Multi-intent determination method and device, storage medium and electronic device

Abstract

The application discloses a multi-intent determination method and device, a storage medium and an electronic device, relates to the technical field of smart homes, and comprises the following steps: performing data preprocessing on interaction information input by a target object to obtain preprocessed target interaction information; performing intent feature extraction on the target interaction information through a deep learning model to obtain a plurality of intent features corresponding to the target interaction information; determining the confidence of an intent recognition result corresponding to each intent feature through a target classifier, and determining a multi-intent recognition result according to the confidence of the intent recognition result corresponding to each intent feature; and determining an intent set corresponding to the interaction information in the multi-intent recognition result according to a logical relationship in the interaction information. By using the technical solution, the problem that the accuracy of an intent recognition technology is not high when the intent recognition technology processes multi-intents in a complex context in the related art is solved.

Description

Technical Field

[0001] This application relates to the field of smart home technology, and more specifically, to a method and apparatus for determining multiple intentions, a storage medium, and an electronic device. Background Technology

[0002] In the field of modern natural language processing, especially in intelligent customer service, virtual assistants, and dialogue systems, multi-intent recognition technology is crucial. When users interact with a system, they often express multiple different needs or information points in a single sentence. Accurately identifying these intents is essential for providing the services users require, answering questions, and facilitating efficient communication. However, existing multi-intent recognition technologies face significant technical challenges, primarily in terms of accuracy and processing efficiency.

[0003] Traditional intent recognition methods are typically designed to identify single intents, which performs well when users express a single need. However, their accuracy drops significantly when dealing with multiple intents in complex contexts. For example, if a user says, "I want to check tomorrow's weather and book a train ticket to Shenzhen," traditional methods may only correctly identify one intent, while the other may be ignored or misunderstood.

[0004] In summary, existing intent recognition technologies are not very accurate when dealing with multiple intents in complex contexts.

[0005] Regarding the issue that intent recognition technology has low accuracy when dealing with multiple intents in complex contexts, no effective solution has yet been proposed. Summary of the Invention

[0006] This application provides a method, apparatus, storage medium, and electronic device for determining multiple intentions, in order to at least solve the problem that the accuracy of intention recognition technology is not high when dealing with multiple intentions in complex contexts in related technologies.

[0007] According to one embodiment of this application, a method for determining multiple intentions is provided, comprising: preprocessing interaction information input by a target object to obtain preprocessed target interaction information; extracting intention features from the target interaction information using a deep learning model to obtain multiple intention features corresponding to the target interaction information; wherein the deep learning model is trained by machine learning using multiple sets of data, each set of data including: sample interaction information and labels of the intention features of the sample interaction information; determining the confidence level of the intention recognition result corresponding to each intention feature using a target classifier, and determining a multiple intention recognition result based on the confidence level of the intention recognition result corresponding to each intention feature; and determining the intention set corresponding to the interaction information in the multiple intention recognition result based on the logical relationship in the interaction information.

[0008] In an exemplary embodiment, intent feature extraction of the target interaction information is performed using a deep learning model to obtain multiple intent features corresponding to the target interaction information. This includes: converting each word in the target interaction information into a word vector feature using a word embedding model, and determining the word frequency-inverse document frequency of each word; weighting the word vector feature and the word frequency-inverse document frequency of each word to determine the target word vector feature of each word; and inputting the target word vector feature of each word into the deep learning model so that the deep learning model extracts intent features of the target interaction information based on the target word vector feature of each word to obtain multiple intent features corresponding to the target interaction information.

[0009] In an exemplary embodiment, converting each word in the target interactive information into word vector features using a word embedding model includes: determining the part-of-speech information of each word in the target interactive information; inputting each word and its part-of-speech information into the word embedding model so that the word embedding model determines the semantic vector features and part-of-speech vector features of each word; and determining the word vector features of each word based on its semantic vector features and part-of-speech vector features.

[0010] In an exemplary embodiment, the target word vector features of each word are input into the deep learning model, so that the deep learning model extracts intent features from the target interaction information based on the target word vector features of each word to obtain multiple intent features corresponding to the target interaction information. This includes: when N meets a first preset value, inputting the target word vector features of the Nth word into the hidden layer of the deep learning model, so that the hidden layer determines the Nth hidden state based on the target word vector features of the Nth word, where N is a positive integer; when N does not meet the first preset value, inputting the target word vector features of the Nth word into the hidden layer of the deep learning model, so that the hidden layer determines the Nth hidden state based on the target word vector features of the Nth word and the (N-1)th hidden state of the hidden layer; when the Nth hidden state is the last hidden state of the hidden layer, inputting the Nth hidden state into the output layer of the deep learning model, so that the output layer outputs multiple intent features corresponding to the target interaction information based on the Nth hidden state, the weights of the output layer, and the bias.

[0011] In an exemplary embodiment, determining the intent set corresponding to the interaction information in the multi-intent recognition result based on the logical relationships in the interaction information includes: determining whether there are connectives in the interaction information; if the connectives are present in the interaction information, determining the logical relationships in the interaction information based on the connectives; if the connectives are absent in the interaction information, determining the grammatical structure information in the interaction information, and determining the logical relationships in the interaction information based on the grammatical structure information; determining the logical type corresponding to the logical relationship, and determining the selection strategy corresponding to the logical type; and determining the intent set corresponding to the interaction information in the multi-intent recognition result based on the selection strategy.

[0012] In one exemplary embodiment, data preprocessing is performed on the interaction information input by the target object to obtain preprocessed target interaction information, including: data cleaning of the interaction information to obtain first interaction information; determining recommendation information that is associated with the first interaction information; and combining the first interaction information and the recommendation information into the target interaction information.

[0013] In an exemplary embodiment, determining the recommendation information associated with the first interaction information includes at least one of the following: determining keywords in the first interaction information and determining a user profile of the target object; determining user preferences of the target object based on the user profile and determining recommendation information corresponding to the keywords based on the user preferences; determining keywords in the first interaction information; and determining recommendation information corresponding to the keywords in an information database, wherein the information database includes information about the target object and other objects besides the target object.

