Natural language query method and device

Abstract

The embodiment of the invention discloses a natural language query method and device. According to the embodiment of the invention, after the natural language query statement is received, the natural language query statement is input into the pre-trained large-scale text embedding model capable of better capturing the semantic and context relationship, and the corresponding intention semantic query vector is obtained; the method comprises the steps that firstly, an intention semantic query vector is obtained, then multiple intention semantic vectors are recalled from a preset intention semantic vector database according to the intention semantic query vector, the similarity between the intention semantic query vector and all the intention semantic vectors is determined to determine a target intention type, and finally feedback information is determined according to a natural language query statement and the target intention type. According to the embodiment of the invention, a large-scale text embedding model is used, the context information can be incorporated into the intention recognition process, and the accuracy of the target intention type is improved. The feedback information is determined according to the natural language query statement and the target intention type, and the accuracy of the feedback information can be improved. And meanwhile, the security of internal data can be ensured without interaction with external data.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular, to a natural language query method and apparatus. Background Art

[0002] With the continuous development of artificial intelligence, intelligent dialogue has become a popular interaction method and has been applied to various service scenarios, such as smart home, intelligent customer service, virtual assistants, or intelligent robots. Intelligent dialogue relies on natural language processing and big data computing processing. Currently, large language models are usually used to determine the corresponding response feedback information. For small and medium-sized enterprises and individual developers, if the intelligent question and answer service relies on a third-party language model, it may cause internal data leakage and it is difficult to ensure the security of internal data. If an enterprise or individual develops a customized language model, large-scale data computing processing and cost investment are a huge burden, and the effect of the trained language model is also difficult to reach a better level. Therefore, how to output more accurate feedback information for the questions raised by users while ensuring the security of internal data is an urgent problem to be solved currently. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a natural language query method and apparatus to more pertinently determine feedback information while ensuring the security of internal data, thereby improving the accuracy rate of the feedback information.

[0004] In a first aspect, a natural language query method is provided. The method includes:

[0005] Obtain a natural language query statement;

[0006] Input the natural language query statement into a pre-trained large-scale text embedding model to determine a corresponding intent semantic query vector;

[0007] Recall at least one intent semantic vector from a preset intent semantic vector database according to the intent semantic query vector;

[0008] Determine a target intent type according to the similarity between the intent semantic query vector and the intent semantic vector;

[0009] Determine query feedback information according to the natural language query statement and the target intent type.

[0010] In a second aspect, a natural language query apparatus is provided. The apparatus includes:

[0011] An obtaining module, configured to obtain a natural language query statement;

[0012] A first determination module, configured to input the natural language query statement into a pre-trained large-scale text embedding model to determine a corresponding intent semantic query vector;

[0013] A recall module, configured to recall at least one intent semantic vector from a preset intent semantic vector database according to the intent semantic query vector;

[0014] A second determination module, configured to determine a target intent type according to the similarity between the intent semantic query vector and the intent semantic vector;

[0015] A third determination module, configured to determine query feedback information according to the natural language query statement and the target intent type.

[0016] In a third aspect, an electronic device is provided, including a memory and a processor, where the memory is configured to store one or more computer program instructions, and wherein the one or more computer program instructions are executed by the processor to implement the method as described in the first aspect above.

[0017] In a fourth aspect, a computer-readable storage medium is provided, where a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, the method as described in the first aspect above is implemented.

[0018] After receiving a natural language query statement input by a user, an embodiment of the present invention inputs it into a pre-trained large-scale text embedding model that can better capture semantic and context relationships to obtain a corresponding intent semantic query vector. Then, according to the intent semantic query vector, at least one intent semantic vector is recalled from a preset intent semantic vector database, and the similarities between the intent semantic query vector and each recalled intent semantic vector are determined respectively, so as to determine the target intent type. Finally, according to the natural language query statement and the target intent type, the feedback information corresponding to the natural language query statement is determined. In the embodiment of the present invention, a large-scale text embedding model is used to convert a natural language query statement into a corresponding intent semantic query vector and then determine the corresponding intent type, which can incorporate context information into the intent recognition process, thereby making the determined intent type more accurate. Then, the feedback information is determined by combining the natural language query with the target intent type, which can improve the accuracy of the feedback information. At the same time, there may be no interaction with external data in the above process, thereby ensuring the security of internal data. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Through the following description of the embodiments of the present invention with reference to the accompanying drawings, the above and other objects, features, and advantages of the present invention will become clearer. In the drawings:

[0020] Figure 1 is a flowchart of the natural language query method according to the embodiment of the present invention;

[0021] Figure 2 Flowchart of the method for determining the intent semantic query vector for the large-scale text embedding model according to the embodiments of the present invention;

[0022] Figure 3 Flowchart of the method for constructing the intent semantic vector database according to the embodiments of the present invention;

[0023] Figure 4 Flowchart of the method for constructing a multi-level tree-like index structure according to the similarity between multiple intent semantic vectors according to the embodiments of the present invention;

[0024] Figure 5 Flowchart of the method for determining the target intent type according to the similarity between the intent semantic query vector and the intent semantic vector according to the embodiments of the present invention;

[0025] Figure 6 Flowchart of the method for determining the target intent type according to the similarity between the intent semantic query vector and the intent semantic vector according to the embodiments of the present invention;

[0026] Figure 7 Flowchart of the method for determining feedback information according to a natural language query statement and a target intent type according to the embodiments of the present invention;

[0027] Figure 8 Flowchart of another method for determining feedback information according to a natural language query statement and a target intent type according to the embodiments of the present invention;

[0028] Figure 9 Data flow diagram of the natural language query method according to the embodiments of the present invention;

[0029] Figure 10 Schematic diagram of the natural language query device according to the embodiments of the present invention;

[0030] Figure 11 Schematic diagram of the electronic device according to the embodiments of the present invention. Detailed implementation manners

[0031] The following describes the present application based on embodiments, but the present application is not limited to these embodiments. In the following detailed description of the present application, some specific details are described in detail. Those skilled in the art can fully understand the present application without the description of these details. To avoid obscuring the essence of the present application, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0032] In addition, those of ordinary skill in the art should understand that the accompanying drawings provided herein are for illustrative purposes only, and the drawings are not necessarily drawn to scale.

