Large-model psychological diagnosis and treatment information analysis enhancement method based on lexical diversity

Abstract

The invention discloses a large-model psychological diagnosis and treatment information analysis enhancement method based on lexical diversity, which comprises the following steps: S1, receiving cross-space-time medical record analysis query input, and retrieving k documents most similar to the query from a document library; s2, analyzing the medical record analysis query statement, extracting key components, and classifying disease query components; s3, the correlation between each component and the retrieval document is evaluated, and then the correlation between the query document and the retrieval document is determined; s4, taking r retrieval documents with the highest correlation as evidence texts, and taking the retrieval document with the lowest correlation as a noise text; and S5, comparing and decoding the evidence text and the noise text through a large language model to obtain an information analysis result. The method can be applied to a medical cross-space-time alignment prediction application scene, the accuracy of cross-space-time medical retrieval is remarkably improved, the perception ability for patient query is enhanced, noise data is effectively processed, the quality of retrieval results is improved, and the method has high practical application value.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing, and specifically relates to a large-model type psychiatric and psychological diagnosis and treatment information analysis enhancement method based on lexical diversity. Background Art

[0002] With the rapid development of large language models, their applications are becoming more and more extensive in various fields, such as conversational assistants, medical analysis, and code generation. However, large language models often suffer from factual errors, outdated information, and low interpretability when dealing with open-domain knowledge tasks. The retrieval-augmented generation method proposed by Lewis et al. improves the generation quality of large language models by integrating external retrieval knowledge. The retrieval-augmented generation method can retrieve and ask questions related knowledge from vector data areas, and significantly improves the performance in open-domain knowledge tasks by incorporating this knowledge into the input of the large model.

[0003] Retrieval-augmented generation methods have shown remarkable effectiveness in knowledge-intensive tasks and have attracted widespread attention in recent years. Existing techniques mainly focus on enhancing the ability of models to retrieve and utilize external knowledge through training. Asai et al. focused on improving retrieval accuracy by training large language models to evaluate the relevance of retrieved documents to the input query. Lin et al. fine-tuned the large language model and retriever, improving their performance in both retrieval and generation tasks. Other methods, such as Xu et al. and Kim et al., aim to summarize the collected documents to reduce the impact of irrelevant content in the documents on the generation process.

[0004] Although the above methods evaluate the relevance of documents and inputs and mitigate the adverse effects of irrelevant content on generation based on document summaries, the relevance evaluations relied on by these methods are usually rough and fail to consider the vocabulary diversity of fine-grained query components. In addition, inaccurate summary compression often leads to the loss of key information. These defects may affect their effectiveness in fine-grained retrieval and generation. Therefore, these methods need to be improved in terms of granular relevance evaluation criteria and comprehensive retrieval and effective use of relevant knowledge. Summary of the invention

[0005] In response to the above-mentioned deficiencies in the prior art, the large-model psychiatric and psychological diagnosis and treatment information analysis enhancement method based on lexical diversity provided by the present invention solves the problem that traditional retrieval methods are difficult to accurately extract key information in the query, resulting in low relevance of the retrieval results and affecting the accuracy of the subsequent analysis process.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a large-model psychiatric diagnosis and treatment information analysis enhancement method based on lexical diversity, comprising the following steps:

[0007] S1, receiving the cross-temporal and spatial medical record analysis query input, using the retriever to retrieve the k documents most similar to the medical record analysis query from the document library according to the embedding similarity;

[0008] S2, guiding the analyzer to conduct in-depth analysis on the medical record analysis query statement, extracting the key components in the query statement, and classifying each disease query component according to the diversity of word meanings;

[0009] S3, set fine-grained evaluation criteria for different types of components, and then evaluate the relevance between each component and the retrieved document, and combine all components to characterize the relevance of the complete medical record analysis query and the retrieved document;

[0010] S4, take the r most relevant retrieval documents as evidence texts, and the least relevant retrieval documents as noise texts;

[0011] S5. Compare and decode the evidence text and the noise text through a large language model to reduce the impact of irrelevant content in the text and obtain information analysis results of cross-temporal and spatial mental health diagnosis and treatment records.

