A text intention recognition method, an electronic device, a storage medium and a program product

Abstract

The application discloses a text intention recognition method, an electronic device, a storage medium and a program product; the method comprises the following steps: obtaining a to-be-recognized user question; if there is no corresponding target natural language expression text of the to-be-recognized user question, determining a target structured language text corresponding to the to-be-recognized user question and a first target prompt text, and the first target prompt text is determined according to the vertical field type corresponding to the to-be-recognized user question; inputting the target structured language text and the first target prompt text into a pre-trained intention recognition large model to perform intention recognition, and determining a target natural language expression text corresponding to the to-be-recognized user question, wherein the intention recognition large model is generated by training according to prompt texts corresponding to different vertical field types, thereby solving the problems of low efficiency and low accuracy of structured language intention recognition; the intention recognition large model is used for intention recognition, the target natural language expression text is accurately determined, and the speed, the efficiency and the cost are low.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a text intent recognition method, electronic device, storage medium, and program product. Background Technology

[0002] The traditional execution process of task-oriented Natural Language Understanding (NLU) involves obtaining a user query and understanding its domain, intent, and slots. These fields have specific specifications, presenting a structured view that is strongly bound to business logic. In the real-world data loop, long-term monitoring of the effects of task-oriented traffic is necessary. One monitoring method is to observe end-to-end effects, such as focusing only on whether error messages are included in the TTS feedback. However, this method has a coarse-grained observation level, making it difficult to monitor fine-grained understanding and easily missing erroneous data. For example, a user query might be "Open the driver's side window," the understood execution might be "Open the passenger side window," and the final TTS feedback might be "Done." Under the first monitoring approach, this type of error is difficult to detect. To improve the accuracy of monitoring results, a monitoring method involving refined manual annotation of intermediate results is proposed. This involves manually identifying the intent of the structured language generated during NLU execution and annotating it as natural language.

[0003] However, manually standardizing and annotating structured languages ​​is extremely difficult. Due to the poor readability of formatted fields, understanding these fields requires a long familiarization process and a deep understanding of the business logic, resulting in low efficiency and accuracy, and very high costs. Summary of the Invention

[0004] This invention provides a text intent recognition method, electronic device, storage medium, and program product to solve the problem of low efficiency and accuracy in structured language intent recognition.

[0005] According to one aspect of the present invention, a text intent recognition method is provided, comprising:

[0006] Obtain the question of the user to be identified;

[0007] If the user question to be identified does not have a corresponding target natural language expression text, determine the target structured language text and the first target prompt text corresponding to the user question to be identified. The first target prompt text is determined according to the vertical domain type corresponding to the user question to be identified.

[0008] The target structured language text and the first target prompt text are input into a pre-trained intent recognition model for intent recognition to determine the target natural language expression text corresponding to the user question to be identified. The intent recognition model is trained and generated according to the prompt text corresponding to different vertical domain types.

[0009] According to another aspect of the present invention, a vehicle is provided, the vehicle comprising:

[0010] At least one processor, and a memory communicatively connected to said at least one processor;

[0011] The memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to perform the text intent recognition method according to any embodiment of the present invention.

[0012] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the text intent recognition method according to any embodiment of the present invention.

[0013] According to another aspect of the present invention, a computer program product is provided, the computer program product comprising a computer program that, when executed by a processor, implements the text intent recognition method according to any embodiment of the present invention.

[0014] The technical solution of this invention involves obtaining a user question to be identified; if the user question does not have a corresponding target natural language expression text, determining the target structured language text and a first target prompt text corresponding to the user question, wherein the first target prompt text is determined based on the vertical domain type corresponding to the user question; inputting the target structured language text and the first target prompt text into a pre-trained intent recognition model for intent recognition, thereby determining the target natural language expression text corresponding to the user question, wherein the intent recognition model is trained and generated based on prompt texts corresponding to different vertical domain types, thus solving the problem of low efficiency and accuracy of structured language intent recognition. The system determines whether a corresponding target natural language expression text exists for the user's question to be identified. If not, it determines the corresponding target structured language text and the corresponding first target prompt text based on the corresponding vertical domain type. The target structured language text and the first target prompt text are then input into a pre-trained intent recognition model for intent recognition. This accurately determines the target natural language expression text corresponding to the target structured language text. The intent recognition model is trained and generated based on the prompt texts corresponding to different vertical domain types, enabling precise intent recognition of structured language texts of different vertical domain types. This improves the accuracy of the recognition results. Furthermore, the intent recognition model offers fast recognition speed, high efficiency, and low cost.

