An Automatic Questionnaire Completion Method Based on Progressive Divide and Conquer and Multi-Question Fusion Collaboration

Abstract

This invention proposes an automated questionnaire completion method based on progressive divide-and-conquer and multi-question fusion collaboration. It constructs a dedicated professional model for each question in the questionnaire; uses questions and options as queries to filter relevant clauses from the user's historical text to form a candidate set; performs independent one-stage training on each professional model based on the candidate clauses to obtain several expert-level models; sets a symptom-aware attention layer, freezes the parameters of the expert-level models, and then fuses their output encoding results and inputs them into the symptom-aware attention layer for two-stage collaborative training; finally, it uses the trained attention layer and the corresponding expert-level models to obtain the predicted results of the questions. This invention achieves accurate modeling of the degree correlation between questions through a "divide-and-conquer" strategy, then mines potential dependencies between questions through a "collaborative" fusion mechanism to alleviate information gaps, and uses task selection gates to avoid negative transfer effects, significantly improving the accuracy, robustness, and interpretability of automated questionnaire completion.

Description

Technical Field

[0001] This invention relates to the field of computer natural language processing, and particularly to a method for automatically filling out questionnaires based on progressive divide-and-conquer and multi-question fusion collaboration. Background Technology

[0002] Questionnaire completion, as a structured and standardized data collection and analysis tool, is one of the core data sources for empirical research. Its core value lies not only in the systematic acquisition of basic information but also extends to multiple dimensions such as uncovering patterns in phenomena and providing quantitative support for scientific decision-making. It plays an indispensable role in academic research, business surveys, social governance, and individual status assessment. However, traditional questionnaire collection methods often face the dual constraints of high time and labor costs. Not only is the data collection cycle long, but it is also limited by objective factors such as geographical coverage and the accessibility of the target population, making it difficult to efficiently collect large-scale samples. Although online questionnaire collection methods have, to some extent, broken through geographical limitations and significantly improved the efficiency of sample collection, this model is prone to data fraud problems such as "questionnaire padding" and "repeated completion," thereby reducing the reliability of the collected data and the dependability of subsequent analysis results.

[0003] With the widespread adoption of internet technology and the extensive penetration of social media platforms, users have generated and accumulated massive amounts of text data in cyberspace. This text encompasses diverse content such as sharing life experiences, expressing opinions, and venting emotions, providing a new, non-intrusive data collection channel for data research. If Natural Language Processing (NLP) technology can be used to automate the completion of questionnaires based on this large-scale user-generated text data, it will not only significantly improve the efficiency and reliability of questionnaire data collection and reduce the time and labor costs of traditional research methods, but also effectively promote the deep integration of Artificial Intelligence (AI) technology with fields such as mental health assessment and social research, providing innovative paradigms and technical support for cross-domain applications.

[0004] Applying natural language processing (NLP) models to automated questionnaire completion essentially involves using the model's understanding of text semantics to automate responses to questionnaire questions. Reading comprehension, a classic NLP task, aims to infer the correct answer to a given question based on the given text, question, and candidate answers. While this task shares similarities with automated psychological questionnaire completion (e.g., both require extracting factual information related to the question from the text), they differ fundamentally in their task paradigms. First, regarding input settings, reading comprehension tasks strictly assume the correct answer is embedded in the given text, thus avoiding the "information gap" problem; whereas in automated questionnaire completion, information gaps are a common challenge. Second, regarding option characteristics, reading comprehension options typically exhibit strict factual exclusivity, meaning that, apart from the correct answer, the remaining options are significantly inconsistent with or irrelevant to the textual facts. This requires the model to possess the ability to discern and logically reason about certain facts. In contrast, questionnaire options are often semantically highly correlated with textual facts; the difference lies primarily in the varying degrees of expression of the same fact, rather than mutually exclusive contradictions. Therefore, the core of the task is to rank or score the degree of correlation between the text and each option. Given the above differences, traditional reading comprehension models that focus on exclusive fact judgments are difficult to apply directly to automated questionnaire completion tasks because they lack the key ability to model degree relationships.

