Voice navigation model training method and device, navigation method and vehicle-mounted navigation system

Abstract

The invention discloses a voice navigation model training method and device, a navigation method and a vehicle-mounted navigation system. The method comprises the following steps: constructing a language database comprising a point-of-interest core word library, an easily confused word mapping list and a corrected sentence pattern template library for correcting navigation points of interest; calling a specified model to generate an annotation data set based on a pre-constructed cue word project and a language database; inputting the labeled data set into a pre-constructed navigation model for iterative training to obtain a target navigation model; the target navigation model is used for performing initial interest point recognition and initial interest point correction on a voice navigation instruction output by a user and determining a target interest point. Therefore, the core word library provides a point-of-interest vocabulary basis, and the easy-to-confuse word mapping list effectively solves the problem of recognition errors caused by non-standard pronunciation of a user and the like. Meanwhile, due to the introduction of the corrected sentence pattern template library, the interest points can be corrected, the navigation failure caused by the input error of the user is reduced, and the navigation precision is improved.

Description

Technical Field

[0001] The present application relates to the field of speech recognition technology, and in particular to a speech navigation model training method, device, navigation method and vehicle-mounted navigation system. Background Art

[0002] With the rapid development of speech recognition and navigation technologies, voice navigation, which combines speech recognition and navigation technologies, is widely used in many scenarios in life. For example, in vehicles, to ensure driving safety, in-vehicle voice navigation can determine the user's navigation point of interest (POI) for navigation.

[0003] In the actual use of voice navigation, when users' voice navigation commands are mispronounced, have an accent, or mispronounce a word in the navigation point of interest (POI), the in-vehicle navigation system often fails to accurately identify the user's navigation point of interest, resulting in a high error rate. Furthermore, if users initiate a second round of voice navigation commands based on the first round to correct errors, rewrite the POI, or supplement their intentions, the in-vehicle navigation system often fails to effectively understand and output satisfactory navigation point of interest results, severely impacting the user's navigation experience.

[0004] Therefore, how to solve the problem that users cannot accurately determine the user's navigation POI due to inaccurate pronunciation, accent, etc., effectively understand the user's intention to rewrite or supplement the navigation POI, and improve navigation accuracy and experience, is an urgent problem to be solved by technical personnel in this field. Summary of the invention

[0005] In view of this, one aspect of the present application provides a method for training a voice navigation model, the method comprising:

[0006] Constructing a language database; the language database includes a core vocabulary of points of interest, a mapping list of easily confused words, and a correction sentence template library for correcting errors in the navigation points of interest;

[0007] Based on the pre-built prompt word project and the language database, calling a specified model to generate a labeled data set;

[0008] The annotated data set is input into a pre-built navigation model for iterative training to obtain a target navigation model; the target navigation model is used to process the voice navigation instructions output by the user; the instruction processing includes initial interest point recognition, initial interest point correction and determination of target interest points.

[0009] Optionally, the constructing of the language database includes:

[0010] collecting navigation point of interest information from at least one mapping service database;

[0011] Extract keywords from the navigation point of interest information; and construct the core thesaurus of the point of interest according to the keywords;

[0012] Extract target words from the core thesaurus of the point of interest; there are confusing words for the target words, and the confusing words include at least one of polyphonic words, homophonic words, synonyms, front-back nasal sound confusing words, and flat-tip tongue sound confusing words;

[0013] Construct the mapping relationship between the target word and the confusing word to obtain the confusing word mapping list.

[0014] Optionally, the construction of the language database includes:

[0015] Obtain a pre-constructed basic correction sentence pattern;

[0016] Generalize the basic correction sentence pattern by adding preset qualifiers to obtain a first generalized sentence pattern, and / or generalize the basic correction sentence pattern through a preset large model to obtain a second generalized sentence pattern;

[0017] Construct the correction sentence pattern template library with the basic correction sentence pattern, the first generalized sentence pattern, and the second generalized sentence pattern.

[0018] Optionally, the data generation rules in the prompt engineering include rules for generating the labeled data set based on a preset data format, and the preset data format consists of the original point of interest, the correction prompt, and the corrected point of interest;

[0019] The original point of interest is generated according to the core thesaurus of the point of interest;

[0020] The correction prompt is generated according to the confusing word mapping list and the correction sentence pattern template library.

[0021] Optionally, the constraint conditions in the prompt engineering include at least one of the following:

[0022] At least one of the target words is included in the original point of interest;

[0023] The replacement word is specified in the correction prompt; the replacement word is used to replace the target word, and the replacement word belongs to the confusing word;

[0024] The replacement word and the target word are homophonic but with different tones;

[0025] The corrected point of interest conforms to the preset place name structure.

[0026] Optionally, before inputting the labeled data set into a pre-constructed navigation model for iterative training, it further includes at least one of the following steps:

[0027] Remove the data in the labeled dataset that does not conform to the specified format through regular expressions;

[0028] Verify the pinyin of the labeled dataset through a pinyin conversion library to remove data with incorrect pinyin;

[0029] Set core confusing words for the target words in the labeled dataset; the core confusing words belong to the confusing words;

[0030] Remove the data in the labeled dataset where the length of the original point of interest is not within the preset length range;

[0031] Remove the data in the labeled dataset with incorrect content; the incorrect content includes non-Chinese, non-conventional place names, and incorrect codes.

[0032] Optionally, before inputting the labeled dataset into a pre-constructed navigation model for iterative training to obtain a target navigation model, at least one of the following steps is further included:

[0033] Remove the data in the labeled dataset where the target word is located at the first position of the original point of interest;

[0034] Remove the data where the difference in the number of pinyin letters between the replacement word and the target word is greater than 1;

[0035] Remove the data in the correction prompt where the correction sentence pattern does not conform to the preset language sentence pattern structure.

[0036] Optionally, before inputting the labeled dataset into a pre-constructed navigation model for iterative training to obtain a target navigation model, at least one of the following steps is further included:

[0037] Expand the target word;

[0038] Insert interfering words into the correction sentence pattern of the correction prompt to expand the correction sentence pattern;

[0039] Remove the data with special characters through regular expressions;

[0040] Calculate the semantic similarity between the data in the labeled dataset; and select any one of the data combinations with the semantic similarity greater than the threshold.