[0012] Furthermore, the step S1 includes the following sub-steps:

[0013] S11, receiving cross-temporal and spatial medical record analysis query input, and using embedded technology to convert the input medical record documents and documents in the document library into high-dimensional vectors;

[0014] S12. Compare the high-dimensional vectors of the medical record document and the document in the document library by cosine similarity, and select the k documents with the highest cosine similarity as the retrieved most similar documents.

[0015] Furthermore, the step S2 includes the following sub-steps:

[0016] S21, setting category set A;

[0017] A={ <invariant>,〈Variant>,〈Supplementary>}

[0018] In the formula, <invariant>Indicates the component categories that are directly extracted from the medical record analysis query statement and cannot be expressed with synonyms; <variant>It indicates the component categories extracted directly from the medical record query sentence and expressed by synonyms; <supplementary>represents the potential component categories derived from the information of the medical record query statement;

[0019] S22. Use the constructed query statement data set to fine-tune the large prediction model as the analysis model, and then extract the key components in the medical record query analysis statement, and divide the extracted components into <invariant>and <variant>kind;

[0020] S23. Construct a supplementary component dataset and use it to fine-tune the analysis model, guiding the fine-tuned analysis model to generate <supplementary>Class components.

[0021] Furthermore, the step S3 includes the following sub-steps:

[0022] S31, determining the relevance scores between different types of components in the medical record analysis query and the retrieved documents;

[0023] S32. Accumulate the correlation scores between the components in the medical record analysis query and the documents as the correlation scores between the case analysis query and the retrieved documents.

[0024] Furthermore, in step S31:

[0025] When the ingredients are <invariant>When the component is explicitly mentioned in the search document, its corresponding relevance score is 1, otherwise it is 0;

[0026] When the ingredients are <variant>Class and <supplementary>Class, use the big prediction model to score, get the correlation between the retrieved document and the query statement, and generate a correlation score between 0 and 1.

[0027] Furthermore, the step S5 includes the following sub-steps:

[0028] S51. Construct a retrieval enhancement generation large prediction model based on evidence text;

[0029] S42. Use noise text to simulate irrelevant noise in evidence text, generate a large prediction model based on retrieval enhancement, and perform comparative decoding on evidence text and noise text to eliminate the influence of irrelevant noise on the generation of retrieval enhanced text, and obtain information analysis results of cross-temporal and spatial psychiatric diagnosis and treatment medical records.

[0030] Furthermore, in step S51, the retrieval enhancement generation large prediction model is expressed as:

[0031]

[0032] In the formula, y represents the model output, x represents the input query, and t + represents the knowledge related to the query in the evidence text, t - represents the noise in the evidence text that is irrelevant to the query, is the large language model parameter.

[0033] Furthermore, in step S52, the formula for comparison decoding is specifically:

[0034]

[0035] In the formula, γ represents the standard weight, t ref Represents noisy text.

[0036] The beneficial effects of the present invention are:

[0037] (1) The present invention can be applied to medical cross-temporal and spatial alignment prediction application scenarios, significantly improving the accuracy of cross-temporal and spatial medical retrieval, enhancing the perception of patient queries, and effectively processing noise data to improve the quality of retrieval results, thus having high practical application value.

[0038] (2) The method of the present invention is based on the principle of lexical diversity perception and uses multi-level semantic analysis and contrast decoding technology to significantly improve retrieval relevance and optimize generation quality.

[0039] (3) The present invention can be applied to tasks such as medical cross-temporal and spatial alignment prediction. The setting of the present invention is close to the real scene, has high practical application value, and has achieved excellent performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A flow chart of the large-model psychiatric diagnosis and treatment information analysis method based on lexical diversity provided by the present invention. DETAILED DESCRIPTION

[0041] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0042] The embodiment of the present invention provides a method for enhancing analysis of mental health diagnosis and treatment information based on a large model of lexical diversity, such as Figure 1 As shown, the following steps are included:

[0043] S1, receiving the cross-temporal and spatial medical record analysis query input, using the retriever to retrieve the k documents most similar to the medical record analysis query from the document library according to the embedding similarity;