[0014] In an exemplary embodiment, determining the confidence level of the intent recognition result corresponding to each intent feature using a target classifier, and determining a multi-intent recognition result based on the confidence level of the intent recognition result corresponding to each intent feature, includes: determining the distance between each intent feature and a preset hyperplane using the target classifier; determining the intent recognition result corresponding to each intent feature and the confidence level of the intent recognition result corresponding to each intent feature based on the distance between each intent feature and the preset hyperplane; determining the relationship between the confidence level of each intent recognition result and a preset confidence level; determining the intent feature with a confidence level greater than the preset confidence level based on the relationship, and determining the intent recognition result with a confidence level greater than the preset confidence level as the multi-intent recognition result.

[0015] According to another embodiment of this application, a multi-intent determination device is also provided, comprising: a preprocessing module, configured to preprocess interaction information input by a target object to obtain preprocessed target interaction information; a feature extraction module, configured to extract intent features from the target interaction information using a deep learning model to obtain multiple intent features corresponding to the target interaction information, wherein the deep learning model is trained using multiple sets of data through machine learning, and each set of data includes: sample interaction information and labels of intent features of the sample interaction information; an intent classification module, configured to determine the confidence level of the intent recognition result corresponding to each intent feature using a target classifier, and determine a multi-intent recognition result based on the confidence level of the intent recognition result corresponding to each intent feature; and a determination module, configured to determine the intent set corresponding to the interaction information in the multi-intent recognition result based on the logical relationship in the interaction information.

[0016] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer program, and the computer program is configured to execute the above-described method for determining multiple intentions when it is run.

[0017] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the above-described method for determining multiple intentions through the computer program.

[0018] In this embodiment, the interactive information input by the target object is preprocessed to obtain preprocessed target interactive information; intent features are extracted from the target interactive information using a deep learning model to obtain multiple intent features corresponding to the target interactive information; the confidence level of the intent recognition result corresponding to each intent feature is determined by a target classifier, and a multi-intent recognition result is determined based on the confidence level of the intent recognition result corresponding to each intent feature; the intent set corresponding to the interactive information is determined from the multi-intent recognition result based on the logical relationship in the interactive information; in this embodiment, preprocessing enhances data quality, a deep learning model extracts rich features, confidence analysis filters highly credible intents, and logical relationships integrate the intent set. This not only handles complex user language input but also intelligently understands and responds to the user's real needs, thus achieving a significant improvement in accuracy for multi-intent recognition in complex contexts. The above technical solution solves the problem of low accuracy in intent recognition technology when handling multiple intents in complex contexts. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the hardware environment for a method for determining multiple intentions according to an embodiment of this application;

[0022] Figure 2 This is a flowchart of a method for determining multiple intents according to an embodiment of this application;

[0023] Figure 3 This is a schematic diagram of a method for determining multiple intentions according to an embodiment of this application;

[0024] Figure 4 This is a schematic diagram of a preprocessing method according to an embodiment of this application;

[0025] Figure 5This is a schematic diagram of the part-of-speech tagging results according to an embodiment of this application;

[0026] Figure 6 This is a schematic diagram of a recurrent neural network according to an embodiment of this application;

[0027] Figure 7 This is a structural block diagram of a multi-intent determination device according to an embodiment of this application. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application 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 this application 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.

[0030] According to one aspect of the embodiments of this application, a method for determining multiple intentions is provided. This method for determining multiple intentions is widely applicable to whole-house intelligent digital control application scenarios such as smart homes, smart home ecosystems, and intelligence house ecosystems. Optionally, in this embodiment, the above-mentioned method for determining multiple intentions can be applied to, for example... Figure 1 The hardware environment shown consists of terminal device 102 and server 104. For example... Figure 1 As shown, server 104 is connected to terminal device 102 via a network and can be used to provide services (such as application services) to the terminal or clients installed on the terminal. A database can be set up on the server or independently of the server to provide data storage services for server 104. Cloud computing and / or edge computing services can be configured on the server or independently of the server to provide data processing services for server 104.

[0031] The aforementioned network may include, but is not limited to, at least one of the following: wired network, wireless network. The aforementioned wired network may include, but is not limited to, at least one of the following: wide area network, metropolitan area network, local area network. The aforementioned wireless network may include, but is not limited to, at least one of the following: Wi-Fi (Wireless Fidelity), Bluetooth. The terminal device 102 may not be limited to PC, mobile phone, tablet computer, smart air conditioner, smart range hood, smart refrigerator, smart oven, smart stove, smart washing machine, smart water heater, smart washing equipment, smart dishwasher, smart projector, smart TV, smart clothes rack, smart curtains, smart audio-visual equipment, smart socket, smart speaker, smart speaker box, smart fresh air equipment, smart kitchen and bathroom equipment, smart bathroom equipment, smart robot vacuum cleaner, smart window cleaning robot, smart mopping robot, smart air purifier, smart steam oven, smart microwave oven, smart water heater, smart air purifier, smart water dispenser, smart door lock, etc.

[0032] This embodiment provides a method for determining multiple intentions, applied to the aforementioned terminal device or server. Figure 2 This is a flowchart of a method for determining multiple intents according to an embodiment of this application, the process including the following steps:

[0033] Step S202: Perform data preprocessing on the interactive information input by the target object to obtain preprocessed target interactive information;

[0034] In step S202, the interactive information input by the target object is preprocessed. Preprocessing includes, but is not limited to, basic operations such as noise removal, text standardization, and word segmentation. It may also involve advanced processing such as syntactic analysis and semantic enhancement to ensure that the deep learning model receives the clearest and most relevant data. Preprocessing reduces the training difficulty of the model, improves its ability to understand complex contexts, and thus enhances the accuracy of intent recognition.

[0035] Step S204: Extract intent features from the target interaction information using a deep learning model to obtain multiple intent features corresponding to the target interaction information. The deep learning model is trained using multiple sets of data through machine learning. Each set of data includes: sample interaction information and a label of the intent feature of the sample interaction information.

[0036] Deep learning models, such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs), can capture dependencies in long sequences of text, making them particularly effective for identifying the logical connections between multiple intentions within a sentence. The intention features extracted by deep learning models are richer and more detailed, encompassing various forms of user intent, thus improving the accuracy of identifying multiple intentions in complex contexts.