[0005] Generally, automated questionnaire completion tasks face several challenges: 1. Relevant information filtering. Users post a large number of messages online, not to answer questionnaire questions. Therefore, much of the content is irrelevant to the questionnaire. Effectively identifying the implicit information in these posts that reflects the questionnaire answers is crucial for accurate prediction. 2. Degree characterization. Unlike traditional reading comprehension which relies solely on factual information, questionnaire responses are more complex. Answers are not only rooted in relevant facts but also involve comparisons and judgments of degree. 3. Information gaps. The original intent of users generating massive amounts of text does not match the information requirements of the questionnaire prediction task, often resulting in text that fails to comprehensively and adequately address all questions. This inherent contradiction leads to a natural information gap in the user-generated text for some questions, constituting one of the core data bottlenecks for this type of task. 4. Balancing accuracy and complexity. Questionnaires typically have dozens of related questions. Designing a classifier for each question is the most basic approach; however, due to the information gap problem, some classifiers suffer significant performance degradation due to a lack of effective input. While using multi-task learning for joint training can alleviate the information gap problem by leveraging the correlation between problems, it may also trigger a negative transfer effect, impairing the accuracy of some sub-tasks. Furthermore, jointly training dozens of classifiers also incurs extremely high computational costs.

[0006] In summary, automated questionnaire completion represents an emerging task paradigm in the field of natural language processing. Its core objective is to automate the completion of standardized questionnaires based on user-generated natural language text. Compared to traditional factual reading comprehension, this task paradigm represents a key shift: its evaluation focus has moved from definitive factual matching between text and options to the assessment and analysis of degree-related relationships. Simultaneously, the task model must be able to handle partial information gaps and achieve an effective balance between overall state evaluation and fine-grained state analysis, which constitutes the core challenge of this task. Summary of the Invention

[0007] In view of the above situation, the main objective of this invention is to propose an automatic questionnaire filling method based on progressive divide-and-conquer and multi-question fusion collaboration, so as to solve the above-mentioned technical problems.

[0008] This invention proposes an automatic questionnaire completion method based on progressive divide-and-conquer and multi-question fusion collaboration, the method comprising the following steps:

[0009] Step 1: Given a questionnaire, obtain the questions in the questionnaire and the corresponding options for each question, build a dedicated professional model for each question, and configure an independent classification head for each professional model;

[0010] Step 2: Using the questions and corresponding options as query terms, search for each user's historical social text information to form a set of candidate clauses for the user;

[0011] Step 3: Concatenate each candidate clause in the user's candidate clause set with the corresponding question and each option of the question to obtain several candidate options for each question. Input the several candidate options for each question into the corresponding professional model and perform a one-stage training on the corresponding professional model and classification head to obtain several symptom-oriented expert models and optimized classification heads.

[0012] Step 4: Set up a symptom-aware attention layer. The symptom-aware attention layer outputs a text representation of the options for each question at a time.

[0013] Step 5: Freeze the parameters of several symptom-oriented expert models, use several symptom-oriented expert models to output the encoding results of all questions and all options, fuse them, and input them into the symptom perception attention layer. Perform two-stage training on the symptom perception attention layer and the optimized classification head. After training, the trained symptom perception attention layer and the trained classification head are obtained.

[0014] Step 6: Utilize the trained symptom-aware attention layer to output the option-aware text representation of the current question, and input the option-aware text representation of the current question into the trained classification head of the corresponding question to obtain the final prediction result of the current question.

[0015] This invention addresses the modeling challenges of complex tasks through a three-stage progressive design: First, it filters relevant information for each question in the questionnaire, extracting targeted information from massive amounts of user-generated text to lay a data foundation for subsequent modeling. Second, based on the "divide and conquer" approach, the automated questionnaire completion task is broken down into multiple independent sub-tasks, building a dedicated professional model for each question. This model focuses on learning the differences in expression between options, deeply mining the degree correlation between user text and each option, achieving precise matching of degree perception, thereby ensuring the singularity and accuracy of modeling for each sub-task. Finally, an innovative collaborative fusion device is designed. Its core function is not simply learning the "trust weight allocation of each professional model," but rather deeply mining and learning the potential dependencies and interaction patterns between professional models. By creatively fusing the representations of multiple professional models, it captures deep semantic connections existing at the intersection of questions that cannot be covered from a single question perspective.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] 1. Unlike classic reading comprehension models that focus on matching deterministic facts, this invention addresses the characteristic that questionnaire answers often reflect differences in degree. During the training phase of the specialized question model, it innovatively integrates a differential attention mechanism and an option-level comparison layer. The differential attention mechanism guides the model to focus on highly discriminative keywords (such as degree modifiers) between clauses and options, rather than broad semantic relationships; the option-level comparison layer explicitly models and reinforces the contrastive features between different options. Working together, these two mechanisms enable the model to accurately perceive and rank the complex relationships between text and options, thereby significantly improving answer prediction accuracy in scenarios with strong degree dependence, such as mental health assessments.

