A method, system and device for redirecting services within a mobile application

Abstract

The present invention discloses a method, system, and device for redirecting services within a mobile application, relating to the field of data processing technology. The method comprises the following steps: obtaining a language instruction, obtaining a semantic vector and a context feature vector based on a standard business scenario specification and context information; obtaining an intent to be processed based on the expression type; extracting entities and actions in the intent to be processed, associating the entities and corresponding actions with business modules, mapping the business modules to URL-Scheme parameters in a preset URL-Scheme scenario specification, locating the target business module based on the mapping relationship, and executing the action based on the business entity and the corresponding business action; if no mapping relationship exists, redirecting to a large cloud model integrated within the application for processing, and returning the result to the mobile terminal to complete the interaction. The method enables users to quickly and intelligently complete redirection and interaction of mobile services through voice instructions.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method, system and device for redirecting services within a mobile application. Background Art

[0002] As mobile applications become increasingly complex, users are increasingly eager to quickly locate their target page. Traditional in-app navigation relies on manually clicking through hierarchical menus or entering input in search boxes. This requires manual switching and navigating between different functional modules within the app, resulting in long, cumbersome, and inefficient operation paths. While mainstream mobile smart assistants support question-and-answer or message card formats, users still need to click multiple times. For example, manually entering text or clicking on voice input will return a business card or response based on the user's input, requiring the user to manually click on the business card to navigate to the relevant business system.

[0003] While existing voice-driven redirection solutions simplify operations, they suffer from significant deficiencies in accuracy, adaptability, and user experience:

[0004] (1) Semantic matching ignores contextual relevance: Existing technologies mostly use static semantic matching (such as direct mapping of keywords to URL-Scheme), which does not take into account the user's current page status and historical operation habits. As a result, ambiguous instructions frequently trigger erroneous jumps and fail to provide users with a personalized and intelligent service experience.

[0005] (2) Insufficient robustness of intent recognition: User intent is determined only through single text matching, without considering high-frequency operation characteristics (such as users frequently checking "attendance records" at 9 o'clock every day), resulting in delayed response to "habitual instructions". For example, if a user says "check yesterday's", the system cannot accurately locate the intention based on the user's frequent behavior of checking "yesterday's performance report".

[0006] (3) The processing mechanism when the voice command cannot be recognized is single: When the voice command cannot be recognized, it directly returns an error prompt or stays on the current page. No integrated remedial path is provided, forcing users to switch to manual operation, interrupting the interaction continuity.

[0007] (4) Rigid mechanism for handling ambiguous instructions: Existing voice jump solutions use a unified recognition logic for ambiguous and clear statements, relying mostly on single semantic matching or fixed rules, without differentiated processing based on the degree of ambiguity. For ambiguous instructions, if the intent is forced to be selected only through keyword matching, the error rate will be high; for clear instructions, deep context analysis will not be fully utilized, resulting in wasted computing power or delayed recognition.

[0008] As a result, the above problems lead to high error rates and high user abandonment rates in existing voice jump solutions, making it difficult to meet the efficient interaction requirements of complex applications. Summary of the Invention

[0009] In order to solve the above problems, the present invention proposes a method, system and device for mobile terminal in-application service jump, which combines dynamic context correction to improve the accuracy and flexibility of voice command-driven page jump, enabling users to quickly and intelligently complete mobile service jump and interaction through voice commands.

[0010] In order to achieve the above object, the present invention adopts the following technical solutions:

[0011] In a first aspect, the present invention provides a method for redirecting a service within a mobile application, comprising:

[0012] Obtain user voice commands and convert them into text commands;

[0013] Semantically match text instructions with pre-set standard business scenario specifications, perform semantic corrections based on historical operations within the current page and the set time period within the application, and obtain a semantic vector.

[0014] Construct a context feature vector based on the current page, historical operations within the application during a set period, and high-frequency operations with an operation frequency greater than a set threshold within the set period;

[0015] If the semantic vector is an ambiguous expression, the semantic vector is matched with the preset intent library, and the candidate intents are scored according to the context feature vector to obtain the intent to be processed. Otherwise, the context feature vector and the semantic vector are concatenated, and the trained large model is used to identify the intent to be processed.

[0016] Extract the entities and actions in the intent to be processed, associate the entities and corresponding actions with the business modules, map the business modules to the URL-Scheme parameters in the preset URL-Scheme scenario specification, locate the target business module based on the mapping relationship, and execute the actions based on the business entities and corresponding business actions;

[0017] If there is no mapping relationship, it will jump to the cloud-based large model integrated in the application for processing, and return the results to the mobile terminal to complete the interaction.

[0018] As an optional implementation method, samples including fuzzy expressions, incomplete expressions and diversified expressions are used to pre-build standard business scenario specifications, where diversified expressions refer to different expression forms corresponding to one business operation.

[0019] As an optional implementation, the process of scoring the obtained candidate intents based on the context feature vector includes:

[0020] Preset the basic weights for the current page, historical operations, and high-frequency operations, and adjust the basic weights based on the type of the current page, the completeness of the historical operation sequence, and the total frequency of high-frequency operations;

[0021] The relevance of the candidate intent to the current page, the matching degree of historical operations, and the contribution of high-frequency operations are evaluated respectively, and the score of each candidate intent is obtained by weighted summation based on the adjusted weights. The candidate intent with the highest score is the intent to be processed.

