessment model, in the question - answering direction guidance analysis layer, the question - answering direction guidance features output by the dialogue generation model in all previous rounds are arranged in chronological order to construct a chronological set of question - answering direction guidance features. The question - answering direction analysis network is established based on the ConvLSTM model, including a ConvLSTM layer and a fully - connected layer. The ConvLSTM layer is used to perform chronological analysis on the chronological set of question - answering direction guidance features through the question - answering direction analysis network, and the fully - connected layer is used to perform a fully - connected operation on the output of the ConvLSTM layer to construct question - answering direction features; In the decoder of the psychological assessment model, a corresponding value vector and key vector are constructed based on the output of the last encoder in the psychological assessment model, a corresponding query vector is constructed based on the question - answering direction features, and the self - attention mechanism is executed; The question - answering direction features can reflect the psychological changes of the user during the psychological assessment process, making the results of the psychological assessment closer to the actual psychological situation of the user; Training the dialogue generation model specifically includes the following steps: Obtain several dialogue generation training samples. The dialogue generation training samples include the question - answering records during consecutive psychological assessment processes, and here the question - answering records are from the psychological assessment processes actually carried out by professional psychologists. All dialogue generation training samples are combined into a dialogue generation training set, and the dialogue generation model is trained unsupervised through the dialogue generation training set. During the unsupervised training process, the question text of the dialogue generation training samples is randomly masked, and the masked question text is used as the training target. It should be noted here that during the unsupervised training process, the set training conditions are generally to reach a certain number of training times; Training the psychological assessment model specifically includes the following steps: Obtain several psychological assessment training samples. The psychological assessment training samples include the Q&A records and the Q&A direction guiding features corresponding to each round in the psychological assessment process executed by the trained dialogue generation model. With the assistance of the trained dialogue generation model, volunteers execute actual psychological assessment experiments and obtain the corresponding Q&A records and Q&A direction guiding features. Label the psychological assessment training samples with the psychological assessment results. Here, the labeled psychological assessment results are the final psychological assessment results obtained by professional psychologists. Combine all the labeled psychological assessment training samples to form a psychological assessment training set, and perform supervised training on the psychological assessment model with the psychological assessment training set. During the training process, use the labeled psychological assessment results as the training target.

[0017] Judge whether the output psychological assessment results tend to be stable, which specifically includes the following steps: Taking the currently output psychological assessment result as the end, obtain the previous N - 1 output psychological assessment results to form a psychological assessment result set. N is generally 5. Conduct cluster analysis on the psychological assessment result set. Generally, K-Means is used for cluster analysis to determine the central psychological assessment result. Calculate the sum of the squares of the similarities between all the psychological assessment results in the psychological assessment result set and the central psychological assessment result, which is denoted as the stability value. The similarity calculation uses the cosine similarity algorithm. Judge whether the stability value is higher than the stability threshold, which is set manually. If the stability value is higher than the stability threshold, it is considered that the continuously output psychological assessment results tend to be stable; otherwise, the continuously output psychological assessment results tend to be stable.

