Intention recognition method and device, storage medium, program product and computer device

Abstract

The application discloses an intention recognition method and device, a storage medium, a program product and a computer device. The method comprises the following steps: obtaining to-be-recognized user data of a target user; performing intention recognition based on the to-be-recognized user data and preset first intention recognition auxiliary information to obtain an initial intention recognition result set; in the case where the initial intention recognition result set meets a first condition, updating the first intention recognition auxiliary information based on historical scene data of an intention recognition scene to obtain second intention recognition auxiliary information, and performing intention recognition based on the second intention recognition auxiliary information and the to-be-recognized user data to obtain a first final intention recognition result, wherein the first condition at least comprises that the discrimination between initial recognition results of any two different categories in the initial intention recognition result set is greater than a preset discrimination threshold. In this way, the accuracy of intention recognition can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an intent recognition method, apparatus, storage medium, program product, and computer equipment. Background Technology

[0002] In related technologies, keyword matching or general semantic analysis models can usually be used to identify the intent of users in the data to be identified (such as voice data, text data or other types of data).

[0003] However, the accuracy of such technologies is usually not high, and they are prone to misjudgment or missed judgment. Summary of the Invention

[0004] To address the aforementioned technical problems, embodiments of this application propose an intent recognition method, apparatus, storage medium, program product, and computer device, which can improve the accuracy of intent recognition.

[0005] In a first aspect, embodiments of this application provide an intent recognition method, including: Obtain the target user data to be identified; Based on the user data to be identified and the preset first intent recognition auxiliary information, intent recognition is performed to obtain an initial intent recognition result set; When the initial intent recognition result set satisfies the first condition, the first intent recognition auxiliary information is updated based on the historical scene data of the intent recognition scenario to obtain the second intent recognition auxiliary information, and intent recognition is performed based on the second intent recognition auxiliary information and the user data to be identified to obtain the first final intent recognition result. The first condition includes at least the following: the discriminant degree between any two different initial recognition results in the initial intent recognition result set is greater than a preset discriminant degree threshold.

[0006] Optionally, the method further includes: If the initial intent recognition result set satisfies the second condition, the first intent recognition auxiliary information is updated based on historical scene data of the intent recognition scenario to obtain second intent recognition auxiliary information; and, the main semantic analysis result set is determined based on the user data to be identified and the second intent recognition auxiliary information; and Intent recognition is performed based on the main semantic analysis result set, the user data to be identified, and the second intent recognition auxiliary information to obtain the second final intent recognition result; The second condition includes: at least one of the discrimination scores is less than or equal to the preset discrimination threshold.

[0007] Optionally, determining the main semantic analysis result set based on the user data to be identified and the second intent recognition auxiliary information includes: The text to be analyzed is obtained based on the user data to be identified; Based on the second intent recognition auxiliary information, the first type of sentences in the text to be analyzed are analyzed to obtain the main semantic analysis result set.

[0008] Optionally, the step of performing intent recognition based on the main semantic analysis result set, the user data to be identified, and the second intent recognition auxiliary information to obtain a second final intent recognition result includes: Based on the first type of statement and the second type of statement in the text to be analyzed, the second intention recognition auxiliary information is adjusted to obtain the third intention recognition auxiliary information. The text to be analyzed is obtained according to the user data to be identified. In the text to be analyzed, the semantic importance of the first type of statement is higher than that of the second type of statement. The second type of statement is analyzed based on the third intent recognition auxiliary information to obtain a set of auxiliary semantic analysis results; Based on the main semantic analysis result set and the auxiliary semantic analysis result set, a second semantic analysis result set to be identified is generated; Intent recognition is performed based on the second set of semantic analysis results to be identified and the user data to be identified, to obtain the second final intent recognition result.

[0009] Optionally, the first intent recognition auxiliary information includes a first semantic sentiment weight parameter set; The intention recognition is performed based on the user data to be identified and the preset first intention recognition auxiliary information to obtain an initial intention recognition result set, including: Determine the test text corresponding to the user data to be identified; Based on the first semantic sentiment weight parameter set, semantic analysis is performed on the text to be tested to obtain the first set of semantic analysis results to be identified. Key semantic elements are extracted from each initial identification result contained in the first set of semantic analysis results to be identified to obtain an intent feature tag set, wherein the intent feature tag set includes multiple tags corresponding to multiple intent categories; The initial identification results are matched with the intent feature label set to determine the intent category to which each initial identification result belongs; Based on each initial identification result and its respective intent category, the initial identification result set is determined, wherein the initial identification results of different categories refer to the initial identification results belonging to different intent categories.

[0010] Optionally, the method further includes: Determine the category weights corresponding to each of the multiple intent categories; For each of the plurality of intent categories, determine the number of initial recognition results belonging to each intent category; The intention weight percentage of each intention category is obtained by weighting the number of results and the category weight corresponding to each of the multiple intention categories. in, The first condition is that the discriminant difference between any two initial recognition results of different classes is greater than a preset discriminant difference threshold, and the intention weight ratio of at least one intention category among the plurality of intention categories is less than a preset intention determination threshold; and / or, The second condition is that at least one of the distinguishability scores is less than or equal to the preset distinguishability threshold, and the intention weight ratio of at least one intention category among the plurality of intention categories is less than the preset intention determination threshold.

[0011] Optionally, the historical scene data includes historical semantic and sentiment data, the first intent recognition auxiliary information includes a first semantic and sentiment weight parameter set, and the second intent recognition auxiliary information includes a second semantic and sentiment weight parameter set. The process of updating the first intent recognition auxiliary information based on historical scene data of the intent recognition scenario to obtain the second intent recognition auxiliary information includes: Based on the aforementioned historical semantic sentiment data, a semantic error factor is determined; The first semantic sentiment weight parameter set is calibrated based on the semantic error factor to obtain the second semantic sentiment weight parameter set.

[0012] Optionally, the historical semantic sentiment data includes actual adaptation weight parameters and simulated weight parameters for each of multiple sets of historical texts, wherein the actual adaptation weight parameters are used to characterize the real weight values ​​of the corresponding set of historical texts, and the simulated weight parameters are used to characterize the simulated weight values ​​of the corresponding set of historical texts obtained through intent recognition. The determination of semantic error factors based on the historical semantic sentiment data includes: The error factor for each group of historical texts is determined based on the ratio between the actual adaptation weight parameters and the simulated weight parameters. Based on the error factors of each of the multiple sets of historical texts, the semantic error factor is obtained through statistical analysis.

[0013] Secondly, embodiments of this application provide an intent recognition device, including: The data acquisition module is used to acquire the user data to be identified for the target user. The first identification module is used to perform intent identification based on the user data to be identified and the preset first intent identification auxiliary information to obtain an initial intent identification result set. The second identification module is used to update the first intention identification auxiliary information based on historical scene data of the intention identification scenario to obtain the second intention identification auxiliary information when the initial intention identification result set satisfies the first condition, and to perform intention identification based on the second intention identification auxiliary information and the user data to be identified to obtain the first final intention identification result. The first condition includes at least the following: the discriminant degree between any two different initial recognition results in the initial intent recognition result set is greater than a preset discriminant degree threshold.

[0014] Thirdly, embodiments of this application provide a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in any of the above-mentioned embodiments.

[0015] Fourthly, embodiments of this application provide a computer program product, including computer instructions that, when executed by a processor, implement the steps of the method described in any of the above-described embodiments.

[0016] Fifthly, embodiments of this application provide a computer device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the steps of the method described in any of the preceding claims.

