A structured information retrieval method and system based on large language model

Abstract

The present invention relates to the field related to language information processing, and discloses a structured information retrieval method and system based on a large language model, which are used in speech recognition information retrieval based on the large language model to realize personalized recognition model iteration effect, achieve the purpose of language understanding with high-precision recognition rate, obtain correctly matched retrieved keyword groups through retrieval feedback based on conventional big data language models, and obtain individual pronunciation features based on the deviation amount between the user voice of these phrases and the standard retrieval features provided by the large language model, and use them for personalized correction of the big data language model to realize a language retrieval scheme with high recognition rate based on individuals.

Description

Technical Field

[0001] The present invention relates to the field of language information processing, and in particular to a structured information retrieval method and system based on a large language model. Background Art

[0002] The implementation of structured information retrieval methods based on large language models usually includes multiple steps of language recognition and demand conversion. The language recognition process depends on the language input method, which can be text input and voice input, respectively, suitable for different working scenarios.

[0003] In the context of the rapid development of AI technology, voice input is an efficient implementation solution that can simplify user operations. Therefore, the recognition accuracy of input voice is very important. However, because the language characteristics and pronunciation characteristics of different users are different, the use of a unified benchmark big data language recognition model will lead to large recognition errors when used for different users, resulting in uncontrollable recognition accuracy and affecting the promotion of language retrieval products. Summary of the invention

[0004] The purpose of the present invention is to provide a structured information retrieval method and system based on a large language model to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A structured information retrieval method based on a large language model, comprising:

[0007] Recognize user speech based on the big data language model, obtain multiple groups of recognition retrieval results for output according to the matching similarity, and obtain corresponding user feedback information, wherein the user feedback information is used to characterize the accuracy of the multiple groups of recognition retrieval results;

[0008] Based on the user feedback result, the recognition search result is structurally split to obtain a plurality of structural phrases, and the structural phrases are screened according to the corresponding relevance to obtain calibration data of the user object, wherein the calibration data includes user voice features corresponding to the structural phrases and recognition matching voice features;

[0009] Performing a personalized bias feature evaluation based on the calibration data to obtain a user personalized correction feature, wherein the personalized feature evaluation is used to characterize a similarity evaluation process of the user's voice;

[0010] The user personalized correction feature is used as a correction vector to adjust the recognition matching parameters of the big data language model to obtain an individual preference language model, and the recognition matching parameter adjustment is used to represent the additional assignment of matching values ​​to the recognition matching voice features.

[0011] As a further solution of the present invention: the step of recognizing the user's voice based on the big data language model and obtaining multiple groups of recognition retrieval results for output according to the matching similarity specifically includes:

[0012] Collect and record the user's search request voice to obtain the user's voice;

[0013] Matching and identifying the user's voice based on the big data language model to obtain a number of corresponding phrase matching results including recognition matching degrees, each user voice segment corresponds to a plurality of phrase matching results, and the matching recognition degree is used to characterize the feature overlap ratio between the voice and the matching phrase in the current big data language model;

[0014] Randomly combining different phrase matching results of different user voice segments based on the position order of the user voice segments in the user voice, and screening based on phrase relevance to obtain multiple groups of voice conversion results, wherein the phrase relevance is used to characterize the combination rules of different phrases based on parts of speech, and the voice conversion results are text data types;

[0015] Keyword information retrieval is performed based on the speech conversion results to obtain corresponding retrieval information content to generate multiple groups of recognition retrieval results. Each speech conversion result can correspond to multiple retrieval information contents, and different retrieval information contents correspond to different keywords of the speech conversion results.

[0016] As a further solution of the present invention: the step of structurally splitting the recognition search results based on the user feedback results, obtaining multiple structural phrases, and screening according to the corresponding relevance of the structural phrases to obtain the calibration data of the user object includes:

[0017] Acquire matching recognition search results and speech conversion results based on the user feedback results, and split the speech conversion results based on phrases, parts of speech, and sentence phrase structures of the speech conversion results to obtain multiple structural phrases;

[0018] The relevance of the plurality of structural phrases to the recognition search results is determined respectively, and a number of structural phrases with high relevance are selected in descending order based on the relevance sorting, and calibration data is correspondingly established according to the corresponding user voice features and the recognition matching voice features.