[0022] As an optional implementation, the process of evaluating the relevance of the candidate intent to the current page, the matching degree of historical operations, and the contribution of high-frequency operations is as follows:

[0023] Define high-relevance intent, medium-relevance intent, and low-relevance intent for each page ID in advance, and set the relevance degree for each. Based on the candidate intent, match the corresponding intent under the corresponding page ID to obtain the relevance degree of the current page.

[0024] Extract the intention sequence from the historical operations and calculate the similarity between the candidate intention and the last set step operation in the intention sequence. At the same time, if the time period of the historical operation is less than the minimum time threshold, multiply the similarity by the first weight. If it is greater than the maximum time threshold, multiply it by the second weight. The first weight is greater than 1 and the second weight is less than 1. This is used to obtain the matching degree of the historical operation.

[0025] Determine the frequency of occurrence of candidate intentions in high-frequency operations, and take the quotient of the candidate intention frequency and the total high-frequency frequency as the basic score. If the occurrence time of the high-frequency operation matches the current time, multiply the basic score by the third weight, and the third weight is greater than 1. Otherwise, the basic score remains unchanged, thereby obtaining the contribution of the high-frequency operation.

[0026] As an optional implementation, the process of adjusting the basic weight includes:

[0027] If the current page is a function-specific page, increase the basic weight of the current page's relevance and reduce the basic weight of high-frequency operations; if the current page is a general page, increase the basic weight of high-frequency operations and reduce the basic weight of the current page;

[0028] If the number of steps in the intended sequence of historical operations per unit time is greater than the set step threshold, the basic weight of the historical operation is increased; otherwise, the basic weight of the historical operation is reduced and replaced by the basic weight of the high-frequency operation;

[0029] If the total frequency of high-frequency operations is less than the set threshold, the basic weight of high-frequency operations is reduced.

[0030] As an optional implementation method, if the difference between the highest score and the second highest score is less than the set difference threshold, a secondary check is triggered. The historical operation success rates of the candidate intents corresponding to the highest score and the second highest score are combined, and the candidate intent with the higher historical operation success rate is selected as the intent to be processed;

[0031] If the highest score is lower than the set minimum score threshold, the intention is determined to be ambiguous and feedback is given to the user for secondary confirmation.

[0032] As an optional implementation, the mapping process includes:

[0033] Pre-build a knowledge graph of entities and relationships, extract keywords for entities and corresponding actions, and then match corresponding business scenarios based on the relationships between entities, actions, and business modules in the knowledge graph.

[0034] According to the association between business modules and URL-Scheme parameters and entities and URL parameters in the knowledge graph, the extracted keywords are filled as URL-Scheme parameters.

[0035] As an optional implementation, candidate intents are generated by calculating the cosine similarity between the semantic vector and each intent in the preset intent library.

[0036] In a second aspect, the present invention provides a mobile terminal in-application service jump system, comprising:

[0037] An initialization module is configured to obtain user voice commands and convert them into text commands;

[0038] The semantic module is configured to semantically match text instructions with preset standard business scenario specifications, perform semantic correction based on the current page and historical operations within a set period within the application, and obtain a semantic vector;

[0039] A context module is configured to construct a context feature vector based on the current page, historical operations within a set period of time within the application, and high-frequency operations with an operation frequency greater than a set threshold within the set period of time;

[0040] The recognition module is configured to match the semantic vector with a preset intent library if the semantic vector is an ambiguous expression, and score the candidate intents obtained based on the context feature vector to obtain the intent to be processed; otherwise, the context feature vector is spliced ​​with the semantic vector and the trained large model is used to identify the intent to be processed;

[0041] A mapping module is configured to extract entities and actions from the intent to be processed, associate the entities and corresponding actions with business modules, map the business modules to URL-Scheme parameters in a preset URL-Scheme scenario specification, locate the target business module based on the mapping relationship, and execute the action based on the business entity and the corresponding business action;

[0042] The cloud module is configured to jump to the cloud model integrated in the application for processing if there is no mapping relationship, and return the results to the mobile terminal to complete the interaction.

[0043] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] The present invention proposes a method, system, and device for business redirection within a mobile application. The method obtains a language instruction and obtains a semantic vector and a context feature vector based on standard business scenario specifications and context information. The method processes the intent to be processed based on the expression type. The method extracts the entities and actions in the intent to be processed, associates the entities and corresponding actions with the business module, maps the business module to the URL-Scheme parameters in the preset URL-Scheme scenario specification, locates the target business module based on the mapping relationship, and executes the action based on the business entity and the corresponding business action. If there is no mapping relationship, the method redirects to a large cloud model integrated in the application for processing, and returns the result to the mobile terminal to complete the interaction. Through ambiguous hierarchical processing, multi-dimensional context fusion, dynamic mapping, and cloud-based collaborative mechanisms, the accuracy, adaptability, and interaction continuity of business redirection within mobile applications are significantly improved.

[0046] The method of the present invention adopts a differentiated processing strategy for ambiguous expressions. In the case of ambiguity, the candidate intent is scored by the context feature vector, and the weights of the current page, historical operations, and high-frequency operations are integrated to avoid misjudgment caused by a single match. In addition, the context and semantic vectors are spliced ​​together and deeply analyzed by the trained large model to fully utilize the semantic understanding ability of the large model.

[0047] The context feature vector in the method of the present invention integrates three-dimensional information of "current page + historical operations in a set time period + high-frequency operations", so that intention recognition can dynamically adapt to user behavior habits and scene characteristics, enhancing scene adaptability and personalized matching.