[0018] Embodiment 2, a psychological assessment system based on semantic analysis, see Figure 1 , including: A user answer module, which is used to obtain the answer text of the user to the question text of the psychological assessment agent; A question text generation module, which is used to construct a Q&A text sequence data by arranging all the historical question texts of the psychological assessment agent and all the historical answer texts of the user in the Q&A order, send the Q&A text sequence data into the dialogue generation model for processing, and output the question text output by the psychological assessment agent in the next round; A psychological assessment module, which is used to send the Q&A text sequence data into the psychological assessment model for processing. The psychological assessment model is established based on the Transformer model, generates the psychological assessment result corresponding to each round, and when the output psychological assessment results tend to be stable, outputs the psychological assessment result output by the psychological assessment model last time as the final psychological assessment result; A psychological assessment result guidance layer and a Q&A direction guidance layer are set in the dialogue generation model. The psychological assessment result guidance layer is used to reconstruct the Q&A text sequence data based on the psychological assessment results output in the previous round, so that the generation of the question text pays more attention to the psychological assessment results corresponding to the user's answers. The Q&A direction guidance layer is used to query the psychological knowledge graph based on the Q&A text sequence data to construct Q&A direction guidance features. The Q&A direction guidance features are used to construct corresponding query vectors when the decoder in the dialogue generation model executes the self-attention mechanism. The Q&A direction guidance features can determine the psychological assessment direction corresponding to the Q&A records between the psychological assessment agent and the user based on the psychological knowledge graph, such as anxiety or depression, so that the questions generated by the psychological assessment agent are more targeted. A Q&A direction guidance analysis layer is set in the psychological assessment model. The Q&A direction guidance analysis layer is used to analyze the Q&A direction guidance features of all previous rounds to construct Q&A direction features. The Q&A direction features are used to construct corresponding query vectors when the decoder in the psychological assessment model executes the self-attention mechanism, so that the psychological changes of the user are concerned during the psychological assessment process, and more accurate psychological assessment results can be obtained.

[0019] It should be understood that those of ordinary skill in the art can make improvements or transformations according to the above description, and all such improvements and transformations shall fall within the protection scope of the appended claims of the present invention. The parts not described in detail in this specification belong to the prior art well-known to those skilled in the art.

Claims

1. A psychological assessment method based on semantic analysis, characterized in that: include: Set the psychological assessment agent to output question texts through the dialogue generation model, construct question-answer text sequence data based on all historical question texts and all historical answer texts of the user in the order of questions and answers, conduct multiple rounds of dialogues between the psychological assessment agent and the user, process the question-answer text sequence data through the psychological assessment model, generate psychological assessment results, and use the psychological assessment results that tend to be stable as the final psychological assessment results; A psychological assessment result guidance layer and a question-answering direction guidance layer are set in the dialogue generation model. The psychological assessment result guidance layer is used to reconstruct the question-answering text sequence data based on the psychological assessment results output in the previous round; the question-answering direction guidance layer is used to query the psychological knowledge graph based on the question-answering text sequence data to construct the question-answering direction guidance feature. The question-answering direction guidance feature is used to construct the corresponding query vector when the decoder in the dialogue generation model executes the self-attention mechanism; A question-answering direction guidance analysis layer is set in the psychological assessment model. The question-answering direction guidance analysis layer is used to analyze the question-answering direction guidance features of all previous rounds and construct question-answering direction features. The question-answering direction features are used to construct the corresponding query vector when the decoder in the psychological assessment model executes the self-attention mechanism.

2. A psychological assessment method based on semantic analysis according to claim 1, characterized in that: The dialogue generation model specifically includes the following: In the psychological evaluation result guidance layer, the psychological evaluation results output by the previous round of psychological evaluation model are obtained, and the psychological evaluation results output by the previous round of psychological evaluation model are added at the end of each line of the question and answer text sequence data to construct the reconstructed question and answer text sequence data, and the reconstructed question and answer text sequence data is input into the encoder of the subsequent dialogue generation model; In the question-answer direction guidance layer, based on the words corresponding to each line of the question-answer text sequence data, a query is performed in the psychological knowledge graph to obtain all related psychological knowledge graph triples, all the obtained psychological knowledge graph triples are spliced ​​from top to bottom, and psychological knowledge graph features are constructed. Based on the psychological knowledge graph features, a corresponding query vector is constructed, and based on the reconstructed question-answer text sequence data, a corresponding value vector and key vector are constructed, and a self-attention mechanism is executed to construct question-answer direction guidance features; In the decoder in the dialogue generation model, the corresponding value vector and key vector are constructed based on the output of the last encoder in the dialogue generation model, the corresponding query vector is constructed based on the question-answer direction guidance features, and the self-attention mechanism is executed.