[0018] 2. Existing naive methods suffer from drastic performance drops when modeling each question independently due to a lack of relevant textual support for some questions; while simply using multi-task joint training is prone to negative transfer and computationally expensive. This invention first employs a divide-and-conquer strategy, training a dedicated "expert model" for each questionnaire question, focusing on learning the question-specific expressions from selected relevant clauses, ensuring optimal performance when information is sufficient. Subsequently, in the collaborative retraining phase, the latent semantic and structural associations between different questions are adaptively mined and integrated through symptom-aware attention mechanisms and a learnable task relevance matrix. This allows the model to reference clues and information from other relevant questions when answering a question, effectively compensating for information gaps from a single question perspective and improving the overall stability and accuracy of predictions.

[0019] 3. Existing multi-task joint training frameworks struggle to balance the relationships between tasks when there are numerous sub-tasks, easily leading to negative transfer and performance conflicts. This invention introduces a task selection gate mechanism during the collaborative retraining phase. This mechanism controls parameter optimization and output to be performed only on a single problem at a time during training, thus fully utilizing cross-problem correlation information for representation learning and fusion while strictly isolating gradient interference between output layers of different problems. This design enables models for dozens of problems to undergo collaborative enhancement training within a unified framework, achieving the perceptual dimension expansion benefits of multi-task learning while fundamentally avoiding the risk of negative transfer, realizing a synergistic effect of "1+1>2".

[0020] 4. Addressing the challenge of user-generated text topics being diverse and not directly matching questionnaire questions, this invention introduces a pre-trained model for mental health based on domain enhancement, with an information filtering stage. This stage uses questionnaire questions and options as queries, intelligently retrieving the most relevant set of clauses from the user's text. This not only provides clean, highly relevant input for subsequent professional models, significantly reducing noise interference, but also makes the model's decision-making process more traceable (based on the filtered clauses). Simultaneously, fine-tuning the general pre-trained model on a mental health domain dataset significantly improves its understanding of domain-specific semantics and expressions, laying a solid foundation for the effectiveness of the entire pipeline.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0022] Figure 1 This is a flowchart of the questionnaire auto-filling method based on progressive divide-and-conquer and multi-question fusion and collaboration proposed in this invention;

[0023] Figure 2This is an architecture diagram of the questionnaire auto-filling method based on progressive divide-and-conquer and multi-question fusion and collaboration proposed in this invention. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0025] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0026] Please see Figure 1 and Figure 2 This embodiment provides a method for automatically completing questionnaires based on progressive divide-and-conquer and multi-question fusion collaboration. The method includes the following steps:

[0027] Step 1: Given a questionnaire, obtain the questions in the questionnaire and the corresponding options for each question, build a dedicated professional model for each question, and configure an independent classification head for each professional model;

[0028] Step 2: Using the questions and corresponding options as query terms, search for each user's historical social text information to form a set of candidate clauses for the user;

[0029] In this step, since users post on social media platforms to share their lives and express their emotions, rather than to respond to specific standard questionnaires, the posts generated by users are not only diverse in topic but also contain a large amount of information irrelevant to the questionnaire questions. To eliminate the noise caused by a large amount of irrelevant information, this invention first filters out a set of clauses relevant to each question.

[0030] The core of relevance assessment lies in understanding the semantics of the text. However, there is a fundamental difference between user-generated text and the textual expression of standardized questionnaires: the questions and options in questionnaires are standardized and uniform, while the expression in user-generated text is more casual and diverse. Therefore, only by exploring the semantic expressions of these two types of text in the dimension of mental health can an effective framework for relevance assessment be established.