[0015] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a text intent recognition method provided in Embodiment 1 of the present invention;

[0018] Figure 2 This is a flowchart of a text intent recognition method provided in Embodiment 2 of the present invention;

[0019] Figure 3 This is an example diagram illustrating the implementation of monitoring task-based NLU results according to Embodiment 2 of the present invention;

[0020] Figure 4This is a schematic diagram of the structure of a text intent recognition device according to Embodiment 3 of the present invention;

[0021] Figure 5 This is a schematic diagram of the structure of an electronic device that implements the text intent recognition method of this invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] Example 1

[0025] Figure 1 This is a flowchart of a text intent recognition method provided in Embodiment 1 of the present invention. This embodiment is applicable to situations requiring accurate recognition of the intent of structured text. The method can be executed by a text intent recognition device, which can be implemented in hardware and / or software and can be configured in a vehicle. Figure 1 As shown, the method includes:

[0026] S101, Obtain the question of the user to be identified.

[0027] In this embodiment, the user question to be identified can be understood as a user question with an intent recognition requirement. The user can interact with the vehicle system to ask the system a question. For example, the user can ask the question "Please open the rear window" via voice.

[0028] After the vehicle's infotainment system is started, users can input various questions via voice, touch, or other means. Upon receiving a user's question, the system can either treat all questions as potential user questions or analyze and judge the question to determine if it meets the requirements. If it does, the system will treat it as a potential user question. For example, it might determine if the user's question is a task-oriented message. If so, it will be considered a potential user question. Task-oriented messages refer to data that controls the vehicle's hardware and software to perform corresponding operations based on the user's instructions. For example, "Please open the rear window" or "Play a TV series" are task-oriented messages, while "What day of the week is it today?" or "What is the temperature today?" are question-and-answer messages.

[0029] S102. If there is no corresponding target natural language expression text for the user question to be identified, determine the target structured language text and the first target prompt text corresponding to the user question to be identified. The first target prompt text is determined according to the vertical domain type corresponding to the user question to be identified.

[0030] In this embodiment, the target natural language expression text can be understood as data text formed by natural language that describes the structured language text corresponding to the user's question to be identified. The first target prompt text can be understood as text used to provide prompts to the large model, i.e., a kind of prompt. The vertical domain type can be media, navigation, vehicle control, telephone, etc.

[0031] Different questions and their corresponding natural language expressions are pre-saved. Upon receiving a user question to be identified, the system checks whether a corresponding target natural language expression exists based on the pre-saved data. If not, the user question is structured to determine its corresponding target structured language text. The system then analyzes the user question, determining its corresponding vertical domain type based on the type and information of the hardware and software to be controlled. Different prompt texts corresponding to different vertical domain types are pre-set. The pre-generated prompt texts are filtered according to the vertical domain type of the user question to be identified, and the prompt text matching the user question is selected as the first target prompt text.

[0032] S103. Input the target structured language text and the first target prompt text into the pre-trained intent recognition model for intent recognition, determine the target natural language expression text corresponding to the user question to be recognized, and train the intent recognition model to generate prompt texts corresponding to different vertical domain types.

[0033] In this embodiment, the intent recognition big model can be understood as a big model for recognizing intent in structured language, such as the multimodal cognitive big model MindGPT, the Chat Generative Pre-trained Transformer (ChatGPT), etc.

[0034] Different prompt texts are generated by pre-filtering data based on different vertical domain types. Data with distinctive or representative characteristics can be selected during the filtering process to allow the large-scale model to learn better knowledge and improve its recognition accuracy. For example, frequently occurring questions within a specific vertical domain type can be selected, and prompt texts can be generated based on the structured language expression of these questions. By generating different prompt texts for different vertical domain types, a large amount of training data covering different domains, businesses, and scenarios can be obtained. The large-scale model is then trained using these different prompt texts, and fine-tuned using a large language model. Model parameters are optimized based on the training data, enabling the model to learn knowledge specific to particular business scenarios. After training, an intent recognition large-scale model capable of accurately recognizing the intent expressed in structured language text can be obtained. The target structured language text and the first target prompt text are input into the trained intent recognition large-scale model, which performs intent recognition to obtain the target natural language expression text.

[0035] The text intent recognition method of this invention solves the problem of low efficiency and accuracy in structured language intent recognition. It determines whether a corresponding target natural language expression text exists for the user's question to be identified. If not, it determines the corresponding target structured language text and the corresponding first target prompt text based on the corresponding vertical domain type. The target structured language text and the first target prompt text are input into a pre-trained intent recognition model for intent recognition. This accurately determines the target natural language expression text corresponding to the target structured language text. The intent recognition model is trained and generated based on prompt texts corresponding to different vertical domain types, enabling precise intent recognition of structured language text of different vertical domain types, improving the accuracy of the recognition results. Furthermore, the intent recognition model offers fast recognition speed, high efficiency, and low cost.