[0037] It should be noted that obtaining sample interaction information includes, but is not limited to, one of the following methods:

[0038] 1. Historical interaction information: Records of previous conversations between the user and the system, including the user's queries, requests, commands, etc. in different scenarios, as well as the corresponding responses and feedback.

[0039] 2. Corpus: This includes publicly available text resources such as Wikipedia, web page text, and news information.

[0040] 3. Internal text and user feedback: This can be internal text data of a specific company or platform, such as product descriptions, user manuals, FAQs, etc., as well as feedback information collected from users.

[0041] 4. Diverse contextual data: In order for the model to handle a variety of different scenarios, the training data should include sentences from different domains, different language styles and different complexities, as well as related contextual information.

[0042] Step S206: Determine the confidence level of the intent recognition result corresponding to each intent feature through the target classifier, and determine the multi-intent recognition result based on the confidence level of the intent recognition result corresponding to each intent feature;

[0043] A target classifier is used to determine the confidence level corresponding to each intent feature, and the multi-intent recognition result is determined based on the confidence level. The confidence level can assess the reliability of each intent; intents with high confidence are more likely to be the user's true intent. When dealing with complex contexts, the classifier can effectively identify primary and secondary intents, avoiding treating all possible interpretations the same, thereby improving the accuracy and practicality of the recognition results.

[0044] Step S208: Determine the intent set corresponding to the interaction information in the multi-intent recognition result according to the logical relationship in the interaction information.

[0045] Step S208 analyzes the logical relationships in the interaction information to determine a reasonable set of intentions from the multi-intention recognition results, integrating multiple possible intentions into a coherent response. Logical relationships, such as parallel, adversative, and conditional relationships, can reveal the correlation or priority between different intentions, helping the system understand the full picture of the user's intentions. For example, a user might say, "I want to buy a washing machine, but I need the size to fit my small apartment." By identifying parallel and adversative logic, the system can correctly treat buying a washing machine and the size restriction as two related intentions, rather than treating them in isolation. This determination of the intention set based on logical relationships makes the system's recognition results more consistent with the natural logic of human language, thereby significantly improving the accuracy of multi-intention recognition in complex contexts.

[0046] Through the above steps, the interactive information input by the target object is preprocessed to obtain preprocessed target interactive information; intent features are extracted from the target interactive information using a deep learning model to obtain multiple intent features corresponding to the target interactive information; the confidence level of the intent recognition result corresponding to each intent feature is determined by a target classifier, and a multi-intent recognition result is determined based on the confidence level of the intent recognition result corresponding to each intent feature; the intent set corresponding to the interactive information is determined from the multi-intent recognition result based on the logical relationship in the interactive information; in this embodiment, preprocessing enhances data quality, a deep learning model extracts rich features, confidence analysis filters highly credible intents, and logical relationships integrate the intent set. This not only handles complex user language input but also intelligently understands and responds to the user's real needs, thus achieving a significant improvement in accuracy for multi-intent recognition in complex contexts. The above technical solution solves the problem of low accuracy in intent recognition technology when handling multiple intents in complex contexts.

[0047] Optionally, step S204 above can also be implemented in the following way: converting each word in the target interaction information into word vector features through a word embedding model, and determining the word frequency-inverse document frequency of each word; weighting the word vector features of each word and the word frequency-inverse document frequency of each word to determine the target word vector features of each word; inputting the target word vector features of each word into the deep learning model, so that the deep learning model extracts intent features from the target interaction information based on the target word vector features of each word, to obtain multiple intent features corresponding to the target interaction information.

[0048] It's important to note that word embedding models, such as Word2Vec, GloVe, or BERT, convert each word in a text into a fixed-length vector. This vector captures the semantic information of the words. For example, in the sentence "I want to book a flight to Shanghai tomorrow and check the weather there," the word embedding model would convert "flight" and "weather" into a vector containing the semantic information of "flight" and "weather," such as their relevance to concepts like travel, booking, and price.

[0049] TF-IDF is used to evaluate the importance of a word to a document or a document in a corpus. Term frequency (TF) refers to how often a word appears in a document, while inverse document frequency (IDF) is the reciprocal of the frequency of that word in the documents in a set of documents, used to penalize common words. The higher the TF-IDF value, the more important the word is in the current document. For example, in the sentence above, "flight tickets" and "weather" might have high TF-IDF values ​​because they are crucial for identifying two different intentions (booking a flight ticket).

[0050] By combining word vector features with TF-IDF values, the system can more accurately capture the importance and semantic information of words. For example, if the word "ticket" has strong word vector features and a high TF-IDF value, then after weighting, it will have a greater weight in the model, helping the model to more accurately identify the intent to "book a ticket".

[0051] The weighted word vector features are then fed into deep learning models, such as LSTM or GRU, which can understand the sequence information and contextual relationships of words. In the example above, the model not only recognizes the importance of the word "ticket," but also understands the relationship between "book" and "ticket," as well as the structure of the entire sentence, thus more accurately identifying the user's intention to book a ticket.

[0052] Deep learning models, based on weighted word vector features, can extract multiple intent-related features. For example, for the sentence "I want to book a flight to Shanghai tomorrow and check the weather there," the model might extract features for the intents "book a flight" and "check the weather."

[0053] For example, suppose there is a sentence: I want to buy a Haier washing machine, but I need to compare prices.

[0054] Each word is converted into a vector using the Word2Vec word embedding model:

[0055] “Purchase”: [0.2,0.6,-0.3,...]; “Haier”: [0.5,-0.1,0.7,...]; “Washing Machine”: [0.8,0.4,-0.2,...]; “Price”: [-0.4,0.3,0.7,...].

[0056] Calculate the TF-IDF value for each word:

[0057] "Purchase": TF-IDF = 0.7; "Haier": TF-IDF = 0.8; "Washing Machine": TF-IDF = 0.9; "Price": TF-IDF = 0.6.