[0031] Therefore, this embodiment, based on data augmentation theory, introduces a mental health status recognition task to enhance the semantic understanding ability of the pre-trained language model BERT in the field of mental health. This invention uses the shared task dataset from CLPsych 2017 to fine-tune BERT. The goal of this task is to automatically identify the mental health status of forum posters. The BERT model fine-tuned for this task shows a significant improvement in text understanding in the field of mental health, and this paper names it the Mental-BERT pre-trained model for mental health.

[0032] To improve the screening effect, this invention uses all questions and options in BDI-II as query terms to filter candidate clause sets for specific questions. First, the text (questions and options) and all clauses in the questionnaire are fed into a pre-trained model for mental health for semantic encoding, obtaining their corresponding semantic vectors. Then, the cosine similarity between different query terms (questions and options) and the semantic vectors of each clause is calculated to obtain pairwise matching scores. Finally, for each query term, all clauses are ranked according to this score, and a Top-Ranked list is selected for each user. N Clauses, constructing a set of candidate clauses for the user. , ,in, , Indicates the user's corresponding number i Symptoms The former N A collection of notes and sentences. Represents the set of clauses Total number of clauses in Represents a set The first in n Items This indicates the total number of questions in the questionnaire.

[0033] Step 3: Concatenate each candidate clause in the user's candidate clause set with the corresponding question and each option of the question to obtain several candidate options for each question. Input the several candidate options for each question into the corresponding professional model and perform a one-stage training on the corresponding professional model and classification head to obtain several symptom-oriented expert models and optimized classification heads.

[0034] In this embodiment, the professional model comprises four layers: an input layer, a clause encoding layer, a multi-clause fusion layer, and an option-level comparison layer. The specific details of each layer and the training process of this professional model are as follows:

[0035] For the input layer:

[0036] This embodiment follows common practices in reading comprehension. Since each target clause needs to be concatenated with the corresponding question and all its candidate options, each clause will be expanded into... The following are examples of input samples:

[0037] ;

[0038] CLS and SEP represent special identifiers for BERT. Indicates the first n The input representation after concatenating the clause and the first candidate option. It is the first n Clause and Section The input representation after concatenating the candidate options. This indicates the number of the BDI-II scale. i One question, yes All corresponding candidate options, express The first corresponding candidate option, express The corresponding number (1 candidate options) This represents text concatenation. These input samples are fed into a pre-trained model for mental health to obtain initial encodings, i.e., for the question... i candidate subset Obtain its initial encoding , .in, This indicates the total number of candidate options for the symptom question. Indicates the first n An initial encoded vector, yes The initial vector matrix of all corresponding clauses.

[0039] For the clause encoding layer:

[0040] This invention takes the initial encoding result as input and outputs it through a multi-layer Transformer clause encoder based on differential attention. The aim is to uncover the correlation between each clause and the responses of different options. Since the expressions in the responses to the questionnaire questions exhibit a high degree of semantic overlap, with differences primarily reflected in degree rhetoric, when applying an attention mechanism to help the model identify the best relevance, this mechanism should guide the model to focus on a few highly discriminative words in the candidate clauses and options, avoiding distraction by semantically broadly related but low-discriminative words. Therefore, this invention introduces a differential attention mechanism, designed to concentrate attention weights on certain key features rather than dispersing them across a large number of related words.

[0041] This invention first projects the initial encoding result to obtain the corresponding first query vector, first key vector, and first value vector. The corresponding process has the following relationship:

[0042] ;

[0043] in, This means splitting the first query vector in half along its feature dimension to obtain two queries. This means splitting the first key vector in half along the feature dimension to obtain two keys. Represents the first value vector. These represent the learnable parameters corresponding to the first query vector, the first key vector, and the first value vector, respectively. This represents the initial encoding result. This indicates vector concatenation;

[0044] Differential attention is calculated on the first query vector, the first key vector, and the first value vector to obtain the differential attention output. The corresponding process has the following relationship:

[0045] ;

[0046] in, This represents the enhanced encoding representation after differential attention mechanism. express The size of the vector dimension, This indicates the differential attention computation operation. The trainable parameters representing differential attention. This indicates the transpose operation. This represents the normalized exponential function. It's worth noting that, to achieve the goals of multi-dimensional feature capture and diverse relationship modeling, the differential attention mechanism retains a multi-head architecture in its design.