[0041] Another aspect of the present application provides a navigation method, which is applied to a navigation system including a target navigation model, and the target navigation model is obtained through the above-mentioned voice navigation model training method. The navigation method includes:

[0042] Obtain the voice navigation instruction output by the user;

[0043] Identify the voice navigation instruction to obtain the recognition result of the initial point of interest;

[0044] Determine whether the initial point of interest is correct;

[0045] If it is correct, obtain the feedback instruction of the user; and determine whether the user is satisfied with the initial point of interest according to the feedback instruction. If satisfied, use the initial point of interest as the target point of interest; if not satisfied, obtain the supplementary intention of the user; and generate the target point of interest according to the supplementary intention of the user;

[0046] If it is incorrect, obtain the instruction for rewriting the point of interest of the user; and rewrite the initial point of interest according to the instruction for rewriting the point of interest to obtain the target point of interest;

[0047] Generate a navigation path according to the target point of interest.

[0048] Another aspect of the present application provides a voice navigation model training device, and the device includes:

[0049] A language database construction module for constructing a language database; the language database includes a core vocabulary library of points of interest, a mapping list of easily confused words, and a correction sentence pattern template library for correcting the navigation points of interest;

[0050] A labeled data set generation module for generating a labeled data set by calling a specified model based on a pre-constructed prompt engineering and the language database;

[0051] An iterative training module for inputting the labeled data set into a pre-constructed navigation model for iterative training to obtain a target navigation model; the target navigation model is used for instruction processing of the voice navigation instruction output by the user; the instruction processing includes initial point of interest recognition, correction of the initial point of interest, determination of the target point of interest, and generation of a navigation path.

[0052] Another aspect of the present application provides a vehicle-mounted navigation system, and the vehicle-mounted navigation system includes:

[0053] Each module in the voice navigation model training device described above;

[0054] A voice navigation instruction acquisition module for acquiring the voice navigation instruction output by the user;

[0055] A model calling module for calling the target navigation model so that the target navigation model determines the target point of interest according to the voice navigation instruction;

[0056] A navigation path generation module for generating a navigation path according to the target point of interest.

[0057] A method and device for training a voice navigation model, a navigation method, and a vehicle-mounted navigation system provided by this application have the following beneficial effects: Thus, an annotated data set is generated based on a language database, and a target navigation model trained through the annotated data set can accurately identify points of interest in a user's voice command. Among them, the core vocabulary provides a clear vocabulary basis for points of interest, and the list of mappings for easily confused words can effectively solve the problem of recognition errors caused by users' inaccurate pronunciation, accent differences, or different vocabulary usage habits, thereby improving the accuracy of point-of-interest recognition. At the same time, the introduction of the corrected sentence pattern template library can correct points of interest when the recognition of points of interest is incorrect, reducing navigation failures caused by users' input errors and improving navigation accuracy and the user's navigation experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a schematic flowchart of a method for training a voice navigation model provided by an embodiment of this application;

[0059] Figure 2 is a schematic diagram of the principle of a method for training a voice navigation model provided by an embodiment of this application;

[0060] Figure 3 is a schematic flowchart of a navigation method provided by an embodiment of this application;

[0061] Figure 4 is a schematic structural diagram of a device for training a voice navigation model provided by an embodiment of this application;

[0062] Figure 5 is a schematic structural diagram of a vehicle-mounted navigation system provided by an embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0063] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms "a", "the", and "said" used in this application and the appended claims are also intended to include the plural forms unless the context clearly dictates otherwise. It should also be understood that the term "and / or" as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0064] It should be understood that although the terms first, second, third, etc. may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to a determination".

[0065] Figure 1 This is a schematic flowchart of a method for training a voice navigation model provided by an embodiment of the present application. As Figure 1 shown, the method includes:

[0066] S10: Construct a language database; the language database includes a core keyword library for points of interest, a mapping list of easily confused words, and a correction sentence pattern template library for correcting navigation points of interest;

[0067] In order to accurately identify the user's navigation points of interest and at the same time solve the recognition errors caused by problems such as the user's accent and unstandardized pronunciation, in an optional embodiment, first, a language database about navigation points of interest is constructed. The language database includes a core keyword library for points of interest, a mapping list of easily confused words, and a correction sentence pattern template library.

[0068] Among them, the core keyword library for points of interest is the core component of the language database. The core keyword library for points of interest contains common navigation point of interest vocabulary and its related information, which is the basis for understanding and identifying the user's voice navigation instructions during navigation.

[0069] In an optional embodiment, the core keyword library for points of interest may include, but is not limited to, the name of the point of interest, category label, and geographical location information. Among them, the name of the point of interest includes the names of various landmarks, places, merchants, scenic spots, etc., such as the Palace Museum and Shanghai Hongqiao Airport. The category label refers to the category to which each point of interest is labeled, such as dining, shopping, transportation facilities, tourist attractions, hospitals, etc. The geographical location information refers to the geographical location information of the point of interest, such as the city and region where the point of interest is located.

[0070] In navigation voice interaction, due to situations such as similar pronunciation, accent differences, or different user expression habits, some words may be misrecognized. Therefore, in a specific embodiment, the language database also includes a mapping list of easily confused words. In specific navigation, the mapping list of easily confused words can be used to handle the diverse expressions and accent differences of users, improving the overall robustness and reliability of the system.

[0071] The mapping list of easily confused words can be pairs of confused words, such as "Beijing" and "Beiping". It can also be homophones, synonyms, easily confused front and back nasal sounds, easily mispronounced flat and curled tongue sounds, etc. In fact, the mapping list of easily confused words can include the mapping relationships of all words that are easily mispronounced in points of interest, and the present application does not make specific limitations on this.

[0072] The corrected sentence pattern library includes sentence patterns for correcting points of interest. Based on the recognition of the corrected sentence patterns, the intention of the user to modify the points of interest can be quickly determined, so as to accurately obtain the user's target points of interest. The corrected sentence pattern, for example, is "A" instead of "B", where A is the replacement word and B is the original incorrect word.

[0073] It should be noted that in a specific embodiment, the language database can be updated regularly according to actual needs, so as to continuously improve the navigation reliability.