[0033] Unless the context clearly requires otherwise, words such as "including" and "comprising" in the entire application document shall be construed in an inclusive sense rather than an exclusive or exhaustive sense; that is, it means "including but not limited to".

[0034] In the description of this application, it should be understood that terms such as "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance. In addition, in the description of this application, unless otherwise specified, the meaning of "a plurality of" is two or more than two.

[0035] Figure 1 It is a flowchart of the natural language query method according to an embodiment of the present invention. As Figure 1 shown, the natural language query method includes the following steps:

[0036] In step S100, a natural language query statement is obtained.

[0037] Among them, the natural language query statement is a question, topic, instruction, keyword, etc. put forward by the user through natural language or other multimodal means. For example, the natural query statement can be "What are the takeaway platforms on the market?", "Find restaurants with high ratings nearby" or "Monthly active index", etc. The format of the natural language query statement can be text or audio. The method for obtaining the natural language query statement can include receiving the text of the natural language query statement input by the user, collecting the audio containing the natural language query statement issued by the user, or receiving the instruction issued by the user and determining the natural language query statement based on the pre-set mapping relationship between the instruction and the natural language query statement.

[0038] In a possible implementation manner, when the obtained natural language query statement is audio, it is necessary to convert the audio into text through audio recognition technology.

[0039] In step S200, the natural language query statement is input into a pre-trained large-scale text embedding model to determine the corresponding intention semantic query vector.

[0040] Among them, the large-scale text embedding model is a general text embedding model obtained by using a large-scale dataset for training and utilizing the Massive Text Embedding (MTE) technology. The large-scale text embedding model is used to convert input data into corresponding vectors. Through training with a large-scale dataset, the large-scale text embedding model has the ability to capture semantic and context relationships. Therefore, when converting a natural language query statement into a corresponding intent semantic query vector, the context information of the natural language query statement is incorporated into the intent recognition process of the natural language query statement, making the output intent semantic query vector more in line with the actual scenario and the real needs of users, thus avoiding the need for domain fine-tuning of the model. Even when multiple natural language query statements input by the user belong to different domains, or when the output requirements for the feedback information corresponding to the natural language query statement are different, the same large-scale text embedding model can be used for processing. For example, different domains can be the e-commerce domain, the food delivery domain, the logistics and distribution domain, or the academic knowledge query domain, etc. The output requirements for the feedback information corresponding to the natural language query statement can be set by the user or determined according to the actual conversation scenario. For example, when the actual conversation scenario is a chatting scenario, the emotional requirements for the feedback information are relatively high, and when the actual conversation scenario is a knowledge query scenario, the requirements for the source and accuracy of the feedback information are relatively high. In the embodiment of the present invention, using the large-scale text embedding model to determine the intent semantic query vector corresponding to the natural language query statement does not require collecting and annotating a large amount of data in each new domain, reducing the data annotation cost and workload.

[0041] Specifically, the large-scale text embedding model can be the Moka Massive Text Embedding model (M3E). M3E is a text embedding model trained through a large-scale Chinese sentence pair dataset and can convert natural language query statements into dense vectors. The training of M3E on the large-scale sentence pair dataset includes multiple domains such as Chinese encyclopedias, finance, medicine, law, news, academia, e-commerce, and logistics. Through the training of datasets in different domains, M3E has the ability to capture the context relationships of natural language query statements belonging to each domain, and thus can adapt to the conversion of natural language query statements belonging to various domains.

[0042] Figure 2 This is a flowchart of the method for determining the intent semantic query vector by the large-scale text embedding model according to the embodiment of the present invention. As Figure 2 shown, the method for the large-scale text embedding model to determine the intent semantic query vector includes the following steps:

[0043] In step S210, determine the format type of the natural language query statement, where the format type includes at least text and audio.

[0044] In step S220, input the natural language query statement into the encoder corresponding to the format type to determine the corresponding intent semantic query vector.

[0045] Specifically, the large-scale text embedding model contains multiple encoders, such as text encoders and audio encoders. The text encoder is responsible for converting the text information corresponding to the natural language query statement input by the user into a fixed-length intent semantic query vector representation. Commonly used text encoders include Recurrent Neural Network (RNN) and Transformer, etc. These encoders can capture the context information in the text and extract the emotional features therein. The audio encoder is mainly used for feature extraction of the audio signal input by the user. Commonly used audio encoders include Convolutional Neural Networks (CNN) and Long Short-Term Memory networks (LSTM), etc. These encoders can capture the context information in the text from the audio signal, extract the spectral features of the sound, and then capture the emotional information therein.

[0046] In a possible implementation, after inputting the natural language query statement into the pre-trained large-scale text embedding model, the large-scale text embedding model processes the multi-modal data of the natural language query statement. For text data, operations such as word segmentation and stop word removal can be performed. For audio data, conversion and preprocessing of the audio signal can be performed. After the preprocessing is completed, the preprocessed data is input into the corresponding encoder for feature extraction, that is, the text data of the natural language query statement is input into the text encoder, and the audio data of the natural language query statement is input into the audio encoder. Each encoder converts the input data into a fixed-length vector. If the natural language query statement is single-modal data, that is, the natural language query statement is text data or audio data, then the vector output by the corresponding encoder is directly determined as the intent semantic query vector. If the natural language query statement is multi-modal data, that is, the natural language query statement is text data and audio data, then the features of different modalities are fused to obtain the fused intent semantic query vector. Simple weighted fusion or other more complex fusion strategies, such as multi-modal fusion networks, can be used. The fused features can comprehensively utilize the context information or emotional information of different modalities to improve the accuracy of the intent semantic query vector.

[0047] Optionally, if the lengths of the vectors output by the encoder are different, the vectors need to be normalized for multimodal vector fusion and subsequent similarity calculation between the intent semantic query vector and the intent semantic vector.