[0048] When no URL-Scheme mapping exists locally, the app automatically redirects to the large cloud-based model integrated within the app for processing, leveraging cloud computing power and a richer knowledge base to resolve complex instructions and returning the results to the mobile device to complete the interaction. This local + cloud collaborative mechanism avoids the interruptions of traditional solutions and ensures continuous and scalable interaction.

[0049] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0051] Figure 1 This is a flow chart of the mobile terminal in-application service jump method provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0053] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0054] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "comprise" and any variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0055] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0056] Example 1

[0057] This embodiment provides a method for jumping to a business within a mobile application, simplifies the user operation process, and quickly opens a business module through voice input. When the voice input cannot match the relevant business operation, the intelligent assistant is opened to enable the system to adapt to various business scenarios.

[0058] like Figure 1 As shown, specifically including:

[0059] Obtain user voice commands and convert them into text commands;

[0060] Semantically match text instructions with pre-set standard business scenario specifications, perform semantic corrections based on historical operations within the current page and the set time period within the application, and obtain a semantic vector.

[0061] Construct a context feature vector based on the current page, historical operations within the application during a set period, and high-frequency operations with an operation frequency greater than a set threshold within the set period;

[0062] If the semantic vector is an ambiguous expression, the semantic vector is matched with the preset intent library, and the candidate intents are scored according to the context feature vector to obtain the intent to be processed. Otherwise, the context feature vector and the semantic vector are concatenated, and the trained large model is used to identify the intent to be processed.

[0063] Extract the entities and actions in the intent to be processed, associate the entities and corresponding actions with the business modules, map the business modules to the URL-Scheme parameters in the preset URL-Scheme scenario specification, locate the target business module based on the mapping relationship, and execute the actions based on the business entities and corresponding business actions;

[0064] If there is no mapping relationship, it will jump to the cloud-based large model integrated in the application for processing, and return the results to the mobile terminal to complete the interaction.

[0065] In this embodiment, after receiving the user's voice command, the user's voice command is converted into a text command. This step can filter out invalid audio through local endpoint detection to solve the problem of irrelevant voice interference. It can accurately recognize the voice content in different environmental noise conditions, improving the accuracy and robustness of speech recognition.

[0066] In this embodiment, in combination with business scenarios within the application, such as clocking in, meetings, schedules, etc., samples including fuzzy expressions, incomplete expressions and diversified expressions are used to pre-build standard business scenario specifications; diversified expressions mean that one business operation corresponds to different expression forms, such as clocking in can be expressed as registering for get off work and leaving, signing in today, etc.

[0067] Secondly, the text instructions are semantically matched with standard business scenario specifications, and semantic corrections are made based on the historical operations (i.e., contextual information) within the current page and the set time period within the application to obtain a semantic vector.

[0068] For example, for diverse expressions, such as "checking in and out of get off work" and "signing in today" and other text instructions with diverse expressions are matched with "punch in".

[0069] For ambiguous and incomplete expressions, semantic completion is performed based on historical operations (i.e., contextual information) within the current page and a set period within the application.

[0070] For example, if the user does not finish a text instruction ("Help me add tomorrow's...") or the instruction is incomplete, the user can complete the instruction by combining the contextual information of the current page and the historical operations within the set time period in the application. If the current page is the "Schedule Homepage", the completion is "Help me add tomorrow's schedule"; if the operation history shows "A meeting was created 3 minutes ago", the completion is "Help me add tomorrow's meeting".

[0071] Then, a context feature vector is constructed based on the current page, historical operations within the set time period within the application, and high-frequency operations with an operation frequency greater than a set threshold within the set time period; through the fusion of context features, the accuracy of intent matching is improved.

[0072] Among them, the current page is such as the check-in list page, the meeting details page, etc.; historical operations within a set time period in the application, such as checking Zhang San’s contact information 10 minutes ago; high-frequency operations with an operation frequency greater than the set threshold within the set time period, such as users performing check-in operations an average of 3 times a day.

[0073] If the text instruction is ambiguous, incomplete, or diverse, the context feature vector and semantic vector are concatenated and the trained large model is used to identify the intent to be processed.

[0074] For example, if a user says "find him," combined with the context "the current page is Zhang San's chat history," it is parsed as "view Zhang San's details" (mapped to URL-Scheme: ecm-contact: / / {Zhang San id}); if the context is "currently on the meeting list page," it is parsed as "view the participants of the meeting."

[0075] If the text instruction is ambiguous, the semantic vector is matched with the preset intent library (such as clocking in, creating a meeting, checking the schedule, etc.);

[0076] Among them, the cosine similarity between the semantic vector and each intention in the preset intent library is calculated to generate candidate intentions. For example, opening a meeting may correspond to viewing the meeting list, creating a meeting, and joining a meeting; then the obtained candidate intentions are scored according to the context feature vector to obtain the intention to be processed.

[0077] The scoring scheme for candidate intents needs to combine the multi-dimensional information of the context feature vector (current page, historical operations, and high-frequency operations) to achieve accurate ranking through quantitative feature correlation and dynamic weight allocation.

[0078] Specifically:

[0079] 1. Scoring dimensions and quantification rules: Based on the three core dimensions of the context feature vector, the scoring indicators are broken down and the quantification method is defined.

[0080] (1) The relevance of the current page (weight W1 is set to 40%); the current page directly reflects the user's immediate operation scenario, and the matching degree with the candidate intent is the core scoring basis.