[0047] For multi-clause fusion layers:

[0048] For each question, different users will select... NThese clauses are used to predict the response to the question. Although these clauses may describe the response to the question from multiple perspectives, it is necessary to ensure... N The prediction logic of each clause is consistent, that is, by comprehensively considering the core information of each clause, so that... N Each clause forms a consensus expression of the response to the question. Therefore, this invention... The input is fed into a clause fusion processor based on a single-layer Transformer. This processor captures the features inherent in different clauses and the intrinsic relationships between them, thereby completing the fusion of the encoding vectors of each independent clause. Specifically, for different options, this fusion processor can construct... N The expressions of each clause for this option are aggregated into a unified, comprehensive vector, and the corresponding process has the following relation:

[0049] ;

[0050] in, The result is represented by the fusion of multiple clauses representing question i.

[0051] For option-level comparison layers:

[0052] Unlike traditional factual reading comprehension, in the context of answering psychological questionnaire questions, users typically compare and analyze multiple candidate options, ultimately selecting the option that best matches their actual situation as their response. Therefore, based on fully exploring the factual correlations between each clause and different option responses, this invention introduces an option-level comparison layer to further capture the comparative features between different options, thereby improving prediction accuracy.

[0053] First, the relationship between the target option and other options is constructed by calculating the attention weights among the options. To eliminate the attention of an option to itself, a mask matrix is ​​introduced. , The corresponding process has the following relationship:

[0054] ;

[0055] ;

[0056] ;

[0057] in, express The Middle f Vector representation of each candidate option express The Middle g Vector representation of each candidate option express and The original correlation score, Represents the mask matrix of the first f OK g The element values ​​of the column, This represents the correlation score after adjustment using the mask matrix. Representing vectors For vectors Option-level attention weights This indicates two different candidate options. This indicates the number of candidate options.

[0058] After inclusion, the information from other options is incorporated into the target option using a weighted summation method based on option-level attention weights, resulting in an enhanced option vector. The corresponding process has the following relationship:

[0059] ;

[0060] in, Represents the enhanced option vector;

[0061] The average vector difference between the target option and the remaining options is calculated as the option-level contrast feature. The enhanced option vector is concatenated with the option-level contrast feature to obtain the option-level option-specific vector representation of the current problem. The corresponding process has the following relationship:

[0062] ;

[0063] ;

[0064] in, Indicates options f The average difference between the options other than itself. This represents the specificity vector of the f-th candidate option, and several specificity vectors of candidate options constitute the current problem. i Option-specific vector representation matrix , .

[0065] Finally, let's address the previous question. i The option-specific vector representation matrix is ​​fed into a classification head consisting of multiple linear layers for final decision-making (e.g., prediction problems). i (What is the final answer?) Each symptom question's classifier is trained independently, using cross-entropy as the loss function during training. After training, an expert-level model oriented towards symptoms can be obtained. The formula for the first cross-entropy loss function is as follows:

[0066] ;

[0067] in, This represents the first cross-entropy loss function. This indicates the number of users participating in the training. Indicates user k Regarding the question i The true label, Indicates user k Regarding the question i The prediction results.

[0068] Step 4: Set up a symptom-aware attention layer. The symptom-aware attention layer outputs a text representation of the options for each question at a time.

[0069] Step 5: Freeze the parameters of several symptom-oriented expert models, use several symptom-oriented expert models to output the encoding results of all questions and all options, fuse them, and input them into the symptom perception attention layer. Perform two-stage training on the symptom perception attention layer and the optimized classification head. After training, the trained symptom perception attention layer and the trained classification head are obtained.

[0070] Because independent symptom question classifiers have limited perceptual scope—specifically, each classifier can only focus on text clauses directly related to its specific symptom question—and different symptom questions in psychological questionnaires have potential correlations, this invention proposes a method to expand the perceptual dimension of independent symptom classifiers by introducing other questions related to the target symptom question, thereby improving the prediction accuracy of responses to the target symptom question. Therefore, this invention performs collaborative fusion retraining on a pre-trained expert-level symptom-oriented classifier.