[0074] S11: Based on the pre-constructed prompt engineering and language database, call a specified model to generate an annotated data set;

[0075] Furthermore, based on the language database and the pre-constructed prompt (Prompt) engineering, call a specified model to generate an annotated data set. It can be understood that the Prompt engineering is used to design and optimize the prompt (Prompt) to guide the language model to generate outputs that meet specific requirements. In the embodiment of the present application, the pre-constructed Prompt engineering is a prompt that can guide the navigation model to accurately identify and determine the user's target points of interest.

[0076] The language database is the basis for generating the annotated data set. It includes a core vocabulary library of points of interest, a mapping list of confusing words, and a corrected sentence pattern library. These contents provide rich context information and an error correction mechanism for the model, making the generated annotated data more accurate and reliable.

[0077] It can be understood that selecting the appropriate specified model is the key to generating high-quality annotated data. In an optional embodiment, the specified model can be a pre-trained language model. Specifically, it can include, but is not limited to, models based on the Transformer architecture, such as the GPT series and BERT, etc. In an optional embodiment, the specified model can be the Qwen-72B large model. In a specific embodiment, the specified model has strong language understanding and generation capabilities. Therefore, the specified model combined with the Prompt engineering can generate a high-quality annotated data set.

[0078] It should be noted that in an optional embodiment, when generating the annotated data set, a batch generation strategy is adopted, and a preset number of data are generated in each batch. For example, 200 data are generated in each batch. And a preset cooling duration is set between adjacent batches. For example, the cooling interval is set to 5 minutes, thereby avoiding the degradation of the specified model and ensuring the quality and stability of the generated data.

[0079] Thus, the labeled dataset generated based on the pre-constructed Prompt engineering and language database can provide high-quality and highly targeted training data for the training of the navigation model. This labeled dataset can cover various actual usage scenarios and user expression methods, improving the speech recognition and interaction capabilities of the navigation system.

[0080] S12: Input the labeled dataset into the pre-constructed navigation model for iterative training to obtain a target navigation model; the target navigation model is used to process the voice navigation instructions output by the user; the instruction processing includes initial point of interest recognition, initial point of interest correction, and determination of the target point of interest.

[0081] Furthermore, after obtaining a high-quality labeled dataset, input the labeled dataset into the pre-constructed navigation model for iterative training, thereby obtaining a target navigation model that can perform point of interest recognition, correction, and determination. It can be understood that the obtained labeled dataset is a Supervised Fine-Tuning (SFT) dataset, and the SFT dataset can enable the target navigation model to show higher accuracy and efficiency in the point of interest correction task.

[0082] In an alternative embodiment, before inputting the labeled dataset into the navigation model, the labeled dataset can be cleaned and preprocessed to ensure the accuracy and consistency of the labeled dataset. It should be noted that the navigation model can be a model based on the Transformer architecture, and this application does not make any limitations in this regard. In addition, it should also be noted that during the training process, as an alternative embodiment, the labeled dataset can be input into the navigation model in batches for training.

[0083] Through continuous iterative training until the preset iteration conditions are met, the target navigation model is obtained. The obtained target navigation model can be used in various fields such as in-vehicle navigation and mobile phone navigation. In a specific embodiment of using the target navigation model, the target navigation model can recognize and parse the voice navigation instructions output by the user, thereby determining an initial point of interest. If the initial point of interest is incorrect, the initial point of interest can be modified according to the point of interest modification instructions output by the user, thereby determining the target point of interest.

[0084] During the navigation model training process, as an optional embodiment, full parameter tuning is performed to improve the target navigation performance. Specifically, by combining cosine annealing and Early Stopping to implement self-usage of the learning rate, when the validation set perplexity does not decrease for 3 consecutive times, it is determined that the preset iteration condition is reached and the training stops. Among them, a part of the data in the labeled dataset is used as the validation set. In addition, the cumulative step size is dynamically adjusted according to the video memory occupancy to achieve gradient product optimization. The training process is accelerated and the memory occupancy is reduced by mixed precision control, while trying to maintain the training accuracy of the model. Specifically, in an optional embodiment, BF16 precision (Ampere architecture) is preferentially used and degraded to FP16 precision.

[0085] In an optional embodiment, all gradients are scaled proportionally by gradient clipping so that their norms do not exceed the threshold, thereby preventing the gradient explosion problem and making the training process more stable. In addition, through weight initialization, the activation values and gradients of neurons are avoided from being too large or too small, thereby improving the convergence speed and performance of the model. At the same time, through regularization techniques, overfitting of the navigation model is prevented and the generalization ability of the model is improved.

[0086] Figure 2 The following is a schematic diagram of the principle of a voice navigation model training method provided by an embodiment of the present application. For ease of understanding, the following will be combined with Figure 2 to further illustrate. As Figure 2 shown, in a specific embodiment, the language database consists of a point of interest core vocabulary, a list of confusing word mappings, and a corrected sentence pattern template library. Based on the language database and the pre-constructed Prompt project, a labeled dataset is generated by specifying a model. Further, the labeled dataset generated by the specified model is used as the input of the navigation model, and the navigation model is iteratively trained. Finally, the pre-constructed initial navigation model is trained to obtain the target navigation model.

[0087] Therefore, the voice navigation model training method provided by the embodiment of the present application can accurately identify the points of interest in the user's voice commands based on the language database and the target navigation model trained by the labeled dataset. Among them, the core vocabulary provides a clear vocabulary basis for points of interest, and the list of confusing word mappings can effectively solve the recognition error problems caused by users' inaccurate pronunciation, accent differences, or different vocabulary usage habits, thereby improving the accuracy of point of interest recognition. At the same time, the introduction of the corrected sentence pattern template library can correct the points of interest when the point of interest recognition is incorrect, reduce the navigation failure caused by user input errors, and improve the navigation accuracy and the user's navigation experience.

[0088] In an optional embodiment, building a language database includes:

[0089] Collect navigation point-of-interest information from at least one map service database;

[0090] Extract keywords from the navigation point-of-interest information; and construct a core thesaurus of point-of-interest based on the keywords;

[0091] Extract target words from the core thesaurus of point-of-interest; there are confusing words for the target words, and the confusing words include at least one of polyphonic words, homophonic words, synonyms, words with confusion between front and back nasal sounds, and words with confusion between flat and retroflex sounds;

[0092] Construct a mapping relationship between the target words and the confusing words to obtain a confusing word mapping list.

