POI (Point of Interest) slot extraction model training method and device, navigation method, system and equipment

Abstract

The invention discloses a POI slot extraction model training method and device, and a navigation method, system and equipment. The training method comprises the following steps: acquiring a first data set and a second data set; the first data set comprises question and answer pairs formed between the user navigation instruction and the corresponding expected output; the second data set comprises a fuzzy navigation instruction; generating a target data set according to the first data set and the second data set; the target data set comprises a corresponding relation among question answer pairs, fuzzy navigation instructions and POI slot values; inputting the target data set into a pre-constructed target model for iterative training to obtain a POI slot extraction model; the POI slot extraction model is used for carrying out POI slot extraction on the navigation instruction output by the user. Therefore, the first data set provides the basis of context semantic understanding, and the second data set provides the basis of anaphora resolution, so that the POI slot extraction model has the capabilities of context semantic understanding and fuzzy navigation instruction analysis at the same time, the POI precision slot extraction is realized, and the navigation precision is improved.

Description

Technical Field

[0001] The present application relates to the technical field of navigation, and in particular, to a method and device for training a POI slot extraction model, a navigation method, a system, and a device. Background Art

[0002] With the rapid development of intelligent navigation, users have higher and higher requirements for navigation accuracy. Especially in multi-turn dialogue scenarios, how to efficiently and accurately perform slot extraction on the points of interest (POIs) in navigation input instructions, that is, POI slot extraction, is one of the crucial research directions in navigation technology.

[0003] Currently, in the navigation scenario of multi-turn dialogue, the navigation system mainly performs navigation by simply splicing the context content of the navigation dialogue or processing the dialogue round by round, and can only process simple instructions during the processing, and cannot understand fuzzy instructions including reference relationships (such as he and she, etc.) and ordinal numbers (such as the first, the second, etc.). It can be seen that the current navigation system lacks context semantic reasoning ability and fuzzy reference resolution ability, and it is difficult to achieve coherent navigation requirement understanding, resulting in low POI slot extraction accuracy, that is, low accuracy in determining the user's target POI, thus seriously affecting the user's navigation experience.

[0004] Therefore, how to perform context semantic reasoning, achieve reference resolution, accurately understand fuzzy instructions in multi-turn dialogue, and improve POI slot extraction accuracy, thereby improving navigation accuracy, is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, one aspect of the present application provides a method for training a POI slot extraction model, and the method includes:

[0006] Obtain a first data set and a second data set; the first data set includes question-answer pairs formed by user navigation instructions and corresponding expected outputs; the second data set includes fuzzy navigation instructions;

[0007] Generate a target data set according to the first data set and the second data set; the target data set includes the corresponding relationship between the question-answer pairs, the fuzzy navigation instructions, and POI slot values;

[0008] Input the target data set into a pre-constructed target model for iterative training to obtain a POI slot extraction model; the POI slot extraction model is used to perform POI slot extraction on the navigation instructions output by the user to obtain target POI slot values.

[0009] Optionally, the obtaining the first data set and the second data set includes:

[0010] Obtain the pre-built Q&A prompt engineering and fuzzy instruction prompt engineering;

[0011] Construct a core keyword library for points of interest;

[0012] Based on the Q&A prompt engineering and the core keyword library for points of interest, call a specified model to generate the first dataset; based on the fuzzy instruction prompt engineering, call the specified model to generate the second dataset.

[0013] Optionally, generating a target dataset according to the first dataset and the second dataset includes:

[0014] Obtain an annotated dataset after annotating the corresponding relationship and a pre-built target prompt engineering;

[0015] Fine-tune and train a specified model through the annotated dataset to obtain a desired model;

[0016] Based on the target prompt engineering, input the first dataset and the second dataset into the desired model in batches to generate the target dataset; in the batch input, a preset time interval is set between the current batch input and the previous batch input.

[0017] Optionally, the data generation rule in the target prompt engineering includes: outputting the POI slot value in the form of a specified slot structure;

[0018] The specified slot structure includes a city-level slot, a regional-level slot, and a specific point of interest slot;

[0019] The city-level slot is used to output city-level fields, the regional-level slot is used to output regional-level fields, and the specific point of interest slot is used to output specific point of interest fields.

[0020] Optionally, the target POI slot value includes at least one of the city-level field, the regional-level field, and the specific point of interest field;

[0021] Among them, the geographical range corresponding to the city-level field is larger than the geographical range corresponding to the regional-level field, and the geographical range corresponding to the regional-level field is larger than the geographical range corresponding to the specific point of interest field.

[0022] Optionally, the constraint condition in the target prompt engineering includes: meeting a preset slotting rule; the preset slotting rule includes at least one of the following:

[0023] When the POI slot value is empty, enter the chatting mode and / or generate a feedback instruction;

[0024] In the question-answer pair, when there are multiple answers and the navigation intention of the fuzzy navigation instruction is not clear, the answer at the first specified position in the answer output order is used as the POI slot value;

[0025] When the navigation intention is clear and the number of navigation intentions is multiple, any answer corresponding to one of the navigation intentions is used as the POI slot value;

[0026] When the navigation intention is clear and the navigation intention does not exist in the answer, the answer at the second specified position in the answer output order is used as the POI slot value.

[0027] Another aspect of the present application provides a navigation method, which is applied to a navigation system including a POI slotting model. The POI slotting model is obtained by training through the above-mentioned POI slotting model training method. The method includes:

[0028] Obtain the navigation instruction output by the user;

[0029] Perform POI slotting on the navigation instruction to obtain a target POI slot value;

[0030] Generate a navigation path according to the target POI slot value; and perform navigation based on the target path selected by the user.

[0031] Optionally, performing POI slotting on the navigation instruction includes:

[0032] Determine whether the current is a deep POI navigation mode according to the navigation instruction; the deep POI navigation mode is a mode in which the target POI slot value cannot be determined according to the navigation instruction output by the user for the first time;

[0033] If so, obtain multi-round navigation instructions; when there is a network search requirement for the multi-round navigation instructions, generate a search result; and perform POI slotting on the multi-round navigation instructions based on the search result.

[0034] Another aspect of the present application provides a POI slotting model training device, and the device includes:

[0035] A data set acquisition module, configured to acquire a first data set and a second data set; the first data set includes a question-answer pair formed by a user navigation instruction and a corresponding expected output; the second data set includes fuzzy navigation instructions;

[0036] A target data set generation module, configured to generate a target data set according to the first data set and the second data set; the target data set includes the corresponding relationship between the question-answer pair, the fuzzy navigation instruction and the POI slot value;

[0037] An iterative training module for iteratively training the target dataset using a pre-constructed target model to obtain a POI slotting model; the POI slotting model is used to perform POI slotting on the navigation instructions output by the user to obtain target POI slot values.