[0081] Quantification rules:

[0082] Preset page-intent association matrix: define high-correlation intent, medium-correlation intent, and low-correlation intent for each page ID, and set the correlation degree for each. According to the candidate intent, match the corresponding intent under the corresponding page ID to obtain the correlation degree of the current page.

[0083] For example: for high-relevance intent, the meeting details page is highly correlated with joining a meeting, and the homepage is highly correlated with viewing the meeting list, and the correlation is set to 0.8-1.0; for medium-relevance intent, the meeting details page and creating a meeting are highly correlated, and the correlation is set to 0.4-0.7; for low-relevance intent, the personal center page and joining a meeting are highly correlated, and the correlation is set to 0-0.3.

[0084] Example: If the current page is the "Meeting List Page", the relevance of viewing the meeting list = 0.9, creating a meeting = 0.6, and joining a meeting = 0.3.

[0085] (2) The matching degree of historical operations (weight W2 is set to 30%); the historical operations within a set time period (such as the past 1 hour) reflect the user's short-term goals, and the matching degree between the candidate intention and the historical operation sequence needs to be calculated.

[0086] Quantification rules:

[0087] Extract the intent sequence from historical operations (such as opening the meeting list → filtering today's meetings → clicking on meeting A), calculate the similarity between the candidate intent and the last set number of steps (such as the last 3 steps) in the intent sequence (the smaller the value, the more similar), and convert it into a score between 0 and 1 (such as 1.0 when the edit distance is 0, and 0 when the distance is greater than 3).

[0088] Introduce a time decay factor: the matching degree of recent operations (such as within 5 minutes) is multiplied by the first weight (such as 1.2 times the weight), and the matching degree of long-term operations (such as 30-60 minutes ago) is multiplied by the second weight (such as 0.8 times) to enhance the impact of timeliness; among them, the first weight is greater than 1 and the second weight is less than 1.

[0089] For example, if the historical operation is creating a meeting → setting the time, the historical matching degree for creating a meeting is 0.9 (edit distance 0), and viewing the meeting list is 0.3 (edit distance 2).

[0090] (3) Contribution of high-frequency operations (weight W3 is set to 30%); high-frequency operations within a set period (such as operation frequency in the past 7 days > threshold) reflect user habits, and the consistency between candidate intentions and high-frequency behaviors needs to be measured.

[0091] Quantification rules:

[0092] Count the frequency of occurrence of each intent in high-frequency operations, and calculate the frequency of candidate intents divided by the total high-frequency frequency as the basic score (for example, viewing the meeting list accounts for 60% of high-frequency operations, then the basic score of this intent is 0.6).

[0093] Superimposed scenario adaptation coefficient: If the scenario of high-frequency operations (such as 9:00 on weekdays) matches the current time / scenario, the base score is multiplied by the third weight (such as 1.3 times). For example, if the user frequently checks the meeting list at 9:00 every day, the score will be increased when the current time is 9:00). Otherwise, the score remains unchanged. The third weight is greater than 1.

[0094] 2. Dynamic weight adjustment mechanism: To avoid scenario adaptation deviation caused by fixed weights, the weights of each dimension are adjusted in real time according to the business scenario priority and feature credibility.

[0095] (1) Scenario priority:

[0096] If the current page is a function-specific page (such as a meeting editing page, order settlement page, etc.), the basic weight of the current page's relevance will be increased (for example, by 50%), while the basic weight of high-frequency operations will be reduced (for example, to 20%), because the exclusive page is highly scene-specific.

[0097] If the current page is a general page (such as the homepage, settings, personal center, etc.), the basic weight of high-frequency operations will be increased (for example, by 40%), and the basic weight of the current page will be reduced (for example, to 30%). This is because it is necessary to rely on user habits for auxiliary judgment.

[0098] (2) Feature credibility:

[0099] If the number of steps in the intention sequence of historical operations within a unit time (such as 1 hour) is greater than the set step threshold, the intention sequence is considered complete (such as ≥ 3 steps in 1 hour), and the basic weight of the historical operation is increased (such as by 10%); otherwise, the basic weight of the historical operation is reduced (such as by 10%), and the basic weight of the high-frequency operation is used to fill the gap.

[0100] If the total frequency of high-frequency operations is less than the set threshold (for example, less than 5 times), the basic weight of high-frequency operations is reduced (for example, to 20%) to avoid misjudgment caused by data sparsity.

[0101] It is understandable that the specific numerical settings for increasing and decreasing the above-mentioned weights are not limited and can be adaptively adjusted according to actual applications.

[0102] 3. Finally, the final score of each candidate intent is the weighted sum of the scores of each dimension: final score = (current page relevance × W1) + (historical operation matching × W2) + (high-frequency operation contribution × W3).

[0103] Arrange the candidate intents in descending order by the final score, and take the highest score as the intent to be processed.

[0104] 4. Complex scene fault tolerance and correction (solving fuzzy and ambiguous problems).

[0105] If the difference between the highest and second-highest scores is less than the set difference threshold, that is, the intent discrimination is low, a second verification is triggered: combined with the historical operation success rates of the candidate intents corresponding to the highest and second-highest scores, such as the average success rate of creating a meeting is 80%, and the average success rate of joining a meeting is 95%, the candidate intent with the higher success rate is given priority.

[0106] If the highest score is lower than the set minimum score threshold, the intent is determined to be ambiguous and needs to be combined with real-time user feedback, such as a pop-up window asking "Do you want to view the meeting list or create a meeting?"; or prompting the user for confirmation, such as "You may want to: 1. Create a new meeting, 2. View today's meetings? Please say the number (voice confirmation is supported)."

