Question and answer intention recognition and classification method for infant care robot

Abstract

The invention discloses a question and answer intention recognition and classification method for an infant care robot. A question rewriting node, an intention recognition node and a branch node are included. The input data of the question rewriting node are questions input by a user, the training data of the question rewriting node are phrases for mothers, infants and child rearing scenes, the model parameters of the question rewriting node are obtained through adjustment based on the training data, and the question rewriting node processes the input data according to the model parameters of the question rewriting node and outputs the data; the intention recognition node obtains data output by the problem rewriting node and recognizes a user intention and an expected result; the branch nodes provide a plurality of large language models in mother and child breeding scenes, and the branch nodes select the corresponding large language models to execute data processing according to output data of the intention recognition nodes. Original user input is clarified through standardization and rewriting, then the accuracy and response efficiency of subsequent intention recognition nodes are remarkably improved, and finally the overall user experience is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a question and answer intention recognition and classification method for a baby care robot. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, intelligent assistants play an increasingly important role in daily life, especially in the field of parenting support, and their potential value is increasingly prominent. The growing demand of parents for instant, scientific and personalized parenting advice provides a broad application prospect for AI-driven baby care robots.

[0003] However, traditional question and answer systems face many challenges in dealing with complex and emotionally rich parenting problems. Keyword-based or simple rule-based systems often struggle to accurately understand user intent, especially when dealing with the inherent ambiguity, incompleteness or colloquial expressions of natural language. For example, users may make spelling mistakes, use keywords instead of complete sentences, or ask ambiguous questions. The complexity of such language often leads to inaccurate or contextually relevant answers. In addition, the parenting field covers multiple professional categories such as medical health, emotional support, daily care, early education, entertainment interaction, etc. Traditional systems usually lack the ability to integrate multi-domain professional knowledge, which means they are difficult to provide detailed and professional responses to diverse parenting topics, resulting in overly general or even inaccurate answers, which reduces user satisfaction and the practicality of the system.

[0004] These limitations directly lead to poor user experience, for example, when the system fails to understand the user's real needs, the user will feel frustrated and unable to obtain effective help.

[0005] Therefore, there is an urgent need for a question and answer intention recognition and classification method for a baby care robot that can solve one or more of the above problems. SUMMARY

[0006] To solve one or more problems in the prior art, the present application provides a question and answer intention recognition and classification method for a baby care robot. The technical solution adopted by the present application to solve the above problems is: a question and answer intention recognition and classification method for a baby care robot, comprising: a question rewriting node, an intention recognition node and a branch node.

[0007] The problem rewriting node is a large language model, input data of the problem rewriting node is a user input question, training data of the problem rewriting node is: original question sentences input by a user group related to mother and baby and child rearing, standard and non-standard questions and question methods in the mother and baby and child rearing scene, and specific type of vocabulary in the mother and baby and child rearing scene, model parameters of the problem rewriting node are obtained by adjusting based on the training data, and the problem rewriting node processes the input data according to the model parameters and outputs data;

[0008] The intent recognition node is a large language model, and the intent recognition node obtains data output by the problem rewriting node and recognizes user intent and expected results;

[0009] The branch node provides a plurality of large language models in the mother and baby and child rearing scene, including: a professional medical question and answer model, a daily casual chat model, a network search model, a poetry picture book model and a music playing model, and the branch node selects a corresponding large language model to execute data processing according to output data of the intent recognition node.

[0010] In some embodiments, the processing steps performed by the problem rewriting node on the input data and the training data include: question semantic analysis and understanding, question rewriting, and professional expression.

[0011] Further, in the question rewriting step: the original fuzzy question in the input data is completed and reconstructed, and it is judged according to the model parameters whether the original fuzzy question and the completion and reconstruction are within the consistency range and whether additional information is introduced.

[0012] Further, in the professional expression step: the data output by the question rewriting is converted into a standard question format according to the model parameters.

[0013] In some embodiments, the intent recognition training data of the intent recognition node is: standard user query and question information in the mother and baby and child rearing scene, a structured prompt word database, standardized label information of the branch node, intent annotation information of the user query and question information, and data output by the problem rewriting node, and the standardized label information of the branch node corresponds to standardized labels of a plurality of large language models included in the branch node.

[0014] The model parameters of the intent recognition node are adjusted based on the intent recognition training data. The intent recognition node classifies keywords, cleanses text, judges context, understands semantics, recognizes intent, and associates labels according to the model parameters. The intent recognition node associates the data output by the question rewriting node with one or more large language models in the branch node through label association according to the recognized user intent and expected result.