[0058] The word vector features and TF-IDF values ​​of each word are weighted:

[0059] The weighted word vector features for “purchase” are: [0.14, 0.42, -0.21, ...] (0.2*0.7, 0.6*0.7, -0.3*0.7, ...); the weighted word vector features for “Haier” are: [0.4, -0.08, 0.56, ...]; the weighted word vector features for “washing machine” are: [0.72, 0.36, -0.18, ...]; the weighted word vector features for “price” are: [-0.24, 0.18, 0.42, ...].

[0060] The weighted feature vectors are fed into an LSTM deep learning model, which identifies two main intentions based on these features: "buying a washing machine" and "comparing prices." The model analyzes the relationships and context between words to determine that the user has two needs: on the one hand, they want to buy a Haier washing machine, and on the other hand, they want to compare washing machine prices to make a more informed purchasing decision.

[0061] Through the embodiments of this application, deep learning models can better understand multiple intentions of users in complex contexts, thereby improving the accuracy of intention recognition.

[0062] Optionally, each word in the target interactive information is converted into word vector features through a word embedding model, including: determining the part-of-speech information of each word in the target interactive information; inputting each word and its part-of-speech information into the word embedding model so that the word embedding model determines the semantic vector features and part-of-speech vector features of each word; and determining the word vector features of each word based on its semantic vector features and part-of-speech vector features.

[0063] In natural language processing, word embedding models convert each word in the target interaction information into word vector features. This not only captures the semantic information of words but also enhances the comprehensive understanding of words by combining part-of-speech information, thereby improving the accuracy of intent recognition. Specifically:

[0064] Part-of-speech (POS) information refers to the grammatical function of words in a sentence, such as noun (N), verb (V), adjective (A), etc. Using POS tagging tools, such as Stanford POS Tagger or jieba word segmentation, the POS of each word in a sentence can be determined. For example, in the sentence "I want to buy a Haier washing machine, but I need to compare prices," "buy" is tagged as a verb (V), while "washing machine" and "price" are tagged as nouns (N).

[0065] Word embedding models consider not only the semantic information of the words themselves but also their part-of-speech (POS) information. Traditional word embedding models such as Word2Vec or GloVe typically only consider the contextual information of words to generate word vector features. However, by introducing POS information, the model can understand the meaning of words more precisely, because synonyms may have different meanings under different POS. For example, "buy" (verb) and "purchase" (noun) may have different meanings in certain contexts.

[0066] The word embedding model generates two vectors for each word: semantic vector features and part-of-speech vector features. Semantic vector features focus on describing the semantic meaning of the word, while part-of-speech vector features focus on describing the grammatical function of the word. For example, the word "washing machine" might have the following vector features: Semantic vector features of "washing machine": [0.8, 0.1, -0.5, ...] Part-of-speech vector features of "washing machine": [0.6, -0.2, 0.1, ...] Here, it is assumed that the semantic vector features of "washing machine" capture its information as a household appliance, while the part-of-speech vector features emphasize its identity as a noun.

[0067] The final word vector features are composed of semantic vector features and part-of-speech vector features combined in some way (such as weighted summation, concatenation, or multiplication) to provide a more comprehensive word representation. For example, for "washing machine", the final word vector features can be a weighted sum of semantic and part-of-speech features: word vector features of "washing machine": [0.8+0.6,0.1-0.2,-0.5+0.1,...].

[0068] Optionally, this application embodiment provides a process for extracting intent features from target interaction information using a deep learning model, particularly recurrent neural networks (RNNs). Specifically: when N meets a first preset value, the target word vector features of the Nth word are input into the hidden layer of the deep learning model, so that the hidden layer determines the Nth hidden state based on the target word vector features of the Nth word, where N is a positive integer; when N does not meet the first preset value, the target word vector features of the Nth word are input into the hidden layer of the deep learning model, so that the hidden layer determines the Nth hidden state based on the target word vector features of the Nth word and the (N-1)th hidden state of the hidden layer; when the Nth hidden state is the last hidden state of the hidden layer, the Nth hidden state is input into the output layer of the deep learning model, so that the output layer outputs multiple intent features corresponding to the target interaction information based on the Nth hidden state, the weights of the output layer, and the bias.

[0069] In this embodiment, the weighted word vector features obtained by the word embedding model for each word are input into the deep learning model. The hidden layer is the part of the deep learning model responsible for processing the input information and extracting features. For recurrent neural networks, such as LSTM or GRU, the hidden layer updates the hidden state of the current time step based on the input word vector features and the hidden state of the previous time step.

[0070] When N (the position of the word) equals the model's preset starting position (e.g., 1), the model determines the current hidden state of the hidden layer solely based on the word vector features of the current word. This is because the model has not processed any words at the beginning, so it initializes the hidden state based only on the information of the current word.

[0071] For other words in the sentence (N>1), the model considers not only the word vector features of the current word but also the hidden state of the previous time step. When the last word in the sentence has been processed (the Nth hidden state is the last hidden state of the hidden layer), the model inputs the last hidden state into the output layer. The output layer outputs multiple intent features corresponding to the target interaction information based on the last hidden state, its own weights and biases.

[0072] The output layer can transform the features extracted by the hidden layer into interpretable intent features. These intent features are typically a set of numerical values ​​representing the probability or confidence of each possible intent.

[0073] This application's embodiments, through the hidden and output layers of a recurrent neural network, can effectively capture and understand multiple intentions within a sentence, even those with complex grammatical and semantic relationships. This approach enables a more accurate determination of the user's true needs, even amidst complex interactive information. This processing method is particularly suitable for long sentences or multi-intention scenarios because it retains a memory of the sentence's historical information, thereby better understanding the overall meaning and intent of the sentence.

[0074] Optionally, determining the intent set corresponding to the interaction information in the multi-intent recognition result based on the logical relationship in the interaction information includes: determining whether there is a connector in the interaction information; if the connector is present in the interaction information, determining the logical relationship in the interaction information based on the connector; if the connector is absent in the interaction information, determining the grammatical structure information in the interaction information, and determining the logical relationship in the interaction information based on the grammatical structure information; determining the logical type corresponding to the logical relationship, and determining the selection strategy corresponding to the logical type; and determining the intent set corresponding to the interaction information in the multi-intent recognition result based on the selection strategy.