[0019] As a further solution of the present invention: the step of performing personalized bias feature evaluation according to the calibration data to obtain the user's personalized correction feature specifically includes:

[0020] Acquire multiple matching syllables by identifying matching speech features in the calibration data, wherein the matching syllables are used to represent standard pronunciation constituent units of corresponding phrases;

[0021] Performing pronunciation feature bias judgment on the user syllable and the matching syllable corresponding to the user voice feature in the calibration data, and respectively obtaining feature expression parameters of the user syllable and the matching syllable, wherein the feature expression parameters are used to characterize the overlap ratio of the corresponding syllable pronunciation feature and multiple sound feature dimensions;

[0022] The deviation ratio between the characteristic expression parameters of the user syllable and the matching syllable is calculated to obtain a personalized correction feature, wherein the personalized correction feature is used to characterize the characteristic deviation amount between the user pronunciation and the standard recognition pronunciation.

[0023] As a further solution of the present invention: it also includes a voice system and feature optimization step, including:

[0024] The personalized correction features are searched and matched through the cloud to obtain multiple dialect features whose overlap reaches a preset value, and auxiliary judgment is performed based on the user's multiple sets of personalized correction features to obtain the judgment result;

[0025] If the user meets the corresponding dialect characteristics, the word preference of the corresponding dialect is obtained through the cloud, and the big data language model is updated based on the word preference.

[0026] The embodiment of the present invention aims to provide a structured information retrieval system based on a large language model, comprising:

[0027] A basic feedback module is used to recognize user speech based on a big data language model, obtain multiple groups of recognition retrieval results for output according to matching similarity, and obtain corresponding user feedback information, wherein the user feedback information is used to characterize the accuracy of the multiple groups of recognition retrieval results;

[0028] A splitting and calibration module, used to perform structural splitting on the recognition search results based on the user feedback results, obtain multiple structural phrases, and screen the structural phrases according to their corresponding relevance to obtain calibration data of the user object, wherein the calibration data includes user voice features corresponding to the structural phrases and recognition matching voice features;

[0029] A personalized calibration module, used to perform personalized bias feature evaluation based on the calibration data to obtain a user's personalized correction feature, wherein the personalized feature evaluation is used to characterize a similarity evaluation process of the user's voice;

[0030] The user embedding module is used to use the user personalized correction feature as a correction vector to adjust the recognition matching parameters of the big data language model to obtain an individual preference language model, and the recognition matching parameter adjustment is used to represent the additional assignment of matching values ​​to the recognition matching speech features.

[0031] As a further solution of the present invention: the basic feedback module includes:

[0032] A data collection unit, used to collect and record the user's search request voice to obtain the user's voice;

[0033] A text conversion unit, used for matching and identifying the user's voice based on a big data language model to obtain a plurality of corresponding phrase matching results including recognition matching degrees, wherein each user's voice segment corresponds to a plurality of phrase matching results, and the matching recognition degree is used to characterize the feature overlap ratio between the voice and the matching phrase in the current big data language model;

[0034] A text screening unit, used for randomly combining different phrase matching results of different user voice segments based on the position order of the user voice segments in the user voice, and screening based on phrase relevance to obtain multiple groups of voice conversion results, wherein the phrase relevance is used to characterize the combination rules of different phrases based on parts of speech, and the voice conversion results are text data types;

[0035] The information retrieval unit is used to perform keyword information retrieval according to the speech conversion result, obtain corresponding retrieval information content, and generate multiple groups of recognition retrieval results accordingly. Each of the speech conversion results can correspond to multiple retrieval information contents, and different retrieval information contents correspond to different keywords of the speech conversion result.