[0038] Another aspect of the present application provides a vehicle-mounted navigation system, including:

[0039] Any one of the modules in the POI slotting model training device described above;

[0040] A navigation instruction acquisition module for acquiring navigation instructions output by the user;

[0041] A model calling module for calling the POI slotting model to process the navigation instructions to determine target POI slot values;

[0042] A navigation initiation module for generating a navigation path based on the target POI slot values; and performing navigation based on the target path selected by the user.

[0043] Another aspect of the present application provides an electronic device, including a memory and a processor, where a computer program that can run on the processor is stored on the memory, and when the processor executes the program, it implements the steps of the POI slotting model training method and / or the steps of the navigation method described above.

[0044] The beneficial effects of the POI slotting model training method, device, navigation method, system, and device provided by the present application are as follows: The first dataset provides a data basis for context semantic understanding, and the second dataset provides a data basis for anaphora resolution, thereby ensuring that the target dataset simultaneously has the capabilities of context semantic understanding and fuzzy instruction understanding. As a result, the POI slotting model trained using the target dataset simultaneously has the capabilities of context semantic understanding and fuzzy navigation instruction parsing. When the user performs navigation based on the POI slotting model, there is no need to precisely express the point of interest. Even if a fuzzy instruction is issued, accurate POI slotting can still be performed, thereby improving navigation accuracy and enhancing the convenience and experience of the user during the navigation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic flowchart of a POI slotting model training method provided by an embodiment of the present application;

[0046] Figure 2 It is a schematic flowchart of a navigation method provided by an embodiment of the present application;

[0047] Figure 3 It is a schematic diagram of the principle of a navigation method provided by an embodiment of the present application;

[0048] Figure 4 A structural schematic diagram of a POI slot extraction model training device provided by an embodiment of the present application;

[0049] Figure 5 A structural schematic diagram of a vehicle navigation system provided by an embodiment of the present application;

[0050] Figure 6 A structural schematic diagram of an electronic device provided by an embodiment of the present application.

[0051] The reference numerals are as follows: 60 is a memory, 61 is a processor, 62 is a display screen, 63 is an input / output interface, 64 is a communication interface, 65 is a power supply, 66 is a communication bus, 601 is a computer program, 602 is an operating system, and 603 is data. Detailed implementation manners

[0052] The terms used in the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "the", and "said" used in the present 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 or all possible combinations of one or more of the associated listed items.

[0053] It should be understood that although the terms first, second, third, etc. may be used in the present application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present 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 determining".

[0054] Figure 1 A flowchart of a POI slot extraction model training method provided by an embodiment of the present application, as Figure 1 shown, the method includes:

[0055] S10: Obtain a first data set and a second data set; the first data set includes question-answer pairs formed between user navigation instructions and corresponding expected outputs; the second data set includes fuzzy navigation instructions;

[0056] In a specific embodiment, in order to obtain a POI slot extraction model with context semantic understanding ability and fuzzy instruction understanding ability, the data set for training the model is crucial. In an optional embodiment, a first data set including question-answer pairs is obtained, where the question refers to the user navigation instructions output by the user, and the answer is the expected output corresponding to the user navigation instructions.

[0057] It is understandable that the question-answer pairs in the first dataset can provide a data basis for the model's context semantic understanding. That is, during the training process, the model can learn the semantic relationship between the question and the answer, so that the trained model has the ability of context semantic understanding.

[0058] In addition, in a specific embodiment, a second dataset including fuzzy navigation instructions needs to be obtained. The second dataset can provide a data basis for the model to parse fuzzy instructions. Among them, fuzzy navigation instructions refer to instructions with unclear points of interest and / or ambiguous meanings. Fuzzy navigation instructions include, but are not limited to, reference instructions, ordinal instructions, incomplete instructions, and polysemous instructions. Among them, reference instructions refer to instructions including pronouns such as "you", "I", "he", "she", "it", "this", "this", "that", and "that". For example, "Go to this store" is a reference instruction. Ordinal instructions refer to instructions including ordinal numbers. For example, the instruction "Go to the first store" includes the ordinal number "first". Incomplete instructions refer to instructions for which the specific point of interest cannot be determined. For example, "Go to the airport". Polysemous instructions refer to instructions with multiple meanings in the same sentence. For example, "Go to the center" may refer to different places such as "shopping center" and "sports center".

[0059] It should be noted that in an optional embodiment, the first dataset and the second dataset can be obtained through training with a large language model, and the large language model selected in this application is not limited. Of course, it can also be obtained through manual annotation. In another optional embodiment, the actual navigation operation records of the user can be obtained as the first dataset and the second dataset, or the data generated from the simulated navigation scenario can be sorted out to obtain the first dataset and the second dataset. For the second dataset, a fuzzy instruction generation algorithm can also be designed to simulate and generate fuzzy navigation instructions, so as to form the second dataset. This application does not limit the acquisition methods of the first dataset and the second dataset.

[0060] S12: Generate a target dataset according to the first dataset and the second dataset; the target dataset includes the corresponding relationship between question-answer pairs, fuzzy navigation instructions, and POI slot values;

[0061] Furthermore, generating a target dataset according to the obtained first dataset and second dataset, it is understandable that based on the context semantic understanding basis of the first dataset and the fuzzy navigation instructions of the second dataset, a rich target dataset including question-answer pairs, fuzzy navigation instructions, and POI slot values can be generated.

[0062] Specifically, for the question-answer pairs in the first dataset, analyze the key information in the navigation instructions and extract the slot values related to POIs. For example, in the instruction "go to the nearest hot pot restaurant", "hot pot restaurant" is the POI slot value. For the fuzzy navigation instructions in the second dataset, try to determine the possible POI slot values or mark the POI categories that the instruction may involve through methods such as context analysis and subsequent clarification instructions from the user. Associate these POI slot values with the question-answer pairs and fuzzy navigation instructions to form the target dataset.

[0063] It should be noted that the generation of the target dataset can also be achieved through large language models. For example, it can be generated through the Qwen-72B large model. The application does not limit the method for generating the target dataset.

[0064] S13: Input the target dataset into a pre-constructed target model for iterative training to obtain a POI slot extraction model; the POI slot extraction model is used to perform POI slot extraction on the navigation instructions output by the user to obtain the target POI slot value.