[0075] Connectives reveal the logical relationships within or between statements, such as parallelism, cause and effect, and contrast. When processing interactive information, it is first necessary to identify the presence of connectives such as "but," "at the same time," "and," and "or," as logical words can indicate the logical relationships between user intentions.

[0076] Once the connecting words are identified, the logic behind them can be interpreted. For example, "but" usually indicates a contrast, suggesting that there is some kind of opposition or complement between the two intentions; "at the same time" indicates a parallel relationship, suggesting that the user may be concerned with or want to perform multiple intentions at the same time.

[0077] In the absence of explicit conjunctions, the system can infer logical relationships by analyzing grammatical structure. For example, the order of clauses in a complex sentence, such as the relative and main clauses, can determine the relationships between intents. Through grammatical analysis, potential parallel or subordinate relationships can be identified, even if no conjunctions are explicitly present.

[0078] Different logical relationships correspond to different processing strategies. Logical relationships include, but are not limited to:

[0079] Parallel Relationship: If the user uses parallel conjunctions such as "and" or "at the same time", multiple identified intents can be merged into a single intent set, indicating that the user has multiple needs that need to be addressed simultaneously.

[0080] Selection Relationship: If a user uses selection conjunctions such as "or", it may mean that the user is choosing one of multiple intents. In this case, these intents can be regarded as a set, but only one may be executed in the end.

[0081] Transitional relationships: If a user uses transitional conjunctions such as "but", it may indicate that the user has opposite or complementary needs, and this complexity needs to be reflected in the final set of intentions.

[0082] Conditional relationships: If the user submits a conditional statement, such as "if...then...", then the execution of a certain intent can be determined based on whether the condition is met.

[0083] Finally, based on the established logical relationships and selection strategies, a final set of intents is selected from the results of multi-intent recognition.

[0084] For example, if the system recognizes that a user says, "I want to book a flight to Shanghai tomorrow and check the weather there," it will logically merge the two recognized intentions of "booking a flight" and "checking the weather." The final set of intentions may be: {"booking a flight", "checking the weather"}.

[0085] The embodiments of this application enable a more accurate understanding and capture of complex user intentions, especially when interactive information contains multiple intentions that require logical integration. This improves the accuracy and practicality of multi-intention recognition, allowing intelligent systems to better respond to and meet user needs.

[0086] Optionally, data preprocessing is performed on the interaction information input by the target object to obtain preprocessed target interaction information, including: data cleaning of the interaction information to obtain first interaction information; determining recommendation information that is related to the first interaction information; and combining the first interaction information and the recommendation information into the target interaction information.

[0087] Data cleaning includes, but is not limited to, removing noise from text, such as redundant punctuation, HTML tags, and special characters. Furthermore, data cleaning may also involve correcting spelling errors and standardizing text formatting to ensure data consistency and quality. For example, cleaning "I want to buy a washing machine." into "I want to buy a washing machine."

[0088] The purpose of determining recommended information is to enhance the intelligence and user-friendliness of the interaction by analyzing the context of user input and providing additional information or suggestions that may be useful to the user. Recommended information can be generated based on the user's historical behavior, preferences, and the context of the current statement. For example, if a user mentions "washing machine," the system might recommend additional information related to washing machine brands, models, price ranges, and purchase locations. For instance, for "I want to book a flight to Shanghai tomorrow," it could be expanded to include two clauses: "I want to book a flight to Shanghai tomorrow" and a big data recommendation: "While you're at it, check the weather there."

[0089] After data cleaning and recommendation information generation, these two parts of information are recombined to form a richer and more informative target interaction information.

[0090] Through this series of preprocessing steps, the original user input is transformed into clearer, richer, and easier-to-understand target interaction information. This helps subsequent intent recognition or recommendation systems to respond more accurately to user needs and provide more valuable information or suggestions.

[0091] Optionally, determining the recommended information associated with the first interaction information includes at least one of the following: determining keywords in the first interaction information and determining a user profile of the target object; determining user preferences of the target object based on the user profile and determining recommended information corresponding to the keywords based on the user preferences; determining keywords in the first interaction information; determining recommended information corresponding to the keywords in an information database, wherein the information database includes information about the target object and other objects besides the target object.

[0092] Keywords are words with specific semantic meanings extracted from user input. They are usually the core parts that express user needs, preferences, or intentions. For example, if a user says, "I want to book a flight to Shanghai tomorrow and check the weather there," the keywords that can be extracted might include "Shanghai," "flight," and "weather."

[0093] User profiles are descriptions of user characteristics built upon data such as historical behavior, preferences, and demographic information. They help systems understand users more deeply, thereby providing more personalized services. For example, a user's purchase history, browsing history, and search keywords can all be used to build user profiles.

[0094] User preferences are a crucial component of user profiles, reflecting a user's tendencies in a particular area, such as sports preferences, price range, and functional needs. By analyzing user profiles, the system can more accurately determine which recommended information a user might be interested in, and then make recommendations accordingly.

[0095] Alternatively, in the absence of a user profile, the system can search and generate relevant recommendations from a database based on keywords in the initial interaction information. For example, given the keywords "Shanghai," "flight tickets," and "weather," the system can find information on food and attractions in Shanghai from the database.

[0096] An information database is a collection of preprocessed data, including not only information about the target user (the current user) but also information about other users and data collected by the system on various products and services. Such a database provides the system with a rich source of data for generating personalized recommendations.

[0097] Optionally, step S206 above can be implemented in the following way: determining the distance between each intent feature and the preset hyperplane through the target classifier; determining the intent recognition result corresponding to each intent feature and the confidence level of the intent recognition result corresponding to each intent feature based on the distance between each intent feature and the preset hyperplane; determining the relationship between the confidence level of each intent recognition result and the preset confidence level; determining the intent features with a confidence level greater than the preset confidence level based on the relationship, and determining the intent recognition results with a confidence level greater than the preset confidence level as the multi-intent recognition results.