[0048] In step S300, at least one intent semantic vector is retrieved from a preset intent semantic vector database according to the intent semantic query vector.

[0049] The intent semantic vector database is a pre-constructed database containing multiple intent semantic vectors.

[0050] Figure 3 It is a flowchart of the method for constructing the intent semantic vector database according to an embodiment of the present invention. As Figure 3 shown, the method for constructing the intent semantic vector database includes the following steps:

[0051] In step S010, multiple preset intent semantic data are determined.

[0052] The preset intent semantic data are used to determine the corresponding intent semantic vectors. The preset intent semantic data at least include intent types, intent descriptions corresponding to the intent types, and multiple intent examples corresponding to the intent types. The intent types at least include knowledge Q&A, data query, and other intents. The classification of intent types can be determined according to user needs. The more refined the user needs are, the more intent types are obtained. The intent type description is used to explain the intent type. For example, the intent type description corresponding to knowledge Q&A can be "query the meaning of proper nouns", the intent type description corresponding to data query can be "specific data queries that can be converted into structured query language", and the intent type description corresponding to other intents can be "intents other than knowledge Q&A and data query". The intent examples corresponding to the intent types are natural language query statement examples under the intent types, which are used to provide more context relationships for the intent types, so that the matching between the intent semantic query statement and the intent semantic vectors corresponding to other intent semantic data is more accurate. For example, the intent example corresponding to knowledge Q&A can be "What is beef noodles?", the intent example corresponding to data query can be "What is the gold price this year?", and the intent example corresponding to other intents can be "What songs are nice to listen to?".

[0053] It should be noted that the preset intent semantic data in the embodiments of the present invention can be set and modified according to the actual use effect and the user's requirements for the accuracy rate of feedback information. Specifically, the intent type, the description of the intent type, and the examples of the intent type can all be adjusted to improve the accuracy rate of the feedback information. For example, new intent types can be added to further subdivide the intent type to which the natural language query statement belongs, or the description of the intent type and / or the examples of the intent type corresponding to the existing intent types can be modified to more precisely indicate the intent type, so that the target intent type corresponding to the subsequent determined natural language query statement is more in line with the user's true intent. The finer the classification granularity of the intent type, the clearer and more accurate the description of the intent type, and the more examples of the intent type, the more accurate the subsequent determined target intent type, but the corresponding processing time may become longer. The preset intent semantic data can be adjusted according to the accuracy rate of the feedback data and the processing duration.

[0054] Since the definitions of the same word or instruction in different fields are different, the preset intent semantic data can also be set and adjusted according to the actual application scenario. For example, the description of the intent type corresponding to each intent type can be modified to limit the applicable field and the actual scenario. Taking the example of a merchant in the food delivery field who needs intelligent dialogue services, if the intent type is knowledge Q&A, the corresponding description of the intent type is adjusted to "query the meaning of proprietary terms in the food delivery field or retrieve indicators, data tables, data calibers, etc. within the food delivery field", and the corresponding examples of the intent type can also be adjusted accordingly, modified to "What is the monthly active user index?", "What is the occupancy rate?", "What is the recommended position?". If the intent type is data query, the corresponding description of the intent type is adjusted to "specific data queries within the food delivery field that can be converted into structured query language", and the corresponding examples of the intent type can also be adjusted accordingly, modified to "What is the total transaction amount of our store in May?" or "What is the month-on-month growth rate of the sales volume of the hot pot category on this food delivery platform in May?", etc. If the intent type is other intents, the corresponding description of the intent type is adjusted to "other intents unrelated to the food delivery field", and the corresponding examples of the intent type are modified to "What is the music with the highest playback volume?", etc. Further, by modifying the preset intent semantic data, the information query range can be limited to a specified data source or database. For example, the description of the intent type corresponding to the intent type knowledge Q&A can be modified to "query the custom interpretation of the proprietary terms of the platform or company or retrieve indicators, data tables, data calibers, etc. within the platform (or merchant)". The description of the intent type corresponding to the intent type data query can be modified to "specific data queries within this platform or this company that can be converted into structured query language". The description of the intent type corresponding to the other intents can be modified to "other intents unrelated to the data within this platform or this company". It is easy to understand that when the limited range changes, the examples of the intent type corresponding to each intent type can also be adjusted accordingly, which will not be listed here.

[0055] In summary, the preset intent semantic data can be modified and adjusted to meet the actual requirements of various application scenarios, different fields, and various platforms or companies. Specifically, the query range can be limited by adjusting the preset intent semantic data to narrow the query range and speed up the query speed, or the final feedback information can be made more targeted.

[0056] In step S020, the preset intent semantic data is input into a pre-trained large-scale text embedding model to determine the corresponding intent semantic vector.

[0057] Specifically, the method for determining the intent semantic vector is the same as the method for determining the intent semantic query vector in step S220 above, and will not be elaborated here.

[0058] In step S030, a multi-level tree-like index structure is constructed according to the similarity between multiple intent semantic vectors, where each level of the index structure includes at least one group of intent semantic vectors or one group of intent index vectors.

[0059] If the similarity search and recall between vectors are performed one by one, the efficiency of vector recall will be very low. In the embodiment of the present invention, a multi-level tree-like index structure is constructed based on the similarity between intent semantic vectors, so that when vector recall is performed, search is carried out in the form of hierarchical indexing, realizing the improvement of vector recall efficiency.

[0060] Among them, the similarity can be cosine similarity, similarity determined based on Euclidean distance, or similarity determined based on dot product. The multi-level tree-like index structure includes intent index vectors and first-level intent index vectors, second-level intent index vectors,..., N-level intent index vectors aggregated based on the intent index vectors. Wherein, N is a positive integer, and N can be a preset value or determined according to the number of intent index vectors that each level of the index structure can contain. That is to say, the number of index levels in the multi-level tree-like index structure can be a preset number. When the preset intent semantic data increases rapidly, each level of the index structure can add new intent index vectors, but the total number of index levels of the index structure is always N. Or, the number of index levels in the multi-level tree-like index structure can change. When the preset intent semantic data increases rapidly, each level of the index structure can add new intent index vectors. When the number of intent index vectors at the top layer of the current tree-like structure exceeds the preset number, a new index structure is added on the basis of the index structure at the top layer of the current tree-like structure. At this time, the number of N increases by one. Therefore, the number of N can also be variable.