[0017] In summary, the embodiments of this application have at least the following beneficial effects: In this embodiment, user data to be identified for a target user is acquired; intent recognition is performed based on the user data to be identified and preset first intent recognition auxiliary information to obtain an initial intent recognition result set; if the initial intent recognition result set meets a first condition, the first intent recognition auxiliary information is updated based on historical scene data of the intent recognition scenario to obtain second intent recognition auxiliary information; and intent recognition is performed based on the second intent recognition auxiliary information and the user data to be identified to obtain a first final intent recognition result. The first condition includes at least: the distinguishability between any two different initial recognition results in the initial intent recognition result set is greater than a preset distinguishability threshold. Thus, the user data to be identified and the first intent recognition auxiliary information can be combined to complete preliminary intent recognition and obtain an initial intent recognition result set. If the initial intent recognition result set meets the first condition, it is determined that the recognition accuracy corresponding to the initial intent recognition result set is insufficient, and a more accurate intent recognition is required. The second intent recognition auxiliary information is updated using historical scene data of the intent recognition scenario to improve the recognition accuracy of the obtained first final intent recognition result. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the intent recognition method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of the intent recognition device provided in the embodiments of this application; Figure 3 This is a schematic diagram of the computer device provided in the embodiments of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments / examples are only a part of the embodiments / examples of this application, and not all of the embodiments / examples. Based on the embodiments / examples in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "multiple" means two or more. In the description of this application, the term "comprising" and its variations are open-ended, meaning "including but not limited to." The term "based on" means "at least partially based on." The term "according to" means "at least partially according to." The term "one embodiment / example" means "at least one embodiment / example"; the term "another embodiment / example" means "at least one additional embodiment / example"; the term "some embodiments / examples" means "at least some embodiments / examples."

[0021] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0022] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this application is for the purpose of describing specific embodiments only and is not intended to limit the application. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0023] In some cases, taking customer intent recognition in service hotlines as an example, related technologies often employ a single intent recognition mode, such as relying solely on keyword matching or simple semantic analysis, which is insufficient to handle complex and ambiguous customer intent expressions. When multiple intents are intertwined in a customer's statement, or when the intent expression is not direct enough, related technologies are prone to misjudgment or omission, resulting in inaccurate traffic redirection after customer intent recognition in the service hotline, thereby affecting the customer experience.

[0024] On the other hand, intent recognition typically requires parameters to assist in the process, but the adjustment of these parameters in related technologies lacks scientific rigor and dynamism. For example, semantic and sentiment weight parameters used in intent recognition often rely on manual experience or are adjusted based on fixed rules, failing to fully utilize historical service data. This makes it impossible to dynamically optimize parameters according to different business scenarios, customer groups, and time changes, resulting in difficulty in guaranteeing recognition accuracy when facing new business needs or changes in customer intent. Furthermore, related technologies have poor adaptability in complex scenarios, lacking effective strategies for handling situations with overlapping or ambiguous customer intents, often only able to recognize simple and clear intents, thus limiting the intelligence level and service scope of service hotlines.

[0025] In view of the above, embodiments of this application provide an intent recognition method, apparatus, storage medium, program product, and computer device, which aim to at least partially solve the shortcomings existing in the above-mentioned related technologies.

[0026] Firstly, see [the following] Figure 1 The diagram shows a flowchart of an intent recognition method provided in an embodiment of this application. The intent recognition method can be applied to a computer device with data processing capabilities. The method includes S101-S103, as detailed below.

[0027] S101, Obtain the user data to be identified for the target user.

[0028] In some examples, the user data to be identified may include the target user's text data to be identified, and / or, the user data to be identified may include the target user's voice data to be identified, wherein the text data to be identified can be obtained by converting the voice data to be identified into text form. For example, in a service hotline scenario, the voice data to be identified may be the voice data generated by the target user when dialing the hotline.

[0029] S102, based on the user data to be identified and the preset first intent recognition auxiliary information, intent recognition is performed to obtain an initial intent recognition result set.

[0030] In some examples, a large model can be used to perform intent recognition on the user data to be identified. For example, the user data to be identified and the first intent recognition auxiliary information can be input into the large model to obtain the initial intent recognition result set output by the large model. The first intent recognition auxiliary information can be used as auxiliary information to prompt the large model to perform intent recognition on the user data to be identified.

[0031] In some examples, the first intent recognition auxiliary information can be information used to assist in intent recognition, such as prompt words used in the large model mentioned above.

[0032] In some examples, the initial intent recognition result set may contain at least one initial intent recognition result. Taking the text data to be recognized as an example, a user may express at least one user intent in a piece of text, and for each user intent, the corresponding initial intent recognition result can be obtained through intent recognition.

[0033] In this embodiment, by introducing a first semantic sentiment weight parameter set as auxiliary information for first intent recognition, the emotional tendency and semantic intensity in the user's expression can be quantified into a computable parameter system. This allows intent recognition to not only rely on keyword matching or surface syntactic structure, but also to deeply capture the semantic sentiment features behind the user's true needs, significantly improving the accuracy and granularity of recognition.

[0034] S103, if the initial intent recognition result set satisfies the first condition, the first intent recognition auxiliary information is updated based on historical scene data of the intent recognition scenario to obtain second intent recognition auxiliary information. Furthermore, intent recognition is performed based on the second intent recognition auxiliary information and the user data to be identified to obtain a first final intent recognition result. The first condition includes at least: the distinguishability between any two different classes of initial recognition results in the initial intent recognition result set is greater than a preset distinguishability threshold.

[0035] In some examples, the discriminant between any two different initial identification results in the initial identification result set is greater than a preset discriminant threshold. Discriminant can be used to characterize the degree of distinguishability between two initial identification results. A discriminant greater than the preset discriminant threshold indicates that the degree of distinguishability between two initial identification results is obvious enough. The discriminant can be obtained by calculating the similarity between two initial identification results. For example, the initial identification results can be semantic analysis results, and the similarity can be obtained by calculating the semantic similarity between two initial identification results.

[0036] In some examples, the initial intent recognition result set may include multiple initial recognition results, which may correspond to multiple intent categories, with each initial recognition result belonging to its corresponding intent category. It is easy to understand that intent recognition can classify intents, for example, through an intent classification model.

[0037] It is understood that specific embodiments for intent recognition using second intent recognition auxiliary information and user data to be identified can refer to the specific embodiments for intent recognition using first intent recognition auxiliary information and user data to be identified in this application, and will not be repeated here.

[0038] In this embodiment, the mechanism for updating auxiliary information is triggered only when the initial recognition result has high discriminative power. This avoids noise interference caused by blind updates in low-confidence scenarios and ensures that the auxiliary information is continuously optimized based on historical scenario data, achieving adaptive evolution of the intent recognition system. Furthermore, the two-stage recognition process (initial recognition and re-recognition based on updated auxiliary information) improves the robustness of the final intent determination while maintaining efficiency, making it particularly suitable for practical applications such as customer service where high accuracy in intent-driven traffic is required.

[0039] In one optional implementation, the first intent recognition auxiliary information includes a first semantic sentiment weight parameter set; The intention recognition is performed based on the user data to be identified and the preset first intention recognition auxiliary information to obtain an initial intention recognition result set, including: Determine the test text corresponding to the user data to be identified; Based on the first semantic sentiment weight parameter set, semantic analysis is performed on the text to be tested to obtain the first set of semantic analysis results to be identified. Key semantic elements are extracted from each initial identification result contained in the first set of semantic analysis results to be identified to obtain an intent feature tag set, wherein the intent feature tag set includes multiple tags corresponding to multiple intent categories; The initial identification results are matched with the intent feature label set to determine the intent category to which each initial identification result belongs; Based on each initial identification result and its respective intent category, the initial identification result set is determined, wherein the initial identification results of different categories refer to the initial identification results belonging to different intent categories.