[0093] It can be understood that the open-source map service database provides rich geographical information data. Therefore, in a specific embodiment, navigation point-of-interest information can be collected from the open-source map service database. The present application does not limit the quantity and specific selection of the open-source map service database.

[0094] Furthermore, perform seed data screening. Specifically, extract keywords from the navigation point-of-interest information, and integrate and construct all the keywords to obtain a core thesaurus of point-of-interest. Among them, the keywords may include, but are not limited to, prefix words, core words, modifier words, function words, location words, attribute words, industry words, and suffix words, and these words together constitute a complete description of a place name or other noun.

[0095] For the sake of easy understanding, take the suffix words as an example for illustration. For example, extract suffix words such as "building / hotel / square" from the navigation point-of-interest information, so as to ensure that the original place name can conform to the geographical entity naming specification and lay a good foundation for subsequent data generation.

[0096] It should be noted that the extraction of keywords can be carried out using a word segmentation tool or a large model. The present application does not limit the method for extracting keywords. In addition, after obtaining the keywords, when constructing the core thesaurus of point-of-interest, the keywords can be sorted and classified, and stored according to different classification results of the point-of-interest. For example, it can be divided into categories such as catering, shopping, transportation, and medical treatment, which is convenient for subsequent management and invocation. The present application also does not limit the classification method and the specific types of classification.

[0097] In an alternative embodiment, in order to ensure the quality of the subsequent generated annotation dataset, after obtaining the core thesaurus of point-of-interest, the core thesaurus of point-of-interest can be optimized. The optimization process may include, but is not limited to, removing incorrect keywords, duplicate keywords, and supplementing missing important keywords.

[0098] Further, target words are extracted from the core keyword library of points of interest. The extracted target words are those that have corresponding confusing words, where the confusing words include at least one of polyphonic characters, homophonic words, near-synonyms, words with confused front and back nasal sounds, and words with confused flat and retroflex sounds. For ease of understanding, examples will be given below.

[0099] For example, the target word is "steel", and there is a homophonic word "just", and "just" is the confusing word corresponding to the target word. For example, the target word is "enter", and the corresponding confusing word is "near", and the two are confused in terms of front and back nasal sounds. For example, the target word is "teacher", and the corresponding confusing word is "thought", and the two are confused in terms of flat and retroflex sounds.

[0100] In a specific embodiment, after determining the target word and its corresponding confusing word, for each target word, list its corresponding confusing word and establish a mapping relationship to form a confusing word mapping list.

[0101] It can be understood that in a specific embodiment, whether the user has an accent or an inaccurate pronunciation, a wrong character will be generated. As long as the confusing word mapping list constructed based on this application is used, the point of interest can be corrected to obtain the correct target point of interest.

[0102] Thus, through the voice navigation model training method provided by the embodiments of this application, a language database including a core keyword library of points of interest and a confusing word mapping list can be constructed, providing strong support for speech recognition and instruction processing in the navigation system and improving the accuracy of navigation.

[0103] In an alternative embodiment, constructing the language database includes:

[0104] Obtain a pre-constructed basic correction sentence pattern;

[0105] Generalize the basic correction sentence pattern by adding preset determiners to obtain a first generalized sentence pattern, and / or generalize the basic correction sentence pattern through a preset large model to obtain a second generalized sentence pattern;

[0106] Construct a correction sentence pattern template library with the basic correction sentence pattern, the first generalized sentence pattern, and the second generalized sentence pattern.

[0107] In a specific embodiment, the correction sentence pattern template library is used to provide sentence patterns for correcting points of interest. When constructing the correction sentence pattern template library, the pre-constructed basic correction sentence pattern is the core of the template library. After obtaining the basic correction sentence pattern, generalization is performed on the basis of the basic correction sentence pattern, and thus rich correction sentence pattern templates can be obtained.

[0108] For ease of understanding, the following gives examples of the basic correction patterns. For example, the basic correction patterns may include: 1. "It is A rather than B"; 2. "It is not A but B"; 3. "Change A to B"; 4. "Change the Xth character to B"; 5. "What I want to say is A, not B". Here, A and B are replacement words or misspelled words, and X is a number.

[0109] Furthermore, the basic correction patterns are generalized by adding preset determiners. Specifically, in order to make the correction patterns more flexible and diverse, the first generalized pattern can be generated by adding preset determiners to the basic correction patterns. It should be noted that the addition of preset determiners does not change the points of interest, and the specific content added is not limited in this application.

[0110] For example, "It is A rather than B" can be regarded as the basic correction pattern, and "What I want to say is A, not B" can be regarded as the first generalized pattern after adding preset determiners, where the preset determiners include "What I want to say" and "rather than".

[0111] In addition, in an alternative embodiment, the basic correction patterns can also be generalized through a preset large model. It can be understood that the large model can generate more natural and diverse patterns based on context and semantic relationships. The preset large model in the embodiments of this application is not limited.

[0112] Furthermore, the basic correction patterns, the first generalized patterns, and the second generalized patterns are integrated together to construct a complete correction pattern template library. In an alternative embodiment, the correction pattern template library can also be classified and stored, and the correction patterns in the correction pattern template library can be continuously expanded and optimized according to user feedback and actual usage in the specific usage process.

[0113] It can be understood that the method provided by the embodiments of this application can, for recognition errors caused by homophones, accents, and confusion between front and back nasal sounds, combine the context semantics, and based on the correction patterns, subtly replace the misspelled words in the original points of interest, thereby generating new points of interest and improving the navigation reliability.

[0114] In an alternative embodiment, the data generation rules in the prompt engineering include the rules for generating an annotated data set based on a preset data format, and the preset data format consists of the original point of interest, the correction prompt, and the corrected point of interest;

[0115] The original point of interest is generated according to the core word library of the point of interest;

[0116] The correction prompt is generated according to the mapping list of easily confused words and the correction pattern template library.