[0061] Specifically, the Faiss (Facebook AI Similarity Search) technology can be used to establish a multi-level tree-like index structure. The Faiss technology supports massive data (dense vectors) in a high-dimensional space, provides efficient and reliable similarity clustering and retrieval methods, and can support the clustering and search of vectors at the billion level.

[0062] Figure 4 FIG. is a flowchart of a method for constructing a multi-level tree-like index structure according to the similarity between multiple intent semantic vectors in an embodiment of the present invention. As Figure 4 shown, the method for constructing a multi-level tree-like index structure according to the similarity between multiple intent semantic vectors includes the following steps:

[0063] In step S031, multi-level clustering is performed according to the similarity between multiple said intent semantic vectors to determine a multi-level clustering result, where each level of clustering result includes multiple groups of intent semantic vectors or multiple groups of intent index vectors.

[0064] In step S032, the intent index vectors corresponding to each group of intent semantic vectors and the next-level intent index vectors corresponding to each group of intent index vectors are determined respectively.

[0065] In step S033, a multi-level tree-like index structure is constructed according to the intent semantic vectors and the intent index vectors at each level.

[0066] Specifically, first-level clustering is performed according to the similarity between multiple said intent semantic vectors to determine a first-level clustering result, which contains multiple groups of intent semantic vectors. The intent index vectors corresponding to each group of intent semantic vectors are determined. Specifically, for each group of clustered intent semantic vectors, the vector corresponding to the clustering center is extracted and determined as the intent index vector corresponding to this group of clustering. After using the above method to determine the intent index vectors corresponding to each group of intent semantic vectors in the first-level clustering result, the intent index vectors in the first-level clustering result are clustered to obtain a second-level clustering result, which contains multiple groups of intent index vectors. The next-level intent index vectors corresponding to each group of intent index vectors are determined. Specifically, for each group of clustered intent index vectors in the second-level clustering result, the vector corresponding to the clustering center is extracted and determined as the next-level (third-level) intent index vector corresponding to this group of clustering. The above method is repeated to obtain the clustering results at each level. Thus, a multi-level tree-like index structure is constructed. It should be noted that the clustering requirements at each level can be the same or different. The clustering requirements at each level are the clustering similarity thresholds, and the clustering similarity thresholds corresponding to the index structures at each level can be the same or different.

[0067] Among them, the index vector can be various types of indexes, such as IndexFlatL2, IndexFlatIP, IndexLSH, IndexPQ, or IndexHNSWFlat, etc. Different types of indexes are applicable to different recall scenarios. For example, IndexFlatL2 is calculated based on the Euclidean distance, IndexFlatIP is calculated based on the dot product. Both IndexFlatL2 and IndexFlatIP are applicable to small datasets, but IndexFlatIP has a faster recall speed than IndexFlatL2. IndexLSH is applicable to low-dimensional data, and IndexHNSWFlat has better recall quality but occupies more memory. The appropriate index can be selected according to actual needs.

[0068] In step S040, based on the intent semantic vector, the intent index vector, and the multi-level tree-like index structure, the intent semantic vector database is constructed.

[0069] Specifically, filling the intent semantic vector and the intent index vector into the corresponding positions in the multi-level tree-like index structure can construct the intent semantic vector database.

[0070] After constructing the intent semantic vector database by the above method, at least one intent semantic vector that meets the predetermined recall condition can be recalled from the constructed intent semantic vector database based on the intent semantic query vector output by the large-scale text embedding model.

[0071] In step S400, the target intent type is determined according to the similarity between the intent semantic query vector and the intent semantic vector.

[0072] In a possible implementation manner, when the number of intent semantic vectors in the intent semantic vector database is small, a brute-force search method can be used to determine the target intent type to ensure high accuracy of intent recognition.

[0073] Figure 5 This is a flowchart of the method for determining the target intent type according to the similarity between the intent semantic query vector and the intent semantic vector in the embodiment of the present invention. As Figure 5 shown, the method for determining the target intent type according to the similarity between the intent semantic query vector and the intent semantic vector includes the following steps:

[0074] In step S411, the highest similarity between the intent semantic query vector and the intent semantic vector is determined.

[0075] Specifically, the similarities between the intent semantic query vector and each intent semantic vector in the intent semantic vector database are respectively determined, and the highest similarity is determined by comparison.

[0076] In step S412, it is determined whether the highest similarity is greater than a preset similarity.

[0077] If it is greater, go to step S413; if it is not greater, go to step S414.

[0078] In step S413, in response to the highest similarity being greater than the preset similarity, the target intent type corresponding to the natural language query statement is determined as the intent type of the intent semantic vector corresponding to the highest similarity.

[0079] Among them, the preset similarity is determined according to the actual situation. The higher the preset similarity, the more accurate the target intent type obtained by intent recognition. The preset similarity can be 0.7, 0.8, 0.9, etc.

[0080] In step S414, in response to the highest similarity not being greater than the preset similarity, the target intent type corresponding to the natural language query statement is determined as other intents.

[0081] That is, when there is no intent semantic vector in the intent semantic vector database whose similarity with the intent semantic query vector meets the similarity requirement, the target intent type corresponding to the intent semantic query vector is determined as other intents.

[0082] In a possible implementation manner, when the number of intent semantic vectors in the intent semantic vector database is large and the cost required for brute-force search is high, a multi-level index matching method can be used to determine the target intent type to improve the efficiency of intent recognition.

[0083] Figure 6 This is a flowchart of the method for determining the target intent type according to the similarity between the intent semantic query vector and the intent semantic vector in the embodiment of the present invention. As Figure 6 shown, the method for determining the target intent type according to the similarity between the intent semantic query vector and the intent semantic vector includes the following steps:

[0084] In step S421, the similarity between the intent semantic query vector and the intent index vectors at each level is determined step by step.