[0044] S2, guiding the analyzer to conduct in-depth analysis on the medical record analysis query statement, extracting the key components in the query statement, and classifying each disease query component according to the diversity of word meanings;

[0045] S3, set fine-grained evaluation criteria for different types of components, and then evaluate the relevance between each component and the retrieved document, and combine all components to characterize the relevance of the complete medical record analysis query and the retrieved document;

[0046] S4, take the r most relevant retrieval documents as evidence texts, and the least relevant retrieval documents as noise texts;

[0047] S5. Compare and decode the evidence text and the noise text through a large language model to reduce the impact of irrelevant content in the text and obtain information analysis results of cross-temporal and spatial mental health diagnosis and treatment records.

[0048] Step S1 of the embodiment of the present invention includes the following sub-steps:

[0049] S11, receiving cross-temporal and spatial medical record analysis query input, and using embedded technology to convert the input medical record documents and documents in the document library into high-dimensional vectors;

[0050] In a specific example of the present invention, the description of symptoms in the cross-temporal and spatial medical records and the documents in the Wikipedia document library are converted into high-dimensional vectors;

[0051] S12. Compare the high-dimensional vectors of the medical record document and the document in the document library by cosine similarity, and select the k documents with the highest cosine similarity as the retrieved most similar documents.

[0052] Step S2 of the embodiment of the present invention includes the following sub-steps:

[0053] S21, setting category set A;

[0054] A={(Invariant>,〈Variant>, <supplementary>}

[0055] In the formula, <invariant>Indicates the component category that is directly extracted from the medical record analysis query statement and cannot be expressed by synonyms; <Variant> indicates the component category that is directly extracted from the medical record query statement and expressed by synonyms; <supplementary>represents the potential component categories derived from the information of the medical record query statement;

[0056] Specifically, <invariant>Classes, such as specific symptoms or disease names (such as depression and anxiety), are often clearly directed to the query; <variant>Class, with high synonymy (such as low mood or depression), this kind of components has a greater impact on the search results and needs to be handled flexibly; <supplementary>This type of component is not explicitly present in the query, but is a potential component derived from the information in the query. For example, a supplementary component related to "psychological therapy" may be derived from the query "anxiety disorder". This type of component is used to supplement and strengthen the evaluation of relevance.

[0057] S22. Use the constructed query statement data set to fine-tune the large prediction model as the analysis model, and then extract the key components in the medical record query analysis statement, and divide the extracted components into <invariant>and the <Variant> class;

[0058] S23. Construct a supplementary component dataset and use it to fine-tune the analysis model, guiding the fine-tuned analysis model to generate components of the <Supplementary> class according to the medical record query analysis statement.

[0059] Step S3 of the embodiment of the present invention includes the following sub-steps:

[0060] S31, determining the relevance scores between different types of components in the medical record analysis query and the retrieved documents;

[0061] Specifically, when a component is of the <Invariant> class, if the component is explicitly mentioned in the search document, its corresponding relevance score is 1, otherwise it is 0;

[0062] When the components are of the <Variant> class and the <Supplementary> class, the large prediction model is used for scoring to obtain the degree of association between the retrieved document and the query statement, generating a relevance score between 0 and 1.

[0063] S32. Accumulate the correlation scores between the components in the medical record analysis query and the documents as the correlation scores between the case analysis query and the retrieved documents.

[0064] Step S5 of the embodiment of the present invention includes the following sub-steps:

[0065] S51. Construct a retrieval enhancement generation large prediction model based on evidence text;

[0066] Among them, the retrieval enhancement generation model is expressed as:

[0067]

[0068] In the formula, y represents the model output, x represents the input query, and t + represents the knowledge related to the query in the evidence text, t - represents the noise in the evidence text that is irrelevant to the query, is the large language model parameter;

[0069] S42. Use noise text to simulate irrelevant noise in evidence text, generate a large prediction model based on retrieval enhancement, and perform comparative decoding on evidence text and noise text to eliminate the influence of irrelevant noise on the generation of retrieval enhanced text, and obtain information analysis results of cross-temporal and spatial psychiatric diagnosis and treatment medical records.