[0040] In some examples, the first semantic sentiment weight parameter set can be pre-defined based on the semantic complexity of the text being tested. Taking a scenario where a customer inquires about package pricing in a service hotline as an example, a corresponding semantic sentiment weight parameter set can be pre-defined. First, at least one core semantic dimension is identified, such as at least one of the following: degree of confusion regarding pricing, degree of skepticism regarding pricing, and urgency of requesting an explanation. For each dimension, multiple levels and their corresponding weight parameter values ​​can be set. For example, in the dimension of degree of confusion regarding pricing, the weight parameter values ​​from slightly confused to somewhat confused to very confused are set to 2, 4, and 6 respectively, with a difference of 2 between any two consecutive weight parameter values; in the dimension of degree of skepticism regarding pricing, the weight parameter values ​​from slightly skeptical to somewhat skeptical to very skeptical are set to 3, 6, and 9 respectively, with a difference of 3; in the dimension of urgency of requesting an explanation, the weight parameter values ​​from generally urgent to somewhat urgent to very urgent are set to 1, 3, and 5 respectively, with a difference of 2. It is worth noting that the difference between any two consecutive weight parameter values ​​in the semantic sentiment weight parameter set can be set to be the same, thereby ensuring the linear interpretability and computational consistency of the weight system.

[0041] In some examples, when the text to be transcribed contains expressions such as "Why is your package pricing so complicated? I don't understand it at all. Explain it to me clearly right now!", semantic judgment can be performed according to preset rules: "Level of confusion about pricing" is judged as "extremely confused" due to the expression "I don't understand it at all," corresponding to a weight parameter value of 6; "Level of doubt about pricing" is judged as "extremely doubtful" due to the expression "Why is it so complicated?", corresponding to a weight parameter value of 9; "Level of urgency for explanation" is judged as "extremely urgent" due to the expression "Explain it to me clearly right now," corresponding to a weight parameter value of 5. By preset and comprehensively calculating the weight parameters of these dimensions, the weighted scores of the text to be tested on each semantic sentiment dimension can be obtained, thus forming the first set of semantic analysis results to be identified (i.e., the preprocessed semantic analysis set).

[0042] In some examples, if the initial intent recognition result set meets a first condition, an auxiliary information update and re-recognition process can be executed. The first condition includes at least: the discriminant strength between any two initial recognition results of different classes in the initial intent recognition result set is greater than a preset discriminant strength threshold. When this first condition is met, it indicates that the current recognition result has high confidence and inter-class separability. Therefore, the first intent recognition auxiliary information can be updated based on historical scene data of the intent recognition scenario to obtain second intent recognition auxiliary information. Subsequently, intent recognition can be performed again based on the second intent recognition auxiliary information and the user data to be identified, ultimately obtaining the first final intent recognition result.

[0043] In some examples, the first semantic sentiment weighting parameter set can be pre-defined based on the semantic complexity of the text to be tested. Taking a customer feedback scenario of "product malfunction" in a service hotline as an example, when applying sentiment weighting, multiple semantic sentiment dimensions can be determined first, such as the degree of impact of the malfunction, the customer's anxiety level, and the urgency of the need. Assuming the weight ratio of the degree of impact of the malfunction is set to 40%, the customer's anxiety level to 30%, and the urgency of the need to 30%, when the text to be tested is "My device suddenly crashed, and all the important data on it was lost. You must help me solve it now!", the degree of impact of the malfunction can be assigned a higher weighting score because it involves the expression "important data loss"; the customer's anxiety level can also be assigned a higher score because of the expressions "suddenly crashed" and "important data was not saved"; and the urgency of the need can also be assigned a high score because of the expression "you must help me solve it now".

[0044] Next, the scores for each dimension can be weighted according to the set weight ratios. For example, a score of 80 for the degree of impact of the fault is multiplied by 40% to get 32 ​​points; a score of 90 for the degree of customer anxiety is multiplied by 30% to get 27 points; and a score of 95 for the degree of urgency of the need is multiplied by 30% to get 28.5 points. Finally, these scores are added together to obtain the comprehensive score of the preprocessed semantic analysis.

[0045] For example, overall score It can be calculated using the first formula Calculated Where n represents the number of semantic sentiment dimensions. This represents the matching score for the i-th dimension. This represents the weight ratio of the i-th dimension. This formula can be used to perform a weighted summation of multi-dimensional semantic sentiment scores, thereby generating a quantifiable and comparable comprehensive score, which can then be used as input for subsequent intent recognition.

[0046] In some examples, the weighting ratios can be adjusted based on historical data (historical scenario data) and the characteristics of the intent recognition scenario. For example, in a "complaint" scenario, the weight of customer dissatisfaction will be increased; in a "consultation" scenario, the weight of information demand will be emphasized. Furthermore, it can be continuously optimized based on actual recognition results: if a certain dimension has high recognition accuracy and is crucial for lead generation decisions, its weight ratio can be increased. This dynamic adjustment mechanism allows the first semantic sentiment weight parameter set to adapt to the intent recognition needs of different business scenarios, improving overall recognition performance.

[0047] In some examples, after completing semantic sentiment weighted analysis, a first set of semantic analysis results to be identified is obtained. Next, key semantic elements can be extracted from each initial identification result in the first set of semantic analysis results to obtain an intent feature label set. Specifically, the intent feature labels belonging to each semantic analysis result in the preprocessed semantic analysis set can be statistically analyzed, and these labels can be integrated to obtain an intent feature label set. Combining the intent feature label set, customer intent identification is performed on each semantic analysis result in the preprocessed semantic analysis set, thus obtaining an initial intent identification result set. Taking the preprocessed semantic analysis set corresponding to the transcribed customer voice from a service hotline as an example, key semantic elements, such as "product malfunction," "fee inquiries," "service complaints," and "business processing," can be extracted from each semantic analysis result in the preprocessed semantic analysis set. Using Natural Language Processing (NLP) technology, these semantic elements are classified and statistically analyzed, grouping semantic analysis results with the same core intent into the same category, and then extracting the corresponding intent feature labels. The intent feature label set can contain multiple intent feature labels. For example, when multiple semantic analysis results all revolve around "poor mobile signal, requesting troubleshooting and repair", the intent feature label "network fault reporting" can be statistically identified. By analogy, the intent label set can be constructed.

[0048] In some examples, after obtaining the intent feature label set, for each semantic analysis result in the preprocessed semantic analysis set, techniques such as text matching and semantic similarity calculation can be used to compare and match the semantic analysis result with each label in the intent feature label set. For example, if a semantic analysis result is "My broadband keeps disconnecting, affecting my work, please help me solve it," this semantic analysis result can be compared one by one with labels such as "network fault repair" and "service complaint." By analyzing the core demands and sentiment in the text, it can be determined that it has the highest matching degree with the "network fault repair" label. Then, based on this matching relationship, combined with preset intent recognition rules (such as matching degree threshold, multi-label priority, etc.), the semantic analysis result is classified to determine its intent category. After summarizing the recognition results of all semantic analysis results, an initial intent recognition result set can be formed (the initial recognition result can refer to the semantic analysis result, and the initial intent recognition result set can contain all initial recognition results and their respective intent categories).

[0049] Furthermore, if a customer says, "I want to check my phone bill this month; it seems a bit high, and I'd like to see the specific charges," during the semantic sentiment weighted analysis, the core semantic parts like "checking the phone bill charges" and the emotional expression conveyed by "seems a bit high" can be assigned corresponding weights, forming a preprocessed semantic analysis set containing this semantic and sentiment information. The intent feature tags for each semantic analysis result in the preprocessed semantic analysis set are then statistically analyzed and integrated to obtain an intent feature tag set. The customer's intent is then identified by combining the intent feature tag set with the various semantic analysis results in the preprocessed semantic analysis set, resulting in an initial intent identification result set. Taking this customer's statement as an example, the identified intent feature tags might include "phone bill inquiry" and "charge question," thus forming an initial intent identification result set containing these intents.

[0050] In this embodiment, the difference between every two consecutive weight parameter values ​​in the semantic sentiment weight parameter set remains consistent, thus constructing a linear, interpretable, and easy-to-maintain weight system. This is beneficial for the stability of model training and the manual calibration of business rules, while also facilitating subsequent dynamic adjustments based on historical data.