[0085] Specifically, first, the similarity between the intent semantic query vector and each intent index vector in the N-level clustering result is determined, and the target N-level intent index vector with the highest similarity to the intent semantic query vector is determined, so as to determine multiple N-1-level intent index vectors corresponding to the target N-level intent index vector. Then, the similarity between the intent semantic query vector and each intent index vector in the N-1-level clustering result is determined, and the target N-1-level intent index vector with the highest similarity to the intent semantic query vector is determined. Then, the above steps are repeated until the target first-level intent index vector is determined.

[0086] In step S422, a target intent semantic vector group is determined according to the similarity. The target intent semantic vector group is a group of intent semantic vectors with the highest similarity to the intent semantic query vector.

[0087] Specifically, the intent semantic vector group corresponding to the target first-level intent index vector is determined as the target intent semantic vector group.

[0088] In step S423, the highest similarity is determined according to the intent semantic query vector and each intent semantic vector in the target intent semantic vector group.

[0089] In step S424, it is determined whether the highest similarity is greater than a preset similarity.

[0090] If it is greater, go to step S425; if it is not greater, go to step S426.

[0091] In step S425, in response to the highest similarity being greater than the preset similarity, the target intent type corresponding to the natural language query statement is determined as the intent type of the intent semantic vector corresponding to the highest similarity.

[0092] In step S426, in response to the highest similarity not being greater than the preset similarity, the target intent type corresponding to the natural language query statement is determined as other intents.

[0093] Through Figure 5 or Figure 6 the method shown can determine the target intent type. The method for determining the target intent type can be selected according to actual needs. If a higher accuracy requirement for intent recognition is required, then use Figure 5 the method shown; if a higher efficiency requirement for intent recognition is required, then use Figure 6 the method shown.

[0094] In the above steps S100 - S400, an intent semantic vector database is constructed, and vector recall is used for intent matching, achieving accurate intent recognition. Specifically, a large-scale text embedding model is used to determine the intent semantic vectors in the intent semantic vector database and the intent semantic vectors corresponding to natural language query statements, which can fully consider semantic and context relationships and make vector matching more accurate. Constructing the intent semantic vector database based on a multi-level tree-like index structure can improve the efficiency of intent matching. Therefore, the embodiments of the present invention can improve the efficiency of intent matching on the premise of ensuring the accuracy of intent matching.

[0095] In step S500, query feedback information is determined according to the natural language query statement and the target intent type.

[0096] Among them, the feedback information is the response information determined for the natural language query statement.

[0097] In a possible implementation manner, the query range can be limited according to the target intent type, such as a specified database. That is, the queryable databases or other query ranges corresponding to different intent types are different. For example, for knowledge Q&A, the query range is a common database that records noun interpretations and data calibers. The database corresponding to data query is a custom database of the platform or company, which is used to store specific data that can be converted into SQL. The database corresponding to other intents is a common database. The above query range can also be set according to the application scenario and the actual needs of the user.

[0098] Figure 7 It is a flowchart of a method for determining feedback information according to a natural language query statement and a target intent type in an embodiment of the present invention. As Figure 7 shown, the method for determining feedback information according to the natural language query statement and the target intent type includes the following steps:

[0099] In step S510, determine the corresponding target database according to the target intent type.

[0100] In step S520, determine the feedback information corresponding to the natural language query statement from the target database.

[0101] Furthermore, different weights can be set for data from different databases, so as to determine the feedback information by weighting. When setting weights, the data from the custom database of the platform or company can be set with a higher weight, so that the finally obtained feedback information better meets the actual needs of the user and improves the accuracy of the feedback information.

[0102] In a possible implementation manner, the natural language query statement and the corresponding target intent type can also be combined, and then the feedback information corresponding to the natural language query statement is determined according to the combined data.

[0103] Figure 8 It is a flowchart of another method for determining feedback information according to a natural language query statement and a target intent type in an embodiment of the present invention. As Figure 8 shown, the method for determining feedback information according to the natural language query statement and the target intent type includes the following steps:

[0104] In step S530, construct a corresponding query feature vector according to the natural language query statement and the target intent type.

[0105] Optionally, the natural language query statement can be respectively converted into a corresponding natural language query vector, and the target intent type can be converted into a corresponding target intent type vector, and then the natural language query vector and the target intent type vector are concatenated to obtain a query feature vector.

[0106] Optionally, first perform text concatenation on the natural language query statement and the target intent type, and then determine the query feature vector corresponding to the text obtained after concatenation.

[0107] In step S540, according to the similarity between the query feature vector and each feature vector in the database, the feedback information corresponding to the natural language query statement is determined.

[0108] Specifically, the cosine similarity between the query feature vector and each feature vector can be calculated, and the information corresponding to the feature vector with the highest cosine similarity is determined as the feedback information.

[0109] Optionally, a feature vector similarity threshold is preset in advance. If there is a feature vector with a similarity greater than the feature vector similarity threshold, then sort according to the similarity corresponding to the feature vector to determine the target feature vector, and then determine the feedback information. If there is no feature vector with a similarity greater than the feature vector similarity threshold, it is determined that the intent of the natural language query statement cannot be recognized, and then other processing is performed. Other processing can be to return a fixed answer, such as "Sorry, I can't answer your question", or to return a question, so as to obtain more information in a human-computer interaction manner to assist in generating feedback information. For example, the natural language query statement is "What is the month-on-month growth rate of the hot pot category?" Using the above steps S100 - S400, the target intent type is determined to be data query. When combining the natural language query statement and the target intent type to query the feedback information, the similarity between each feature vector in the database and the query feature vector is not greater than the feature vector similarity threshold. Then, according to the information corresponding to a predetermined number of feature vectors with relatively high similarity, the necessary information that may be missing in the natural language query statement, such as time or location, can be determined. For example, the query statement "May I ask which month's data you want to know?" or "Do you want to query the data of Wuhua District this month?" etc. can be output. More information is obtained through multiple human-computer interactions until the feedback information can be determined.