[0036] Example 2

[0037] Figure 2 This is a flowchart of a text intent recognition method provided in Embodiment 2 of the present invention. This embodiment is a refinement based on the above embodiments. Figure 2 As shown, the method includes:

[0038] S201, Question about obtaining the user to be identified.

[0039] S202. Obtain cached text, which includes at least one set of user intent expression text, including user questions, structured language text, and natural language expression text.

[0040] In this embodiment, cached text can be understood as pre-cached text data. The cached text includes at least one set of user intent expression text, which includes user questions, structured language text, and natural language expression text. User questions are pre-obtained, structured to obtain corresponding structured language text, and then intent recognition is performed on the structured language text to determine its corresponding natural language expression text. The user questions, structured language text, and natural language expression text are then combined into a set of user intent expression text. Multiple sets of user intent expression text are generated based on multiple user questions, and all user intent expression texts are cached in a file or text storage to obtain the cached text. The cached text can be saved in specified memory, database, cloud, or other storage spaces. During text intent recognition, the cached text is directly retrieved based on the cache path, etc.

[0041] S203. Match the user question to be identified with the user questions in the cached text. If the match fails, proceed to S204; otherwise, proceed to S207.

[0042] Identify all user questions in the cached text. Match the user question to be identified with the user questions in the cached text. For example, calculate the similarity between the user question to be identified and each user question in the cached text, and compare the magnitudes of each similarity. Determine the highest similarity. If this similarity is greater than a set similarity threshold, then the user question corresponding to this similarity is determined to match the user question to be identified, and the match is successful. Proceed to step S207. If this similarity is not greater than the set similarity threshold, then the user question to be identified does not match any user questions in the cached text, and the match is unsuccessful. Proceed to step S204.

[0043] S204. Determine that the user question to be identified does not have a corresponding target natural language expression text.

[0044] S205. Determine the target structured language text corresponding to the user question to be identified and the first target prompt text. The first target prompt text is determined according to the vertical domain type corresponding to the user question to be identified.

[0045] S206. Input the target structured language text and the first target prompt text into the pre-trained intent recognition model for intent recognition, determine the target natural language expression text, and the intent recognition model is trained and generated according to the prompt text corresponding to different vertical domain types.

[0046] The intent recognition model is used to identify the intent of structured language, obtain the corresponding target natural language expression text, and end the intent recognition of the structured language for the user's question.

[0047] S207. Determine that the user question to be identified has a corresponding target natural language expression text.

[0048] S208. The natural language expression text corresponding to the user question that matches the user question to be identified in the cached text is determined as the target natural language expression text.

[0049] When a user question that matches the user question to be identified exists in the cached text, the natural language expression text corresponding to the user question that matches the user question to be identified is determined, and this natural language expression text is determined as the target natural language expression text.

[0050] By pre-generating cached text to store different structured language texts and natural language expressions, during the intent recognition process, if there is a user question in the cached text that matches the user question to be identified, the natural language expression text corresponding to the structured language text of the user question to be identified can be directly obtained from the cached text. Pre-caching can reduce some inference costs and service pressure.

[0051] Optionally, the steps for generating cached text include steps A1-A5:

[0052] A1. Obtain the structured language text corresponding to at least one user question.

[0053] Pre-collect user questions generated from interactions between users and the vehicle's infotainment system, as well as the corresponding structured language text for these user questions.

[0054] Users can interact with the vehicle's infotainment system in real time after startup, inputting user questions. Upon receiving a user question, the system processes it in a structured format, generating corresponding structured language text. This structured text is then input into a large model for recognition, and the system controls the vehicle to perform the appropriate actions. The user questions and structured language text involved in this process can be saved. When generating cached text, the saved user questions and structured language text can be directly retrieved.

[0055] In this embodiment of the application, when obtaining multiple user questions, if the user questions are duplicated, the user questions can be deduplicated to obtain unique user questions and their corresponding structured language text.

[0056] A2. Remove duplicates from the structured language text to obtain the structured language text to be identified.

[0057] In this embodiment, the structured language text to be identified can be understood as structured language text with intent recognition requirements.

[0058] Different user questions may correspond to the same structured language text. Therefore, when performing structured language text recognition, it is necessary to deduplicate the structured language text. By comparing the structured language texts using a preset algorithm or rules, it is determined whether they are identical. If they are identical, only one structured language text needs to be retained, and the redundant structured language texts are deleted, thus completing the deduplication process. The remaining structured language texts after deduplication are used as the structured language texts to be recognized. For example, if there are 100 structured language texts before deduplication, 90 unique structured language texts are obtained after deduplication. These 90 structured language texts are used as the structured language texts to be recognized, with one structured language text corresponding to each user question.