[0051] In an optional implementation, the method further includes: Determine the category weights corresponding to each of the multiple intent categories; For each of the plurality of intent categories, determine the number of initial recognition results belonging to each intent category; The intention weight percentage of each intention category is obtained by weighting the number of results and the category weight corresponding to each of the multiple intention categories. in, The first condition is that the discriminant difference between any two initial recognition results of different classes is greater than a preset discriminant difference threshold, and the intention weight ratio of at least one intention category among the plurality of intention categories is less than a preset intention determination threshold; and / or, The second condition is that at least one of the distinguishability scores is less than or equal to the preset distinguishability threshold, and the intention weight ratio of at least one intention category among the plurality of intention categories is less than the preset intention determination threshold.

[0052] In some examples, if the initial intent recognition result set meets the third condition, the initial intent recognition result set can be used as the final intent recognition result. The third condition can refer to the fact that the intent weight percentages of multiple intent categories are all greater than or equal to a preset intent determination threshold.

[0053] In some examples, a second calculation formula can be used. The percentage of intent results for the j-th type of intent was calculated. ,in, This represents the number of recognition results for the j-th type of intent, and m represents the total number of intent categories.

[0054] The weight percentage of the j-th type of intent Then it can be calculated using the third formula. The calculation yielded that, represents the base weight for the j-th type of intent.

[0055] Taking a customer intent recognition scenario for a service hotline as an example, assume the initial intent recognition result set includes three intent results: "fault reporting," "billing inquiries," and "service processing." First, count the number of each intent result. For example, there are 20 "fault reporting" results, 10 "billing inquiries," and 5 "service processing" results, for a total of 35. Then, the percentage of each intent result is calculated as follows: "fault reporting" accounts for approximately 20 ÷ 35 ≈ 57.14%, "billing inquiries" for approximately 10 ÷ 35 ≈ 28.57%, and "service processing" for approximately 5 ÷ 35 ≈ 14.29%.

[0056] Next, we calculate the weight percentage of intents. Assuming we pre-set a base weight of 4 for "Fault Reporting," 3 for "Pricing Inquiry," and 2 for "Service Processing," then the weight percentage for "Fault Reporting" is (20×4)÷(20×4+10×3+5×2)=80÷(80+30+10)=80÷120≈66.67%; for "Pricing Inquiry," it's (10×3)÷120=30÷120=25%; and for "Service Processing," it's (5×2)÷120=10÷120≈8.33%. Through this statistical calculation, we can obtain the percentage of intent results and the weight percentage of intents.

[0057] In some examples, an intent determination threshold can be pre-set, say 80%. Then, identical results are extracted from the initial intent identification set; these identical results are the intent results to be tested. For example, if most identification results are for call charges, then call charges would fall under the intent results to be tested, corresponding to the majority of identification results. Next, the quantity weight percentage of the intent results to be tested among all results can be calculated to obtain the intent weight percentage. If this intent weight percentage is greater than or equal to the pre-set intent determination threshold, such as reaching 85%, then a final intent result can be generated based on the customer intent identification result corresponding to this intent result. Based on this final intent result, a standard referral strategy is executed, which involves transferring the customer to the dedicated agent responsible for call charges inquiries. A comprehensive referral report can also be output, detailing relevant information from the entire customer intent identification and referral process.

[0058] In some examples, the first condition includes at least: the discriminant degree between any two different initial recognition results in the initial intent recognition result set is greater than a preset discriminant degree threshold, and the intention weight ratio of at least one intent category among the plurality of intent categories is less than a preset intent determination threshold; and / or, the second condition is: at least one of the discriminant degrees is less than or equal to the preset discriminant degree threshold, and the intention weight ratio of at least one intent category among the plurality of intent categories is less than the preset intent determination threshold.

[0059] When the first condition is met, the first intention recognition auxiliary information can be updated based on the historical scene data of the intention recognition scenario to obtain the second intention recognition auxiliary information. Then, the intention recognition is performed again based on the second intention recognition auxiliary information and the user data to be identified to obtain the first final intention recognition result.

[0060] Specifically, the historical scene data includes historical semantic sentiment data, the first intent recognition auxiliary information includes a first semantic sentiment weight parameter set, and the second intent recognition auxiliary information includes a second semantic sentiment weight parameter set. Updating the first intent recognition auxiliary information based on the historical scene data of the intent recognition scene to obtain the second intent recognition auxiliary information includes: determining a semantic error factor based on the historical semantic sentiment data; and performing error calibration on the first semantic sentiment weight parameter set based on the semantic error factor to obtain the second semantic sentiment weight parameter set.

[0061] The historical semantic sentiment data includes actual and simulated weight parameters for multiple sets of historical texts. The actual weight parameters characterize the true weight values ​​of a corresponding set of historical texts (e.g., optimal weights confirmed by manual annotation or business verification), while the simulated weight parameters characterize the simulated weight values ​​obtained through intent recognition for a corresponding set of historical texts (e.g., weights predicted by the current model). Determining the semantic error factor based on the historical semantic sentiment data includes: determining the error factor for each set of historical texts based on the ratio between the actual and simulated weight parameters; and obtaining the semantic error factor through statistical analysis (e.g., calculating the mean, median, and / or weighted average) based on the error factors of each set of historical texts.

[0062] For example, if the actual weight parameter for a set of historical texts in the "pricing question" scenario is 6, while the simulated weight parameter is 4, then its error factor is 6 / 4 = 1.5. After calculating the error factor for a large number of such samples, the average value is taken as the overall semantic error factor. The semantic error factor can be used to scale or offset the parameters in the first semantic sentiment weight parameter set or the parameters that need error calibration, thereby generating a second semantic sentiment weight parameter set that is closer to real business needs.

[0063] In this embodiment, by introducing a first semantic sentiment weight parameter set as auxiliary information for first intent recognition, the emotional tendency and semantic intensity in user expression are quantified into a computable parameter system. This allows intent recognition to not only rely on keyword matching or surface syntactic structure, but also to deeply capture the semantic sentiment features behind the user's true needs, significantly improving the accuracy and granularity of recognition. Furthermore, by employing multi-dimensional semantic sentiment analysis combined with weighted calculation, different semantic sentiment dimensions are scored separately and comprehensively calculated according to weight ratios. This achieves refined modeling of the user's emotional state and the urgency of their needs, enhancing the ability to judge user intent in complex contexts. Additionally, this embodiment introduces an error calibration mechanism based on historical semantic sentiment data: by comparing the actual adapted weight parameters with the simulated weight parameters, a semantic error factor is calculated, and the first semantic sentiment weight parameter set is calibrated accordingly to generate a second semantic sentiment weight parameter set. This mechanism enables closed-loop optimization of the intent recognition auxiliary information, giving the intent recognition scheme continuous learning and self-correction capabilities.

[0064] In one optional implementation, the historical scene data includes historical semantic sentiment data, the first intent recognition auxiliary information includes a first semantic sentiment weight parameter set, and the second intent recognition auxiliary information includes a second semantic sentiment weight parameter set. The process of updating the first intent recognition auxiliary information based on historical scene data of the intent recognition scenario to obtain the second intent recognition auxiliary information includes: Based on the aforementioned historical semantic sentiment data, a semantic error factor is determined; The first semantic sentiment weight parameter set is calibrated based on the semantic error factor to obtain the second semantic sentiment weight parameter set.

[0065] In one optional implementation, the historical semantic sentiment data includes actual adaptation weight parameters and simulated weight parameters for each of multiple sets of historical texts, wherein the actual adaptation weight parameters are used to characterize the real weight values ​​of the corresponding set of historical texts, and the simulated weight parameters are used to characterize the simulated weight values ​​of the corresponding set of historical texts obtained through intent recognition. The determination of semantic error factors based on the historical semantic sentiment data includes: The error factor for each group of historical texts is determined based on the ratio between the actual adaptation weight parameters and the simulated weight parameters. Based on the error factors of each of the multiple sets of historical texts, the semantic error factor is obtained through statistical analysis.