[0071] The potential correlation between symptom-related issues is manifested at both the content structure and content expression levels. The structural correlation stems from the design logic of psychological questionnaires—questionnaires typically design multiple questions to examine a respondent's psychological traits in a certain dimension, and these questions naturally possess structural correlation. The content expression-level correlation, on the other hand, is determined by the characteristics of user-generated text; that is, the information presented by users in the text can create correlations between certain questions, resulting in strong individual differences in content expression-level correlation.

[0072] To address this, this invention constructs an attention mechanism for symptom perception with a task relevance matrix. In this mechanism, the core of self-attention computation lies in quantifying the degree of correlation between symptom questions based on content representation features; simultaneously, a learnable task relevance matrix is ​​introduced. , The initial value is the identity matrix, which effectively guides the model to learn specific structural dependencies.

[0073] Effective learning of the potential dependencies among all symptom problems requires collaborative training of all problems, and multi-task joint training frameworks are currently the mainstream technical solution for handling this type of problem. Multi-task joint training aims to expand the perceptual dimension of each sub-task through the collaborative learning of multiple related sub-tasks, thereby improving the performance of individual sub-tasks. However, due to the constraint of negative transfer effect, existing multi-task joint training methods can usually only integrate a small number of highly related tasks, a limitation that makes it difficult to meet the requirement of the number of sub-tasks in this embodiment.

[0074] Therefore, during training, a task selection gate is used (its main function is to control the output of only the predicted label of a single question during retraining). Optimization training is performed on only a single symptom question at a time (i.e., the task selection gate ensures that the third-stage training outputs only the predicted value of a single symptom question each time), avoiding negative transfer. It should be noted that when performing symptom fusion collaborative retraining, the optimization of these parameters is limited to the multi-question fusion collaborative retraining stage. The specific steps of multi-question fusion collaborative retraining are as follows:

[0075] The parameters of several symptom-oriented expert models are frozen, and the option-specific vector representation matrix of all questions is output using several symptom-oriented expert models.

[0076] Based on the number of questions, a task relevance matrix is ​​set, and the identity matrix is ​​used as the initial value of the task relevance matrix to obtain a learnable task relevance matrix.

[0077] The option-specific vector representation matrices of all questions are input into the symptom-perception attention layer. An attention mechanism is used to fuse the option-specific vector representations of different questions. During the fusion process, a learnable task relevance matrix guides the model to learn specific structural dependencies, resulting in an enhanced option-specific vector representation for each question. The corresponding process follows the following relationship:

[0078] ;

[0079] ;

[0080] in, This represents the enhanced option-specific vector representation for each question. The option-specific vector representation matrix represents all questions. These represent the second query vector, the second key vector, and the second value vector, respectively. These represent the learnable parameters corresponding to the second query vector, the second key vector, and the second value vector, respectively.

[0081] Select one of the enhanced option-specific vector representations corresponding to each question and input it into the multi-layer linear layer of the corresponding symptom-oriented expert model to obtain the prediction result of the second current question;

[0082] Using the prediction results of the second current problem, a second cross-entropy loss function is constructed. This second cross-entropy loss function is then used to train the symptom perception attention layer. After training, the trained symptom perception attention layer is obtained. In this embodiment, the second cross-entropy loss function has the same form as the first cross-entropy loss function, and will not be described in detail here.

[0083] Step 6: Utilize the trained symptom-aware attention layer to output the option-aware text representation of the current question, and input the option-aware text representation of the current question into the symptom-oriented expert model of the corresponding question to obtain the final prediction result of the current question.