[0065] In an optional embodiment, after obtaining the target dataset, in order to improve the training effect of the POI slot extraction model, match the initially generated target dataset with the Json format using regular expressions, that is, unify the format of the target dataset, thereby screening out the data that meets the format requirements and providing a standardized data basis for subsequent processing. In addition, since there may be dirty data in the target dataset, after obtaining the initial target data, the data can also be integrated and cleaned to improve the data quality of the training dataset. Data cleaning can include, but is not limited to, removing duplicate data, deleting data with missing important fields, standardizing the format, unifying the fields, and manual verification.

[0066] After obtaining the preprocessed target dataset, in order to obtain a POI slot extraction model with context semantic understanding ability and fuzzy navigation instruction understanding ability, a target model is pre-constructed. The target model can be a natural language processing model based on deep learning. For example, a model based on the Transformer architecture, or it can also be the Qwen-72B large model. The application does not limit the target model.

[0067] Furthermore, input the target dataset into the target model for iterative training until the iteration condition is reached, thereby obtaining the POI slot extraction model. It can be understood that during the training process of the target model, the model will continuously learn the language features in the navigation instructions in the target dataset, the extraction rules of POI slot values, and how to process fuzzy navigation instructions. Through continuous iterative training and optimizing the parameters of the model, the final POI slot extraction model can accurately extract the POI slot values from various navigation instructions.

[0068] It should be noted that the POI slot extraction model training method and the generated POI slot extraction model in the embodiments of the present application can both be applied to vehicles to achieve accurate POI slot extraction for vehicle navigation. Among them, vehicles include but are not limited to sedans, sport utility vehicles (SUVs), multi-purpose vehicles (MPVs), off-road vehicles, pickup trucks, or other power-driven non-railborne vehicles.

[0069] When the POI slot extraction model is applied to actual vehicle navigation, when the user outputs a navigation instruction, the POI slot extraction model can quickly and accurately perform POI slot extraction to obtain the target POI slot value, thereby providing a more accurate navigation service for the user.

[0070] In an optional embodiment, when training the target model, by introducing a low-rank matrix and using distributed training to fine-tune the target model, a trained POI slot extraction model is obtained. Specifically, the DeepSpeed distributed training framework is used and the ZeRO-3 optimizer is enabled to implement a multi-dimensional parallel training strategy. This framework effectively alleviates the video memory pressure in large-scale training through optimizer state partitioning, gradient sharing, and parameter sharding techniques. Combining a hybrid strategy of data parallelism, optimizer parallelism, and model parallelism, a training environment is established on an 8*A100 GPU cluster, and cross-node communication optimization is achieved through the AllReduce algorithm.

[0071] During the training process of the target model, the parameters are continuously adjusted. Specifically, based on the LoRA method, incremental updates are implemented for the selected weight matrices (Q_proj or V_proj). The incremental update formula is ΔW = BA, where W is the incremental update matrix, B is one of the low-rank matrices, and B ∈ R^(d×r), and R^(d×r) is used to represent the shape of matrix B. d is the number of rows of the incremental update matrix W, corresponding to the feature dimension. r is the rank of matrix B, which is a hyperparameter used to control the complexity of low-rank decomposition. In an optional embodiment, the rank r = 8. A is another low-rank matrix, and A ∈ R^(r×k), and R^(r×k) is used to represent the shape of matrix A. k is the number of columns of the incremental update matrix W, corresponding to the output dimension. While freezing the original parameter W ∈ R^(d×k), only the low-rank matrix with a new parameter proportion of about 0.1% needs to be trained, significantly reducing the video memory consumption.

[0072] The AdamW optimizer is used in combination with BF16 mixed-precision training, and the micro-batch size is set to 8 / GPU. The strategy of initializing B with a zero matrix and initializing A with a random Gaussian distribution ensures training stability. At the same time, gradient clipping with max_grad_norm = 1.0 is applied to prevent gradient explosion.

[0073] After the above initialization configuration is completed, the target model performs forward and backward propagation. Before each propagation, with the help of the all-gather algorithm AllReduce, the weights are collected from each GPU and concatenated into complete parameters to ensure model consistency. After the parameter combination is completed, each GPU independently performs forward propagation calculation and backward propagation calculation based on local data and the complete model parameters, and outputs independent gradients. In addition, using the all-gather algorithm, the gradients calculated by each GPU are transmitted to other GPUs, and the average value of the gradients of all GPUs is calculated by aggregation. Each GPU updates the local model parameters according to this gradient mean value.

[0074] In iterative training, before each forward and backward propagation, the parameters are aggregated again, and this cycle of iteration continues until a preset training stop condition is reached, for example, the loss converges or the preset number of iterations is reached. After training is completed, a weight merging operation is performed, adding the parameters of matrix A and matrix B to W0, that is, on the basis of the original weight matrix W0, adding the parameters of matrix A and matrix B to complete the fine-tuning training of the target model and obtain a POI slotting model with optimized performance and efficient inference.

[0075] After training is completed, weight fusion is performed: W' = W0 + αBA (scaling factor α = 1 / r), merging the low-rank adapter parameters into the base model. In a specific embodiment, no additional calculation branches are required during the inference process, obtaining a fine-tuning effect while maintaining the original inference speed. For ease of understanding, the following will illustrate the output results of the target POI slot values of the POI slotting model during multiple rounds of user navigation conversations, where History is the historical conversation content, that is, the question-answer pair, and Query is the fuzzy navigation instruction.

[0076] For example, {"History": ["What are the delicious foods in Ningbo", "The following are the delicious places in Ningbo:\n\n1. Gangyagou: A century-old brand, very popular. Its glutinous rice balls are made from high-quality raw materials and are secret-made by traditional techniques, with a soft and glutinous texture and a sweet but not greasy filling.\n\n2. Yuyuan Ningbo Glutinous Rice Ball Shop: The creative 'Snake Turns Fortune' glutinous rice balls are cute and delicious. In addition, there are also Four Seasons Peace dumplings, vegetarian 'big red envelopes' and other delicacies."], "Query": "Navigate to this place"}. Target POI slot values: {"City": "Ningbo", "CenterLocation": "", "Keyword": "Gangyagou"}

[0077] Therefore, for the POI slotting model training method provided by the embodiments of this application, the first dataset provides a data basis for context semantic understanding, and the second dataset provides a data basis for anaphora resolution. Furthermore, it ensures that the target dataset has the capabilities of both context semantic understanding and fuzzy instruction understanding. As a result, the POI slotting model trained through the target dataset has both the context semantic understanding ability and the fuzzy navigation instruction parsing ability. When a user navigates based on the POI slotting model, there is no need to precisely express the point of interest. Even if a fuzzy instruction is issued, accurate POI slotting can still be performed, thereby improving navigation accuracy and enhancing the convenience and experience of the user during the navigation process.