[0066] In some examples, the first semantic sentiment weight parameter set includes multiple semantic sentiment weight parameters, which correspond to each initial recognition result in the initial intent recognition result set; The step of calibrating the first semantic sentiment weight parameter set based on the semantic error factor to obtain the second semantic sentiment weight parameter set includes: From the first set of semantic sentiment weight parameters, at least one first parameter to be calibrated and / or at least one group of second parameters to be calibrated are determined. Different first parameters to be calibrated correspond to different intent categories. Each first parameter to be calibrated is selected from the semantic sentiment weight parameters corresponding to the initial recognition results belonging to the same intent category. Each group of second parameters to be calibrated includes at least two second parameters to be calibrated. The at least two second parameters to be calibrated correspond to at least two initial recognition results. The at least two initial recognition results belong to different intent categories, and the intent weight percentage of the intent category to which each of the at least two initial recognition results belongs is less than a preset intent determination threshold. Based on the semantic error factor, error calibration is performed on the at least one first parameter to be calibrated and / or at least one second parameter to be calibrated, so as to update the first semantic sentiment weight parameter set to the second semantic sentiment weight parameter set.

[0067] In an optional implementation, the method further includes: If the initial intent recognition result set satisfies the second condition, the first intent recognition auxiliary information is updated based on historical scene data of the intent recognition scenario to obtain second intent recognition auxiliary information; and, the main semantic analysis result set is determined based on the user data to be identified and the second intent recognition auxiliary information; and Intent recognition is performed based on the main semantic analysis result set, the user data to be identified, and the second intent recognition auxiliary information to obtain the second final intent recognition result; The second condition includes: at least one of the discrimination scores is less than or equal to the preset discrimination threshold.

[0068] In one optional implementation, determining the main semantic analysis result set based on the user data to be identified and the second intent recognition auxiliary information includes: The text to be analyzed is obtained based on the user data to be identified; Based on the second intent recognition auxiliary information, the first type of sentences in the text to be analyzed are analyzed to obtain the main semantic analysis result set.

[0069] In one optional implementation, the step of performing intent recognition based on the main semantic analysis result set, the user data to be identified, and the second intent recognition auxiliary information to obtain a second final intent recognition result includes: Based on the first type of statement and the second type of statement in the text to be analyzed, the second intention recognition auxiliary information is adjusted to obtain the third intention recognition auxiliary information. The text to be analyzed is obtained according to the user data to be identified. In the text to be analyzed, the semantic importance of the first type of statement is higher than that of the second type of statement. For example, the first type of statement can be the main semantic statement in the text to be analyzed, and the second type of statement can be the auxiliary semantic statement in the text to be analyzed. The second type of statement is analyzed based on the third intent recognition auxiliary information to obtain a set of auxiliary semantic analysis results; Based on the main semantic analysis result set and the auxiliary semantic analysis result set, a second semantic analysis result set to be identified is generated; Intent recognition is performed based on the second set of semantic analysis results to be identified and the user data to be identified, to obtain the second final intent recognition result.

[0070] In some examples, the first type of statement can refer to statements that are more semantically important than the second type of statements in the text to be analyzed, such as core appeal statements like "I want to subscribe to a 5G package" or "I want to check my phone bill".

[0071] In some examples, if the intent weight percentage is less than the intent determination threshold, and the different intent results are only clearly distinguishable, the difference in intent to be tested is obtained by extracting the different results from the initial intent identification result set. Taking a customer's inquiry about mobile services in a service hotline as an example, assuming the initial intent identification result set contains two different results: "subscribe to a data plan" and "cancel a voice plan," these two different results can be extracted to obtain the difference in intent to be tested. For instance, if the customer mentions both "wanting to subscribe to a large data plan" and "feeling that the voice plan is useless and wanting to cancel it," by analyzing the initial intent identification result set, the two different results "subscribe to a data plan" and "cancel a voice plan" are extracted to form the difference in intent to be tested, which is then used for subsequent operations such as determining their semantic sentiment weight parameter relationship.

[0072] If it is determined that the semantic sentiment weight parameters of each result in the same intent test are continuous, preprocessing case one can be output.

[0073] Taking customer feedback on product satisfaction in a service hotline as an example, if the parameters related to satisfaction in the semantic sentiment weight parameter set range from "slightly satisfied" to "somewhat satisfied" to "very satisfied," with corresponding weight parameter values ​​of 2, 4, and 6 respectively, and the difference between any two consecutive parameter values ​​is the same and shows a continuous increasing change; or if the parameters related to dissatisfaction range from "slightly dissatisfied" to "somewhat dissatisfied" to "very dissatisfied," with weight parameter values ​​of 3, 6, and 9 respectively, and the difference is fixed. Here, we can determine whether there is a continuous relationship between these semantic sentiment weight parameters by detecting whether the numerical changes of these parameters show a continuous, consistent increasing or decreasing pattern. For example, if the semantic sentiment weight parameters of each result in the same test intent change continuously in value, and the difference between adjacent parameters is the same, then a continuous relationship can be determined.

[0074] In some examples, historical semantic sentiment data of service hotlines from different periods can be obtained. The ratio between the actual adaptation weight parameters of historical text in the historical semantic sentiment data and the simulated weight parameters of historical text in the historical semantic sentiment data can be calculated to obtain the semantic error factor. The actual adaptation weight parameters and the simulated weight parameters of historical text each belong to the same text containing customer semantic expression and / or sentiment characteristics.

[0075] In some examples, during the customer intent recognition process for a service hotline, if the intent weight percentage is less than a pre-set intent judgment threshold, and there are clearly distinguishable differences between the different results of the customer intent, these different results can be extracted from the initial intent recognition result set to obtain the intent difference results to be tested. If it is determined that the semantic sentiment weight parameters corresponding to each result of the same intent to be tested exhibit a continuous relationship, then preprocessing scenario one can be output.

[0076] In some examples, after obtaining historical semantic sentiment data generated by the service hotline in a historical period, the actual weight parameters and simulated weight parameters of the historical text can be found in this historical semantic sentiment data, and the semantic expression and / or sentiment tendency characteristics of the customer in the text to which these two parameters belong are the same.

[0077] For example, a customer calls a mobile operator's customer service hotline, expressing an intention to both inquire about their phone bill and learn about new service plans. After initial intent identification, the weight percentages of these two intents do not reach the judgment threshold, but they are clearly distinguishable. Therefore, these two different intents can be extracted as the difference in intents to be tested. If it is determined that the semantic sentiment weight parameters of these two intents are continuous, then preprocessing scenario one is output. Next, historical semantic sentiment data of customers simultaneously expressing intents to inquire about phone bills and learn about service plans can be retrieved. The actual and simulated weight parameters for these historical texts can be found. Since the semantic expression and sentiment tendency of customers are the same in these historical texts, their ratio can be calculated to obtain the semantic error factor, which prepares for more accurate identification of customer intents and subsequent lead generation.

[0078] In some examples, under the above preprocessing case one, the maximum value is selected from the semantic sentiment weight parameters to which the same intention results belong to obtain the first parameter to be calibrated (i.e. the first parameter to be calibrated), and the semantic sentiment weight parameters to which the intention difference results belong to are extracted from the set of semantic sentiment weight parameters to obtain the second parameter to be calibrated (i.e. the second set of parameters to be calibrated). The set of parameters to be calibrated may include the first parameter to be calibrated and the second parameter to be calibrated.