[0117] In a specific embodiment, the pre-constructed Prompt project consists of a task description, a data generation rule, a standard example, and a constraint condition. Among them, the task description refers to the task description of rewriting the point of interest, the data generation rule is used to guide the format and content of the generation of the labeled data set, the standard example refers to an example of the labeled data set, and the constraint condition is used to limit the content of the labeled data set.

[0118] Among them, the data generation rule includes a rule for generating a labeled data set based on a preset data format, and the preset data format consists of an original point of interest, a correction prompt, and a corrected point of interest.

[0119] Among them, the original point of interest is generated according to the core vocabulary of the point of interest constructed in the above embodiment, and includes the names of the points of interest that the user may mention. The correction prompt is generated according to the homophone mapping list and the correction sentence pattern template library, and is used to guide the correction of the original point of interest. The corrected point of interest refers to the point of interest obtained by correcting the original point of interest according to the correction prompt.

[0120] In a specific embodiment, the constraint condition is used to limit the generation scope of the Prompt, that is, to limit the content of the labeled data set, so that the generated labeled data set adapts to the navigation scenario of the point of interest modification. In an optional embodiment, the constraint condition includes at least one of the following:

[0121] At least one target word is included in the original point of interest;

[0122] The replacement word is specified in the correction prompt; the replacement word is used to replace the target word, and the replacement word belongs to the homophone;

[0123] The replacement word and the target word have the same pronunciation but different tones;

[0124] The corrected point of interest conforms to the preset place name structure.

[0125] In fact, when dealing with the above constraint conditions, in an optional embodiment, other constraint conditions such as prohibiting the output of serial numbers and redundant explanations may also be included. The application does not make specific limitations on the number and specific content of the constraint conditions. For the sake of understanding, the pre-constructed Prompt project will be illustrated by examples below.

[0126] For example, in an optional embodiment, the pre-constructed Prompt project is as follows:

[0127] Task description: You are a dialect phonetics expert. In view of the recognition errors caused by homophones in the point of interest, combined with the context semantics and matching the characters in the original point of interest, perform replacement to generate the corrected point of interest. Please generate according to the following data generation rules:

[0128] 1. Only process the front and back nasal vowels (-n / -ng) and homophonic homographs.

[0129] 2. Focus on the following pronunciation categories: front nasal sounds: an / en / in; back nasal sounds: ang / eng / ing; easily confused combinations: in / ing, en / eng, an / ang.

[0130] 3. The generated data format is:

[0131] {

[0132] "original":"original point of interest";

[0133] "phonetic_clue":"Revision Tip";

[0134] "modified": "Modified points of interest"

[0135] }

[0136] Generalize the sentence structure based on the basic modified sentence structure:

[0137] Basic revised sentence: Change "lian" to "lian";

[0138] Generalized sentence pattern: [help me / I want / give me]+[put / will]+[this]+(connecting link)+(change into / replace into / change into / rewrite into / modify into)+(connecting link).

[0139] Standard example:

[0140] 1. {“original”:“Guangzhou Audio and Video City”, “phonetic_clue”:“Please help me change the sound of this music to “Impression””, “modified”:“Guangzhou Impression City”};

[0141] 2. {"original":"Yun Ding Restaurant","phonetic_clue":"I want to change the word 'Yun' in 'luck' to 'Yun' in 'allowed',"modified":"Yun Ding Restaurant"};

[0142] 3. {“original”:“珍味坊”,”phonetic_clue”:“Change the “zhen” in this pearl to “zhishizhen”,”modified”:“真味坊”}.

[0143] Constraints include:

[0144] 1. The original interest point contains at least one replaceable nasal vowel Chinese character;

[0145] 2. Pronunciation cues should clearly point to the target pronunciation (e.g., using signature words: "北京" refers to the front nasal sound, "光明" refers to the back nasal sound);

[0146] 3. Maintain a reasonable place name structure (e.g., common combinations such as "XX Park" and "XX Hotel");

[0147] 4. The replacement word has the same pronunciation as the target word but a different tone (tone changes are allowed);

[0148] 5. Negative sentence examples: “It is not [target word] but [replacement word]”;

[0149] 6. Affirmative sentence examples: “It is [the target word] of [the replacement word]”;

[0150] 7. Prohibit outputting serial numbers and redundant explanations;

[0151] Now please generate a preset set (for example, 30 sets) of annotation data that meets the requirements.

[0152] It should be noted that the format and content of the annotated dataset generated based on the Prompt project and language database are the same as those of the standard instance, that is, the standard dataset consists of original points of interest, revised prompts, and revised points of interest.

[0153] As an optional embodiment, before inputting the labeled dataset into the pre-built navigation model for iterative training, at least one of the following steps is further included:

[0154] Use regular expressions to remove data that does not conform to the specified format in the labeled dataset;

[0155] The pinyin of the annotated dataset is verified using the pinyin conversion library to eliminate data with incorrect pinyin.

[0156] Set core confusion words for the target words in the labeled dataset; core confusion words are easily confused words;

[0157] Eliminate data from the labeled dataset where the length of the original interest point is outside the preset length range;

[0158] Eliminate data with erroneous content in the labeled dataset; erroneous content includes non-Chinese characters, unconventional place names, and incorrect codes.

[0159] In a specific embodiment, after the initial generation of the annotated dataset, the annotated dataset is post-processed. Specifically, to ensure the format consistency of the annotated dataset input into the navigation model, in an optional embodiment, regular expressions are used to match and filter data that conforms to a specific pattern in the annotated dataset. For example, Json format data is matched to provide a standardized data foundation for subsequent processing. In another optional embodiment, the annotated data is verified using a pinyin conversion library to ensure that the replacement operation complies with the easily confused word mapping list, further ensuring the accuracy and rationality of the annotated data.

[0160] To further improve the quality of the annotated dataset, the annotated dataset is screened and filtered. Specifically, in an optional embodiment, to more effectively target easily confused words, a core confused word is set for the target word in the annotated dataset. The core confused word belongs to the easily confused word corresponding to the target word in the easily confused word mapping list. For example, the core confused word for the target word "金" is set to "京".

[0161] In another optional embodiment, the length of the original interest points in the annotated data set is controlled within a preset length range, for example, within a range of 4-8 characters, thereby ensuring that the length of the interest points is moderate, neither too short to lack information nor too long to increase unnecessary complexity.