[0078] In an alternative embodiment, obtaining the first dataset and the second dataset includes:

[0079] Obtaining a pre-constructed question-and-answer prompt engineering and a fuzzy instruction prompt engineering;

[0080] Constructing a core keyword library for points of interest;

[0081] Based on the question-and-answer prompt engineering and the core keyword library for points of interest, calling a specified model to generate the first dataset; based on the fuzzy instruction prompt engineering, calling a specified model to generate the second dataset.

[0082] For the generation of a complex target dataset, to ensure the instructions of the target dataset, this application decomposes the task of constructing the dataset based on the context prompt learning technology of the chain of thought. Specifically, in a specific embodiment, a core keyword library for points of interest (POI) is constructed. Specifically, POI information of different categories such as food, scenic spots, entertainment, hotels, life entertainment, and sports and fitness can be obtained from an open-source map database, thereby constructing a POI core keyword library.

[0083] At the same time, obtain a pre-constructed question-and-answer prompt (Prompt) engineering and a fuzzy instruction Prompt engineering. The question-and-answer Prompt engineering can be used to generate question-answer pairs between user navigation instructions and corresponding expected outputs. For example, in an alternative embodiment, the question-and-answer Prompt engineering (QA-Prompt) is as follows:

[0084] QA-Prompt = """#Role

[0085] You are a professional travel Q&A expert. You are familiar with scenic spots, food, and entertainment places across the country and answer questions according to the user's questions:

[0086] Constraints:

[0087] 1. Only answer the user's questions and refrain from answering redundant content;

[0088] 2. If you don't know the user's city, the default city is randomly selected from Beijing, Shanghai, Nanjing, Ningbo, Xi'an, Hefei, Hangzhou, and Suzhou for answering;

[0089] 3. It is forbidden to output content that does not include navigation points of interest information. """

[0090] Example returns:

[0091] ## Example 1:

[0092] Q(Question): Are there any delicious Xiaolongbao restaurants in Shanghai?

[0093] A (Answer): There are many delicious xiaolongbao restaurants in Shanghai. Here are some recommendations: \n Nanxiang Mantou Restaurant: Located near the Lake Pavilion of Chenghuang Temple, it is a xiaolongbao restaurant with a long history. The xiaolongbao here has thin skin and large fillings, delicious soup, and a very authentic taste. \u200b\n Guqi Garden Restaurant: Located in Nanxiang Guqi Garden, its xiaolongbao is famous for its thin skin, rich juice, and delicious taste. The environment of the restaurant is also very beautiful, and it is a good place to taste xiaolongbao. \n Shanshan Xiaolongguan: There are many branches in Shanghai. The xiaolongbao here is small, thin, and the soup is sweet and fragrant. It is deeply loved by diners. \u200b\n Wanshouzhai: It is a time-honored xiaolongbao restaurant. Its xiaolongbao tastes delicious, the soup is rich, and the price is affordable. It is a memory of many Shanghainese. \n Jiajia Tangbao: This store has a wide variety of xiaolongbao, including fresh meat xiaolongbao, crab meat xiaolongbao, etc. It tastes salty and fresh, not sweet at all, which is completely different from the seasoning style of this area. \u200b\nLai Lai Xiao Long Bao: specializes in crab meat dumplings, with regular skin, stuffing and seasoning, and stable quality, enjoying a high reputation in the Shanghai Xiao Long Bao industry. \nThe above are some of the more famous Xiao Long Bao restaurants in Shanghai, of course there are many other small restaurants worth trying. You can choose according to your taste and preference.

[0094] ## Example 2:

[0095] Q: Please introduce Shanghai City God Temple.

[0096] A: Shanghai City God Temple is located at No. 249, Fangbang Middle Road, Huangpu District, Shanghai. It is an important Taoist temple in Shanghai with the following characteristics: It was first built during the Yongle period of the Ming Dynasty (1403-1424) and was rebuilt from the original Jinshan Temple to enshrine Qin Yubo, the City God of Shanghai. It has a rich history and has undergone many repairs and reconstructions as well as various events in different periods. For example, it was constantly damaged by wars after the Daoguang period of the Qing Dynasty, and has been restored and opened many times in modern times. It has a history of nearly 600 years.

[0097] Further, based on the Q&A prompt engineering and the POI core vocabulary, a specified model is called to generate a first dataset. In a specific embodiment, the specified model will generate question-answer pairs between the user navigation instructions and the corresponding expected outputs according to the information in the Q&A Prompt engineering and the POI core vocabulary, forming a first dataset.

[0098] It should be noted that generating the first dataset based on the point of interest core vocabulary ensures that the place names in the first data conform to the geographical entity naming specification, laying a good foundation for the subsequent generation of target data. In an alternative embodiment, the specified model may be a pre-trained language model. Specifically, it may include, but is not limited to, models based on the Transformer architecture, such as the GPT series and BERT, etc. In an alternative embodiment, the specified model may be the Qwen-72B large model.

[0099] For the fuzzy instruction Prompt engineering, it is a Prompt designed for the fuzzy navigation instructions that users may issue, used to guide the model to understand and generate possible intents or clarification questions related to the fuzzy instructions. For example, in an alternative embodiment, the pre-constructed fuzzy instruction Prompt engineering (Instruction-Prompt) is:

[0100] Instruction-Prompt = """Please generate multiple short fuzzy navigation instructions, requiring concise language, conforming to the usage habits of daily language expressions, and being able to clearly express the user's operation intent for fuzzy navigation information in the navigation scenario, such as: "I want to go to the first one", "Go to this one", "Go to that one", and "Navigate to the first one", etc."""

[0101] Further, based on the pre-constructed fuzzy instruction Prompt engineering, similarly, a second dataset is generated through the specified model. Thus, the obtained first dataset and second dataset provide a data basis for the subsequent generation of a POI slot model with context semantic understanding and fuzzy instruction parsing.