[0036] As a further solution of the present invention: the splitting and calibration module includes:

[0037] A sentence splitting unit, used to obtain matching recognition search results and speech conversion results based on user feedback results, and split the speech conversion results based on phrases, parts of speech, and sentence phrase structures of the speech conversion results to obtain multiple structural phrases;

[0038] The relevance judgment unit is used to judge the relevance of the multiple structural phrases and the recognition search results respectively, and select several highly relevance structural phrases in descending order based on the relevance sorting, and establish calibration data according to the corresponding user voice features and recognition matching voice features.

[0039] As a further solution of the present invention: the personalized calibration module includes:

[0040] A syllable splitting unit, used to obtain a plurality of matching syllables by identifying matching speech features in the calibration data, wherein the matching syllables are used to represent standard pronunciation constituent units of corresponding phrases;

[0041] A bias judgment unit, used to perform pronunciation feature bias judgment on the user syllable and the matching syllable corresponding to the user voice feature in the calibration data, and obtain feature expression parameters of the user syllable and the matching syllable respectively, wherein the feature expression parameters are used to characterize the overlap ratio of the corresponding syllable pronunciation feature and multiple sound feature dimensions;

[0042] The personalized judgment unit is used to calculate the deviation ratio between the user syllable and the matching syllable feature expression parameter to obtain a personalized correction feature, wherein the personalized correction feature is used to characterize the feature deviation amount between the user pronunciation and the standard recognition pronunciation.

[0043] As a further solution of the present invention: it also includes a vocabulary optimization module, including:

[0044] The dialect judgment unit is used to search and match the personalized correction features through the cloud, obtain multiple dialect features whose overlap reaches a preset value, and perform auxiliary judgment based on multiple sets of personalized correction features of the user to obtain a judgment result;

[0045] The dialect optimization unit is used to obtain the word preference of the corresponding dialect through the cloud if the user meets the corresponding dialect characteristics, and update the big data language model based on the word preference.

[0046] Compared with the prior art, the beneficial effects of the present invention are: it is used in speech recognition information retrieval based on a large language model to achieve a personalized recognition model iteration effect, achieve the purpose of language understanding with a high-precision recognition rate, obtain correctly matched retrieved keyword groups through retrieval feedback based on a conventional big data language model, and obtain individual pronunciation features based on the deviation between the user voice of these phrases and the standard retrieval features provided by the large language model, and use it for personalized correction of the big data language model to achieve a language retrieval solution based on individuals with a high recognition rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flowchart of a structured information retrieval method based on a large language model.

[0048] Figure 2 A flowchart of the process of obtaining identification retrieval results in a structured information retrieval method based on a large language model.

[0049] Figure 3 A block diagram of a structured information retrieval system based on a large language model. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0051] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0052] like Figure 1 The structured information retrieval method based on a large language model provided by an embodiment of the present invention comprises the following steps:

[0053] S10, recognizing the user's voice based on the big data language model, obtaining multiple groups of recognition retrieval results for output according to the matching similarity, and obtaining corresponding user feedback information, wherein the user feedback information is used to characterize the accuracy of the multiple groups of recognition retrieval results;

[0054] S20, structurally splitting the recognition search result based on the user feedback result to obtain a plurality of structural phrases, and screening the structural phrases according to the corresponding relevance to obtain calibration data of the user object, wherein the calibration data includes user voice features corresponding to the structural phrases and recognition matching voice features;

[0055] S30, performing personalized bias feature evaluation according to the calibration data to obtain a personalized correction feature of the user, wherein the personalized feature evaluation is used to characterize a similarity evaluation process of the user's voice;

[0056] S40, using the user personalized correction feature as a correction vector to adjust the recognition matching parameters of the big data language model to obtain an individual preference language model, wherein the recognition matching parameter adjustment is used to represent an additional assignment of a matching value to the recognition matching speech feature.