[0162] In another optional embodiment, data with erroneous content is eliminated from the annotated data set. For example, data with erroneous content such as non-Chinese characters, unconventional place names, and incorrect codes are filtered out to avoid data processing difficulties and navigation model understanding deviations caused by erroneous content.

[0163] Therefore, the voice navigation model training method provided in the embodiment of the present application can improve the quality of the labeled data set through the above data verification and filtering methods, provide high-quality data support for the iterative training of the navigation model, and thus improve the accuracy and robustness of the target navigation model.

[0164] Based on the above embodiment, as an optional embodiment, before inputting the labeled dataset into the pre-built navigation model for iterative training to obtain the target navigation model, at least one of the following steps is further included:

[0165] Eliminate the data where the target word is located at the first position of the original interest point in the labeled dataset;

[0166] Eliminate data where the difference in the number of pinyin letters between the replacement word and the target word is greater than 1;

[0167] Eliminate data in the correction prompt that does not conform to the preset language sentence structure.

[0168] In order to further improve the reliability of the annotated dataset, in an optional embodiment, the data in the annotated dataset where the target word is located at the first position of the original point of interest is filtered, thereby avoiding changing the semantics of the place name due to replacing the target word at the first position of the original point of interest, thereby ensuring that the corrected point of interest can still be correctly understood and identified.

[0169] In another optional embodiment, to ensure that the revised POI maintains a high degree of phonetic similarity with the original POI, consistent with actual phonetic confusion, data where the difference in the number of pinyin letters between the replacement word and the target word is greater than 1 is filtered out. For example, data for "an→ang" can be retained, while data for "an→ou" will be filtered out.

[0170] In addition, to further improve the accuracy of the annotated dataset, in an optional embodiment, data in the correction prompt that do not conform to the preset language sentence structure is deleted. It should be noted that the preset language sentence structure refers to the sentence structure that conforms to the user's language description order and habits. Therefore, the annotated data corresponding to the sentence structure that does not conform to the normal language description order and habits is deleted.

[0171] Based on the above embodiment, as an optional embodiment, before inputting the labeled dataset into the pre-built navigation model for iterative training to obtain the target navigation model, at least one of the following steps is further included:

[0172] Expand the target word;

[0173] Insert interference words into the revised sentence pattern of the revision prompt to expand the revised sentence pattern;

[0174] Use regular expressions to remove data containing special characters;

[0175] Calculate the semantic similarity between the data in the labeled dataset; and filter out any data combination whose semantic similarity is greater than the threshold.

[0176] In a specific embodiment, in order to further improve the richness of the annotated dataset and thus further improve the navigation accuracy of the target navigation model, data enhancement processing is performed on the annotated dataset before inputting it into the navigation model for training.

[0177] Specifically, in an optional embodiment, the target word is further expanded, that is, the target word is expanded twice. For example, the target word is "金", and the secondary expanded words are "京" and "荆". Thus, the annotation data set is further expanded, and the adaptability of the target navigation model to different writing methods of homophones is increased.

[0178] In addition, in another optional embodiment, an interference word is inserted into the revised sentence pattern of the correction prompt to perform a secondary expansion of the revised sentence pattern. Wherein, the inserted interference word does not change the point of interest. For example, the revised sentence pattern before the interference word is inserted is "Help me change Lianhua to Chenghua", and the revised sentence pattern after the interference word is inserted is: "I'm talking about you. Help me change Lianhua to Chenghua." Wherein, the interference word is "I'm talking about you."

[0179] It should be noted that when expanding the target word and the corrected sentence pattern, it can be expanded manually or through a large model. This application does not make any limitations in this regard. To balance time consumption and expansion accuracy, it can be expanded through a large model. For example, the Qwen-3B large language model can be selected for expansion.

[0180] In an alternative embodiment, before the labeled dataset processed through the above filtering, data augmentation, etc. is input into the navigation system, a final data cleaning is performed. Specifically, special characters can be filtered through regular expressions. In addition, the semantic similarity between the data in the labeled dataset is calculated, and any one of the data combinations with a semantic similarity greater than the threshold is selected. Specifically, BERTScore can be used for semantic similarity deduplication. At the same time, a Json Schema validator is designed for format verification to ensure field integrity and type correctness.

[0181] Furthermore, after obtaining the labeled dataset, the labeled dataset is divided into a training set, a validation set, and a test set according to a preset ratio for iterative training of the navigation model. For example, 12,000 labeled data are divided according to the ratio of 8:1:1 to obtain 9,600 training data, 1,200 validation data, and 1,200 test data.

[0182] In the above embodiment, the voice navigation model training method is described in detail. This application also provides an embodiment corresponding to a navigation method. The navigation method is applied to a navigation system, which includes a target navigation model. The target navigation model is obtained by training through the voice navigation model training method in the above embodiment.

[0183] Figure 3 The flowchart of a navigation method provided by an embodiment of this application is as Figure 3 shown. The method includes:

[0184] S30: Obtain the voice navigation instruction output by the user;

[0185] S31: Recognize the voice navigation instruction to obtain the recognition result of the initial point of interest;

[0186] S32: Determine whether the initial point of interest is correct; if correct, go to step S33; if incorrect, go to step 36;

[0187] S33: Obtain the feedback instruction of the user; and determine whether the user is satisfied with the initial point of interest according to the feedback instruction; if satisfied, go to step S34; if not satisfied, go to step S35;

[0188] S34: Use the initial point of interest as the target point of interest;

[0189] S35: Obtain the user's supplementary intention; and generate a target point of interest according to the user's supplementary intention.

[0190] S36: Obtain the user's point-of-interest rewriting instruction; and rewrite the initial point of interest according to the point-of-interest rewriting instruction to obtain the target point of interest.

[0191] S37: Generate a navigation path according to the target point of interest.

[0192] In a specific embodiment, after parsing and recognizing the voice navigation instruction initiated by the user, if the initial point of interest initiated by the user is accurately recognized, it is necessary to further obtain a feedback instruction to determine whether the user is satisfied with the currently recognized initial point of interest. If satisfied, the initial point of interest is used as the target point of interest for navigation. If not satisfied, it is necessary to further obtain the user's supplementary intention and generate a target point of interest according to the user's supplementary intention to achieve further screening of the point of interest. Specifically, the navigation system will expand the search scope or adjust the search parameters according to the user's supplementary intention to find a result that better meets the user's needs.