[0102] In an alternative embodiment, according to the first dataset and the second dataset, a target dataset is generated, including:

[0103] Obtain an annotated dataset after annotating the corresponding relationship and a pre-constructed target prompt engineering;

[0104] Fine-tune and train the specified model through the annotated dataset to obtain an expected model;

[0105] Based on the target prompt engineering, input the first dataset and the second dataset into the expected model in batches to generate a target dataset; in the batch input, there is a preset time interval between the current batch input and the previous batch input.

[0106] In an alternative embodiment, the target data set can be generated by specifying a model to learn the first data set and the second data set. Similarly, the specified model can be the Qwen-72B large model, and the present application does not limit the specified model. In order to improve the quality of the target data set, in an alternative embodiment, after fine-tuning and training the specified model with the labeled data set, a high-quality desired model can be obtained. It can be understood that the labeled data set is an important basis for generating the target data set and provides accurate supervision signals for model training.

[0107] Specifically, a high-quality labeled data set can be obtained by manually annotating the correspondence between question-answer pairs, fuzzy navigation instructions, and POI slot values. Further, the specified model is fine-tuned and trained with the labeled data set to meet the requirements of specific tasks.

[0108] In an alternative embodiment, after fine-tuning and training the specified model with the labeled data set, a strict performance evaluation is performed on the specified model. When the model exceeds the threshold (e.g., 98%) accuracy rate on the validation set, it is determined that the model meets the expectations, thereby triggering a batch automatic annotation process for the unlabeled data set. Among them, the threshold accuracy rate ensures that the specified model has sufficient reliability, thereby reducing the workload of subsequent manual correction. In a specific embodiment, it can be set according to actual business requirements.

[0109] In order to further ensure the accuracy of the target data set generated by the specified model in the future, after the model meets the threshold accuracy rate requirement and completes the automatic annotation of the unlabeled data set, a random sampling check is performed on the automatic annotation results of the model to ensure that the annotation quality meets the expected standard, further guaranteeing the data quality and providing feedback information for the further optimization of the specified model.

[0110] Thus, the desired model is obtained by organically combining manual annotation and model annotation, making full use of the high efficiency of the model and the accuracy of manual annotation, reducing both the economic cost of manual annotation and ensuring the high-quality construction of the data set, laying a solid foundation for subsequent model training and optimization based on these data.

[0111] In an alternative embodiment, a target Prompt engineering for generating the target data set is pre-constructed. In a specific embodiment, the target Prompt engineering can guide the desired model to more accurately extract and generate the POI slot values and their corresponding relationships when processing navigation instructions.

[0112] Further, the first data set and the second data set are input into the desired model in batches to generate the target data set. For example, 200 pieces of data are input in each batch. Inputting in batches can control the load and resource consumption of data processing, and also helps the desired model to better process large-scale data sets. In the batch input, a preset time interval is set between the current batch input and the previous batch input. For example, the interval is 5 minutes. Setting the preset time interval of this cooling interval can effectively prevent the desired model from degrading, and ensure the quality and stability of the generated target data set. In addition, the preset time interval can be adjusted according to the processing capacity and resource limitations of the model. For example, if the processing capacity of the desired model is strong, the interval time can be appropriately shortened; if the resources are limited, the interval time can be appropriately extended to ensure the stable operation of the desired model.

[0113] Thus, a target data set containing question-answer pairs, fuzzy navigation instructions, and the corresponding relationship of POI slot values can be generated. The target data set not only contains rich navigation instruction samples, but also covers the accurate extraction of POI slot values and the processing logic of fuzzy instructions, providing high-quality data support for the subsequent development and optimization of the navigation system.

[0114] In an optional embodiment, the data generation rules in the target Prompt engineering include: outputting the POI slot value in the form of a specified slot structure; wherein, the specified slot structure includes a city-level slot, a regional-level slot, and a specific interest point slot; and the city-level slot is used to output city-level fields, the regional-level slot is used to output regional-level fields, and the specific interest point slot is used to output specific interest point fields.

[0115] In a specific embodiment, in order to solve problems such as incomplete POI elements in the process of POI slot extraction, a three-level address system including a city-level slot (represented by City), a regional-level slot (represented by CenterLocation), and a specific interest point slot (represented by Keyword) is constructed. Thus, when the POI slot extraction model outputs the target POI slot value, the extracted city-level fields (such as Shanghai, Beijing, Hangzhou) are output in the city-level slot, the regional-level fields (such as No. 234 Anlu Road) are output in the regional-level slot, and the specific interest point fields (such as People's Hospital) are output in the specific interest point slot.

[0116] It should be noted that the city-level slot is used to identify the larger geographical area or administrative area mentioned in the user's conversation, generally locations above the district level. For example, "Ningbo City" is a prefecture-level city, and "Hangzhou Bay New Area" is its subordinate administrative region, and "Ningbo Hangzhou Bay New Area" constitutes a complete description of the city-level field. The city-level slot helps the model determine the large-scale area where the user is located, and then provide more accurate local services and information.

[0117] The regional slot is used to record specific streets, house numbers or small geographic locations, usually places below the district level, such as streets, communities, buildings, etc. For example, "No. 234, Gui'an Road" is a specific address, which may point to a shop, residence or other type of building. The detailed information of the regional field in the regional slot is crucial for navigation, positioning services or finding a specific place.

[0118] The specific POI slot is used to capture specific destinations or POIs mentioned in user conversations, covering various places that users may want to visit or learn about. The specific POI field can be the name of a public facility, commercial organization, tourist attraction, hospital, restaurant, etc. For example, "People's Hospital" is a place that provides medical services, and "Gobeis Italian Restaurant" is a specific dining location. By identifying the specific POI field, the specific POI slot extraction model can grasp the user's intention and provide relevant information and services such as route planning and reservation services.

[0119] Therefore, correspondingly, in an optional embodiment, the target POI slot value finally output by the POI slot model obtained based on the target prompt word project includes: at least one of: a city-level field, a region-level field, and a specific point of interest field. Among them, the geographical range corresponding to the city-level field is larger than the geographical range corresponding to the region-level field, and the geographical range corresponding to the region-level field is larger than the geographical range corresponding to the specific point of interest field. For example, in an optional embodiment, the target prompt project (Poi-Prompt) is:

[0120] Poi-prompt = """#Character

[0121] You are a professional and efficient POI expert, good at accurately extracting POI information from historical conversations. You can deeply analyze the semantics of user conversations to ensure that the output POI information is accurate. You firmly put an end to fabrication and only extract the real information in the user's questions.

[0122] ##Constraints:

[0123] 1. It is forbidden to extract information that does not exist in the historical conversation.