[0057] In this embodiment, a structured information retrieval method based on a large language model is provided, which is used in speech recognition information retrieval based on a large language model to achieve a personalized recognition model iteration effect and achieve the purpose of language understanding with a high-precision recognition rate. Through the retrieval feedback based on the conventional big data language model, the correctly matched retrieved keyword phrases are obtained, and the individual pronunciation characteristics are obtained based on the deviation between the user voice of these phrases and the standard retrieval features provided by the large language model, and used for personalized correction of the big data language model to achieve a language retrieval scheme based on individuals with a high recognition rate. In the prior art, the implementation of the structured information retrieval method based on the large language model often includes multiple steps of language recognition and demand conversion, wherein the process of language recognition depends on the input method of the language, which can be multiple types of text input and voice input. In the context of the rapid development of current AI technology, voice input is an efficient implementation that can simplify user operations. Therefore, the recognition accuracy of the input speech is very important, but because the language characteristics and pronunciation characteristics of different users are different, the use of a unified benchmark big data language recognition model will lead to large recognition errors when used for different users, resulting in uncontrollable recognition accuracy and a certain negative impact on the product's reputation. The solution in this embodiment is: in the initial use of the user, in order to ensure that the retrieval process can feed back the results required by the user, multiple similar recognition results are matched according to the big data language model, and the results are retrieved and fed back to the user, and the user selects the correct required result, and the feature deviation between the user pronunciation of the relevant phrases of the correct result and the model standard pronunciation feature is quantitatively calculated, thereby realizing the acquisition of personalized correction features and realizing personalized optimization of the big data language model, so that when used by different users, it has a higher recognition accuracy.

[0058] like Figure 2 As shown, as another preferred embodiment of the present invention, the step of recognizing the user's voice based on the big data language model and obtaining multiple groups of recognition retrieval results for output according to the matching similarity specifically includes:

[0059] S11, collecting and recording the user's search request voice to obtain the user's voice;

[0060] S12, matching and identifying the user voice based on the big data language model to obtain a plurality of corresponding phrase matching results including recognition matching degrees, wherein each user voice segment corresponds to a plurality of phrase matching results, and the matching recognition degree is used to characterize the feature overlap ratio between the voice and the matching phrase in the current big data language model;

[0061] S13, randomly combining different phrase matching results of different user voice segments based on the position order of the user voice segments in the user voice, and screening based on phrase relevance to obtain multiple groups of voice conversion results, wherein the phrase relevance is used to characterize the combination rules of different phrases based on parts of speech, and the voice conversion results are text data types;

[0062] S14, performing keyword information retrieval according to the speech conversion result, obtaining corresponding retrieval information content, so as to generate multiple groups of recognition retrieval results accordingly, each of the speech conversion results can correspond to multiple retrieval information contents, and different retrieval information contents correspond to different keywords of the speech conversion result.

[0063] In this embodiment, step S10 is further expanded. Because in the initial state, there is a lack of the user's data for personalized correction, there may be problems with low recognition accuracy or even recognition errors. Therefore, in the initial processing, when a certain pronunciation of the user is different from the pronunciation of the standard model, multiple possible results are selected based on the similarity. In this way, each pronunciation can correspond to a variety of characters and meanings (for example, "xihu" can have different meanings such as "West Lake, tin pot, washing and care", and there are also differences in the part of speech). Therefore, the whole sentence can be combined into a variety of different required meanings. Of course, in this process, the combination of different consecutive words in the sentence still needs to be screened and judged. For example, different phrases have different parts of speech, and there are restrictions on the parts of speech that can be used at different positions in the sentence, as well as restrictions on the coordination of parts of speech between adjacent words.

[0064] As another preferred embodiment of the present invention, the step of structurally splitting the recognition search results based on the user feedback results, obtaining multiple structural phrases, and screening the structural phrases according to the corresponding relevance to obtain the calibration data of the user object includes:

[0065] Acquire matching recognition search results and speech conversion results based on the user feedback results, and split the speech conversion results based on phrases, parts of speech, and sentence phrase structures of the speech conversion results to obtain multiple structural phrases;

[0066] The relevance of the plurality of structural phrases to the recognition search results is determined respectively, and a number of structural phrases with high relevance are selected in descending order based on the relevance sorting, and calibration data is correspondingly established according to the corresponding user voice features and the recognition matching voice features.