[0110] It should be noted that the above steps S100 - S400 are used for intent recognition to determine the target intent type. During this process, external data does not need to be accessed, which can ensure the security of internal data. Step S500 is used to determine feedback information. During the determination process, the data source can be customized. The customized data source can include an internal database and an external database. The intent semantic query vector can be matched with the intent semantic vector according to the intent semantics, and then the current query data source can be determined according to the intent semantic data corresponding to the corresponding intent semantic vector. For the query of sensitive data, it can be carried out in the internal database to better protect the security of internal data and avoid being monitored by other non - internal servers for accessing external data operations.

[0111] After receiving the natural - language query statement input by the user, the method according to the embodiment of the present invention inputs it into a pre - trained large - scale text embedding model that can better capture semantic and context relationships, obtains the corresponding intent semantic query vector, and then recalls at least one intent semantic vector from the preset intent semantic vector database according to the intent semantic query vector, determines the similarity between the intent semantic query vector and each recalled intent semantic vector respectively, so as to determine the target intent type. Finally, according to the natural - language query statement and the target intent type, the feedback information corresponding to the query natural - language query statement is determined. The embodiment of the present invention uses a large - scale text embedding model to convert the natural - language query statement into a corresponding intent semantic query vector and then determines the corresponding intent type, which can incorporate context information into the intent recognition process, so that the determined intent type is more accurate. Then, by combining the natural - language query with the target intent type to determine the feedback information, the accuracy of the feedback information can be improved. At the same time, there can be no interaction with external data in the above process, thus ensuring the security of internal data.

[0112] Figure 9 This is the data - flow diagram of the natural - language query method according to the embodiment of the present invention. As Figure 9 shown, the data flow of the natural - language query method is as follows:

[0113] In step S901, a natural - language query statement input by the user is obtained.

[0114] In step S902, the natural - language query statement is input into a pre - trained large - scale text embedding model.

[0115] In step S903, the large - scale text embedding model outputs the corresponding intent semantic query vector.

[0116] In step S904, according to the intent semantic query vector, at least one intent semantic vector is recalled from the preset intent semantic vector database.

[0117] Among them, the intention semantic vector database stores intention semantic vectors in a multi-level tree-like index structure. When recalling vectors, the similarity between the intention semantic query vector and intention index vectors at all levels can be determined step by step, and a target intention semantic vector group can be determined according to the similarity. The target intention semantic vector group is a group of intention semantic vectors with the highest similarity to the intention semantic query vector. Recall multiple intention semantic vectors in the target semantic vector group.

[0118] In step S905, calculate the similarity between the intention semantic query vector and each recalled intention semantic vector respectively, compare multiple similarities, and determine the highest similarity.

[0119] In step S906, determine the target intention type according to the highest similarity.

[0120] Specifically, in response to the highest similarity being greater than the preset similarity, determine the intention type of the natural language query statement as the intention type of the intention semantic vector corresponding to the highest similarity. In response to the highest similarity not being greater than the preset similarity, determine the intention type of the natural language query statement as other intentions.

[0121] In step S907, determine the corresponding target database according to the target intention type.

[0122] In step S908, determine the feedback information corresponding to the natural language query statement from the target database.

[0123] In step S909, construct a corresponding query feature vector according to the natural language query statement and the target intention type.

[0124] In step S910, calculate the similarity between the query feature vector and each feature vector in the database.

[0125] In step S911, determine the feedback information according to the similarity between the query feature vector and each feature vector in the database.

[0126] After receiving a natural language query statement input by a user, the method according to an embodiment of the present invention inputs it into a pre-trained large-scale text embedding model that can better capture semantic and context relationships to obtain a corresponding intent semantic query vector. Then, according to the intent semantic query vector, at least one intent semantic vector is recalled from a preset intent semantic vector database, and the similarities between the intent semantic query vector and each recalled intent semantic vector are determined respectively, so as to determine the target intent type. Finally, according to the natural language query statement and the target intent type, the feedback information corresponding to the query natural language query statement is determined. In the embodiment of the present invention, a large-scale text embedding model is used to convert a natural language query statement into a corresponding intent semantic query vector and then determine the corresponding intent type, and context information can be incorporated into the intent recognition process, so that the determined intent type is more accurate. Then, the feedback information is determined by combining the natural language query and the target intent type, which can improve the accuracy of the feedback information. At the same time, there may be no interaction with external data in the above process, so as to ensure the security of internal data.

[0127] Figure 10 It is a schematic diagram of a natural language query device according to an embodiment of the present invention. As Figure 10 shown, the natural language query device includes:

[0128] An acquisition module 1001, configured to acquire a natural language query statement.

[0129] A first determination module 1002, configured to input the natural language query statement into a pre-trained large-scale text embedding model to determine a corresponding intent semantic query vector.

[0130] A recall module 1003, configured to recall at least one intent semantic vector from a preset intent semantic vector database according to the intent semantic query vector.

[0131] A second determination module 1004, configured to determine a target intent type according to the similarity between the intent semantic query vector and the intent semantic vector.

[0132] A third determination module 1005, configured to determine query feedback information according to the natural language query statement and the target intent type.

[0133] After receiving a natural language query statement input by a user, the device according to an embodiment of the present invention inputs the natural language query statement into a large-scale text embedding model that has been pre-trained and can better capture semantic and context relationships, to obtain a corresponding intent semantic query vector. Then, according to the intent semantic query vector, at least one intent semantic vector is recalled from a preset intent semantic vector database, and the similarities between the intent semantic query vector and each recalled intent semantic vector are determined respectively, so as to determine a target intent type. Finally, according to the natural language query statement and the target intent type, feedback information corresponding to the natural language query statement is determined. In the embodiment of the present invention, a large-scale text embedding model is used to convert a natural language query statement into a corresponding intent semantic query vector and then determine a corresponding intent type, so that context information can be incorporated into the intent recognition process, thereby making the determined intent type more accurate. Then, by combining the natural language query and the target intent type to determine the feedback information, the accuracy of the feedback information can be improved. At the same time, there may be no interaction with external data in the above process, thereby ensuring the security of internal data.