[0193] If the initially recognized point of interest initiated by the user is inaccurate, the user will output a user's point-of-interest rewriting instruction, that is, give a second-round instruction to correct the initial point of interest. Specifically, the navigation system rewrites the initial point of interest according to the point-of-interest rewriting instruction to obtain the target point of interest to provide a more accurate result.

[0194] Based on the user's secondary point-of-interest correction mechanism, a closed-loop learning is constructed by capturing the user's behavior. When the first recognition fails, the user's rewriting behavior is automatically recorded and targeted training data is generated to achieve continuous optimization based on the real scenario, enabling the error correction mechanism to have the ability of dynamic evolution.

[0195] Breaking through the static limitation of the traditional hot word library, by systematically constructing confusing voice interference items, a labeled data set covering more than 95% of common speech recognition traps is constructed, and a target navigation system is trained through the labeled data set. Based on the target navigation model, context reasoning is realized. Experimental data shows that the navigation method provided by the embodiments of the present application improves the point-of-interest recognition accuracy by 27.8% in complex scenarios, and especially improves the correction success rate by 41.5% when dealing with compound speech errors (such as homophones, mixed front and back nasal sounds).

[0196] Therefore, the navigation method provided by the embodiments of the present application ensures the accuracy of interest point recognition through a multi-level verification mechanism (such as the confirmation of the initial interest point, the analysis of user feedback instructions, etc.). When the initial interest point is incorrect, it can be corrected through the rewriting instructions provided by the user, thereby avoiding navigation deviations caused by speech recognition errors or unclear user expressions. In addition, when the user is not satisfied with the initial interest point, the system can obtain the supplementary intention of the user and generate a target interest point that better meets the user's needs. It effectively solves the problems of complex user needs or inaccurate initial recognition, and further improves the navigation reliability.

[0197] In the above embodiments, the method for training a voice navigation model is described in detail. The present application also provides an embodiment corresponding to a voice navigation model training device.

[0198] Figure 4 As shown in the structure diagram of a voice navigation model training device provided by the embodiments of the present application, Figure 4 as shown, the device includes:

[0199] A language database construction module 40 for constructing a language database; the language database includes a core interest point vocabulary, a list of confused word mappings, and a correction sentence pattern template library for correcting navigation interest points;

[0200] A labeled dataset generation module 41 for generating a labeled dataset by calling a specified model based on a pre-constructed prompt engineering and language database;

[0201] An iterative training module 42 for iteratively training the labeled dataset by inputting it into a pre-constructed navigation model to obtain a target navigation model; the target navigation model is used for processing voice navigation instructions output by the user; the instruction processing includes initial interest point recognition, initial interest point correction, determination of the target interest point, and generation of a navigation path.

[0202] In addition, the voice navigation model training device provided by the embodiments of the present application further includes:

[0203] A navigation interest point information collection module for collecting navigation interest point information from at least one map service database;

[0204] A keyword extraction module for extracting keywords from the navigation interest point information; and constructing a core interest point vocabulary based on the keywords;

[0205] A target word extraction module for extracting target words from the core interest point vocabulary; there are confused words for the target words, and the confused words include at least one of polyphonic words, homophonic words, synonyms, front and back nasal sound confused words, and flat and rolled tongue sound confused words;

[0206] A mapping relationship construction module for constructing a mapping relationship between a target word and a confusing word to obtain a confusing word mapping list.

[0207] A basic correction sentence pattern acquisition module for acquiring a pre-constructed basic correction sentence pattern;

[0208] A sentence pattern generalization module for generalizing the basic correction sentence pattern by adding preset qualifiers to obtain a first generalized sentence pattern, and / or generalizing the basic correction sentence pattern through a preset large model to obtain a second generalized sentence pattern;

[0209] A correction sentence pattern template library construction module for constructing a correction sentence pattern template library with the basic correction sentence pattern, the first generalized sentence pattern, and the second generalized sentence pattern.

[0210] A format verification module for removing data that does not conform to the specified format in the labeled dataset through regular expressions;

[0211] A pinyin verification module for verifying the pinyin of the labeled dataset through a pinyin conversion library to remove data with incorrect pinyin;

[0212] A core confusing word setting module for setting core confusing words for the target words in the labeled dataset; the core confusing words belong to the confusing words;

[0213] A length verification module for removing data in the labeled dataset where the length of the original point of interest is not within the preset length range;

[0214] A data content verification module for removing data with incorrect content in the labeled dataset; the incorrect content includes non-Chinese, non-conventional place names, and incorrect codes.

[0215] A target word position verification module for removing data in the labeled dataset where the target word is located at the first position of the original point of interest;

[0216] A pinyin letter number difference verification module for removing data where the difference in the number of pinyin letters between the replacement word and the target word is greater than 1;

[0217] A sentence pattern structure verification module for removing data in the correction prompt where the correction sentence pattern does not conform to the preset language sentence pattern structure.

[0218] A target word expansion module for expanding the target word;

[0219] A correction sentence pattern expansion module for inserting interference words into the correction sentence pattern in the correction prompt to expand the correction sentence pattern;

[0220] A special character verification module for removing data with special characters through regular expressions;

[0221] A similarity verification module is used to calculate the semantic similarity between the data in the labeled dataset; and select any one of the data combinations with a semantic similarity greater than the threshold.

[0222] In the above embodiments, the navigation method has been described in detail. The present application also provides an embodiment corresponding to a vehicle-mounted navigation system.

[0223] Figure 5 As shown in the structure schematic diagram of a vehicle-mounted navigation system provided by an embodiment of the present application, Figure 5 as shown, the device includes:

[0224] Each module in the voice navigation model training device 50 in the above embodiments;

[0225] A voice navigation instruction acquisition module 51 is used to acquire the voice navigation instructions output by the user;

[0226] A model call module 52 is used to call a target navigation model so that the target navigation model determines a target point of interest according to the voice navigation instructions;

[0227] A navigation path generation module 53 is used to generate a navigation path according to the target point of interest.