[0070] Among them, the formula for contrast decoding is specifically:

[0071]

[0072] In the formula, γ represents the standard weight, t ref Represents noisy text.

[0073] In the embodiment of the present invention, examples of the effects of using the method of the present invention and the traditional retrieval enhancement method on the USMLE-mental dataset cross-temporal and spatial medical record analysis task are as follows:

[0074] Patient's description of symptoms across time and space:

[0075] A38-year-old man comes to the physician because of persistent sadness and difficulty concentrating for the past 6 weeks. During this period, he has also had difficulty sleeping. He adds that he has been "feeling down" most of the time since his girlfriend broke up with him 4 years ago. Since then, he has only had a few periods of time when he did not feel that way, but none of these lasted for more than a month. He reports having no problems with appetite, weight, or energy.He does not use illicit drugs or alcohol.Mentalstatus examination shows a depressed mood and constricted affect.Which of the following is the most likely diagnosis?

[0076] The results of the present invention in this embodiment:

[0077] Cyclothymic disorder

[0078] 解释:患者在过去6周内一直有持续的抑郁症状,并且也经历过没有感到抑郁的时间段,但这些都没有持续超过一个月。他报告说食欲、体重或精力没有问题。他不使用非法药物或酒精。精神状态检查显示情绪低落和情感受限。检索到的关于恶劣心境的文本描述了一种情绪障碍,其认知和身体问题与抑郁症相同,但症状持续时间更长,这与患者的症状一致。文本还指出,在《精神疾病诊断与统计手册》第5版中,恶劣心境被持续性抑郁障碍(PDD)所取代,并且患者可能在被诊断之前多年都有症状,这也与患者的情况一致。其他选项,如环性心境障碍、适应障碍、混合情感状态和重度抑郁症,均与患者的症状和病史不符。

[0079] Results of traditional retrieval enhancement methods:

[0080] Major depressive disorder

[0081] Explanation:The patient's symptoms,such as persistent sadness,difficulty concentrating,and difficulty sleeping,are consistent with majordepressive disorder.The patient's history of recurring episodes ofdepression,with periods of time where he did not feel depressed,also supportsthis diagnosis.Additionally,the patient's mental status examination reveals adepressed mood and constricted affect,which is also consistent with majordepressive disorder.The patient's symptoms do not suggest the presence ofmania,hypomania,or mixed affective states,which are characteristic of bipolardisorder.The patient's symptoms are also not severe enough to be consideredpersistent depressive disorder,and the patient's history does not suggest arecent stressor that would be consistent with adjustment disorder.

[0082] In this embodiment, the traditional method improperly analyzes and interprets the symptoms, resulting in an erroneous analysis. The algorithm provided by the present invention can not only make a correct medical record analysis, but also make a reasonable analysis of cross-temporal and spatial symptoms, thereby generating a higher quality cross-temporal and spatial medical record analysis content than the traditional method.

[0083] Based on the principle of lexical diversity perception, the present invention uses multi-level semantic analysis and contrast decoding technology to significantly improve retrieval relevance and optimize generation quality. The present invention can be applied to tasks such as medical cross-temporal and spatial alignment prediction. The setting of the present invention is close to the real scene, has high practical application value, and has achieved excellent performance.

[0084] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0085] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.< / invariant> < / supplementary> < / variant> < / invariant> < / supplementary> < / invariant> < / supplementary> < / supplementary> < / variant> < / invariant> < / supplementary> < / variant> < / invariant> < / supplementary> < / variant> < / invariant> < / invariant>

Claims

1. A large-model psychiatric diagnosis and treatment information analysis enhancement method based on lexical diversity, characterized in that: The following steps are involved: S1, receiving the cross-temporal and spatial medical record analysis query input, using the retriever to retrieve the k documents most similar to the medical record analysis query from the document library according to the embedding similarity; S2, guiding the analyzer to conduct in-depth analysis on the medical record analysis query statement, extracting the key components in the query statement, and classifying each disease query component according to the diversity of word meanings; S3, set fine-grained evaluation criteria for different types of components, and then evaluate the relevance between each component and the retrieved document, and combine all components to characterize the relevance of the complete medical record analysis query and the retrieved document; S4, take the r most relevant retrieval documents as evidence texts, and the least relevant retrieval documents as noise texts; S5. Compare and decode the evidence text and the noise text through a large language model to reduce the impact of irrelevant content in the text and obtain information analysis results of cross-temporal and spatial mental health diagnosis and treatment records.