[0107] In addition, we design personalized model adaptation: for high-frequency users (e.g., monthly usage ≥30 times), we record their voice features (speech speed, common vocabulary, such as users accustomed to saying "report to" instead of "punch in"), and fine-tune the personalized model through federated learning (local storage of model parameters to protect privacy) to improve the accuracy of user intent recognition.

[0108] Example description.

[0109] Taking the candidate intent of "open meeting" (view meeting list / create meeting / join meeting) as an example, the scoring process in different contexts is as follows:

[0110] The current page is the "Meeting List Page", the historical operation is "Filter Today's Meetings", and among the high-frequency operations, "View Meeting List" accounts for 70%; dynamic weights: Current Page W1 = 50%, Historical Operation W2 = 30%, High-Frequency Operation W3 = 20%.

[0111] Intention score:

[0112] View the conference list: Current page relevance = 0.9 → 0.9 × 0.5 = 0.45; Historical matching = 0.8 → 0.8 × 0.3 = 0.24; High-frequency contribution = 0.7 → 0.7 × 0.2 = 0.14 → Total score = 0.83;

[0113] Create a meeting: Current page relevance = 0.4 → 0.4 × 0.5 = 0.2; Historical matching = 0.2 → 0.2 × 0.3 = 0.06; High-frequency contribution = 0.1 → 0.1 × 0.2 = 0.02 → Total score = 0.28;

[0114] Join the meeting: Current page relevance = 0.3→0.3×0.5=0.15; Historical matching = 0.3→0.3×0.3=0.09; High-frequency contribution = 0.2→0.2×0.2=0.04→Total score = 0.28.

[0115] Result: View Meeting List scored highest as a pending intent.

[0116] The method of this embodiment has the following advantages:

[0117] Accuracy: Resolves ambiguity caused by a single context by correlating multi-dimensional features (e.g., differentiating the intent of "opening a meeting" on different pages / times);

[0118] Iterative: Quantization rules for each dimension (such as time decay coefficient and scenario adaptation coefficient) can be optimized through training of user behavior data (such as using a logistic regression model to learn optimal weights).

[0119] Improve the ranking accuracy of candidate intents and significantly reduce jump errors caused by misjudgment of intent.

[0120] In this embodiment, a specific example of a solution for a typical complex scenario is given.

[0121] (1) The user's expression is vague (such as "let's do this"), and combined with the current page (such as the "Daily Information Filling Page"), the completion semantics is "fill in the daily information". If the score is greater than the threshold, it will jump directly.

[0122] (2) For ambiguous expressions (such as “open a meeting”), generate candidate intents (view / create / join a meeting), and combine them with operation history (such as “a meeting was created last week”) to improve the confidence of creating a meeting.

[0123] (3) For unfinished instructions (such as “Help me add tomorrow’s…”), the system completes the instruction to “Add tomorrow’s schedule” based on the current page (such as “Schedule Page”). If the score after completion is greater than the threshold, the system jumps directly.

[0124] (4) Diversified expressions (e.g., “checking in and out” refers to “punching in”). Through a large model, the similarity between the diversified expression vector and the punching intention vector is improved to achieve accurate matching.

[0125] In this embodiment, a URL-Scheme scenario specification is pre-built to identify unique identifiers of different business modules and pages in an application, so as to map the intent to be processed with the URL-Scheme scenario specification.

[0126] URL-Scheme scenario specifications such as:

[0127] View user details mapped as ecm-contact: / / {user id};

[0128] Enter the chat channel mapping as ecc-channel: / / {channel id};

[0129] To call a user, map it to ecm-callphone: / / {user id};

[0130] Create a group mapping based on the group member ID as ecc-channel: / / newgroup?name={group name}&members={member ID};

[0131] Open applications, such as clocking in, reporting work, and traveling. All applications (H5) are mapped to ecc-app: / / {application id};

[0132] Set the text size mapping to cloudplus-system: / / font?size={text size value};

[0133] Switch the Cloud Plus language mapping to cloudplus-system: / language?type={language type};

[0134] Set the application theme color mapping to cloudplus-system: / / theme?type={theme color value};

[0135] Searching for users / groups / chat messages / applications / functions is mapped to cloudplus-function: / / search?content={search content};

[0136] View the calendar homepage mapping as inspur-ecc-native: / / calendar;

[0137] To view the specific schedule details, map it to inspur-ecc-native: / / calendar?{schedule id};

[0138] Check that the task homepage is mapped to inspur-ecc-native: / / to-do;

[0139] View the specific task details mapped to inspur-ecc-native: / / to-do?{task id};

[0140] Check that my meeting information is mapped to inspur-ecc-native: / / webex;

[0141] Check the exchange mailbox list mapping to inspur-ecc-native: / / mail;

[0142] Check the address book module (the address book tab at the bottom) and map it to ecc-contacts: / / index;

[0143] View the group news mapping as ecc-component: / / news.ecc;

[0144] Check the trip information mapping as ecc-component: / / trips / {trip id};

[0145] View the public knowledge document mapping as ecc-component: / / knowledge;

[0146] The unknown mapping is ecc-component: / / iGIXAi?id={intelligent assistant application ID}; = As a fallback solution, the client parses and opens the intelligent assistant (H5 page with integrated large model) to interact with the user.

[0147] It will be understood that the above is only an example, and those skilled in the art may make adaptive adjustments based on the actual business of the mobile application.