[0067] In this embodiment, the process of obtaining calibration data is explained. Calibration data is data used to realize the acquisition of user personalized features. Therefore, it is necessary to clearly determine the correct speech content to be recognized. Therefore, directly using the speech-converted text data corresponding to the accurate recognition retrieval results of user feedback is inaccurate, because in the retrieval process, the part of a complete sentence actually used for result retrieval is ultimately only one or more keywords contained therein. Therefore, in addition to determining that the speech recognition of the keyword corresponding phrase is accurate, the other parts of the sentence cannot effectively determine whether the recognition is accurate. Therefore, when determining the calibration data, it is necessary to split it from the entire complete sentence to obtain the phrase part of the sentence corresponding to the retrieval feedback result content, thereby ensuring the accuracy of the subsequent personalization process and avoiding erroneous recognition results from entering the optimization data and affecting the accuracy of the model.

[0068] As another preferred embodiment of the present invention, the step of performing personalized bias feature evaluation according to the calibration data to obtain the user's personalized correction feature specifically includes:

[0069] Acquire multiple matching syllables by identifying matching speech features in the calibration data, wherein the matching syllables are used to represent standard pronunciation constituent units of corresponding phrases;

[0070] Performing pronunciation feature bias judgment on the user syllable and the matching syllable corresponding to the user voice feature in the calibration data, and respectively obtaining feature expression parameters of the user syllable and the matching syllable, wherein the feature expression parameters are used to characterize the overlap ratio of the corresponding syllable pronunciation feature and multiple sound feature dimensions;

[0071] The deviation ratio between the characteristic expression parameters of the user syllable and the matching syllable is calculated to obtain a personalized correction feature, wherein the personalized correction feature is used to characterize the characteristic deviation amount between the user pronunciation and the standard recognition pronunciation.

[0072] In the present embodiment, this personalized process mainly lies in the process of determining the deviation between the user's pronunciation and the matching pronunciation recorded by the standard model. However, if the judgment is made directly by the degree of coincidence with the matching syllable, the actual effect may have a certain deviation. This is because when different users make the same pronunciation, although the feature matching degree with the standard syllable may be 80%, the pronunciation of two people actually deviates in different directions. Therefore, for a syllable, it is necessary to set multiple dimensions of similar pronunciation. When judging, the degree of coincidence with the pronunciation of multiple dimensions is judged, thereby realizing a three-dimensional feature evaluation. For example, for a certain pronunciation, the syllable of the standard model can be expressed as (0.6A, 0.8B, 0.5C), and the user's expression is (0.3A, 0.8B, 0.7C). Based on this, the pronunciation characteristics of the user in three dimensions can be judged.

[0073] As another preferred embodiment of the present invention, the method further includes a voice system and feature optimization step, including:

[0074] The personalized correction features are searched and matched through the cloud to obtain multiple dialect features whose overlap reaches a preset value, and auxiliary judgment is performed based on the user's multiple sets of personalized correction features to obtain the judgment result;

[0075] If the user meets the corresponding dialect characteristics, the word preference of the corresponding dialect is obtained through the cloud, and the big data language model is updated based on the word preference.

[0076] In this embodiment, the pronunciation of different users may be affected by the dialect system, so there will be certain fixed characteristics in the use of some words. Therefore, after collecting a certain amount of user data, data synchronization can be carried out through the Internet, and data can be updated with each other. Some special pronunciation and language usage methods of users with consistent pronunciation tendencies can be localized and synchronized, making the subsequent big data model recognition more efficient and accurate.