2. The large-model psychiatric diagnosis and treatment information analysis enhancement method based on lexical diversity according to claim 1 is characterized in that: The step S1 comprises the following sub-steps: S11, receiving cross-temporal and spatial medical record analysis query input, and using embedded technology to convert the input medical record documents and documents in the document library into high-dimensional vectors; S12. Compare the high-dimensional vectors of the medical record document and the document in the document library by cosine similarity, and select the k documents with the highest cosine similarity as the retrieved most similar documents.

3. The large-model psychiatric diagnosis and treatment information analysis enhancement method based on lexical diversity according to claim 1 is characterized in that: The step S2 comprises the following sub-steps: S21, setting category set A; A＝{ <invariant> , <variant> , <supplementary>}< / supplementary> < / variant> < / invariant> In the formula, <invariant>Indicates the component categories that are directly extracted from the medical record analysis query statement and cannot be expressed with synonyms; <variant>It indicates the component categories extracted directly from the medical record query sentence and expressed by synonyms; <supplementary> represents the potential component categories derived from the information of the medical record query statement;< / supplementary> < / variant> < / invariant> S22. Use the constructed query statement data set to fine-tune the large prediction model as the analysis model, and then extract the key components in the medical record query analysis statement, and divide the extracted components into <invariant>and <variant> kind;< / variant> < / invariant> S23. Construct a supplementary component dataset and use it to fine-tune the analysis model, guiding the fine-tuned analysis model to generate <supplementary> Class components.< / supplementary> 4. The large-model psychiatric diagnosis and treatment information analysis enhancement method based on lexical diversity according to claim 3 is characterized in that: The step S3 comprises the following sub-steps: S31, determining the relevance scores between different types of components in the medical record analysis query and the retrieved documents; S32. Accumulate the correlation scores between the components in the medical record analysis query and the documents as the correlation scores between the case analysis query and the retrieved documents.

5. The large-model psychiatric diagnosis and treatment information analysis enhancement method based on lexical diversity according to claim 4 is characterized in that: In step S31: When the ingredients are <invariant> When the component is explicitly mentioned in the search document, its corresponding relevance score is 1, otherwise it is 0;< / invariant> When the ingredients are <variant>Class and <supplementary> Class, use the big prediction model to score, get the correlation between the retrieved document and the query statement, and generate a correlation score between 0 and 1.< / supplementary> < / variant> 6. The large-model psychiatric diagnosis and treatment information analysis enhancement method based on lexical diversity according to claim 1 is characterized in that: The step S5 comprises the following sub-steps: S51. Construct a retrieval enhancement generation large prediction model based on evidence text; S42. Use noise text to simulate irrelevant noise in evidence text, generate a large prediction model based on retrieval enhancement, and perform comparative decoding on evidence text and noise text to eliminate the influence of irrelevant noise on the generation of retrieval enhanced text, and obtain information analysis results of cross-temporal and spatial psychiatric diagnosis and treatment medical records.

7. The large-model psychiatric diagnosis and treatment information analysis enhancement method based on lexical diversity according to claim 6 is characterized in that: In step S51, the retrieval enhancement generation large prediction model is expressed as: In the formula, y represents the model output, x represents the input query, and t + represents the knowledge related to the query in the evidence text, t - represents the noise in the evidence text that is irrelevant to the query, is the large language model parameter.

8. The method for enhancing analysis of psychiatric diagnosis and treatment information based on lexical diversity and large models according to claim 7 is characterized in that: In step S52, the comparison decoding formula is specifically: In the formula, γ represents the standard weight, t ref Represents noisy text.

Images