[0059] A3. Based on the structured language text to be identified and the pre-generated prompt text, perform intent recognition to determine the natural language expression text and determine the natural language expression text corresponding to each user question.

[0060] Different prompt texts are pre-generated and combined with the structured language text to be recognized. For example, the structured language text to be recognized can be randomly combined with the prompt text, or combined according to certain rules. Then, intent recognition is performed on the structured language text to be recognized based on the prompt text to determine the corresponding natural language expression text. After determining the natural language expression text, the natural language expression text corresponding to each user question is determined based on the correspondence between the structured language text to be recognized and the natural language expression text, and the correspondence between the structured language text to be recognized and the user question.

[0061] A4. Generate a set of user intent expression texts based on each user question and the corresponding structured language text and natural language expression text.

[0062] After determining the natural language expression text corresponding to each structured language text, based on the correspondence between user questions and structured language texts, the structured language text and natural language expression text corresponding to each user question are determined. Based on all user questions and their corresponding structured language texts and natural language expression texts, a set of expression texts representing user intent is generated.

[0063] A5. Generate cached text based on the expressive text of each group of users' intentions.

[0064] The text expressing the user intent of each group is cached to a text file or file, generating cached text.

[0065] Optionally, intent recognition is performed based on the structured language text to be recognized combined with pre-generated prompt text, including:

[0066] B1. Determine the vertical domain type corresponding to the structured language text to be identified.

[0067] Analyze the structured language text to be identified and determine its corresponding vertical domain type. For example, if the structured language text to be identified includes a Domain, the specific type corresponding to the Domain is taken as the vertical domain type.

[0068] The structured language text to be recognized is called DAS. NLU (Natural Language Understanding), as the semantic understanding module in a task-oriented dialogue system, transforms the text transcribed from user text input / voice input into a structured representation (DAS) that is easy for machines to understand. DAS comprises three main modules: ① Domain Recognition, ② Intent Recognition, and ③ Slot Filling.

[0069] Domain: Also known as a vertical domain or scenario. A domain is the front-end of NLU (Network Logic) and is divided according to business scenarios. For example, vehicle control, music, telephone, and navigation are all different domains. Different domains correspond to different intents and slots, and subsequently different solutions. This division facilitates downstream routing.

[0070] Action / Intent: This refers to the user's current action or stated intention, usually expressed as a verb. For example, "open" and "close" are different Actions. The definition of Action / Intent varies across different Domains.

[0071] Slot: A slot represents important information that the machine needs to focus on during the current dialogue. For example, in the field of vehicle control, a slot can contain in-vehicle entities such as windows, doors, and air conditioners; in the field of navigation, a slot can contain place names such as location 1 and location 2. To facilitate downstream processing, some slot values ​​need to be normalized. For example, "window," "car window," and "window" must be uniformly normalized to "WINDOW" before the edge device can execute correctly. The definition of a slot differs across domains.

[0072] For example, Table 1 is a data table of user questions and structured language text, which is used to illustrate user questions and structured language text:

[0073] Table 1. Data table of user questions and structured language text.

[0074]

[0075] As shown in Table 1, after obtaining the structured language text to be recognized, the vertical domain type can be directly determined based on the domain.

[0076] B2. Filter the pre-generated prompt texts according to the vertical domain type to determine the second target prompt text. The second target prompt text is generated according to the instruction lookup table corresponding to the vertical domain type.

[0077] In this embodiment, the second target prompt text can be understood as text used to provide prompts to the large model during large model recognition. The instruction lookup table can be understood as a data table storing the conversion between different instructions, for example, AC: air conditioner, BLUETOOTH: Bluetooth.

[0078] Beforehand, different instruction data under different vertical domain types are collected to determine the transformed data corresponding to different instructions during intent recognition, forming an instruction lookup table. Corresponding prompt texts are generated based on the instruction lookup table, which assists the large model in intent recognition and accurately identifies user intent. The generated prompt texts are then associated with their corresponding vertical domain types. After determining the vertical domain type corresponding to the structured language text to be recognized, each prompt text is filtered based on the association between the vertical domain type and the prompt text. The prompt text that matches the vertical domain type corresponding to this structured language text to be recognized is selected as the second target prompt text. The second target prompt text obtained at this point is the prompt text generated based on the instruction lookup table corresponding to the vertical domain type.

[0079] B3. Input the structured language text to be recognized and the second target prompt text into the intent recognition model for intent recognition.

[0080] The structured language text to be identified and the second target prompt text are input into the intent recognition model. The intent recognition model performs intent recognition on the structured language text to be identified based on the prompt text of the second target text, and obtains its corresponding natural language expression text.

[0081] Optionally, the prompt text includes a table of instructions corresponding to the vertical domain type, closed datasets, open datasets, supplementary information on conversion rules, and text conversion examples;

[0082] The instruction lookup table includes textual conversion descriptions for different fields, and supplementary information on conversion rules is used to instruct on data conversion for closed and open datasets.