[0079] Based on the previously calculated semantic error factor, the parameter set to be calibrated is calibrated to obtain the first set of secondary semantic simulation parameters (i.e., the second set of semantic sentiment weight parameters). Then, the test text can be subjected to semantic sentiment weighted analysis again based on the first set of secondary semantic simulation parameters to obtain the first set of semantic analysis parameters to be recognized. For example, a customer calls a service hotline expressing two intentions: to inquire about phone bills and to subscribe to a 5G package. Neither of these intentions has reached the set threshold, but they are clearly distinguishable and meet the conditions of preprocessing case one. In this case, the maximum value is selected from the same intentions (such as the weight parameters related to inquiring about phone bills) as the first parameter to be calibrated, and the second parameter to be calibrated is extracted from the intention difference results (the weight parameters for subscribing to a 5G package). These two parameters are combined into the parameter set to be calibrated. The semantic error factor is then used to calibrate this parameter set to obtain the first set of secondary semantic simulation parameters. This parameter set is then used to reanalyze the customer's speech-to-text to obtain the first set of semantic analysis parameters to be recognized. If, after identifying the customer intent in the semantic analysis set to be identified, the weight ratio of identical results in the result set is greater than or equal to the intent determination threshold (e.g., the weight ratio of the intent to subscribe to a 5G package reaches the threshold), then subscribing to a 5G package can be determined as the final intent result. Based on this final intent result, an adaptive referral strategy can be further executed, transferring the customer to a dedicated agent responsible for 5G package subscriptions. A detailed referral report recording the entire customer intent identification and referral process can also be generated.

[0080] In some examples, if the intention weight percentage is less than the intention determination threshold, and there is a fuzzy distinction between different intention results, if it is determined that the semantic sentiment weight parameters of each result in the same intention result to be tested are not continuous, preprocessing case two can be output.

[0081] In the second preprocessing scenario, the semantic and emotional correlation between the core appeal statement and the auxiliary descriptive statement in the text to be tested after processing with any three weight parameters in the semantic and emotional weight parameter set can be statistically determined to obtain the text correlation degree. A correlation threshold is preset, and the text fragments whose correlation degree is greater than or equal to the correlation threshold are extracted from the text to be tested to obtain the text fragments to be analyzed.

[0082] It can be calculated using the sixth formula Calculate the semantic sentiment correlation ,in, This represents the semantic similarity under the i-th parameter dimension. This represents the selected i semantic sentiment weight parameters.

[0083] The semantic weight ratio factor can be obtained by calculating the ratio between the semantic weight of the core statement in the text segment to the semantic weight of the auxiliary descriptive statement in the text segment.

[0084] It can be calculated using the seventh formula The semantic weight ratio factor is calculated, where, The semantic weight of statements indicating core claims This indicates the semantic weight of the auxiliary descriptive statements.

[0085] In some examples, when conducting customer intent recognition for service hotlines, if the intent weight percentage is less than the pre-set intent judgment threshold and there is a fuzzy distinction between different intent results, if it is determined that the semantic sentiment weight parameters of each result in the same intent result to be tested are not continuous, then preprocessing case two will be output.

[0086] In the second preprocessing scenario, the client's test text can be processed. Specifically, any three weight parameters are selected from the set of semantic sentiment weight parameters, and the text is processed using these three parameters respectively. Then, the semantic sentiment correlation between the core appeal statement and the auxiliary descriptive statement after processing is statistically analyzed to obtain the text correlation degree. A correlation threshold is preset, and the text with a correlation degree greater than or equal to the correlation threshold is extracted from the test text. These extracted texts are the text fragments to be analyzed.

[0087] For example, a customer calls the service hotline and says, "I feel my current data plan isn't cost-effective. A friend told me about a cheaper plan with more data, and I'd like to see if I can switch to that. I'd also like to know what the requirements are for switching." Here, wanting to switch plans is the core message, while the friend mentioning a cheaper plan with more data and wanting to know the requirements for switching are secondary descriptive statements. After initial intent recognition, the intent weight percentage did not reach the set threshold, and there was ambiguity in distinguishing between different intents (e.g., simply inquiring about new plan details or switching plans). It was determined that the semantic sentiment weight parameters for the same intent results were not continuous, so preprocessing scenario two can be output. Any three weight parameters are selected from the semantic sentiment weight parameter set, and these three parameters are used to process the customer's speech-to-text to calculate the semantic sentiment correlation between the core message and the secondary descriptive statements. Assuming the preset correlation threshold is 80, the calculated text correlation is 85, which meets the requirement of being greater than or equal to the correlation threshold. Therefore, this part of the text is extracted from the test text as the text segment to be analyzed. Then, the semantic weight ratio factor is obtained by calculating the ratio between the semantic weight of the core appeal statement and the semantic weight of the auxiliary descriptive statement in the text segment to be analyzed.

[0088] Based on the semantic error factor, the parameter set to be calibrated is calibrated to obtain the second semantic simulation parameter set. Based on the second semantic simulation parameter set, the core appeal statements in the text segment to be analyzed are subjected to semantic sentiment weighted analysis to obtain the main semantic analysis set.

[0089] Based on the semantic weight ratio factor, the parameters of the second set of secondary semantic simulation parameters are adjusted to obtain the third set of secondary semantic simulation parameters. Based on the third set of secondary semantic simulation parameters, semantic sentiment weighted analysis is performed on the auxiliary descriptive statements in the text segment to be analyzed to obtain the auxiliary semantic analysis set. The second set of semantic analysis results to be identified (the second set of semantic analysis results to be identified) can be generated, which includes the main semantic analysis set and the auxiliary semantic analysis set.

[0090] If the weight ratio of identical results in the result set after customer intent identification in the semantic analysis set 2 is greater than or equal to the intent determination threshold, the final intent result is determined, and a precise traffic redirection strategy is executed based on the result and a total traffic redirection report is output.

[0091] When accurately identifying and directing customer intent for service hotlines, a second semantic simulation parameter set can be obtained by calibrating the parameter set to be calibrated based on a semantic error factor. This second set of parameter sets is then used to perform semantic sentiment weighted analysis on the core appeal statements in the text segment to be analyzed, resulting in the main semantic analysis set.

[0092] The second set of secondary semantic simulation parameters is adjusted according to the semantic weight ratio factor, resulting in the third set of secondary semantic simulation parameters. Based on the third set of secondary semantic simulation parameters, semantic sentiment weighting is applied to the auxiliary descriptive statements in the text segment to be analyzed, resulting in a secondary semantic analysis set. Subsequently, the primary semantic analysis set and the secondary semantic analysis set are combined to form the second set of semantic analysis parameters to be identified.

[0093] For example, a customer calls a service hotline. The core message of their text segment to be analyzed is that they want to subscribe to a more cost-effective 5G plan. The auxiliary description is that they've heard the new 5G plans offer more data and are cheaper, and they want to know what the requirements are. First, the parameter set to be calibrated is adjusted based on the semantic error factor to obtain a second set of semantic simulation parameters. This parameter set is then used to analyze the core message, resulting in the main semantic analysis set, where the semantic and emotional information related to the core message is highlighted. Next, the second set of semantic simulation parameters is adjusted based on the semantic weight ratio factor to obtain a third set of semantic simulation parameters. This third set is then used to analyze the auxiliary description, resulting in a secondary semantic analysis set, where the semantic and emotional details are accurately captured. After combining the main and secondary semantic analysis sets into the second set of semantic analysis parameters to be identified, customer intent is identified. If the weight ratio of identical results in the identified result set is greater than or equal to the intent determination threshold, such as the intent weight ratio of "apply for a 5G package and understand the application conditions" reaching the threshold, then this intent is determined to be the final intent result. Based on this final intent result, a precise referral strategy can be executed, transferring the customer to a dedicated agent responsible for explaining the 5G package application and conditions, while simultaneously generating a detailed referral report recording the entire process of customer intent identification and referral.

[0094] In some examples, it can be calculated using the fourth formula. The semantic error factor was calculated. Where q represents the number of historical text groups. This represents the actual adaptation weight parameter for the t-th group of historical texts. This represents the simulated weight parameters for the t-th group of historical texts.