[0148] In this embodiment, the entities and actions in the intent to be processed are extracted, the entities and corresponding actions are associated with the business modules, the business modules are mapped to the URL-Scheme parameters in the preset URL-Scheme scenario specification, the target business module is located according to the mapping relationship, and the action is executed according to the business entity and the corresponding business action;

[0149] Specifically:

[0150] Extract entities and actions from the intent to be processed; entities include user name, user ID, business name (such as clocking in), meeting ID, schedule ID, group name, and member ID (when creating a group); actions include viewing user details, entering a chat channel, clocking in, creating a group, filling in daily information, meeting details page, and schedule homepage; for example, "meeting" in "view tomorrow's meetings" and "schedule" and time parameters in "add Wednesday afternoon schedule."

[0151] Define the entities and relationships of the knowledge graph to bind keywords with business jump logic, as shown in Table 1 and Table 2.

[0152] Table 1 Entity and action types;

[0153] .

[0154] Table 2 Relationship types;

[0155] .

[0156] The association establishment process is as follows:

[0157] (1) Entities and corresponding actions are associated with business modules. After extracting keywords, the corresponding business scenarios are matched through the association relationship of entity → action → business module in the knowledge graph.

[0158] For example: when "Zhang San" (user name) is extracted, the related business modules found through knowledge graph query are "View user details" and "Make a call"; when "Meeting" is extracted, it is associated with business modules such as "View meeting details" and "Create meeting".

[0159] (2) Binding with URL-Scheme parameters;

[0160] According to the association relationship between business module → URL-Scheme parameter and entity → URL parameter in the knowledge graph, the extracted keywords are filled as URL-Scheme parameters.

[0161] For example, if the service module "Call User" is bound to the URL-Scheme "ecm-callphone: / / {user id}", and the extracted "Zhang San" is mapped to the user id "1001", then "ecm-callphone: / / 1001" is generated.

[0162] The business module "View Schedule Details" is bound to the URL-Scheme "inspur-ecc-native: / / calendar?{schedule id}". The extracted "Wednesday afternoon schedule" is mapped to the schedule id "s306", and "inspur-ecc-native: / / calendar?s306" is generated.

[0163] (3) Locate the target business module based on the mapping relationship, and execute actions based on the business entity and the corresponding business action; through the URL-Scheme, realize the functions of jumping to the page within the application, opening the smart assistant, etc.

[0164] If the server recognizes the user to clock in, it returns the clock-in URL-Scheme (ecc-app: / / {clock-in application id}) to the client. After receiving it, the client jumps to the clock-in page to complete the automatic clock-in.

[0165] If it is recognized that the user has created a supervision, the URL-Scheme (ecc-app: / / {supervision application id}) will be returned. After receiving it, the client will jump to the supervision page and automatically fill in the user's voice input information and basic information.

[0166] If it is recognized that a chat channel is opened to send a message to someone, the URL-Scheme (ecc-channel: / / {channel id}) is returned. After receiving it, the client parses the URL, opens the relevant channel according to the channel ID, and sends the message.

[0167] If it is recognized that the user calls a contact in the enterprise, the call URL-Scheme (ecm-callphone: / / {user id}) is returned. The client searches for the contact according to the URL-Scheme and automatically dials the number.

[0168] In addition, in this embodiment, precise association can also be performed in combination with the context: context information (current application, operation history) is used as an association correction condition to filter out irrelevant business modules.

[0169] For example, if the user is currently in the "chat app" and "Zhang San" is retrieved, priority will be given to "Send a message to Zhang San" (ecc-channel: / / {channel id}) rather than "View user details"; if the operation history shows "Opened the calendar page 10 minutes ago" and "Meeting" is retrieved, priority will be given to "Meeting details in the calendar" (inspur-ecc-native: / / calendar?{schedule id}).

[0170] In this embodiment, a trained large model is used to perform deep semantic understanding and intent analysis of text instructions. Trained on massive amounts of text data, the large model can understand complex language expressions, identify users' business needs and operational intentions, and incorporate personalized adaptation and fault-tolerance mechanisms to address issues such as vague, ambiguous user expressions and abnormal speech speeds. During intent analysis, contextual information from the user's operations within the mobile app, including the current app, page, and operation history, is used as a reference to ensure more accurate intent understanding and align with actual usage scenarios.

[0171] In this embodiment, when the extracted intent cannot match the pre-calibrated URL-Scheme scenario specification, that is, there is no corresponding business module, it is deemed that the user intent cannot be recognized, and a no-result URL-Scheme (ecc-component: / / iIGXAi?id={smart assistant application ID}) is returned, and the process jumps to the cloud-based large model (smart assistant) integrated in the application for processing. The user session is passed to the smart assistant, which serves as a backup solution to complete the interaction, such as querying the company's business trip policy, etc. It can adapt to various complex business scenarios and improve the versatility and stability of the system.

[0172] In this embodiment, a real-time optimization solution is also provided, specifically:

[0173] (1) Speech recognition interacts through sockets and uses dynamic correction to return recognition results in real time. Each result returned may be an addition to the previous result, or it may replace the result returned previously (i.e., correction).

[0174] (2) During voice input, if the socket interaction recognizes that the user is silent (i.e., no user input) for more than 1.5 seconds, the voice input is terminated and the input voice text is transmitted to the server.

[0175] (3) When the voice input ends, set the service response timeout to 5 seconds. If no response is received from the server before the time ends, the user will be prompted that the voice input cannot be recognized, and the user will be asked to input the voice again and the user's voice input will be monitored again.