[0083] In this embodiment, the closed dataset and the open dataset store one or more fields, each of which can be a specific type of field from the vertical domain and the intent action. The supplementary information for transformation rules can be understood as information describing the transformation rules for different fields during the intent recognition process of the large intent recognition model. This supplementary information instructs on the data transformation of the closed dataset and the open dataset; for example, it instructs that the data in the closed dataset be normalized according to an instruction lookup table, while the data in the open dataset is directly output as literal values. The text transformation examples can be understood as pre-given sample examples of converting structured text into natural language expressions. The number of text transformation examples in the prompt text can be one or more. Setting multiple text transformation examples can better guide the intent recognition model to perform accurate intent recognition.

[0084] It is important to know that since the first target prompt text and the second target prompt text are one type of prompt text, the first target prompt text and the second target prompt text have the same structure as the prompt text, both including: an instruction lookup table corresponding to the vertical domain type, a closed dataset, an open dataset, supplementary information on conversion rules, and a text conversion example; among them, the instruction lookup table includes text conversion descriptions for different fields, and the supplementary information on conversion rules is used to instruct on the data conversion of closed datasets and open datasets.

[0085] For example, Table 2 provides an instruction reference table, whose corresponding vertical domain type is vehicle control. Table 2 shows some of the instructions.

[0086] Table 2 Command Comparison Table

[0087]

[0088] For example, this application provides a prompt text whose corresponding vertical type is media.

[0089] prompt = "You are an expert in evaluating user queries for xx car, capable of converting instructions from closed datasets into natural language text descriptions. I will now provide you with an instruction lookup table and the corresponding instructions. Please convert them into natural language text descriptions that are easy for humans to understand."

[0090] Requirements: Please strictly follow the example format for output and do not output any additional content.

[0091] Instruction Reference Table:

[0092] Closed dataset:

[0093] media

[0094] search=>search

[0095] play => Play

[0096] mediaType => Media type, such as music, album, singing, video, audiobook, radio, news.

[0097] screenName => Screen type and location information, Passenger entertainment screen | Rear screen | Center control screen HU | Passenger | Second row | Rear open dataset:

[0098] Resource slots are named as follows: musicTitle, musicAlbum, videoTitle, audioBooksTitle, ip

[0099] Name slots: singer, actors, director, nickname, personName

[0100] Conversion rule supplement:

[0101] The slot value corresponding to the closed set slot needs to be normalized according to the instruction lookup table, and the selectable values ​​can only be in a finite set.

[0102] The slot values ​​corresponding to open set slots do not need to be normalized; the literal values ​​can be output directly.

[0103] Input: domain=media, action=search, mediaType=listen to music, musicTitle=AAA

[0104] Output: The user wants to search for the song AAA

[0105] Input: domain=media, action=search, mediaType=video, actors=actors1, videoTitle=xx

[0106] Output: The user wants to search for videos featuring actor 1.

[0107] Input: domain=media, action=search, nediaType=video, ip=yy, screenName=PASSENGER

[0108] Output: The user wants to watch videos on the YY topic from the passenger seat.

[0109] Input: domain=media, action=search, mediaType=audiobook, nickname=Art Master 1, audioBookTitle=zzz

[0110] Output: The user wants to listen to an audiobook, authored by Art Master 1, titled zzz.

[0111] Input: ({format_das})

[0112] Output: """

[0113] Optionally, the method further includes:

[0114] C1. Perform correlation analysis between the user's question and the target natural language expression text.

[0115] To perform correlation analysis between the user question to be identified and the target natural language expression text, one can use a large model to score the correlation, or use algorithms or neural network models to calculate the correlation to determine whether the user question to be identified and the target natural language expression text are related.

[0116] C2. Based on the analysis results, determine the meaning of structured language text conversion. Figure 1 Consistency rate.

[0117] In this embodiment, the analysis results can be related / unrelated, or the degree of correlation, such as high correlation / low correlation, or the correlation can be represented by probability, etc.

[0118] The analysis results determine whether the user question to be identified and the target natural language text are related, and then determine whether the intent of the converted user question matches the intent of the user question when it is converted into structured language text. Figure 1 To achieve this, historical data can be combined to determine the meaning of structured language text conversion. Figure 1 Consistency rate; or, after meeting the analysis conditions, perform a unified analysis and calculation on all analysis results to determine the meaning of structured language text conversion. Figure 1 Consistency rate; for example, the analysis condition is that the number of user questions to be identified accumulates to a certain number (e.g., after identifying N user questions, the meaning of structured language text conversion is determined based on the analysis results corresponding to these N user questions). Figure 1 (e.g., consistency rate), or, the analysis condition is that the difference between the current time and the last analysis time exceeds time T (e.g., obtaining the analysis results corresponding to all user questions to be identified from the previous day each day to determine the meaning of structured language text conversion). Figure 1 (e.g., success rate), etc.