[0095] In the technical implementation of the semantic error factor, the first step is to filter out historical text groups from historical semantic sentiment data where customer semantic expression and sentiment characteristics are completely identical. For example, a group of texts might all express strong dissatisfaction with "broadband lag," with a sentiment tendency of "extreme dissatisfaction." Then, for each group of texts, determine the actual adapted weight parameters (e.g., the true weight values ​​derived from semantic sentiment analysis in actual business scenarios) and the simulated weight parameters (e.g., weight values ​​calculated through simulation algorithms). Next, divide the actual adapted weight parameters of each group by the simulated weight parameters; the resulting ratio is the semantic error factor for that group. Finally, perform statistical analysis on the semantic error factors of multiple groups, such as taking the average or determining the final semantic error factor used for calibration based on distribution patterns, thereby achieving accurate calibration of subsequent semantic sentiment weight parameters.

[0096] In some examples, the fifth calculation formula can be used. The quadratic semantic simulation parameters were calculated. ,in, This represents the original parameters in the set of parameters to be calibrated. This represents the semantic error factor. If the weight ratio of identical results in the result set after customer intent identification in the semantic analysis set to be identified is greater than or equal to the intent determination threshold, the final intent result can be determined. Based on this result, an adaptive traffic redirection strategy is executed and a total traffic redirection report is output.

[0097] Secondly, correspondingly, this application also provides an intent recognition device capable of implementing all the processes of the intent recognition method provided in the above embodiments.

[0098] See Figure 2 The diagram shows a schematic representation of the intent recognition device 200 provided in an embodiment of this application. The intent recognition device 200 includes: Data acquisition module 201 is used to acquire the user data to be identified for the target user; The first identification module 202 is used to perform intent identification based on the user data to be identified and the preset first intent identification auxiliary information to obtain an initial intent identification result set. The second identification module 203 is used to update the first intention identification auxiliary information based on historical scene data of the intention identification scenario to obtain the second intention identification auxiliary information when the initial intention identification result set satisfies the first condition, and to perform intention identification based on the second intention identification auxiliary information and the user data to be identified to obtain the first final intention identification result. The first condition includes at least the following: the discriminant degree between any two different initial recognition results in the initial intent recognition result set is greater than a preset discriminant degree threshold.

[0099] In an optional embodiment, the device further includes a third identification module, the third identification module being used for: If the initial intent recognition result set satisfies the second condition, the first intent recognition auxiliary information is updated based on historical scene data of the intent recognition scenario to obtain second intent recognition auxiliary information; and, the main semantic analysis result set is determined based on the user data to be identified and the second intent recognition auxiliary information; and Intent recognition is performed based on the main semantic analysis result set, the user data to be identified, and the second intent recognition auxiliary information to obtain the second final intent recognition result; The second condition includes: at least one of the discrimination scores is less than or equal to the preset discrimination threshold.

[0100] In one optional implementation, determining the main semantic analysis result set based on the user data to be identified and the second intent recognition auxiliary information includes: The text to be analyzed is obtained based on the user data to be identified; Based on the second intent recognition auxiliary information, the first type of sentences in the text to be analyzed are analyzed to obtain the main semantic analysis result set.

[0101] In one optional implementation, the step of performing intent recognition based on the main semantic analysis result set, the user data to be identified, and the second intent recognition auxiliary information to obtain a second final intent recognition result includes: Based on the first type of statement and the second type of statement in the text to be analyzed, the second intention recognition auxiliary information is adjusted to obtain the third intention recognition auxiliary information. The text to be analyzed is obtained according to the user data to be identified. In the text to be analyzed, the semantic importance of the first type of statement is higher than that of the second type of statement. The second type of statement is analyzed based on the third intent recognition auxiliary information to obtain a set of auxiliary semantic analysis results; Based on the main semantic analysis result set and the auxiliary semantic analysis result set, a second semantic analysis result set to be identified is generated; Intent recognition is performed based on the second set of semantic analysis results to be identified and the user data to be identified, to obtain the second final intent recognition result.

[0102] In one optional implementation, the first intent recognition auxiliary information includes a first semantic sentiment weight parameter set; The intention recognition is performed based on the user data to be identified and the preset first intention recognition auxiliary information to obtain an initial intention recognition result set, including: Determine the test text corresponding to the user data to be identified; Based on the first semantic sentiment weight parameter set, semantic analysis is performed on the text to be tested to obtain the first set of semantic analysis results to be identified. Key semantic elements are extracted from each initial identification result contained in the first set of semantic analysis results to be identified to obtain an intent feature tag set, wherein the intent feature tag set includes multiple tags corresponding to multiple intent categories; The initial identification results are matched with the intent feature label set to determine the intent category to which each initial identification result belongs; Based on each initial identification result and its respective intent category, the initial identification result set is determined, wherein the initial identification results of different categories refer to the initial identification results belonging to different intent categories.

[0103] In an optional embodiment, the device further includes an intent weight percentage calculation module, the intent weight percentage calculation module being used for: Determine the category weights corresponding to each of the multiple intent categories; For each of the plurality of intent categories, determine the number of initial recognition results belonging to each intent category; The intention weight percentage of each intention category is obtained by weighting the number of results and the category weight corresponding to each of the multiple intention categories. in, The first condition is that the discriminant difference between any two initial recognition results of different classes is greater than a preset discriminant difference threshold, and the intention weight ratio of at least one intention category among the plurality of intention categories is less than a preset intention determination threshold; and / or, The second condition is that at least one of the distinguishability scores is less than or equal to the preset distinguishability threshold, and the intention weight ratio of at least one intention category among the plurality of intention categories is less than the preset intention determination threshold.

[0104] In one optional implementation, the historical scene data includes historical semantic sentiment data, the first intent recognition auxiliary information includes a first semantic sentiment weight parameter set, and the second intent recognition auxiliary information includes a second semantic sentiment weight parameter set. The process of updating the first intent recognition auxiliary information based on historical scene data of the intent recognition scenario to obtain the second intent recognition auxiliary information includes: Based on the aforementioned historical semantic sentiment data, a semantic error factor is determined; The first semantic sentiment weight parameter set is calibrated based on the semantic error factor to obtain the second semantic sentiment weight parameter set.

[0105] In one optional implementation, the historical semantic sentiment data includes actual adaptation weight parameters and simulated weight parameters for each of multiple sets of historical texts, wherein the actual adaptation weight parameters are used to characterize the real weight values ​​of the corresponding set of historical texts, and the simulated weight parameters are used to characterize the simulated weight values ​​of the corresponding set of historical texts obtained through intent recognition. The determination of semantic error factors based on the historical semantic sentiment data includes: The error factor for each group of historical texts is determined based on the ratio between the actual adaptation weight parameters and the simulated weight parameters. Based on the error factors of each of the multiple sets of historical texts, the semantic error factor is obtained through statistical analysis.

[0106] Thirdly, embodiments of this application provide a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in any of the above-mentioned embodiments.

[0107] Fourthly, embodiments of this application provide a computer program product, including computer instructions that, when executed by a processor, implement the steps of the method described in any of the above-described embodiments.

[0108] Fifthly, embodiments of this application provide a computer device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the steps of the method described in any of the preceding claims.

[0109] See Figure 3 The computer device in this embodiment includes a processor 301, a memory 302, and a computer program, such as an intent recognition program, stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program, it implements the steps in the various intent recognition method embodiments described above, for example... Figure 1 The steps S101-S103 are shown.

[0110] For example, the computer program may be divided into one or more modules / units, which are stored in the memory 302 and executed by the processor 301 to complete this application. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the computer device.

[0111] The computer device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that the schematic diagram is merely an example of a computer device and does not constitute a limitation on the computer device. It may include more or fewer components than shown, or combine certain components, or different components. For example, the computer device may also include input / output devices, network access devices, buses, etc.