[0015] Further, the intent recognition node performs intent feature extraction, intent feature classification, association of intent features with standardized label information of the branch node, confidence checking, and fallback options on the data output by the question rewriting node when processing data.

[0016] Further, the confidence checking includes global confidence checking and branch confidence checking corresponding to the large language models in the branch node.

[0017] The global confidence checking defines the minimum confidence required to trigger the branch confidence checking. The branch confidence checking defines the minimum confidence of the intent feature classification, the association of the intent features with the standardized label information of the branch node, and the fallback options.

[0018] Further, when the global confidence checking determines that the intent recognition node fails to recognize intent or the prediction is ambiguous, the fallback options are used to request the user to provide additional information or ask the user clarifying questions.

[0019] In some embodiments, the professional medical question and answer model provides functions including parenting knowledge advice, emotional support, and hospital recommendations. The daily casual conversation model provides functions including emotional companionship and parenting advice. The network search model provides functions including real-time network search. The poetry picture book model provides functions including searching or creating bedtime stories, fables, and picture books. The music playing model provides functions including matching and playing music according to the current scene.

[0020] In some embodiments, the branch node further includes an enterprise search enhancement generation model for introducing enterprise design, background information, and brand information.

[0021] The technical effects achieved by the present application are: the above-mentioned problem rewriting node directly solves the "noise" and "fuzziness" problems that may exist in the original user query, clarifies the original user input through standardization and rewriting, thereby significantly improving the accuracy and response efficiency of the subsequent intent recognition node, and ultimately improving the overall user experience; the intent recognition node further identifies, classifies, and associates the standardized user problem output by the problem rewriting node, realizes single-problem multi-intent shunt processing, and combines the intent with the downstream response logic call (large language model of branch node), ensuring the efficiency and controllability of the entire system intent recognition to task response process. Through targeted training of the problem rewriting node and the intent recognition node, and coordinated association of the training data, the accuracy of the data output by the model in the mother and baby and parenting scenarios is improved, and the model "thinking" time is reduced. The original user problem is distributed to different function large language models through the branch node to solve the user problem.

[0022] By modularizing the large language model as a node, it can be flexibly integrated and updated with knowledge in different fields without interfering with the core intent recognition process, making the system highly scalable and maintainable, meeting the needs of products such as baby robots that need long-term evolution and adaptation to new knowledge. By having independent professional models, the system can be updated incrementally and efficiently, reducing maintenance costs and enhancing its commercial viability. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is a schematic block diagram of the present application. DETAILED DESCRIPTION

[0024] To make the above-mentioned purposes, features and advantages of the present application more apparent, the specific embodiments of the present application will be described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be implemented in many other ways different from the description, and those skilled in the art can make similar improvements without departing from the scope of the present application, so the present application is not limited to the specific embodiments disclosed below.

[0025] As Figure 1 shown, the present application discloses a question and answer intent recognition and classification method for a baby robot, characterized by comprising: a problem rewriting node, an intent recognition node, and a branch node.

[0026] The problem rewriting node is a large language model, the input data of the problem rewriting node is a user input question, the training data of the problem rewriting node is: original question sentences input by a user group related to mother and baby and child rearing, standard and non-standard questions and their question methods in the mother and baby and child rearing scene, and specific type of vocabulary in the mother and baby and child rearing scene, the model parameters of the problem rewriting node are obtained by adjusting based on the training data, and the problem rewriting node processes the input data according to the model parameters and outputs data;

[0027] The intent recognition node is a large language model, and the intent recognition node obtains the data output by the problem rewriting node and recognizes user intent and expected results;

[0028] The branch node provides a plurality of large language models in the mother and baby and child rearing scene, including: a professional medical question and answer model, a daily casual chat model, a network search model, a poetry picture book model, a music playing model and an enterprise retrieval enhancement generation model, and the branch node selects a corresponding large language model according to the output data of the intent recognition node to execute data processing, wherein the professional medical question and answer model is used in combination with the baby care retrieval enhancement generation model to provide more accurate answers.

[0029] For the enterprise retrieval enhancement generation model, it is used to introduce enterprise design, background information and brand information, and is used in combination with an enterprise question and answer model (a large language model), and the purpose is: "brand correction + accurate fact annotation", which strengthens the brand dissemination function of the robot, and makes it have the dual values of strategic promotion and brand strengthening in the vertical enterprise service field.