[0119] For example, Table 3 provides a comparison table of user questions and natural language expressions.

[0120] Table 3. Comparison of User Questions and Natural Language Expression Text

[0121]

[0122] The embodiments of this application are intended for performing structured language text conversion. Figure 1 When analyzing the consistency rate, analysis can also be performed based on vertical domain type. This involves classifying all user questions to be identified according to vertical domain type. For each vertical domain type, the structured language text conversion intention is determined based on the relevance analysis results of all user questions to be identified under that vertical domain type and the target natural language expression text. Figure 1 Consistency. Through intention Figure 1 Consistency rate represents the consistency between the natural language representation text and the structured language representation of the large model. Through meaning... Figure 1 Accuracy can be improved by adjusting the large model used in the vehicle control process and promptly detecting anomalies in the vehicle control process. For example, by analyzing and determining that users often make erroneous responses when controlling vehicle windows with their voice, the large model can be adjusted accordingly to improve the recognition accuracy of the large model for window control.

[0123] For example, Figure 3 This document provides an example diagram for monitoring task-oriented NLU results. During the data closure process, leveraging the language capabilities of a large-scale model, it transforms structured business fields, which are difficult to understand through manual annotation, into natural language text, enabling refined monitoring of task-oriented NLU results. Monitoring of task-oriented NLU includes online and offline monitoring. Offline monitoring involves pre-recognizing intent using a large-scale intent recognition model, generating cached text, and then querying the cached text to determine the corresponding target natural language expression text offline. Online monitoring involves recognizing intent using the large-scale intent recognition model to determine the target natural language expression text. Online monitoring is typically performed when the target natural language expression text corresponding to the user's question is not pre-saved in the cached text.

[0124] The core of the offline optimization process consists of two parts: fine-tuning and prompt-tuning. Performing prompt-tuning first, followed by fine-tuning, allows for rapid verification and adjustment of the model's performance on a specific task. If the former performs well, further fine-tuning can achieve even better performance. Fine-tuning and prompt-tuning require iterative processing of the data, indicated by the curved arrows in the diagram. The left side of the diagram shows the offline tuning and cache generation process; the right side shows the online inference and recognition process. During the offline process, an intent recognition model can be trained, and cached text can be obtained from this model. The offline intent recognition model and cached text data can then be applied to the online inference and recognition process.

[0125] The offline tuning and cache generation process can be referenced as follows:

[0126] 1. Using the NLU structured documentation, extract the Chinese and English descriptions and mapping relationships of all fields to form sample descriptions, and add the textualized instructions of all closed set instructions to the prompt. Continuously optimize the prompt based on the response effect of large models.

[0127] 2. Based on online data from different fields, select high-quality data and the optimal prompt for splicing and assembly to obtain a large amount of training data from different fields, businesses and scenarios.

[0128] 3. Use a large language model to fine-tune instructions. Based on the high-quality training data obtained in the first two steps, optimize the model parameters so that the model learns knowledge in specific business scenarios.

[0129] 4. Combine the frequently used DAS data online, remove duplicates, and obtain the corresponding natural language expression text for this part of the structured language text offline, storing it in a cache file.

[0130] Combining fine-tuning and prompt-tuning serves two main purposes. Firstly, the target task is a domain-specific structured text-to-natural language conversion task, strongly related to business scenarios. Many instructions are converted from custom rules, representing knowledge not seen during pre-training. This requires a large model for deep understanding and adaptation; therefore, fine-tuning can be used to update the model's knowledge and continuously improve online understanding performance. Secondly, when the NLU structured documentation changes, prompt-tuning allows for rapid intervention, ensuring correct online results are provided immediately after configuration file modifications.

[0131] A large-scale intent recognition model is trained offline to generate cached files. During application, after obtaining the user's question to be recognized, the cached text is queried to determine if it matches the cached file. If it does, the cached text is read to determine the target natural language expression text; otherwise, inference recognition is performed based on the large-scale intent recognition model and the prompt (i.e., the first target prompt text). The target natural language expression text is determined based on the recognition result of the large-scale intent recognition model. The target natural language expression text is analyzed through methods such as manual annotation or automatic scoring to determine the intent of the structured language text conversion. Figure 1 Consistency rate.