[0084] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0085] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0086] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0087] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for automatically completing questionnaires based on progressive divide-and-conquer and multi-question fusion collaboration, characterized in that, The method includes the following steps: Step 1: Given a questionnaire, obtain the questions in the questionnaire and the corresponding options for each question, build a dedicated professional model for each question, and configure an independent classification head for each professional model; Step 2: Using the questions and corresponding options as query terms, search for each user's historical social text information to form a set of candidate clauses for the user; Step 3: Concatenate each candidate clause in the user's candidate clause set with the corresponding question and each option of the question to obtain several candidate options for each question. Input the several candidate options for each question into the corresponding professional model and perform a one-stage training on the corresponding professional model and classification head to obtain several symptom-oriented expert models and optimized classification heads. Step 4: Set up a symptom-aware attention layer. The symptom-aware attention layer outputs a text representation of the options for each question at a time. Step 5: Freeze the parameters of several symptom-oriented expert models, use several symptom-oriented expert models to output the encoding results of all questions and all options, fuse them, and input them into the symptom perception attention layer. Perform two-stage training on the symptom perception attention layer and the optimized classification head. After training, the trained symptom perception attention layer and the trained classification head are obtained. Step 6: Utilize the trained symptom-aware attention layer to output the option-aware text representation of the current question, and input the option-aware text representation of the current question into the trained classification head of the corresponding question to obtain the final prediction result of the current question.

2. The questionnaire auto-filling method based on progressive divide-and-conquer and multi-question fusion and collaboration as described in claim 1, characterized in that, In step 3, several candidate options corresponding to each question are input into the corresponding professional model, and the corresponding professional model is independently trained for one stage to obtain several symptom-oriented expert models. The specific steps include the following: For each question, several candidate options are input into the corresponding professional model, and then encoded using a pre-trained model for mental health to obtain several initial encoding results for each question. Several initial encoding results for each question are input into a multi-layer Transformer clause encoder based on differential attention to perform multi-clause fusion and obtain the multi-clause fusion result for each question. Calculate the correlation between the target option and other options in the multi-clause fusion result for each question to obtain the option-level attention weight; By using option-level attention weights, information from other options is incorporated into the target option in a weighted summation manner to obtain an enhanced option vector; Calculate the average vector difference between the target option and the remaining options as the comparative feature at the option level; The enhanced option vector is concatenated with the option-level contrast features to obtain the option-level option-specific vector representation of the current problem; several option-level option-specific vector representations of the current problem are combined to form the option-specific vector representation matrix of the current problem. The option-specific vector representation matrix of the current problem is input into a classification head composed of multiple linear layers for final decision-making, and the prediction result of the first current problem is obtained. Based on the prediction results of the first current problem, a first cross-entropy loss function is constructed, and the corresponding professional model is trained using the first cross-entropy loss function. After training, an expert-level model for symptoms is obtained. Repeat the above steps to explore all the questions and obtain several expert-level models oriented towards symptoms.

3. The questionnaire auto-filling method based on progressive divide-and-conquer and multi-question fusion and collaboration as described in claim 2, characterized in that, Several initial encoding results for each question are input into a multi-layer Transformer clause encoder based on differential attention to perform multi-clause fusion, obtaining the multi-clause fusion result for each question. The specific steps include the following: The initial encoding result is projected to obtain the corresponding first query vector, first key vector, and first value vector. The corresponding process has the following relationship: ； in, This means splitting the first query vector in half along its feature dimension to obtain two queries. This means splitting the first key vector in half along the feature dimension to obtain two keys. Represents the first value vector. These represent the learnable parameters corresponding to the first query vector, the first key vector, and the first value vector, respectively. This represents the initial encoding result. This indicates vector concatenation; Differential attention is calculated on the first query vector, the first key vector, and the first value vector to obtain the differential attention output. The corresponding process has the following relationship: ； in, This represents the encoded representation after differential attention enhancement. express Dimensions This indicates the differential attention computation operation. The trainable parameters representing differential attention. This indicates the transpose operation. This represents the normalized exponential function; Will The input is fed into a clause fusion processor based on a single-layer Transformer to capture the features contained in different clauses and the inherent relationships between clauses, thereby completing the fusion processing of the encoding vectors of each independent clause and obtaining the multi-clause fusion result for each problem. The corresponding process has the following relationship: ； in, Indicate the problem i The result of merging multiple clauses.

4. The questionnaire auto-filling method based on progressive divide-and-conquer and multi-question fusion and collaboration as described in claim 2, characterized in that, The calculation process of the trainable parameters of differential attention follows the following relationship: ； in, This represents four different trainable parameters. Indicates hyperparameter constants. .