[0228] The embodiment of the present application corresponds to the navigation method in the above embodiments. For specific descriptions, reference can be made to the above embodiments, which will not be elaborated here for the time being.

[0229] It should be noted that although the operations are depicted in a specific order in the drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all the illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. In addition, the separation of various system modules and components in the above embodiments should not be construed as required in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

Claims

1. A method for training a voice navigation model, characterized in that, The method includes: Constructing a language database; the language database includes a core point-of-interest vocabulary, a list of homophone mappings, and a template library of corrected sentences for correcting navigation points of interest; Based on a pre-constructed prompt engineering and the language database, calling a specified model to generate an annotated data set; Inputting the annotated data set into a pre-constructed navigation model for iterative training to obtain a target navigation model; the target navigation model is used to process voice navigation instructions output by a user; the instruction processing includes initial point-of-interest recognition, correction of the initial point of interest, and determination of a target point of interest.

2. The method for training a voice navigation model according to claim 1, wherein The constructing of the language database includes: Collecting navigation point-of-interest information from at least one map service database; Extracting keywords from the navigation point-of-interest information; and constructing the core point-of-interest vocabulary according to the keywords; Extracting target words from the core point-of-interest vocabulary; there are homophones for the target words, and the homophones include at least one of polyphones, homophones, synonyms, words with confused front and back nasal sounds, and words with confused flat and curled tongue sounds; Constructing a mapping relationship between the target words and the homophones to obtain the list of homophone mappings.

3. The voice navigation model training method according to claim 1, wherein, The constructing of the language database includes: Obtaining a pre-constructed basic corrected sentence pattern; Generalizing the basic corrected sentence pattern by adding preset qualifiers to obtain a first generalized sentence pattern, and / or generalizing the basic corrected sentence pattern through a preset large model to obtain a second generalized sentence pattern; Constructing the template library of corrected sentence patterns with the basic corrected sentence pattern, the first generalized sentence pattern, and the second generalized sentence pattern.

4. The method for training a voice navigation model according to claim 2, wherein The data generation rules in the prompt engineering include rules for generating the annotated data set based on a preset data format, and the preset data format consists of an original point of interest, a correction prompt, and a corrected point of interest; The original point of interest is generated according to the core point-of-interest vocabulary; The correction prompt is generated according to the list of homophone mappings and the template library of corrected sentence patterns.

5. The method for training a voice navigation model according to claim 4, wherein The constraints in the prompt engineering include at least one of the following: At least one of the target words is included in the original point of interest; The replacement word is specified in the correction prompt; the replacement word is used to replace the target word, and the replacement word belongs to the homophones; The replacement word and the target word are homophones with different tones; The corrected point of interest conforms to a preset place name structure.

6. The method for training a voice navigation model according to claim 2, wherein Before inputting the annotated data set into a pre-constructed navigation model for iterative training, it also includes at least one of the following steps: Removing data that does not conform to the specified format in the annotated data set through regular expressions; Verifying the pinyin of the annotated data set through a pinyin conversion library to remove data with incorrect pinyin; Setting core homophones for the target words in the annotated data set; the core homophones belong to the homophones; Removing data in the annotated data set where the length of the original point of interest is not within the preset length range; Removing data with incorrect content in the annotated data set; the incorrect content includes non-Chinese, non-conventional place names, and incorrect codes.

7. The method for training a voice navigation model according to claim 5, wherein Before inputting the labeled dataset into a pre-constructed navigation model for iterative training to obtain a target navigation model, at least one of the following steps is further included: Eliminate the data in the labeled dataset where the target word is at the first position of the original point of interest; Eliminate the data where the difference in the number of pinyin letters between the replacement word and the target word is greater than 1; Eliminate the data in the correction prompt where the corrected sentence pattern does not conform to the preset language sentence pattern structure.

8. The method for training a voice navigation model according to claim 5, wherein Before inputting the labeled dataset into a pre-constructed navigation model for iterative training to obtain a target navigation model, at least one of the following steps is further included: Expand the target word; Insert interfering words into the corrected sentence pattern of the correction prompt to expand the corrected sentence pattern; Remove the data with special characters through regular expressions; Calculate the semantic similarity between the data in the labeled dataset; and select any one of the data combinations with the semantic similarity greater than the threshold.

9. A navigation method, characterized in that, Applied to a navigation system including a target navigation model, the target navigation model is obtained by training through the voice navigation model training method according to any one of claims 1 to 8, and the navigation method includes: Obtain the voice navigation instruction output by the user; Recognize the voice navigation instruction to obtain the recognition result of the initial point of interest; Determine whether the initial point of interest is correct; If it is correct, obtain the feedback instruction of the user; and determine whether the user is satisfied with the initial point of interest according to the feedback instruction. If satisfied, use the initial point of interest as the target point of interest; if not satisfied, obtain the supplementary intention of the user; and generate the target point of interest according to the supplementary intention of the user; If it is incorrect, obtain the instruction for rewriting the point of interest of the user; and rewrite the initial point of interest according to the instruction for rewriting the point of interest to obtain the target point of interest; Generate a navigation path according to the target point of interest.

10. A voice navigation model training device, characterized in that, The device includes: A language database construction module for constructing a language database; the language database includes a core vocabulary library of points of interest, a mapping list of easily confused words, and a correction sentence pattern template library for correcting navigation points of interest; A labeled dataset generation module for generating a labeled dataset by calling a specified model based on a pre-constructed prompt engineering and the language database; An iterative training module for inputting the labeled dataset into a pre-constructed navigation model for iterative training to obtain a target navigation model; the target navigation model is used for processing the voice navigation instruction output by the user; the instruction processing includes initial point of interest recognition, correction of the initial point of interest, determination of the target point of interest, and generation of a navigation path.

11. A vehicle-mounted navigation system, characterized in that, The in-vehicle navigation system includes: Each module in the voice navigation model training device according to claim 10; A voice navigation instruction acquisition module for acquiring the voice navigation instruction output by the user; A model calling module for calling the target navigation model so that the target navigation model determines the target point of interest according to the voice navigation instruction; A navigation path generation module for generating a navigation path according to the target point of interest.

Images