[0132] The text intent recognition method of this invention solves the problem of low efficiency and accuracy in structured language intent recognition. It matches the user question to be identified with the user question in the cached text to determine if a corresponding target natural language expression text exists. If it exists, the target natural language expression text stored in the cached text is directly obtained, saving data processing time and resources. If it does not exist, the corresponding target structured language text and the first target prompt text are input into the intent recognition large model for intent recognition, accurately determining the target natural language expression text corresponding to the target structured language text. The intent recognition large model is trained and generated based on prompt texts corresponding to different vertical domain types, enabling accurate intent recognition of structured language text of different vertical domain types, improving the accuracy of the recognition results. Furthermore, the intent recognition large model offers fast recognition speed, high efficiency, and low cost. By analyzing the correlation between the user question to be identified and the identified target natural language expression text, the intent of the structured language text conversion can be determined. Figure 1 Consistency rate.

[0133] Example 3

[0134] Figure 4 This is a schematic diagram of a text intent recognition device provided in Embodiment 3 of the present invention. Figure 4 As shown, the device includes: a question acquisition module 31, a first target text determination module 32, and an intent recognition module 33.

[0135] The question acquisition module 31 is used to acquire the questions of the user to be identified;

[0136] The first target prompt text determination module 32 is used to determine the target structured language text and the first target prompt text corresponding to the user question to be identified if there is no corresponding target natural language expression text for the user question to be identified. The first target prompt text is determined according to the vertical domain type corresponding to the user question to be identified.

[0137] The intent recognition module 33 is used to input the target structured language text and the first target prompt text into a pre-trained intent recognition model for intent recognition, and to determine the target natural language expression text corresponding to the user question to be recognized. The intent recognition model is trained and generated according to the prompt text corresponding to different vertical domain types.

[0138] The text intent recognition method of this invention solves the problem of low efficiency and accuracy in structured language intent recognition. It determines whether a corresponding target natural language expression text exists for the user's question to be identified. If not, it determines the corresponding target structured language text and the corresponding first target prompt text based on the corresponding vertical domain type. The target structured language text and the first target prompt text are input into a pre-trained intent recognition model for intent recognition. This accurately determines the target natural language expression text corresponding to the target structured language text. The intent recognition model is trained and generated based on prompt texts corresponding to different vertical domain types, enabling precise intent recognition of structured language text of different vertical domain types, improving the accuracy of the recognition results. Furthermore, the intent recognition model offers fast recognition speed, high efficiency, and low cost.

[0139] Optionally, the device may also include:

[0140] A cached text acquisition module is used to acquire cached text, which includes at least one set of user intent expression text, including user questions, structured language text, and natural language expression text.

[0141] The matching module is used to match the user question to be identified with the user questions in the cached text. If the match fails, it is determined that the user question to be identified does not have a corresponding target natural language expression text; if the match succeeds, it is determined that the user question to be identified has a corresponding target natural language expression text.

[0142] Optionally, the device may also include:

[0143] The target natural language text determination module is used to determine the natural language text corresponding to the user question that matches the user question to be identified in the cached text as the target natural language text if the user question to be identified has a corresponding target natural language expression text.

[0144] Optionally, the device may also include:

[0145] The structured text acquisition module is used to acquire the structured language text corresponding to at least one user question.

[0146] The deduplication module is used to deduplicatize the structured language text to obtain the structured language text to be recognized.

[0147] The natural language text determination module is used to perform intent recognition based on the structured language text to be identified and the pre-generated prompt text, determine the natural language expression text, and determine the natural language expression text corresponding to each user question;

[0148] The expression text generation module is used to generate a set of expression texts of user intent based on each of the user questions, the structured language text corresponding to the user questions, and the natural language expression text corresponding to the user questions.

[0149] The cached text generation module is used to generate cached text based on the expressive text of each group of users' intentions.

[0150] Optional, natural text determination module, including:

[0151] Vertical domain type determination unit, used to determine the vertical domain type corresponding to the structured language text to be identified;

[0152] The second target prompt text determination unit is used to filter the pre-generated prompt text according to the vertical domain type and determine the second target prompt text. The second target prompt text is generated according to the instruction lookup table corresponding to the vertical domain type.

[0153] The model input unit is used to input the structured language text to be recognized and the second target prompt text into the intent recognition big model for intent recognition.

[0154] Optionally, the prompt text includes an instruction lookup table corresponding to the vertical domain type, closed datasets, open datasets, supplementary information on conversion rules, and text conversion examples;

[0155] The instruction lookup table includes text conversion descriptions for different fields, and the supplementary information on conversion rules is used to instruct on the data conversion between the closed dataset and the open dataset.

[0156] Optionally, the device may also include:

[0157] The correlation analysis module is used to perform correlation analysis on the user question to be identified and the target natural language expression text;

[0158] meaning Figure 1 The consistency determination module is used to determine the consistency of structured language text conversion based on the analysis results. Figure 1 Consistency rate.