[0098] In this embodiment, support vector machines (SVM) or other classifiers with similar decision boundaries are used to determine and filter multi-intent recognition results. Specifically:

[0099] In SVM, the hyperplane is the decision boundary used by the classifier to distinguish different categories. The distance between each input sample (i.e., the intent feature) and the hyperplane reflects the sample's relative position to the classification boundary, and this distance serves as the basis for confidence. The greater the distance, the higher the confidence that the sample is classified into a certain category.

[0100] The decision function output of an SVM classifier can be used to calculate the distance between each intent feature and the hyperplane. This distance value can be converted into a confidence score, typically achieved by calculating the absolute value of the decision function's output. A larger absolute value indicates a higher confidence score, making the sample more likely to belong to the predicted class.

[0101] After determining the confidence score for each intent recognition result, the confidence score of each intent recognition result is compared with a preset confidence level to determine which intent recognition results should be retained and which should be ignored. All intent recognition results with confidence scores higher than the preset confidence level constitute the final multi-intent recognition results.

[0102] Optionally, in this embodiment of the application, decision trees can also be used to determine and filter multi-intent recognition results, specifically:

[0103] For each intent feature, the decision tree proceeds along the path until it reaches a leaf node. The category of the leaf node is then used to determine the intent recognition result corresponding to each intent feature; that is, the category of the leaf node is the predicted intent recognition result. The decision tree calculates confidence based on the proportion of data representing intent recognition results determined at each decision node along the path to the leaf node. For example, if 80% of the data points on the path to a leaf node belong to the target intent recognition result, then 80% can be used as the confidence score for that sample belonging to that category.

[0104] Determine the relationship between the confidence level of each intent recognition result and a preset confidence level; determine the intent features with a confidence level greater than the preset confidence level based on the relationship, and identify the intent recognition results with a confidence level greater than the preset confidence level as the multi-intent recognition results.

[0105] To better understand the process of determining multiple intentions described above, the implementation flow of the method for determining multiple intentions will be further described below with reference to optional embodiments, but this is not intended to limit the technical solutions of the embodiments of this application.

[0106] This embodiment provides a method for determining multiple intentions. Figure 3 This is a schematic diagram of a method for determining multiple intentions according to an embodiment of this application, such as... Figure 3 As shown, it includes:

[0107] Step 1: Sentence segmentation and data preprocessing;

[0108] Specifically, data preprocessing includes... Figure 4 As shown:

[0109] 1. Text standardization.

[0110] 2. Text cleaning: Remove noise and irrelevant characters, such as punctuation marks and special characters.

[0111] 3. Perform word segmentation on the cleaned sentences to divide them into words or phrases.

[0112] 4. Word cleaning.

[0113] 5. Word standardization.

[0114] 6. Spelling errors.

[0115] After data preprocessing, word vector representation is performed on the segmented results, converting each word into a corresponding vector form for subsequent matching by our self-developed algorithm.

[0116] The system utilizes a self-developed machine learning training model algorithm for automatic segmentation and semantic expansion. For example, the sentence "I want to book a flight to Shanghai tomorrow" can be expanded into two clauses: "I want to book a flight to Shanghai tomorrow" and "By the way, check the weather there." The self-developed algorithm is a machine learning-based segmentation algorithm that combines Long Short-Term Memory (LSTM) and Hidden Markov Model (HMM) segmentation algorithms. This model is trained using a large amount of labeled data and employs high probability scores for selection.

[0117] Step 2: Intent feature extraction;

[0118] 1. Word vector generation (using Word2Vec);

[0119] First, the Word2Vec model is used to convert each word in the text into a vector representation. This step is performed before the deep learning model training begins. Word2Vec learns the semantic relationships between words by analyzing the corpus, and the generated word vectors can capture the relationships between synonyms, antonyms, etc., providing a foundation for subsequent feature extraction and intent recognition.

[0120] 2. Sequence labeling (e.g., part-of-speech tagging using LSTM+CRF);

[0121] In some cases, part-of-speech tagging can be used as part of feature extraction to provide auxiliary information for intent recognition. The LSTM+CRF model can tag the part of speech of each word, which helps in understanding syntactic structure and the role of words in a sentence, thereby assisting in intent recognition. Sequence tagging results are shown below. Figure 5 As shown.

[0122] 3. TF-IDF weighted;

[0123] Word2Vec-generated word vectors are combined with TF-IDF values ​​to assign weights to each word vector. TF-IDF is a statistical method used to evaluate the importance of a word to a sentence in a document or corpus. The TF-IDF value for each word is obtained by multiplying the term frequency (TF) and inverse document frequency (IDF); a higher value indicates that the word is more important in the current sentence. Multiplying the TF-IDF value by the word vector yields a weighted word vector, which aims to highlight words that play a key role in the current intent recognition.

[0124] 4. Use deep learning models (such as CNN or RNN) for feature extraction, such as... Figure 6 As shown;

[0125] The weighted word vectors are used as input, and a deep learning model is used to extract higher-level features. CNNs can capture local semantic features and are suitable for phrases or fixed-length sentence segments; while RNNs (such as LSTM or GRU) can process the sequence information of the entire sentence and capture long-distance dependencies between words. Depending on the needs of the task, CNNs, RNNs, or a combination of both can be used to build feature extraction models.

[0126] Step 3: Multi-intent recognition;

[0127] Based on the extracted intent features, a classifier is used to classify the intent. The classifier can employ algorithms such as Support Vector Machine (SVM) or Decision Tree.

[0128] For each classification result, a confidence score is calculated. The higher the confidence score, the greater the likelihood of that intent. Based on a preset threshold, several intents with high confidence scores are selected as the final recognition results.

[0129] After identifying multiple intents, these intents need to be fused and confirmed to determine the final intent set of the entire sentence. Rule-based methods can be used, such as fusion based on semantic relationships and contextual information; alternatively, machine learning algorithms can be used to automatically fuse intents by training a fusion model. For example, if there is a clear logical relationship between two clauses, such as conjunctions like "by the way" or "and," the intents of these two clauses can be considered to belong to the same overall intent set. After fusion and confirmation, the final intent set of the entire sentence "I want to book a flight to Shanghai tomorrow, and by the way, check the weather there." is determined to be {"book flight", "check the weather"}.

[0130] Step 4: Data storage.