[0112] The processor 301 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor 301 can be any conventional processor. The processor 301 is the control center of the computer device, connecting various parts of the entire computer device through various interfaces and lines.

[0113] The memory 302 can be used to store the computer programs and / or modules. The processor 301 implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory 302 and calling the data stored in the memory 302. The memory 302 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 302 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0114] Wherein, if the modules / units integrated into the computer device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a non-transitory computer-readable storage medium. When the computer program is executed by the processor 301, it can implement the steps of the various method embodiments described above. Wherein, the computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0115] In summary, the embodiments of this application have at least the following beneficial effects: In this embodiment, user data to be identified for a target user is acquired; intent recognition is performed based on the user data to be identified and preset first intent recognition auxiliary information to obtain an initial intent recognition result set; if the initial intent recognition result set meets a first condition, the first intent recognition auxiliary information is updated based on historical scene data of the intent recognition scenario to obtain second intent recognition auxiliary information; and intent recognition is performed based on the second intent recognition auxiliary information and the user data to be identified to obtain a first final intent recognition result. The first condition includes at least: the distinguishability between any two different initial recognition results in the initial intent recognition result set is greater than a preset distinguishability threshold. Thus, the user data to be identified and the first intent recognition auxiliary information can be combined to complete preliminary intent recognition and obtain an initial intent recognition result set. If the initial intent recognition result set meets the first condition, it is determined that the recognition accuracy corresponding to the initial intent recognition result set is insufficient, and a more accurate intent recognition is required. The second intent recognition auxiliary information is updated using historical scene data of the intent recognition scenario to improve the recognition accuracy of the obtained first final intent recognition result.

[0116] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary hardware platforms, or it can be implemented entirely by hardware. Based on this understanding, all or part of the technical solutions of this application that contribute to the background technology can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM (Read-Only Memory) / RAM (Random Access Memory), magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0117] The above description is the 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 are also considered to be within the scope of protection of this application.

Claims

1. An intent recognition method, characterized in that, include: Obtain the target user data to be identified; Based on the user data to be identified and the preset first intent recognition auxiliary information, intent recognition is performed to obtain an initial intent recognition result set; When the initial intent recognition result set satisfies the first condition, the first intent recognition auxiliary information is updated based on the historical scene data of the intent recognition scenario to obtain the second intent recognition auxiliary information, and intent recognition is performed based on the second intent recognition auxiliary information and the user data to be identified to obtain the first final intent recognition result. The first condition includes at least the following: the discriminant difference between any two different initial recognition results in the initial intent recognition result set is greater than a preset discriminant difference threshold.

2. The method according to claim 1, characterized in that, The method further includes: If the initial intent recognition result set satisfies the second condition, the first intent recognition auxiliary information is updated based on historical scene data of the intent recognition scenario to obtain second intent recognition auxiliary information; and, the main semantic analysis result set is determined based on the user data to be identified and the second intent recognition auxiliary information; and Intent recognition is performed based on the main semantic analysis result set, the user data to be identified, and the second intent recognition auxiliary information to obtain the second final intent recognition result; The second condition includes: at least one of the discrimination scores is less than or equal to the preset discrimination threshold.

3. The method according to claim 2, characterized in that, The step of determining the main semantic analysis result set based on the user data to be identified and the second intent recognition auxiliary information includes: The text to be analyzed is obtained based on the user data to be identified; Based on the second intent recognition auxiliary information, the first type of sentences in the text to be analyzed are analyzed to obtain the main semantic analysis result set.

4. The method according to claim 2, characterized in that, The process of performing intent recognition based on the main semantic analysis result set, the user data to be identified, and the second intent recognition auxiliary information to obtain a second final intent recognition result includes: Based on the first type of statement and the second type of statement in the text to be analyzed, the second intention recognition auxiliary information is adjusted to obtain the third intention recognition auxiliary information. The text to be analyzed is obtained according to the user data to be identified. In the text to be analyzed, the semantic importance of the first type of statement is higher than that of the second type of statement. The second type of statement is analyzed based on the third intent recognition auxiliary information to obtain a set of auxiliary semantic analysis results; Based on the main semantic analysis result set and the auxiliary semantic analysis result set, a second semantic analysis result set to be identified is generated; Intent recognition is performed based on the second set of semantic analysis results to be identified and the user data to be identified, to obtain the second final intent recognition result.

5. The method according to claim 2, characterized in that, The first intent recognition auxiliary information includes a first semantic sentiment weight parameter set; The intention recognition is performed based on the user data to be identified and the preset first intention recognition auxiliary information to obtain an initial intention recognition result set, including: Determine the test text corresponding to the user data to be identified; Based on the first semantic sentiment weight parameter set, semantic analysis is performed on the text to be tested to obtain the first set of semantic analysis results to be identified. Key semantic elements are extracted from each initial identification result contained in the first set of semantic analysis results to be identified to obtain an intent feature tag set, wherein the intent feature tag set includes multiple tags corresponding to multiple intent categories; The initial identification results are matched with the intent feature label set to determine the intent category to which each initial identification result belongs; Based on each initial identification result and its respective intent category, the initial identification result set is determined, wherein the initial identification results of different categories refer to the initial identification results belonging to different intent categories.

6. The method according to claim 5, characterized in that, The method further includes: Determine the category weights corresponding to each of the multiple intent categories; For each of the plurality of intent categories, determine the number of initial recognition results belonging to each intent category; The intention weight percentage of each intention category is obtained by weighting the number of results and the category weight corresponding to each of the multiple intention categories. in, The first condition is that the discriminant difference between any two initial recognition results of different classes is greater than a preset discriminant difference threshold, and the intention weight ratio of at least one intention category among the plurality of intention categories is less than a preset intention determination threshold; and / or, The second condition is that at least one of the distinguishability scores is less than or equal to the preset distinguishability threshold, and the intention weight ratio of at least one intention category among the plurality of intention categories is less than the preset intention determination threshold.

7. The method according to any one of claims 1-6, characterized in that, The historical scene data includes historical semantic and sentiment data, the first intention recognition auxiliary information includes a first semantic and sentiment weight parameter set, and the second intention recognition auxiliary information includes a second semantic and sentiment weight parameter set. The process of updating the first intent recognition auxiliary information based on historical scene data of the intent recognition scenario to obtain the second intent recognition auxiliary information includes: Based on the aforementioned historical semantic sentiment data, a semantic error factor is determined; The first semantic sentiment weight parameter set is calibrated based on the semantic error factor to obtain the second semantic sentiment weight parameter set.

8. The method according to claim 7, characterized in that, The historical semantic sentiment data includes actual adaptation weight parameters and simulated weight parameters for each of multiple sets of historical texts. The actual adaptation weight parameters are used to characterize the real weight values ​​of the corresponding set of historical texts, and the simulated weight parameters are used to characterize the simulated weight values ​​of the corresponding set of historical texts obtained through intent recognition. The determination of semantic error factors based on the historical semantic sentiment data includes: The error factor for each group of historical texts is determined based on the ratio between the actual adaptation weight parameters and the simulated weight parameters. Based on the error factors of each of the multiple sets of historical texts, the semantic error factor is obtained through statistical analysis.

9. An intent recognition device, characterized in that, include: The data acquisition module is used to acquire the user data to be identified for the target user. The first identification module is used to perform intent identification based on the user data to be identified and the preset first intent identification auxiliary information to obtain an initial intent identification result set. The second identification module is used to update the first intention identification auxiliary information based on historical scene data of the intention identification scenario to obtain the second intention identification auxiliary information when the initial intention identification result set satisfies the first condition, and to perform intention identification based on the second intention identification auxiliary information and the user data to be identified to obtain the first final intention identification result. The first condition includes at least the following: the discriminant difference between any two different initial recognition results in the initial intent recognition result set is greater than a preset discriminant difference threshold.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-8.

11. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the method described in any one of claims 1-8.

12. A computer device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the method of any one of claims 1-8.

Images