[0124] 2. Output strictly according to the given designated slot structure to ensure that the content of each slot meets its defined requirements.

[0125] 3. Only deal with content related to POI and refuse to answer irrelevant questions.

[0126] 4. If the web search fails, the model replies "I'm sorry xx", and the extraction result is empty: {'City':",'CenterLocation':",'Keyword':"}

[0127] 5. Meet the preset slot extraction rules;

[0128] 6. Data generation rules:

[0129] (1) Output in Json dictionary format;

[0130] (1) Output the POI slot values in the following format:

[0131] {

[0132] "City": "", / / **Extract the larger geographical area or administrative region mentioned in the user's conversation, fill in the location above the district level, such as "Beijing", "Ningbo City", etc.**;

[0133] "CenterLocation": "", / / **The specific street, house number or small - scale geographical location, extract the location below the district level, such as streets, communities, buildings, etc.**;

[0134] "Keyword": "", / / **Extract the specific destination or point of interest mentioned in the user's conversation, such as the names of public facilities, commercial institutions, tourist attractions, hospitals, restaurants, etc.**;

[0135] }"

[0136] Thus, by specifying the slot structure, that is, the three - level address system, the POI is granularly parsed, multi - source data integration is achieved, fuzzy address parsing is supported, the problem of traditional navigation systems dealing with complex positioning requirements is solved. At the same time, a fault - tolerance mechanism allowing some fields (city - level fields, regional - level fields, and specific point - of - interest fields) to be missing is adopted to improve the navigation availability in complex scenarios.

[0137] On the basis of the above - mentioned embodiments, as an optional embodiment, the constraint conditions in the target prompt word engineering include: meeting the preset slot extraction rules; the preset slot extraction rules include at least one of the following four items:

[0138] 1. When the POI slot value is empty, enter the chatting mode and / or generate feedback instructions;

[0139] In an optional embodiment, when the POI slot extraction result is empty, that is, the city - level field, the regional - level field, and the specific point - of - interest field cannot be extracted. For example, the navigation conversation is as follows:

[0140] Q: What are the good hot - pot restaurants in Shanghai?

[0141] A: The Chaoniu Hot - Spicy Hot - Pot Restaurant in Wanda Plaza;

[0142] Q: Navigate to this place;

[0143] At this time, the POI slot value is empty, and the fallback chat model can be entered. At the same time, the result of the failed POI slot extraction can be used to generate a feedback instruction and returned to the client.

[0144] 2. In the question-answer pair, when the answer includes multiple items and the navigation intention of the ambiguous navigation instructions is not clear, the answer at the first specified position in the output order of the answers is used as the POI slot value.

[0145] For example, Q: What are the good hot pot restaurants in Shanghai?

[0146] A: 1. Chaoniu Hot Pot Restaurant in Wanda Plaza, 2. Tan Yaxue Hot Pot Restaurant, 3. Nan Hot Pot...

[0147] Q: Navigate to this one.

[0148] POI slot value: Chaoniu Hot Pot Restaurant in Shanghai Wanda Plaza.

[0149] In a specific embodiment, when the obtained answer includes multiple items, but when the user selects the final navigation POI, the instruction is ambiguous and the navigation intention cannot be determined. At this time, the answer at the first specified position can be selected as the POI slot value. In the above example, the answer at the first position is used as the POI slot value. In fact, the present application does not limit the first specified position, that is, it does not limit the default position.

[0150] 3. When the navigation intention is clear and the number of navigation intentions is multiple, any answer corresponding to one of the navigation intentions is used as the POI slot value.

[0151] For example, Q: What are the good hot pot restaurants in Shanghai?

[0152] A: 1. Chaoniu Hot Pot Restaurant in Wanda Plaza, 2. Tan Yaxue Hot Pot Restaurant, 3. Nan Hot Pot...

[0153] Q: Navigate to the second and third ones.

[0154] POI slot value: Shanghai Tan Yaxue Hot Pot Restaurant.

[0155] In a specific embodiment, the user's instruction finally includes multiple intentions. At this time, any one of the multiple intentions can be used as the navigation POI.

[0156] 4. When the navigation intention is clear and the navigation intention does not exist in the answer, the answer at the second specified position in the output order of the answers is used as the POI slot value.

[0157] For example: Q: What are the good hot pot restaurants in Shanghai?

[0158] A: 1. Chaoniu Hot Pot Restaurant in Wanda Plaza, 2. Tan Yaxue Hot Pot Restaurant, 3. Nan Hot Pot;

[0159] Q: Navigate to the tenth one;

[0160] POI slot value: Shanghai Nan Hot Pot.

[0161] In this example, the answer includes 3, but the user's navigation intention is the 10th, which obviously does not exist in the multiple answers. At this time, the answer at the second specified position in the output order of the answers is used as the POI slot value. For example, the last one can be used as the POI slot value. Similarly, this application does not make a specific limitation on the second specified position.

[0162] In the above embodiments, the POI slotting model training method has been described in detail. This application also provides an embodiment corresponding to a navigation method. This navigation method is applied to a navigation system including a POI slotting model, and the POI slotting model is obtained by training through the POI slotting model training method of the above embodiments.

[0163] Figure 2 It is a schematic flowchart of a navigation method provided by an embodiment of this application, as Figure 2 shown, the device includes:

[0164] S20: Obtain the navigation instruction output to the user;

[0165] S21: Perform POI slotting on the navigation instruction to obtain the target POI slot value;

[0166] Figure 3 It is a schematic principle diagram of a navigation method provided by an embodiment of this application, as Figure 3 shown, in specific implementation, after the user outputs a navigation instruction, the navigation system including the POI slotting model obtains the navigation instruction and performs POI slotting on the navigation instruction to determine the target POI slot value.

[0167] Specifically, in an optional embodiment, as Figure 2 shown, performing POI slotting on the navigation instruction includes:

[0168] S210: According to the navigation instruction, determine whether the current is the deep POI navigation mode; the deep POI navigation mode is a mode in which the target POI slot value cannot be determined based on the user's first output navigation instruction; if so, enter step S211;

[0169] In a specific embodiment, as Figure 3As shown in the figure, after the navigation system obtains a navigation instruction, it needs to first determine the current mode, where the modes include a navigation mode and a chat mode. It can be understood that when it is determined according to the instruction output by the user that the user has no navigation requirement or intention, the chat mode is entered at this time; otherwise, the navigation mode is entered. It should be noted that when a place name appears in the instruction, there is often a navigation intention, and the chat mode cannot be entered at this time. For example, "What are the interesting places in Shanghai?" After performing a web search for the user, the user may select one of the addresses for navigation, so the chat mode cannot be entered at this time.