[0077] like Figure 3 As shown, the present invention also provides a structured information retrieval system based on a large language model, which comprises:

[0078] The basic feedback module 100 is used to recognize the user's voice based on the big data language model, obtain multiple groups of recognition retrieval results for output according to the matching similarity, and obtain corresponding user feedback information, wherein the user feedback information is used to characterize the accuracy of the multiple groups of recognition retrieval results;

[0079] The splitting and calibration module 200 is used to perform structural splitting on the recognition search results based on the user feedback results, obtain multiple structural phrases, and screen them according to the corresponding relevance of the structural phrases to obtain calibration data of the user object, wherein the calibration data includes the user voice features corresponding to the structural phrases and the recognition matching voice features;

[0080] A personalized calibration module 300, used to perform personalized bias feature evaluation based on the calibration data to obtain a user personalized correction feature, wherein the personalized feature evaluation is used to characterize a similarity evaluation process of the user's voice;

[0081] The user embedding module 400 is used to use the user personalized correction feature as a correction vector to adjust the recognition matching parameters of the big data language model to obtain an individual preference language model, and the recognition matching parameter adjustment is used to represent the additional assignment of matching values ​​to the recognition matching speech features.

[0082] As another preferred embodiment of the present invention, the basic feedback module includes:

[0083] A data collection unit, used to collect and record the user's search request voice to obtain the user's voice;

[0084] A text conversion unit, used for matching and identifying the user's voice based on a big data language model to obtain a plurality of corresponding phrase matching results including recognition matching degrees, wherein each user's voice segment corresponds to a plurality of phrase matching results, and the matching recognition degree is used to characterize the feature overlap ratio between the voice and the matching phrase in the current big data language model;

[0085] A text screening unit, used for randomly combining different phrase matching results of different user voice segments based on the position order of the user voice segments in the user voice, and screening based on phrase relevance to obtain multiple groups of voice conversion results, wherein the phrase relevance is used to characterize the combination rules of different phrases based on parts of speech, and the voice conversion results are text data types;

[0086] The information retrieval unit is used to perform keyword information retrieval according to the speech conversion result, obtain corresponding retrieval information content, and generate multiple groups of recognition retrieval results accordingly. Each of the speech conversion results can correspond to multiple retrieval information contents, and different retrieval information contents correspond to different keywords of the speech conversion result.

[0087] As another preferred embodiment of the present invention, the splitting and calibration module includes:

[0088] A sentence splitting unit, used to obtain matching recognition search results and speech conversion results based on user feedback results, and split the speech conversion results based on phrases, parts of speech, and sentence phrase structures of the speech conversion results to obtain multiple structural phrases;

[0089] The relevance judgment unit is used to judge the relevance of the multiple structural phrases and the recognition search results respectively, and select several highly relevance structural phrases in descending order based on the relevance sorting, and establish calibration data according to the corresponding user voice features and recognition matching voice features.

[0090] As another preferred embodiment of the present invention, the personalized calibration module includes:

[0091] A syllable splitting unit, used to obtain a plurality of matching syllables by identifying matching speech features in the calibration data, wherein the matching syllables are used to represent standard pronunciation constituent units of corresponding phrases;

[0092] A bias judgment unit, used to perform pronunciation feature bias judgment on the user syllable and the matching syllable corresponding to the user voice feature in the calibration data, and obtain feature expression parameters of the user syllable and the matching syllable respectively, wherein the feature expression parameters are used to characterize the overlap ratio of the corresponding syllable pronunciation feature and multiple sound feature dimensions;

[0093] The personalized judgment unit is used to calculate the deviation ratio between the user syllable and the matching syllable feature expression parameter to obtain a personalized correction feature, wherein the personalized correction feature is used to characterize the feature deviation amount between the user pronunciation and the standard recognition pronunciation.

[0094] As another preferred embodiment of the present invention, a vocabulary optimization module is also included, including:

[0095] The dialect judgment unit is used to search and match the personalized correction features through the cloud, obtain multiple dialect features whose overlap reaches a preset value, and perform auxiliary judgment based on multiple sets of personalized correction features of the user to obtain a judgment result;

[0096] The dialect optimization unit is used to obtain the word preference of the corresponding dialect through the cloud if the user meets the corresponding dialect characteristics, and update the big data language model based on the word preference.

[0097] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0098] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the disclosure in the specification and examples. This application is intended to cover any variations, uses or adaptations of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.