[0176] (4) Anti-noise optimization: Integrated adaptive noise suppression algorithm (spectral subtraction), the recognition accuracy remains ≥92% in ambient noise below 50dB, ensuring real-time performance in complex environments.

[0177] (5) Jump error correction strategy: If the page jumps to an unexpected page, the user can click the back button on the page and try again with voice input; Intelligent error correction: If there are two consecutive jump errors, the intelligent assistant will be automatically triggered (calling URL-Scheme: ecc-component: / / iGIXAi), and the assistant will process the voice input.

[0178] The above method of this embodiment has the following advantages:

[0179] (1) Save operation time and improve usage efficiency: Through voice recognition technology, the principle is to convert user voice input into bit stream in real time, which is parsed into text by the client and sent to the large model server. Users do not need to perform tedious operations manually, but can quickly complete business jumps through voice commands, saving operation time and improving usage efficiency.

[0180] (2) Accurately understand user expressions and intentions: Through the semantic understanding and intent analysis of big model technology, identify user business needs and operational intentions. By utilizing the powerful semantic understanding and intent analysis capabilities of big models, it is possible to accurately understand the user's complex language expressions and operational intentions, reducing the occurrence of misoperations and misunderstandings.

[0181] (3) Compatible with complex business scenarios and supports multiple business scenarios: Provides a fallback strategy. If the server cannot recognize the specified URL-Scheme based on the user input, the mobile terminal will jump to the intelligent assistant (web page) integrated into the application after receiving the response and pass the user input text to the intelligent assistant, which will complete the interaction with the user. For example, if the user wants to query company rules and regulations, reimbursement quotas, etc., it can adapt to various complex business scenarios and improve the versatility and stability of the system.

[0182] (4) Efficient interaction, real-time voice-to-text conversion, and business jumps are completed according to URL-Scheme instructions: users open the voice input interface through the mobile application homepage. Users only need to use voice instructions. The server parses the intention based on the large model and generates a URL-Scheme and returns it to the mobile terminal. The mobile terminal quickly completes the business jump. Users do not need to click multiple times to open a page (some pages have deep entrances and require multiple operations to open), saving operation time and improving usage efficiency.

[0183] Example 2

[0184] This embodiment provides a mobile terminal in-application service jump system, including:

[0185] An initialization module is configured to obtain user voice commands and convert them into text commands;

[0186] The semantic module is configured to semantically match text instructions with preset standard business scenario specifications, perform semantic correction based on the current page and historical operations within a set period within the application, and obtain a semantic vector;

[0187] A context module is configured to construct a context feature vector based on the current page, historical operations within a set period of time within the application, and high-frequency operations with an operation frequency greater than a set threshold within the set period of time;

[0188] The recognition module is configured to match the semantic vector with a preset intent library if the semantic vector is an ambiguous expression, and score the candidate intents obtained based on the context feature vector to obtain the intent to be processed; otherwise, the context feature vector is spliced ​​with the semantic vector and the trained large model is used to identify the intent to be processed;

[0189] A mapping module is configured to extract entities and actions from the intent to be processed, associate the entities and corresponding actions with business modules, map the business modules to URL-Scheme parameters in a preset URL-Scheme scenario specification, locate the target business module based on the mapping relationship, and execute the action based on the business entity and the corresponding business action;

[0190] The cloud module is configured to jump to the cloud model integrated in the application for processing if there is no mapping relationship, and return the results to the mobile terminal to complete the interaction.

[0191] It should be noted that the systems provided in the above embodiments are only illustrated by the division of the above functional modules when implementing their functions. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the systems and method embodiments provided in the above embodiments are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0192] In further embodiments, there is also provided:

[0193] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed by the processor, wherein when the computer instructions are executed by the processor, the method described in Example 1 is performed. For the sake of brevity, no further details are given here.

[0194] It should be understood that in this embodiment, the processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), off-the-shelf field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0195] The memory may include a read-only memory and a random access memory, and provides instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0196] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the method described in Example 1 is performed.

[0197] The method in Example 1 can be directly implemented as a hardware processor, or can be implemented using a combination of hardware and software modules in the processor. The software module can be located in a storage medium mature in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, it will not be described in detail here.

[0198] A computer program product includes a computer program, which implements the method described in embodiment 1 when executed by a processor.

[0199] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions contained in program modules, which are executed in a device on a real or virtual processor of a target to perform the process / method described above. Generally, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided between program modules as needed. The machine-executable instructions for the program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0200] The computer program code for implementing the method of the present invention can be written in one or more programming languages. These computer program codes can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the computer or other programmable data processing device, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on a computer, partially on a computer, as an independent software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0201] In the context of the present invention, computer program code or related data can be carried by any appropriate carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals include electrical, optical, radio, acoustic, or other forms of propagated signals, such as carrier waves, infrared signals, and the like.

[0202] Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with this embodiment can be implemented using electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0203] It should be noted that the collection and processing of relevant data in this invention, when applied in practice, should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of authorization of laws and regulations and the personal information subject.