[0131] In this embodiment, the input statement is preprocessed, including cleaning, word segmentation and word vector representation; a deep learning model is used to extract intent features from the preprocessed statement; based on the extracted intent features, a classifier is used to perform multi-intent recognition, and the confidence score of each intent is calculated. The final multi-intent recognition result is determined according to a preset threshold, which solves the problem that the accuracy of intent recognition technology is not high when dealing with multiple intents in complex contexts.

[0132] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0133] Figure 7 This is a structural block diagram of a multi-intent determination device according to an embodiment of this application; as shown... Figure 7 As shown, it includes:

[0134] The preprocessing module 72 is used to preprocess the interactive information input by the target object to obtain the preprocessed target interactive information;

[0135] The feature extraction module 74 is used to extract intent features from the target interaction information using a deep learning model to obtain multiple intent features corresponding to the target interaction information. The deep learning model is trained by machine learning using multiple sets of data, and each set of data includes: sample interaction information and labels of intent features of the sample interaction information.

[0136] The intent classification module 76 is used to determine the confidence level of the intent recognition result corresponding to each intent feature through the target classifier, and to determine the multi-intent recognition result based on the confidence level of the intent recognition result corresponding to each intent feature;

[0137] The determining module 78 is used to determine the intent set corresponding to the interaction information in the multi-intent recognition result based on the logical relationship in the interaction information.

[0138] The aforementioned device preprocesses the interactive information input by the target object to obtain preprocessed target interactive information; it then extracts intent features from the target interactive information using a deep learning model to obtain multiple intent features corresponding to the target interactive information; a target classifier determines the confidence level of the intent recognition result corresponding to each intent feature, and determines a multi-intent recognition result based on the confidence level of the intent recognition result corresponding to each intent feature; and it determines the intent set corresponding to the interactive information in the multi-intent recognition result based on the logical relationships in the interactive information. In this embodiment, preprocessing enhances data quality, a deep learning model extracts rich features, confidence analysis filters highly credible intents, and logical relationships integrate the intent set. This not only handles complex user language input but also intelligently understands and responds to the user's real needs, thus achieving a significant improvement in accuracy for multi-intent recognition in complex contexts. The above technical solution solves the problem of low accuracy in intent recognition technology when handling multiple intents in complex contexts.

[0139] In an exemplary embodiment, the feature extraction module 74 is configured to convert each word in the target interaction information into a word vector feature using a word embedding model, and to determine the word frequency-inverse document frequency of each word; to weight the word vector feature of each word and the word frequency-inverse document frequency of each word to determine the target word vector feature of each word; and to input the target word vector feature of each word into the deep learning model so that the deep learning model performs intent feature extraction on the target interaction information based on the target word vector feature of each word to obtain multiple intent features corresponding to the target interaction information.

[0140] In an exemplary embodiment, the feature extraction module 74 is used to determine the part-of-speech information of each word in the target interaction information; input each word and the part-of-speech information of each word into the word embedding model, so that the word embedding model determines the semantic vector feature and part-of-speech vector feature of each word; and determine the word vector feature of each word based on the semantic vector feature and part-of-speech vector feature of each word.

[0141] In an exemplary embodiment, the feature extraction module 74 is configured to: input the target word vector feature of the Nth word into the hidden layer of the deep learning model when N meets a first preset value, so that the hidden layer determines the Nth hidden state based on the target word vector feature of the Nth word, where N is a positive integer; input the target word vector feature of the Nth word into the hidden layer of the deep learning model when N does not meet the first preset value, so that the hidden layer determines the Nth hidden state based on the target word vector feature of the Nth word and the (N-1)th hidden state of the hidden layer; and input the Nth hidden state into the output layer of the deep learning model when the Nth hidden state is the last hidden state of the hidden layer, so that the output layer outputs multiple intent features corresponding to the target interaction information based on the Nth hidden state, the weights of the output layer, and the bias.

[0142] In an exemplary embodiment, a determining module 78 is configured to: determine whether a connector exists in the interaction information; if the connector exists in the interaction information, determine a logical relationship in the interaction information based on the connector; if the connector does not exist in the interaction information, determine grammatical structure information in the interaction information, and determine a logical relationship in the interaction information based on the grammatical structure information; determine the logical type corresponding to the logical relationship, and determine a selection strategy corresponding to the logical type; and determine the intent set corresponding to the interaction information in the multi-intent recognition result based on the selection strategy.

[0143] In an exemplary embodiment, the preprocessing module 72 is configured to perform data cleaning on the interaction information to obtain first interaction information; determine recommendation information that is associated with the first interaction information; and combine the first interaction information and the recommendation information into the target interaction information.

[0144] In an exemplary embodiment, the preprocessing module 72 is configured to: determine keywords in the first interaction information and determine a user profile of the target object; determine user preferences of the target object based on the user profile and determine recommendation information corresponding to the keywords based on the user preferences; determine keywords in the first interaction information; and determine recommendation information corresponding to the keywords in an information database, wherein the information database includes information about the target object and other objects besides the target object.

[0145] In an exemplary embodiment, the intent classification module 76 is configured to: determine the distance between each intent feature and a preset hyperplane using the target classifier; determine the intent recognition result corresponding to each intent feature and the confidence level of the intent recognition result corresponding to each intent feature based on the distance between each intent feature and the preset hyperplane; determine the relationship between the confidence level of each intent recognition result and a preset confidence level; determine the intent features with a confidence level greater than the preset confidence level based on the relationship, and identify the intent recognition results with a confidence level greater than the preset confidence level as the multi-intent recognition results.

[0146] Embodiments of this application also provide a storage medium including a stored program, wherein the program executes any of the methods described above when it is run.

[0147] Optionally, in this embodiment, the storage medium may be configured to store program code for performing the following steps:

[0148] S1, perform data preprocessing on the interactive information input by the target object to obtain the preprocessed target interactive information;

[0149] S2, extract intent features from the target interaction information using a deep learning model to obtain multiple intent features corresponding to the target interaction information. The deep learning model is trained using multiple sets of data through machine learning. Each set of data includes: sample interaction information and a label of the intent feature of the sample interaction information.