[0099] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A structured information retrieval method based on a large language model, characterized in that: Include: Recognize user speech based on the big data language model, obtain multiple groups of recognition retrieval results for output according to the matching similarity, and obtain corresponding user feedback information, wherein the user feedback information is used to characterize the accuracy of the multiple groups of recognition retrieval results; Based on the user feedback result, the recognition search result is structurally split to obtain a plurality of structural phrases, and the structural phrases are screened according to the corresponding relevance to obtain calibration data of the user object, wherein the calibration data includes user voice features corresponding to the structural phrases and recognition matching voice features; Performing a personalized bias feature evaluation based on the calibration data to obtain a user personalized correction feature, wherein the personalized feature evaluation is used to characterize a similarity evaluation process of the user's voice; The user personalized correction feature is used as a correction vector to adjust the recognition matching parameters of the big data language model to obtain an individual preference language model, and the recognition matching parameter adjustment is used to represent the additional assignment of matching values ​​to the recognition matching voice features.

2. A structured information retrieval method based on a large language model according to claim 1, characterized in that: The step of recognizing the user's voice based on the big data language model and obtaining multiple groups of recognition retrieval results for output according to the matching similarity specifically includes: Collect and record the user's search request voice to obtain the user's voice; Matching and identifying the user's voice based on the big data language model to obtain a number of corresponding phrase matching results including recognition matching degrees, each user voice segment corresponds to a plurality of phrase matching results, and the matching recognition degree is used to characterize the feature overlap ratio between the voice and the matching phrase in the current big data language model; Randomly combining different phrase matching results of different user voice segments based on the position order of the user voice segments in the user voice, and screening based on phrase relevance to obtain multiple groups of voice conversion results, wherein the phrase relevance is used to characterize the combination rules of different phrases based on parts of speech, and the voice conversion results are text data types; Keyword information retrieval is performed based on the speech conversion results to obtain corresponding retrieval information content to generate multiple groups of recognition retrieval results. Each speech conversion result can correspond to multiple retrieval information contents, and different retrieval information contents correspond to different keywords of the speech conversion results.

3. A structured information retrieval method based on a large language model according to claim 2, characterized in that: The step of structurally splitting the recognition search results based on the user feedback results, obtaining a plurality of structural phrases, and screening the structural phrases according to the corresponding relevance to obtain the calibration data of the user object includes: Acquire matching recognition search results and speech conversion results based on the user feedback results, and split the speech conversion results based on phrases, parts of speech, and sentence phrase structures of the speech conversion results to obtain multiple structural phrases; The relevance of the plurality of structural phrases to the recognition search results is determined respectively, and a number of structural phrases with high relevance are selected in descending order based on the relevance sorting, and calibration data is correspondingly established according to the corresponding user voice features and the recognition matching voice features.

4. A structured information retrieval method based on a large language model according to claim 3, characterized in that: The step of evaluating the personalized bias feature according to the calibration data to obtain the user's personalized correction feature specifically includes: Acquire multiple matching syllables by identifying matching speech features in the calibration data, wherein the matching syllables are used to represent standard pronunciation constituent units of corresponding phrases; Performing pronunciation feature bias judgment on the user syllable and the matching syllable corresponding to the user voice feature in the calibration data, and respectively obtaining feature expression parameters of the user syllable and the matching syllable, wherein the feature expression parameters are used to characterize the overlap ratio of the corresponding syllable pronunciation feature and multiple sound feature dimensions; The deviation ratio between the characteristic expression parameters of the user syllable and the matching syllable is calculated to obtain a personalized correction feature, wherein the personalized correction feature is used to characterize the characteristic deviation amount between the user pronunciation and the standard recognition pronunciation.

5. A structured information retrieval method based on a large language model according to claim 4, characterized in that: It also includes voice system and feature optimization steps, including: The personalized correction features are searched and matched through the cloud to obtain multiple dialect features whose overlap reaches a preset value, and auxiliary judgment is performed based on the user's multiple sets of personalized correction features to obtain the judgment result; If the user meets the corresponding dialect characteristics, the word preference of the corresponding dialect is obtained through the cloud, and the big data language model is updated based on the word preference.