[0170] In the navigation mode, it is determined whether the current navigation is in the deep POI navigation mode according to the navigation instruction. Among them, the deep POI navigation mode refers to the mode in which the navigation system cannot perform POI slotting to determine the POI slot value after the user outputs the navigation instruction for the first time. For example, the navigation instruction output for the first time is "What are the delicious restaurants in Shanghai?" In another alternative embodiment, in the navigation mode, if it is not the deep POI navigation mode, the simple POI navigation mode should be entered. Correspondingly, it can be understood that the simple POI navigation mode refers to the mode in which the navigation system can determine the POI slot value after the user outputs the navigation instruction for the first time. For example, "Navigate to Shanghai Railway Station".

[0171] S211: Obtain multi-round navigation instructions;

[0172] S212: Determine whether there is a network search requirement in the multi-round navigation instructions; if so, go to step S213;

[0173] S213: Generate search results; and based on the search results, perform POI slotting on the multi-round navigation instructions.

[0174] It can be understood that after entering the deep POI navigation mode, it is often necessary for the user to have multi-round conversations. At this time, multi-round navigation instructions are obtained, as Figure 3 shown, and it is determined whether there is a web search in the multi-round navigation instructions. If so, a web search is performed, and the search results are fed back to the POI slotting model. Of course, if there is no web search, the navigation instructions of the multi-round conversation are directly transmitted to the POI slotting model. Thus, POI slotting is achieved through the POI slotting model.

[0175] S22: Generate a navigation path according to the target POI slot value; and perform navigation based on the target path selected by the user.

[0176] Furthermore, after obtaining the target POI slot, the navigation system generates a navigation path based on the target POI slot value and initiates navigation based on the target path finally selected by the user.

[0177] Therefore, the navigation method provided in the embodiment of the present application solves the problem of fuzzy navigation quality of POIs across rounds in multi-round dialogue scenarios, as well as the problem of POI semantic discontinuity in web search scenarios. It supports multi-round dialogue backtracking, realizes cross-round information association through a dynamic attention mechanism, has a more reasonable understanding and association analysis of contextual information, and improves the naturalness and coherence of the dialogue. In addition, data collaboration is more efficient and cross-modal feature fusion is realized, while the prior art processes web search and dialogue data independently to meet the diverse navigation needs of users.

[0178] In the above embodiments, the POI slot model training method is described in detail. The present application also provides a corresponding embodiment of a POI slot model training device.

[0179] Figure 4 A schematic diagram of the structure of a POI slot model training device provided in an embodiment of the present application is shown in FIG. Figure 4 As shown, the device comprises:

[0180] The data set acquisition module 40 is used to acquire a first data set and a second data set; the first data set includes a question-answer pair formed by a user navigation instruction and a corresponding expected output; the second data set includes a fuzzy navigation instruction;

[0181] A target data set generation module 41 is used to generate a target data set according to the first data set and the second data set; the target data set includes a correspondence between question answer pairs, fuzzy navigation instructions and POI slot values;

[0182] The iterative training module 42 is used to input the target data set into the pre-built target model for iterative training to obtain a POI slot model; the POI slot model is used to perform POI slotting on the navigation instructions output by the user to obtain a target POI slot value.

[0183] In addition, the POI slot model training device provided in the embodiment of the present application also includes:

[0184] A prompt word project acquisition module, used to acquire pre-built question and answer prompt word projects and fuzzy instruction prompt word projects;

[0185] A core word library construction module is used to construct a core word library of interest points;

[0186] The data generation module is used to call the specified model to generate the first data set based on the question and answer prompt word project and the interest point core word library; and to call the specified model to generate the second data set based on the fuzzy instruction prompt word project.

[0187] A target acquisition module is used to acquire a labeled data set after the corresponding relationship is labeled and a pre-built target prompt word project;

[0188] The fine-tuning training module is used to fine-tune and train a specified model through an annotated dataset to obtain an expected model;

[0189] The target dataset generation module is further used to input the first dataset and the second dataset into the expected model in batches based on the target prompt engineering to generate a target dataset; in the batch input, a preset time interval is set between the current batch input and the previous batch input.

[0190] In the above embodiments, the navigation method has been described in detail. The present application also provides an embodiment corresponding to an in-vehicle navigation system.

[0191] Figure 5 As shown in the structural schematic diagram of an in-vehicle navigation system provided by an embodiment of the present application, Figure 5 as shown, the device includes:

[0192] Any one of the modules in the POI slot model training device 50 in the above embodiments;

[0193] The navigation instruction acquisition module 51 is used to acquire navigation instructions output to the user;

[0194] The model calling module 52 is used to call the POI slot model to process the navigation instructions to determine the target POI slot value;

[0195] The navigation initiation module 53 is used to generate a navigation path according to the target POI slot value; and perform navigation based on the target path selected by the user.

[0196] In addition, the in-vehicle navigation system provided by the embodiment of the present application further includes:

[0197] The navigation mode determination module is used to determine whether the current is the deep POI navigation mode according to the navigation instructions; the deep POI navigation mode is a mode in which the target POI slot value cannot be determined based on the navigation instructions initially output by the user; if so, call the processing module;

[0198] The processing module is used to acquire multi-round navigation instructions; when there is a network search requirement in the multi-round navigation instructions, generate a search result; and perform POI slotting on the multi-round navigation instructions based on the search result.

[0199] The embodiment of the present application corresponds to the navigation method in the above embodiment. For specific descriptions, reference can be made to the above embodiment, which will not be elaborated here.

[0200] Figure 6 As shown in the structural schematic diagram of an electronic device provided by an embodiment of the present application, Figure 6 as shown, the electronic device includes: a memory 60 for storing a computer program;

[0201] A processor 61, which is configured to implement the steps of the POI slot model training method and / or the navigation method as mentioned in the foregoing embodiments when executing a computer program.

[0202] The electronic device provided in this embodiment may include, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, etc.

[0203] Among them, the processor 61 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 61 may be implemented in at least one hardware form of a digital signal processor (DSP), a field-programmable gate array (FPGA), and a programmable logic array (PLA). The processor 61 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the wake state, also known as a central processing unit (CPU); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 61 may be integrated with a graphics processing unit (GPU), and the GPU is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 61 may further include an artificial intelligence (AI) processor, and the AI processor is used to process computing operations related to machine learning.