[0159] The text intent recognition device provided in the embodiments of the present invention can execute the text intent recognition method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0160] Example 4

[0161] Figure 5A schematic diagram of an electronic device 40 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0162] like Figure 5 As shown, the electronic device 40 includes at least one processor 41 and a memory, such as a read-only memory (ROM) 42 or a random access memory (RAM) 43, communicatively connected to the at least one processor 41. The memory stores computer programs executable by the at least one processor. The processor 41 can perform various appropriate actions and processes based on the computer program stored in the ROM 42 or loaded into the RAM 43 from storage unit 48. The RAM 43 may also store various programs and data required for the operation of the electronic device 40. The processor 41, ROM 42, and RAM 43 are interconnected via a bus 44. An input / output (I / O) interface 45 is also connected to the bus 44.

[0163] Multiple components in electronic device 40 are connected to I / O interface 45, including: input unit 46, such as keyboard, mouse, etc.; output unit 47, such as various types of monitors, speakers, etc.; storage unit 48, such as disk, optical disk, etc.; and communication unit 49, such as network card, modem, wireless transceiver, etc. Communication unit 49 allows electronic device 40 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0164] Processor 41 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 41 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 41 performs the various methods and processes described above, such as text intent recognition methods.

[0165] In some embodiments, the text intent recognition method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 48. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 40 via ROM 42 and / or communication unit 49. When the computer program is loaded into RAM 43 and executed by processor 41, one or more steps of the text intent recognition method described above may be performed. Alternatively, in other embodiments, processor 41 may be configured to perform the text intent recognition method by any other suitable means (e.g., by means of firmware).

[0166] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0167] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0168] This invention provides a computer program product, which includes a computer program that, when executed by a processor, implements the text intent recognition method described in any embodiment of this invention.

[0169] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0170] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0171] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0172] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0173] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0174] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A text intent recognition method, characterized in that, include: Obtain the question of the user to be identified; If the user question to be identified does not have a corresponding target natural language expression text, determine the target structured language text and the first target prompt text corresponding to the user question to be identified. The first target prompt text is determined according to the vertical domain type corresponding to the user question to be identified. The target structured language text and the first target prompt text are input into a pre-trained intent recognition model for intent recognition to determine the target natural language expression text corresponding to the user question to be identified. The intent recognition model is trained and generated according to the prompt text corresponding to different vertical domain types.

2. The method according to claim 1, characterized in that, After obtaining the question of the user to be identified, the following is also included: Obtain cached text, which includes at least one set of text expressing user intent, including user questions, structured language text, and natural language text. The user question to be identified is matched with the user questions in the cached text. If the match fails, it is determined that the user question to be identified does not have a corresponding target natural language expression text; if the match succeeds, it is determined that the user question to be identified has a corresponding target natural language expression text.

3. The method according to claim 2, characterized in that, Also includes: If the user question to be identified has a corresponding target natural language expression text, the natural language expression text corresponding to the user question that matches the user question to be identified in the cached text is determined as the target natural language expression text.

4. The method according to claim 2, characterized in that, Before retrieving the cached text, the following is also included: Obtain the structured language text corresponding to at least one user question; The structured language text is deduplicated to obtain the structured language text to be identified; Intent recognition is performed based on the structured language text to be identified and the pre-generated prompt text to determine the natural language expression text, and the natural language expression text corresponding to each user question is determined; a set of user intent expression texts is generated based on each user question, the structured language text corresponding to the user question, and the natural language expression text corresponding to the user question; Cache text is generated based on the expressive text of each group of users' intentions.

5. The method according to claim 4, characterized in that, The intent recognition based on the structured language text to be recognized combined with pre-generated prompt text includes: Determine the vertical domain type corresponding to the structured language text to be identified; The pre-generated prompt text is filtered according to the vertical domain type to determine the second target prompt text, which is generated according to the instruction lookup table corresponding to the vertical domain type. The structured language text to be identified and the second target prompt text are input into the intent recognition big model for intent recognition.

6. The method according to any one of claims 4-5, characterized in that, The prompt text includes a table of instructions corresponding to the vertical domain type, closed datasets, open datasets, supplementary information on conversion rules, and text conversion examples; The instruction lookup table includes text conversion descriptions for different fields, and the supplementary information on conversion rules is used to instruct on the data conversion between the closed dataset and the open dataset.

7. The method according to claim 1, characterized in that, Also includes: A correlation analysis is performed on the user question to be identified and the target natural language expression text; The intent consistency rate of structured language text conversion is determined based on the analysis results.

8. An electronic device, characterized in that, The electronic device includes: At least one processor, and a memory communicatively connected to said at least one processor; The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the text intent recognition method according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the text intent recognition method according to any one of claims 1-7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the text intent recognition method according to any one of claims 1-7.

Images