3. A psychological assessment method based on semantic analysis according to claim 2, characterized in that: The psychological assessment model specifically includes the following contents: In the question-answer direction guidance analysis layer, the question-answer direction guidance features output by the dialogue generation model of all previous rounds are arranged in chronological order to construct a time series set of question-answer direction guidance features. The question-answer direction analysis network includes a ConvLSTM layer and a fully connected layer. The ConvLSTM layer is used to perform time series analysis on the time series set of question-answer direction guidance features through the question-answer direction analysis network. The fully connected layer is used to perform a fully connected operation on the output of the ConvLSTM layer to construct question-answer direction features. In the decoder of the psychological assessment model, the corresponding value vector and key vector are constructed based on the output of the last encoder in the psychological assessment model, the corresponding query vector is constructed based on the question-answer direction features, and the self-attention mechanism is executed.

4. A psychological assessment method based on semantic analysis according to claim 3, characterized in that: Training the dialogue generation model includes the following steps: A number of dialogue generation training samples are obtained, which include question and answer records in a continuous psychological assessment process. All dialogue generation training samples are combined into a dialogue generation training set. The dialogue generation model is trained unsupervisedly through the dialogue generation training set. During the unsupervised training process, the question text of the dialogue generation training sample is randomly masked, and the masked question text is used as the training target.

5. A psychological assessment method based on semantic analysis according to claim 4, characterized in that: Training the psychological assessment model includes the following steps: A number of psychological assessment training samples are obtained, which include question and answer records in the psychological assessment process performed by the trained dialogue generation model and question and answer direction guidance features of the corresponding rounds. The psychological assessment training samples are labeled according to the psychological assessment results, and all labeled psychological assessment training samples are combined into a psychological assessment training set. The psychological assessment model is supervised trained through the psychological assessment training set. During the training process, the labeled psychological assessment results are used as training targets.

6. A psychological assessment method based on semantic analysis according to claim 5, characterized in that: Judging that the output psychological assessment results tend to be stable, the specific steps include the following: Taking the currently output psychological assessment result as the end, obtain the first N-1 output psychological assessment results to form a psychological assessment result set, perform cluster analysis on the psychological assessment result set, determine the central psychological assessment result, calculate the sum of squares of the similarities between all psychological assessment results in the psychological assessment result set and the central psychological assessment result, record it as the stable value, and judge whether the stable value is higher than the stable threshold. If the stable value is higher than the stable threshold, it is considered that the continuously output psychological assessment results tend to be stable; otherwise, the continuously output psychological assessment results tend to be stable.

7. A psychological assessment system based on semantic analysis, characterized in that: The system applies a psychological assessment method based on semantic analysis as described in any one of claims 1 to 6, including: A user answer module is used to obtain the user's answer text to the question text of the psychological assessment agent; The question text generation module is used to construct question and answer text sequence data from all the historical question texts of the psychological assessment agent and all the historical answer texts of the user in the order of questions and answers, send the question and answer text sequence data to the dialogue generation model for processing, and output the question text to be output by the next round of psychological assessment agent; The psychological assessment module is used to send the question-answer text sequence data to the psychological assessment model for processing. The psychological assessment model is established based on the Transformer model to generate the psychological assessment results corresponding to each round. When the output psychological assessment results tend to be stable, the psychological assessment results output by the psychological assessment model for the last time are output as the final psychological assessment results. A psychological assessment result guidance layer and a question-answering direction guidance layer are set in the dialogue generation model. The psychological assessment result guidance layer is used to reconstruct the question-answering text sequence data based on the psychological assessment results output in the previous round; the question-answering direction guidance layer is used to query the psychological knowledge graph based on the question-answering text sequence data to construct the question-answering direction guidance feature. The question-answering direction guidance feature is used to construct the corresponding query vector when the decoder in the dialogue generation model executes the self-attention mechanism; A question-answering direction guidance analysis layer is set in the psychological assessment model. The question-answering direction guidance analysis layer is used to analyze the question-answering direction guidance features of all previous rounds and construct question-answering direction features. The question-answering direction features are used to construct the corresponding query vector when the decoder in the psychological assessment model executes the self-attention mechanism.

Images