[0134] Figure 11 is a schematic diagram of an electronic device according to an embodiment of the present invention. In this embodiment, the electronic device 1100 includes a server, a terminal, etc. As Figure 11 shown, the electronic device 1100: includes at least one processor 1101; and, a memory 1102 communicatively connected to at least one processor 1101; and, a communication component 1103 communicatively connected to a scanning device, and the communication component 1103 receives and sends data under the control of the processor 1101; wherein, the memory 1102 stores instructions executable by at least one processor 1101, and the instructions are executed by at least one processor 1101 to implement the above natural language query method.

[0135] Specifically, the electronic device includes: one or more processors 1101 and a memory 1102, Figure 11 taking one processor 1101 as an example. The processor 1101 and the memory 1102 may be connected by a bus or other means, Figure 11 taking connection by a bus as an example. As a non-volatile computer-readable storage medium, the memory 1102 can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. The processor 1101 executes various functional applications and data processing of the device by running the non-volatile software programs, instructions, and modules stored in the memory 1102, that is, implements the above natural language query method.

[0136] The memory 1102 may include a program storage area and a data storage area. The program storage area may store an operating system and application programs required for at least one function. The data storage area may store an option list and the like. In addition, the memory 1102 may include a high-speed random access memory and may also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices. In some embodiments, the memory 1102 may optionally include a memory remotely provided with respect to the processor 1101, and these remote memories may be connected to an external device through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0137] One or more modules are stored in the memory 1102 and, when executed by one or more processors 1101, execute the natural language query method in any of the above method embodiments.

[0138] The above product can execute the method provided in the embodiments of the present application, and has corresponding functional modules and beneficial effects for executing the method. For technical details not described in detail in this embodiment, reference may be made to the method provided in the embodiments of the present application.

[0139] After receiving a natural language query statement input by a user, an embodiment of the present invention inputs it into a large-scale text embedding model that has been pre-trained and can better capture semantic and context relationships to obtain a corresponding intent semantic query vector. Then, according to the intent semantic query vector, at least one intent semantic vector is recalled from a preset intent semantic vector database, the similarities between the intent semantic query vector and each recalled intent semantic vector are determined respectively, so as to determine the target intent type. Finally, according to the natural language query statement and the target intent type, feedback information corresponding to the natural language query statement is determined. The embodiment of the present invention uses a large-scale text embedding model to convert a natural language query statement into a corresponding intent semantic query vector and then determine the corresponding intent type, and can incorporate context information into the process of intent recognition, so that the determined intent type is more accurate. Then, by combining the natural language query with the target intent type to determine the feedback information, the accuracy of the feedback information can be improved. At the same time, there may be no interaction with external data in the above process, thus ensuring the security of internal data.

[0140] Another embodiment of the present invention relates to a non-volatile storage medium for storing a computer-readable program, and the computer-readable program is used for a computer to execute the above partial or all method embodiments.

[0141] An embodiment of the present invention discloses A1. A natural language query method, the method comprising:

[0142] Obtain a natural language query statement;

[0143] Input the natural language query statement into a pre-trained large-scale text embedding model to determine the corresponding intent semantic query vector;

[0144] Recall at least one intent semantic vector from a preset intent semantic vector database according to the intent semantic query vector;

[0145] Determine the target intent type according to the similarity between the intent semantic query vector and the intent semantic vector;

[0146] Determine query feedback information according to the natural language query statement and the target intent type.

[0147] A2. According to the method described in A1, the method for constructing the intent semantic vector database includes:

[0148] Determine multiple preset intent semantic data, where the preset intent semantic data at least includes intent types, intent descriptions corresponding to the intent types, and multiple intent examples corresponding to the intent types, and the intent types at least include knowledge Q&A, data query, and other intents;

[0149] Input the preset intent semantic data into a pre-trained large-scale text embedding model to determine the corresponding intent semantic vector;

[0150] Construct a multi-level tree-like index structure according to the similarity between multiple intent semantic vectors, where each level of the index structure at least includes a group of intent semantic vectors or a group of intent index vectors;

[0151] Construct the intent semantic vector database based on the intent semantic vector, the intent index vector, and the multi-level tree-like index structure.

[0152] A3. According to the method described in A2, the step of constructing a multi-level tree-like index structure according to the similarity between multiple intent semantic vectors includes:

[0153] Perform multi-level clustering according to the similarity between multiple intent semantic vectors to determine the multi-level clustering result, where each level of the clustering result includes multiple groups of intent semantic vectors or multiple groups of intent index vectors;

[0154] Respectively determine the intent index vectors corresponding to each group of intent semantic vectors and the next-level intent index vectors corresponding to each group of intent index vectors;

[0155] Construct a multi-level tree-like index structure according to the intent semantic vector and each level of intent index vectors.

[0156] A4. According to the method described in A1, the steps for the large-scale text embedding model to determine the intent semantic query vector include:

[0157] Determine the format type of the natural language query statement, where the format type includes at least text and audio;

[0158] Input the natural language query statement into the encoder corresponding to the format type to determine the corresponding intent semantic query vector.

[0159] A5. According to the method described in A3, determining the target intent type according to the similarity between the intent semantic query vector and the intent semantic vector includes:

[0160] Determine the similarity between the intent semantic query vector and each level of intent index vector step by step;

[0161] Determine the target intent semantic vector group according to the similarity, where the target intent semantic vector group is a group of intent semantic vectors with the highest similarity to the intent semantic query vector;

[0162] Determine the highest similarity according to the intent semantic query vector and each intent semantic vector in the target intent semantic vector group;

[0163] In response to the highest similarity being greater than the preset similarity, determine the target intent type corresponding to the natural language query statement as the intent type of the intent semantic vector corresponding to the highest similarity.

[0164] A6. According to the method described in A1, determining the target intent type according to the similarity between the intent semantic query vector and the intent semantic vector includes:

[0165] Determine the highest similarity between the intent semantic query vector and the intent semantic vector;

[0166] In response to the highest similarity being greater than the preset similarity, determine the target intent type corresponding to the natural language query statement as the intent type of the intent semantic vector corresponding to the highest similarity.