[0030] For the problem rewriting node, the main function is to optimize the original question input by the user into a clearer, more complete and standard expression form, so as to reduce the ambiguity and noise in natural language. Users often make spelling errors, input key words or fragmented expressions, which can easily affect subsequent processing. The node performs text cleaning and standardization through word segmentation, lowercasing, and removing punctuation, which significantly improves the input quality; has a question and answer optimization background, is meticulous and professional, is good at language optimization and the field of mother and baby and child rearing, and serves the mother and baby user group, to ensure that the rewritten content meets the language standard and is consistent with the scene.

[0031] Specifically, the processing steps performed by the problem rewriting node on the input data and the training data include: question semantic analysis and understanding, question rewriting, and professional expression.

[0032] In the question rewriting step: the original fuzzy question in the input data is completed and reconstructed, and it is judged according to the model parameters whether the meaning of the original fuzzy question and the completion and reconstruction is within the consistency range or whether additional information is introduced;

[0033] In the professional expression step: the data output by the problem rewriting node is converted into a standard question format according to the model parameters of the intent recognition node, wherein the standard question format can be a language format with language stylization.

[0034] Specifically, the intent recognition training data of the intent recognition node includes: standardized user queries and question information in the baby and parenting scenarios, a structured prompt word database, standardized label information of branch nodes, intent annotation information of user queries and question information, and data output by the problem rewriting node. The standardized label information of the branch nodes corresponds to the standardized labels of the plurality of large language models contained in the branch nodes.

[0035] The model parameters of the intent recognition node are adjusted based on the intent recognition training data. The intent recognition node implements keyword classification, text cleaning, context judgment, semantic understanding, intent recognition, and label association on the data output by the problem rewriting node according to the model parameters. The intent recognition node associates the data output by the problem rewriting node with one or more large language models in the branch nodes through label association according to the recognized user intent and expected result.

[0036] The intent recognition node performs intent feature extraction, intent feature classification, association of intent features and standardized label information of the branch nodes, confidence checking, and fallback options on the data output by the problem rewriting node when processing data.

[0037] The confidence checking includes: global confidence checking and a plurality of branch confidence checkings corresponding to the large language models in the branch nodes.

[0038] The global confidence checking defines the minimum confidence required to trigger the branch confidence checking. The branch confidence checking defines the minimum confidence of the intent feature classification, the association of the intent features and the standardized label information of the branch nodes, and the fallback options.

[0039] When the global confidence checking determines that the intent recognition node fails to recognize the intent or the prediction is ambiguous, the fallback options are used to request the user to supplement information or ask the user clarifying questions.

[0040] The result output by the intent recognition node is fed back to the training data of the problem rewriting node: standardized and non-standardized questions and their question formats in the baby and parenting scenarios.

[0041] The confidence checking and the fallback options belong to the safety mechanism in the baby robot system, which are used to reduce and avoid inappropriate and even harmful suggestions and answers that may be caused by misinterpretation of intent, and to prevent potential false responses.

[0042] Specifically, the mutual feedback of the training data of the problem rewriting node and the intent recognition node, combined with the externally updated training data, realizes real-time dynamic improvement of the system.

[0043] Specifically, the functions provided by the professional medical question and answer model include: parenting knowledge suggestion, emotional support and hospital recommendation; the functions provided by the daily casual chat model include: emotional accompaniment and parenting suggestion; the functions provided by the networking search model include: real-time networking precise search function; the functions provided by the poetry picture book model include: search or create bedtime stories, fables, picture book functions; the functions provided by the music playing model include: music matching and playing function for the current scene. The output data of the intent recognition node is distributed to the corresponding large language model through the branch node, and then more professional and accurate answer responses are provided. The large language model included in the branch node can be an existing large language model.

[0044] It should be noted that the training steps of the large language model used in the present application include: training data collection, training data preprocessing, feature extraction, model training, model output, output result response generation / process logic generation, output result execution index evaluation and reliability evaluation, output result feedback, parameter adjustment, model optimization.

[0045] In summary, the present application directly solves the "noise" and "ambiguity" problems that may exist in the original user query through the above-mentioned problem rewriting node, clarifies the original user input through standardization and rewriting, thereby significantly improving the accuracy and response efficiency of the subsequent intent recognition node, and ultimately improving the overall user experience; the standardized user problem output by the problem rewriting node is further identified, classified and associated by the intent recognition node, realizing single-problem multi-intent shunt processing, and combining the intent with the downstream response logic call (large language model of the branch node), to ensure the efficiency and controllability of the entire system intent recognition to task response process. Through targeted training of the problem rewriting node and the intent recognition node, and coordinated association of the training data, the accuracy of the data output by the model in the maternal and infant care scene is improved and the "thinking" time of the model is reduced. Through the branch node, the original user problem is distributed to different function large language models to solve the user problem.