[0204] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A method for redirecting services within a mobile application, characterized in that: include: Obtain user voice commands and convert them into text commands; Semantically match text instructions with pre-set standard business scenario specifications, perform semantic corrections based on historical operations within the current page and the set time period within the application, and obtain a semantic vector. Construct a context feature vector based on the current page, historical operations within the application during a set period, and high-frequency operations with an operation frequency greater than a set threshold within the set period; If the semantic vector is an ambiguous expression, the semantic vector is matched with the preset intent library, and the candidate intents obtained are scored according to the context feature vector to obtain the intent to be processed; Otherwise, the context feature vector and the semantic vector are concatenated and the trained large model is used to identify the intent to be processed. Extract the entities and actions in the intent to be processed, associate the entities and corresponding actions with the business modules, map the business modules to the URL-Scheme parameters in the preset URL-Scheme scenario specification, locate the target business module based on the mapping relationship, and execute the actions based on the business entities and corresponding business actions; If there is no mapping relationship, it will jump to the cloud-based large model integrated in the application for processing, and return the results to the mobile terminal to complete the interaction.

2. A mobile terminal application service jump method according to claim 1, characterized in that: Standard business scenario specifications are pre-built using samples including ambiguous expressions, incomplete expressions, and diversified expressions. Diversified expressions refer to different expression forms corresponding to one business operation.

3. A mobile terminal application service jump method according to claim 1, characterized in that: The process of scoring the candidate intents based on the context feature vector includes: Preset the basic weights for the current page, historical operations, and high-frequency operations, and adjust the basic weights based on the type of the current page, the completeness of the intended sequence of historical operations, and the total frequency of high-frequency operations; The relevance of the candidate intent to the current page, the matching degree of historical operations, and the contribution of high-frequency operations are evaluated respectively, and the score of each candidate intent is obtained by weighted summation based on the adjusted weights. The candidate intent with the highest score is the intent to be processed.

4. A mobile terminal in-application service jump method according to claim 3, characterized in that: The process of evaluating the relevance of candidate intent to the current page, the matching degree of historical operations, and the contribution of high-frequency operations is as follows: Define high-relevance intent, medium-relevance intent, and low-relevance intent for each page ID in advance, and set the relevance degree for each. Based on the candidate intent, match the corresponding intent under the corresponding page ID to obtain the relevance degree of the current page. Extract the intention sequence from the historical operations and calculate the similarity between the candidate intention and the last set step operation in the intention sequence. At the same time, if the time period of the historical operation is less than the minimum time threshold, multiply the similarity by the first weight. If it is greater than the maximum time threshold, multiply it by the second weight. The first weight is greater than 1 and the second weight is less than 1. This is used to obtain the matching degree of the historical operation. Determine the frequency of occurrence of candidate intentions in high-frequency operations, and take the quotient of the candidate intention frequency and the total high-frequency frequency as the basic score. If the occurrence time of the high-frequency operation matches the current time, multiply the basic score by the third weight, and the third weight is greater than 1. Otherwise, the basic score remains unchanged, thereby obtaining the contribution of the high-frequency operation.

5. A mobile terminal in-application service jump method according to claim 3, characterized in that: The process of adjusting the base weights includes: If the current page is a function-specific page, increase the basic weight of the current page and reduce the basic weight of high-frequency operations; if the current page is a general page, increase the basic weight of high-frequency operations and reduce the basic weight of the current page; If the number of steps in the intended sequence of historical operations per unit time is greater than the set step threshold, the basic weight of the historical operation is increased; otherwise, the basic weight of the historical operation is reduced and replaced by the basic weight of the high-frequency operation; If the total frequency of high-frequency operations is less than the set threshold, the basic weight of high-frequency operations is reduced.

6. A mobile terminal in-application service jump method according to claim 3, characterized in that: If the difference between the highest score and the second highest score is less than the set difference threshold, a secondary check is triggered. The historical operation success rates of the candidate intents corresponding to the highest score and the second highest score are combined, and the candidate intent with the higher historical operation success rate is selected as the intent to be processed; If the highest score is lower than the set minimum score threshold, the intention is determined to be ambiguous and feedback is given to the user for secondary confirmation.

7. A mobile terminal in-application service jump method according to claim 1, characterized in that: The mapping process includes: Pre-build a knowledge graph of entities and relationships, extract keywords for entities and corresponding actions, and then match corresponding business scenarios based on the relationships between entities, actions, and business modules in the knowledge graph. According to the association between business modules and URL-Scheme parameters and entities and URL parameters in the knowledge graph, the extracted keywords are filled as URL-Scheme parameters.

8. A mobile terminal in-application service jump method according to claim 1, characterized in that: Candidate intents are generated by calculating the cosine similarity between the semantic vector and each intent in the preset intent library.

9. A mobile terminal in-application service jump system, characterized in that: include: An initialization module is configured to obtain user voice commands and convert them into text commands; The semantic module is configured to semantically match text instructions with preset standard business scenario specifications, perform semantic correction based on the current page and historical operations within a set period within the application, and obtain a semantic vector; A context module is configured to construct a context feature vector based on the current page, historical operations within a set period of time within the application, and high-frequency operations with an operation frequency greater than a set threshold within the set period of time; The recognition module is configured to match the semantic vector with a preset intent library if the semantic vector is an ambiguous expression, and score the obtained candidate intents based on the context feature vector to obtain the intent to be processed; Otherwise, the context feature vector and the semantic vector are concatenated and the trained large model is used to identify the intent to be processed. A mapping module is configured to extract entities and actions from the intent to be processed, associate the entities and corresponding actions with business modules, map the business modules to URL-Scheme parameters in a preset URL-Scheme scenario specification, locate the target business module based on the mapping relationship, and execute the action based on the business entity and the corresponding business action; The cloud module is configured to jump to the cloud model integrated in the application for processing if there is no mapping relationship, and return the results to the mobile terminal to complete the interaction.

10. An electronic device, characterized in that: The method comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 8 is completed.

Images