[0167] A7. According to the method described in A5 or A6, the method further includes:

[0168] In response to the highest similarity not being greater than the preset similarity, determine the target intent type corresponding to the natural language query statement as other intent.

[0169] A8. According to the method described in A1, determining the query feedback information according to the natural language query statement and the target intent type includes:

[0170] Determine the corresponding target query feedback information library according to the target intent type;

[0171] Determine the query feedback information corresponding to the natural language query statement from the target query feedback information library.

[0172] A9. According to the method described in A1, the determining the query feedback information according to the natural language query statement and the target intent type includes:

[0173] Construct a corresponding query feature vector according to the natural language query statement and the target intent type;

[0174] Determine the query feedback information corresponding to the natural language query statement according to the similarity between the query feature vector and each feature vector in the query feedback information library.

[0175] B1. A natural language query device, the device includes:

[0176] An acquisition module, configured to acquire a natural language query statement;

[0177] A first determination module, configured to input the natural language query statement into a pre-trained large-scale text embedding model to determine a corresponding intent semantic query vector;

[0178] A recall module, configured to recall at least one intent semantic vector from a preset intent semantic vector database according to the intent semantic query vector;

[0179] A second determination module, configured to determine a target intent type according to the similarity between the intent semantic query vector and the intent semantic vector;

[0180] A third determination module, configured to determine query feedback information according to the natural language query statement and the target intent type.

[0181] C1. An electronic device, including a memory and a processor, the memory is used to store one or more computer program instructions, wherein, the one or more computer program instructions are executed by the processor to implement the method described in any one of A1 - A9.

[0182] D1. A computer-readable storage medium, in which a computer program is stored, and when the computer program is executed by a processor, it implements the method described in any one of A1 - A9.

[0183] That is, those skilled in the art can understand that all or part of the steps in implementing the methods of the above embodiments can be completed by instructing relevant hardware through a program. The program is stored in a storage medium, including several instructions to enable a device (which can be a single-chip microcomputer, a chip, etc.) or a processor to execute all or part of the steps of the methods described in the embodiments of the present application. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0184] The foregoing are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, various modifications and changes can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

Claims

1. A natural language query method, characterized in that, The method includes: Obtaining a natural language query statement; Inputting the natural language query statement into a pre-trained large-scale text embedding model to determine a corresponding intent semantic query vector; Recalling at least one intent semantic vector from a preset intent semantic vector database according to the intent semantic query vector; Determining a target intent type according to the similarity between the intent semantic query vector and the intent semantic vector; Determining query feedback information according to the natural language query statement and the target intent type.

2. The method according to claim 1, characterized in that, The method for constructing the intent semantic vector database includes: Determining a plurality of preset intent semantic data, where the preset intent semantic data at least includes an intent type, an intent description corresponding to the intent type, and a plurality of intent examples corresponding to the intent type, and the intent type at least includes knowledge Q&A, data query, and other intents; Inputting the preset intent semantic data into a pre-trained large-scale text embedding model to determine corresponding intent semantic vectors; Constructing a multi-level tree-like index structure according to the similarity between a plurality of the intent semantic vectors, where each level of the index structure at least includes a group of intent semantic vectors or a group of intent index vectors; Constructing the intent semantic vector database based on the intent semantic vectors, the intent index vectors, and the multi-level tree-like index structure.

3. The method according to claim 2, characterized in that, The constructing a multi-level tree-like index structure according to the similarity between a plurality of the intent semantic vectors includes: Performing multi-level clustering according to the similarity between a plurality of the intent semantic vectors to determine a multi-level clustering result, where each level of the clustering result includes multiple groups of intent semantic vectors or multiple groups of intent index vectors; Respectively determining intent index vectors corresponding to each group of intent semantic vectors and next-level intent index vectors corresponding to each group of intent index vectors; Constructing a multi-level tree-like index structure according to the intent semantic vectors and the intent index vectors at each level.

4. The method according to claim 1, characterized in that, The steps for the large-scale text embedding model to determine an intent semantic query vector include: Determining the format type of the natural language query statement, where the format type at least includes text and audio; Inputting the natural language query statement into an encoder corresponding to the format type to determine a corresponding intent semantic query vector.

5. The method according to claim 3, characterized in that, The determining a target intent type according to the similarity between the intent semantic query vector and the intent semantic vector includes: Determining the similarity between the intent semantic query vector and the intent index vectors at each level step by step; Determining a target intent semantic vector group according to the similarity, where the target intent semantic vector group is a group of intent semantic vectors with the highest similarity to the intent semantic query vector; Determining the highest similarity according to the intent semantic query vector and each intent semantic vector in the target intent semantic vector group; In response to the highest similarity being greater than a preset similarity, determining the target intent type corresponding to the natural language query statement as the intent type of the intent semantic vector corresponding to the highest similarity.

6. The method according to claim 1, characterized in that, The determining a target intent type according to the similarity between the intent semantic query vector and the intent semantic vector includes: Determining the highest similarity between the intent semantic query vector and the intent semantic vector; In response to the highest similarity being greater than a preset similarity, determining the target intent type corresponding to the natural language query statement as the intent type of the intent semantic vector corresponding to the highest similarity.

7. The method according to claim 5 or 6, characterized in that, The method further includes: In response to the highest similarity not being greater than the preset similarity, determining the target intent type corresponding to the natural language query statement as other intent.

8. A natural language query device, characterized in that, The apparatus includes: An acquisition module, configured to acquire a natural language query statement; A first determination module, configured to input the natural language query statement into a pre-trained large-scale text embedding model to determine a corresponding intent semantic query vector; A recall module, configured to recall at least one intent semantic vector from a preset intent semantic vector database according to the intent semantic query vector; A second determination module, configured to determine a target intent type according to the similarity between the intent semantic query vector and the intent semantic vector; A third determination module, configured to determine query feedback information according to the natural language query statement and the target intent type.

9. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1-7 is implemented.

Images