6. A structured information retrieval system based on a large language model, characterized in that: Include: A basic feedback module is used to recognize user speech based on a big data language model, obtain multiple groups of recognition retrieval results for output according to matching similarity, and obtain corresponding user feedback information, wherein the user feedback information is used to characterize the accuracy of the multiple groups of recognition retrieval results; A splitting and calibration module, used to perform structural splitting on the recognition search results based on the user feedback results, obtain multiple structural phrases, and screen the structural phrases according to their corresponding relevance to obtain calibration data of the user object, wherein the calibration data includes user voice features corresponding to the structural phrases and recognition matching voice features; A personalized calibration module, used to perform personalized bias feature evaluation based on the calibration data to obtain a user's personalized correction feature, wherein the personalized feature evaluation is used to characterize a similarity evaluation process of the user's voice; The user embedding module is used to use the user personalized correction feature as a correction vector to adjust the recognition matching parameters of the big data language model to obtain an individual preference language model, and the recognition matching parameter adjustment is used to represent the additional assignment of matching values ​​to the recognition matching speech features.

7. A structured information retrieval system based on a large language model according to claim 6, characterized in that: The basic feedback module includes: A data collection unit, used to collect and record the user's search request voice to obtain the user's voice; A text conversion unit, used for matching and identifying the user's voice based on a big data language model to obtain a plurality of corresponding phrase matching results including recognition matching degrees, wherein each user's voice segment corresponds to a plurality of phrase matching results, and the matching recognition degree is used to characterize the feature overlap ratio between the voice and the matching phrase in the current big data language model; A text screening unit, used for randomly combining different phrase matching results of different user voice segments based on the position order of the user voice segments in the user voice, and screening based on phrase relevance to obtain multiple groups of voice conversion results, wherein the phrase relevance is used to characterize the combination rules of different phrases based on parts of speech, and the voice conversion results are text data types; The information retrieval unit is used to perform keyword information retrieval according to the speech conversion result, obtain corresponding retrieval information content, and generate multiple groups of recognition retrieval results accordingly. Each of the speech conversion results can correspond to multiple retrieval information contents, and different retrieval information contents correspond to different keywords of the speech conversion result.

8. A structured information retrieval system based on a large language model according to claim 7, characterized in that: The splitting and calibration module comprises: A sentence splitting unit, used to obtain matching recognition search results and speech conversion results based on user feedback results, and split the speech conversion results based on phrases, parts of speech, and sentence phrase structures of the speech conversion results to obtain multiple structural phrases; The relevance judgment unit is used to judge the relevance of the multiple structural phrases and the recognition search results respectively, and select several highly relevance structural phrases in descending order based on the relevance sorting, and establish calibration data according to the corresponding user voice features and recognition matching voice features.

9. A structured information retrieval system based on a large language model according to claim 8, characterized in that: The personalized calibration module includes: A syllable splitting unit, used to obtain a plurality of matching syllables by identifying matching speech features in the calibration data, wherein the matching syllables are used to represent standard pronunciation constituent units of corresponding phrases; A bias judgment unit, used to perform pronunciation feature bias judgment on the user syllable and the matching syllable corresponding to the user voice feature in the calibration data, and obtain feature expression parameters of the user syllable and the matching syllable respectively, wherein the feature expression parameters are used to characterize the overlap ratio of the corresponding syllable pronunciation feature and multiple sound feature dimensions; The personalized judgment unit is used to calculate the deviation ratio between the user syllable and the matching syllable feature expression parameter to obtain a personalized correction feature, wherein the personalized correction feature is used to characterize the feature deviation amount between the user pronunciation and the standard recognition pronunciation.

10. A structured information retrieval system based on a large language model according to claim 9, characterized in that: It also includes a vocabulary optimization module, including: The dialect judgment unit is used to search and match the personalized correction features through the cloud, obtain multiple dialect features whose overlap reaches a preset value, and perform auxiliary judgment based on multiple sets of personalized correction features of the user to obtain a judgment result; The dialect optimization unit is used to obtain the word preference of the corresponding dialect through the cloud if the user meets the corresponding dialect characteristics, and update the big data language model based on the word preference.

Images