[0150] S3, determine the confidence level of the intent recognition result corresponding to each intent feature through the target classifier, and determine the multi-intent recognition result based on the confidence level of the intent recognition result corresponding to each intent feature;

[0151] S4, determine the intent set corresponding to the interaction information in the multi-intent recognition result based on the logical relationship in the interaction information.

[0152] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0153] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0154] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0155] S1, perform data preprocessing on the interactive information input by the target object to obtain the preprocessed target interactive information;

[0156] S2, extract intent features from the target interaction information using a deep learning model to obtain multiple intent features corresponding to the target interaction information. The deep learning model is trained using multiple sets of data through machine learning. Each set of data includes: sample interaction information and a label of the intent feature of the sample interaction information.

[0157] S3, determine the confidence level of the intent recognition result corresponding to each intent feature through the target classifier, and determine the multi-intent recognition result based on the confidence level of the intent recognition result corresponding to each intent feature;

[0158] S4, determine the intent set corresponding to the interaction information in the multi-intent recognition result based on the logical relationship in the interaction information.

[0159] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0160] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0161] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0162] The embodiments described herein also provide a computer program that includes computer instructions stored in a computer-readable storage medium; a processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in any of the above method embodiments.

[0163] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0164] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0165] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for determining multiple intentions, characterized in that, include: Data preprocessing is performed on the interactive information input by the target object to obtain preprocessed target interactive information; The intent features of the target interaction information are extracted using a deep learning model to obtain multiple intent features corresponding to the target interaction information; wherein, the deep learning model is trained by machine learning using multiple sets of data, and each set of data includes: sample interaction information and labels of the intent features of the sample interaction information; The confidence level of the intent recognition result corresponding to each intent feature is determined by the target classifier, and the multi-intent recognition result is determined based on the confidence level of the intent recognition result corresponding to each intent feature. Based on the logical relationships in the interaction information, the intent set corresponding to the interaction information is determined in the multi-intent recognition result.

2. The method for determining multiple intentions according to claim 1, characterized in that, The intent features of the target interaction information are extracted using a deep learning model to obtain multiple intent features corresponding to the target interaction information, including: Each word in the target interactive information is converted into a word vector feature through a word embedding model, and the word frequency-inverse document frequency of each word is determined. Each word is converted into a word vector feature and weighted with the word frequency-inverse document frequency of each word to determine the target word vector feature of each word; The target word vector features of each word are input into the deep learning model, so that the deep learning model can extract intent features from the target interaction information based on the target word vector features of each word, thereby obtaining multiple intent features corresponding to the target interaction information.

3. The method for determining multiple intentions according to claim 2, characterized in that, Each word in the target interaction information is converted into a word vector feature using a word embedding model, including: Determine the part-of-speech information of each word in the target interaction information; Each word and its part-of-speech information are input into the word embedding model so that the word embedding model can determine the semantic vector features and part-of-speech vector features of each word. The word vector features of each word are determined based on the semantic vector features and part-of-speech vector features of each word.

4. The method for determining multiple intentions according to claim 2, characterized in that, The target word vector features of each word are input into the deep learning model, so that the deep learning model extracts intent features from the target interaction information based on the target word vector features of each word, thereby obtaining multiple intent features corresponding to the target interaction information, including: When N meets the first preset value, the target word vector features of the Nth word are input into the hidden layer of the deep learning model, so that the hidden layer determines the Nth hidden state of the hidden layer according to the target word vector features of the Nth word, where N is a positive integer; If N does not meet the first preset value, the target word vector feature of the Nth word is input into the hidden layer of the deep learning model, so that the hidden layer determines the Nth hidden state of the hidden layer based on the target word vector feature of the Nth word and the (N-1)th hidden state of the hidden layer. When the Nth hidden state is the last hidden state of the hidden layer, the Nth hidden state is input to the output layer of the deep learning model, so that the output layer outputs multiple intent features corresponding to the target interaction information according to the Nth hidden state, the weights and biases of the output layer.

5. The method for determining multiple intentions according to claim 1, characterized in that, Based on the logical relationships in the interaction information, the set of intents corresponding to the interaction information is determined from the multi-intent recognition results, including: Determine whether there are any connecting words in the interaction information; If the connecting word exists in the interactive information, the logical relationship in the interactive information is determined based on the connecting word; If the connector is not present in the interaction information, determine the grammatical structure information in the interaction information, and determine the logical relationship in the interaction information based on the grammatical structure information; Determine the logical type corresponding to the logical relationship, and determine the selection strategy corresponding to the logical type; determine the intent set corresponding to the interaction information in the multi-intent recognition result according to the selection strategy.

6. The method for determining multiple intentions according to claim 1, characterized in that, Data preprocessing is performed on the interactive information input by the target object to obtain preprocessed target interactive information, including: The interaction information is cleaned to obtain the first interaction information; Determine the recommended information that is related to the first interactive information; The first interaction information and the recommendation information are combined into the target interaction information.

7. The method for determining multiple intentions according to claim 6, characterized in that, The recommended information that is associated with the first interaction information includes at least one of the following: Identify keywords in the first interaction information and determine the user profile of the target object; determine the user preferences of the target object based on the user profile, and determine the recommendation information corresponding to the keywords based on the user preferences; Identify the keywords in the first interactive information; The recommended information corresponding to the keyword is determined in the information database, wherein the information database includes: information on the target object and other objects besides the target object.

8. The method for determining multiple intentions according to claim 1, characterized in that, The confidence level of the intent recognition result corresponding to each intent feature is determined by a target classifier, and a multi-intent recognition result is determined based on the confidence level of the intent recognition result corresponding to each intent feature, including: The distance between each intent feature and the preset hyperplane is determined by the target classifier; The intention recognition result corresponding to each intention feature and the confidence level of the intention recognition result corresponding to each intention feature are determined based on the distance between each intention feature and the preset hyperplane. Determine the relationship between the confidence level of each intent recognition result and the preset confidence level; Based on the size relationship, intent features with confidence levels greater than the preset confidence level are determined, and the intent recognition results with confidence levels greater than the preset confidence level are determined as the multi-intent recognition results.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program, when executed, performs the method according to any one of claims 1 to 8.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 8 through the computer program.

Images