5. The questionnaire auto-filling method based on progressive divide-and-conquer and multi-question fusion and collaboration as described in claim 2, characterized in that, Calculate the association between the target option and other options in the multi-clause fusion result for each question to obtain the option-level attention weights. The corresponding process has the following relationship: ； ； ； in, express The Middle f Vector representation of each candidate option express The Middle g Vector representation of each candidate option express and The original correlation score, Represents the mask matrix of the first f OK g The element values ​​of the column, This represents the correlation score after adjustment using the mask matrix. Representing vectors For vectors Option-level attention weights This indicates two different candidate options. This indicates the number of candidate options.

6. The questionnaire auto-filling method based on progressive divide-and-conquer and multi-question fusion and collaboration as described in claim 5, is characterized in that, By using option-level attention weights to incorporate information from other options into the target option through a weighted summation, an enhanced option vector is obtained. The corresponding process is governed by the following relationship: ； in, This represents the enhanced option vector.

7. The questionnaire auto-filling method based on progressive divide-and-conquer and multi-question fusion and collaboration as described in claim 6, characterized in that, The average vector difference between the target option and the remaining options is calculated as the option-level contrast feature. The enhanced option vector is concatenated with the option-level contrast feature to obtain the option-level option-specific vector representation of the current problem. The corresponding process has the following relationship: ； ； in, Indicates options f The average difference between the options other than itself. This represents the option-specific vector representation matrix for the current problem i, formed by several candidate option-specific vector representations. , .

8. The questionnaire auto-filling method based on progressive divide-and-conquer and multi-question fusion and collaboration as described in claim 7, is characterized in that, Based on the prediction results of the first current problem, the first cross-entropy loss function is constructed, and the corresponding process has the following relationship: ； in, This represents the first cross-entropy loss function. This indicates the number of users participating in the training. Indicates user k Regarding the question i The true label, Indicates user k Regarding the question i The prediction results.

9. The questionnaire auto-filling method based on progressive divide-and-conquer and multi-question fusion and collaboration as described in claim 8, characterized in that, In step 5, the parameters of several symptom-oriented expert models are frozen, and the encoded results of all questions and all options are output from several symptom-oriented expert models, fused, and input into the symptom-aware attention layer. The symptom-aware attention layer is then trained in two stages. After training, the trained symptom-aware attention layer is obtained, which specifically includes the following steps: The parameters of several symptom-oriented expert models are frozen, and the option-specific vector representation matrix of all questions is output using several symptom-oriented expert models. Based on the number of questions, a task relevance matrix is ​​set, and the identity matrix is ​​used as the initial value of the task relevance matrix to obtain a learnable task relevance matrix. The option-specific vector representation matrix of all questions is input into the symptom perception attention layer. The option-specific vector representations of different questions are fused using the attention mechanism. During the fusion process, the learnable task relevance matrix is ​​used to guide the model to learn the structural dependencies, resulting in an enhanced option-specific vector representation for each question. Select one of the enhanced option-specific vector representations corresponding to each question and input it into the multi-layer linear layer of the corresponding symptom-oriented expert model to obtain the prediction result of the second current question; Using the prediction results of the second current problem, a second cross-entropy loss function is constructed. The symptom perception attention layer is trained using the second cross-entropy loss function. After training, the trained symptom perception attention layer is obtained.

10. The questionnaire auto-filling method based on progressive divide-and-conquer and multi-question fusion and collaboration as described in claim 9, characterized in that, The option-specific vector representation matrices of all questions are input into the symptom-perception attention layer. An attention mechanism is used to fuse the option-specific vector representations of different questions. During the fusion process, a learnable task relevance matrix guides the model to learn structural dependencies, resulting in an enhanced option-specific vector representation for each question. The corresponding process follows the following relationship: ； ； in, This represents the enhanced option-specific vector representation for each question. The option-specific vector representation matrix represents all questions. These represent the second query vector, the second key vector, and the second value vector, respectively. These represent the learnable parameters corresponding to the second query vector, the second key vector, and the second value vector, respectively. This represents the learnable task relevance matrix.

Images