[0204] The memory 60 may include one or more computer-readable storage media, and the computer-readable storage media may be non-transitory. The memory 60 may further include a high-speed random access memory and a non-volatile memory, such as one or more disk storage devices and flash storage devices. In this embodiment, the memory 60 is at least used to store the following computer program 601. After the computer program is loaded and executed by the processor 61, it can implement the relevant steps of the POI slot model training method and / or the navigation method disclosed in any of the foregoing embodiments. In addition, the resources stored in the memory 60 may further include an operating system 602 and data 603, etc., and the storage method may be temporary storage or permanent storage. Among them, the operating system 602 may include Windows, Unix, Linux, etc. The data 603 may include, but is not limited to, relevant data involved in the POI slot model training method and / or the navigation method.

[0205] In some embodiments, the electronic device may further include a display screen 62, an input / output interface 63, a communication interface 64, a power supply 65, and a communication bus 66.

[0206] Those skilled in the art can understand that Figure 6 the structure shown in does not constitute a limitation on the electronic device, and it may include more or fewer components than those shown in the figure.

[0207] The electronic device provided by the embodiment of the present application includes a memory and a processor. When the processor executes the program stored in the memory, it can implement the POI extraction model training method and / or the navigation method in the above embodiments.

[0208] It should be noted that although the operations are depicted in a specific order in the drawings, this should not be construed as requiring the operations to be performed in the specific order shown or sequentially, or requiring all of 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 the various system modules and components in the above embodiments should not be understood 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 POI slotting model, characterized in that The method includes: Obtaining a first data set and a second data set; the first data set includes question-answer pairs formed by user navigation instructions and corresponding expected outputs; the second data set includes fuzzy navigation instructions; Generating a target data set according to the first data set and the second data set; the target data set includes the corresponding relationship between the question-answer pairs, the fuzzy navigation instructions, and the POI slot values; Inputting the target data set into a pre-constructed target model for iterative training to obtain a POI slot extraction model; the POI slot extraction model is used to perform POI slot extraction on the navigation instructions output by the user to obtain target POI slot values.

2. The POI extraction model training method according to claim 1, wherein The obtaining of the first data set and the second data set includes: Obtaining a pre-constructed question-and-answer prompt engineering and a fuzzy instruction prompt engineering; Constructing a POI core vocabulary; Based on the question-and-answer prompt engineering and the POI core vocabulary, calling a specified model to generate the first data set; based on the fuzzy instruction prompt engineering, calling the specified model to generate the second data set.

3. The POI extraction model training method according to claim 1, wherein The generating of the target data set according to the first data set and the second data set includes: Obtaining an annotated data set after annotating the corresponding relationship and a pre-constructed target prompt engineering; Fine-tuning and training a specified model through the annotated data set to obtain an expected model; Based on the target prompt engineering, inputting the first data set and the second data set into the expected model in batches to generate the target data set; in the batch input, a preset time interval is set between the current batch input and the previous batch input.

4. The POI slot model training method according to claim 3, wherein, The data generation rule in the target prompt engineering includes: outputting the POI slot value in the form of a specified slot structure; The specified slot structure includes a city-level slot, a regional-level slot, and a specific interest point slot; The city-level slot is used to output city-level fields, the regional-level slot is used to output regional-level fields, and the specific interest point slot is used to output specific interest point fields.

5. The POI extraction model training method according to claim 4, wherein The target POI slot value includes at least one of the city-level field, the regional-level field, and the specific interest point field; Wherein, the geographical range corresponding to the city-level field is larger than the geographical range corresponding to the regional-level field, and the geographical range corresponding to the regional-level field is larger than the geographical range corresponding to the specific interest point field.

6. The POI slot model training method according to claim 3, wherein, The constraint condition in the target prompt engineering includes: meeting a preset slot extraction rule; the preset slot extraction rule includes at least one of the following: When the POI slot value is empty, enter the chatting mode and / or generate a feedback instruction; In the question-answer pair, when there are multiple answers and the navigation intention of the fuzzy navigation instruction is not clear, take the answer at the first specified position in the answer output order as the POI slot value; When the navigation intention is clear and the number of navigation intentions is multiple, take the answer corresponding to any one of the navigation intentions as the POI slot value; When the navigation intention is clear and the navigation intention does not exist in the answer, take the answer at the second specified position in the answer output order as the POI slot value.

7. A navigation method, characterized in that, Applied to a navigation system including a POI slotting model, the POI slotting model is obtained by training through the POI slotting model training method described in any one of claims 1 to 6, and the method includes: Obtain navigation instructions output to the user; Perform POI slotting on the navigation instructions to obtain a target POI slot value; Generate a navigation path according to the target POI slot value; and perform navigation based on the target path selected by the user.

8. The navigation method according to claim 7, wherein, Performing POI slotting on the navigation instructions includes: Determine whether the current is a deep POI navigation mode according to the navigation instructions; the deep POI navigation mode is a mode in which the target POI slot value cannot be determined based on the navigation instructions first output by the user; If so, obtain multiple rounds of navigation instructions; when there is a network search requirement for the multiple rounds of navigation instructions, generate a search result; and perform POI slotting on the multiple rounds of navigation instructions based on the search result.

9. A POI grooving model training device, characterized in that, The device includes: A data set acquisition module for acquiring a first data set and a second data set; the first data set includes question-answer pairs formed between user navigation instructions and corresponding expected outputs; the second data set includes fuzzy navigation instructions; A target data set generation module for generating a target data set according to the first data set and the second data set; the target data set includes the corresponding relationship between the question-answer pairs, the fuzzy navigation instructions, and the POI slot values; An iterative training module for inputting the target data set into a pre-constructed target model for iterative training to obtain a POI slotting model; the POI slotting model is used to perform POI slotting on the navigation instructions output by the user to obtain a target POI slot value.

10. A vehicle-mounted navigation system, characterized in that, Includes: Any one of the modules in the POI slotting model training device described in claim 9; A navigation instruction acquisition module for acquiring navigation instructions output to the user; A model call module for calling the POI slotting model to process the navigation instructions to determine the target POI slot value; A navigation initiation module for generating a navigation path according to the target POI slot value; And perform navigation based on the target path selected by the user.

11. An electronic device, comprising a memory and a processor, wherein a computer program capable of running on the processor is stored on the memory, characterized in that When the processor executes the program, it implements the steps of the POI slotting model training method described in any one of claims 1 to 6, and / or the steps of the navigation method described in claim 7 or 8.

Images