[0046] By modularizing the large language model as a node, it can flexibly integrate and update knowledge in different fields without interfering with the core intent recognition process, making the system highly scalable and maintainable, meeting the product needs of a long-term evolution and adaptation to new knowledge such as a baby robot, and through having independent professional models, the system can be incrementally and efficiently updated, reducing maintenance costs and enhancing its commercial viability.

[0047] The above-described embodiments are merely one or more embodiments of the present application, which are described in more detail and in more specifically, but should not be understood as limiting the patent of the present application. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A question-and-answer intent recognition and classification method for infant care robots, characterized in that, include: Problem rewriting node, intent recognition node, and branch node; The question rewriting node is a large language model. The input data of the question rewriting node is the question input by the user. The training data of the question rewriting node includes: original question segments input by users related to maternal and infant care, normative and non-normative questions and their wording in maternal and infant care scenarios, and specific types of vocabulary in maternal and infant care scenarios. The model parameters of the question rewriting node are adjusted based on the training data. The question rewriting node processes the input data and outputs data according to its model parameters. The intent recognition node is a large language model. The intent recognition node obtains the data output by the question rewriting node and identifies the user's intent and expected result. The branch node provides multiple large language models for maternal and infant and childcare scenarios, including: professional medical question-and-answer model, daily chat model, online search model, poetry and picture book model and music playback model. The branch node selects the corresponding large language model to perform data processing based on the output data of the intent recognition node.

2. The question-and-answer intent recognition and classification method for infant care robots according to claim 1, characterized in that, The processing steps performed by the problem rewriting node on the input data and the training data include: problem semantic parsing and understanding, problem rewriting, and professional expression.

3. The question-and-answer intent recognition and classification method for infant care robots according to claim 2, characterized in that, In the problem rewriting step: the original fuzzy problem in the input data is completed and reconstructed, and the meaning of the original fuzzy problem and the completion and reconstruction are determined according to the model parameters to see if they are within the consistency range and whether additional information is introduced.

4. The question-and-answer intent recognition and classification method for infant care robots according to claim 2, characterized in that, In the professional expression step: the data output from the rewritten question is converted into a standard question format based on the model parameters.

5. The question-and-answer intent recognition and classification method for infant care robots according to claim 1, characterized in that, The intent recognition training data for the intent recognition node includes: standardized user query and question information in maternal and infant and childcare scenarios, a structured prompt word database, standardized label information of branch nodes, intent annotation information of user query and question information, and data output by the question rewriting node. The standardized label information of the branch node corresponds to the standardized labels of multiple large language models contained in the branch node. The model parameters of the intent recognition node are adjusted based on the intent recognition training data. The intent recognition node performs keyword classification, text cleaning, context judgment, semantic understanding, intent recognition, and tag association on the data output by the question rewriting node according to its model parameters. The intent recognition node associates the data output by the question rewriting node with one or more large language models in the branch node through tags based on the identified user intent and expected results.

6. The question-and-answer intent recognition and classification method for infant care robots according to claim 5, characterized in that, When processing data, the intent recognition node performs intent feature extraction, intent feature classification, association of intent features with the standardized label information of the branch node, confidence verification, and rollback options on the data output by the problem rewriting node.

7. The question-and-answer intent recognition and classification method for infant care robots according to claim 6, characterized in that, The confidence verification includes: global confidence verification and branch confidence verification of multiple large language models corresponding to the branch nodes; The global confidence check defines the minimum confidence level required to trigger the branch confidence check, and the branch confidence check defines the intent feature classification, the association between the intent feature and the standardized label information of the branch node, and the minimum confidence level of the fallback option.

8. The question-and-answer intent recognition and classification method for infant care robots according to claim 7, characterized in that, When the global confidence check determines that the current intent recognition node has failed to recognize the intent or the prediction is unclear, the user is requested to provide supplementary information or asked clarification questions through the rollback option.

9. The question-and-answer intent recognition and classification method for infant care robots according to claim 1, characterized in that, The professional medical Q&A model provides functions including: parenting knowledge advice, emotional support, and hospital recommendations; the casual chat model provides functions including: emotional companionship and parent-child advice; the online search model provides functions including: real-time online accurate retrieval; the poetry and picture book model provides functions including: searching for or creating bedtime stories, fables, and picture books; and the music playback model provides functions including: music matching and playback based on the current scene.

10. The question-and-answer intent recognition and classification method for infant care robots according to claim 1, characterized in that, The branch node also includes: an enterprise retrieval enhancement generation model, which is used to introduce enterprise design, background information, and brand information.

Images