Interaction method and device based on artificial intelligence, equipment, medium and product

Abstract

The invention provides an interaction method and device based on artificial intelligence, equipment, a storage medium, a program product and an intelligent agent, and relates to the field of artificial intelligence, in particular to the technical field of deep learning, large models and generative models. According to the specific implementation scheme, input content input by a target object in an interactive interface is received; obtaining a reference knowledge set related to the input content based on a first reference knowledge set determined from a special knowledge base by using a keyword index mode and a second reference knowledge set determined from a general knowledge base by using a vector index mode; inputting the input content and the reference knowledge set into the large model to obtain feedback content for feeding back the input content; and displaying the feedback content on the interactive interface.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence, in particular to the technical field of deep learning, large model and generative model, and specifically to an interaction method and device based on artificial intelligence, electronic equipment, storage medium, program product and agent. BACKGROUND

[0002] With the wide application of large language models (LLM), also known as large models, in the fields of question answering and content generation, the illusion problem existing in the output content of the large language models is increasingly prominent, and gradually becomes a major challenge to the reliability of the large language models. SUMMARY

[0003] The present disclosure provides an interaction method and device based on artificial intelligence, electronic equipment, storage medium, program product and agent.

[0004] According to an aspect of the present disclosure, an interaction method based on artificial intelligence is provided, which includes: receiving input content input by a target object in an interaction interface; obtaining a reference knowledge set related to the input content based on a first reference knowledge set determined from a special knowledge base by using a keyword indexing method and a second reference knowledge set determined from a general knowledge base by using a vector indexing method; inputting the input content and the reference knowledge set into a large model to obtain feedback content for feeding back the input content; and displaying the feedback content in the interaction interface; wherein the knowledge stored in the special knowledge base and the knowledge stored in the general knowledge base are from different sources.

[0005] According to another aspect of the present disclosure, an interaction device based on artificial intelligence is provided, which includes: a receiving module configured to receive input content input by a target object in an interaction interface; a retrieval enhancement module configured to obtain a reference knowledge set related to the input content based on a first reference knowledge set determined from a special knowledge base by using a keyword indexing method and a second reference knowledge set determined from a general knowledge base by using a vector indexing method; a model processing module configured to input the input content and the reference knowledge set into a large model to obtain feedback content for feeding back the input content; and a display module configured to display the feedback content in the interaction interface; wherein the knowledge stored in the special knowledge base and the knowledge stored in the general knowledge base are from different sources.

[0006] According to another aspect of the present disclosure, an artificial intelligence agent is provided, comprising: an input module configured to receive input information; a processing module configured to determine a target task based on the input information received by the input module, determine a large model based on the target task, execute the method described above by calling the large model, and obtain output information; and an output module configured to output the output information obtained by the processing module.

[0007] According to another aspect of the present disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described above.

[0008] According to another aspect of the present disclosure, a non-transitory computer readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to perform the method described above.

[0009] According to another aspect of the present disclosure, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the method described above.

[0010] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the present disclosure, nor to limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0011] The accompanying drawings are used to better understand the present scheme, and do not limit the present disclosure. Among them:

[0012] Figure 1 An exemplary system architecture to which the artificial intelligence-based interaction method and device according to embodiments of the present disclosure can be applied is schematically shown;

[0013] Figure 2 A flowchart of the artificial intelligence-based interaction method according to embodiments of the present disclosure is schematically shown;

[0014] Figure 3 A construction method schematic diagram of the index content set in the specific embodiments of the present disclosure is schematically shown;

[0015] Figure 4 A construction method schematic diagram of the index content set in the specific embodiments of the present disclosure is schematically shown;

[0016] Figure 5 A construction method schematic diagram of the reference knowledge set in the specific embodiments of the present disclosure is schematically shown;

[0017] Figure 6 A schematic diagram of a method for constructing a general knowledge base in an embodiment of the present disclosure is shown schematically.

[0018] Figure 7 A schematic diagram of a method for constructing a special knowledge base in an embodiment of the present disclosure is shown schematically.

[0019] Figure 8 A schematic diagram of a multi-level decision-making method for identifying results of attribution in an embodiment of the present disclosure is shown schematically.

[0020] Figure 9 A block diagram of an interactive device based on artificial intelligence according to an embodiment of the present disclosure is shown schematically.

[0021] Figure 10 A structural block diagram of an agent of artificial intelligence according to an embodiment of the present disclosure is shown schematically.

[0022] Figure 11 A schematic block diagram of an example electronic device 1100 that can be used to implement embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0023] Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, which include various details of embodiments of the present disclosure to assist in understanding, which should be considered in their context only. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Also, descriptions of well-known functions and structures are omitted in the following description for the sake of clarity and brevity.

[0024] With the wide application of large language models (LLM), also known as large models, in the fields of question answering and content generation, the hallucination problem existing in the output content of the large models has become increasingly prominent, and has gradually become a core challenge affecting their reliability.

[0025] The hallucination problem can be understood as the large model generating content that appears reasonable but is actually incorrect or lacks factual basis. The root cause of this problem lies in the fact that the large model is essentially a probabilistic text generation model, lacking the logical reasoning ability based on real knowledge from a real knowledge base. Therefore, when a user uses a large model to process time-sensitive and professional problems such as score lines and registration times in an educational scenario, the hallucination problem of the large model may output incorrect guidance information.

[0026] For the hallucination problem of large models, current methods mainly include post-processing verification, model fine-tuning, traditional retrieval augmented generation (RAG), etc. Among them, the post-processing verification method usually adds a verification link after the generation model output, and corrects errors through fact checking, confidence evaluation and consistency checking. This method increases the response delay and computing cost, and is difficult to cover complex error types. The model fine-tuning method includes domain fine-tuning, reinforcement learning and adversarial training, etc. Although it can locally optimize the generation quality, it relies on a large amount of labeled data and cannot break through the knowledge boundary of the model. The traditional retrieval augmented generation method usually uses a single retrieval method such as keyword retrieval or vector retrieval. This method cannot balance recall rate and precision, resulting in missing of key information or introduction of noise. In addition, the traditional retrieval augmented generation method directly inputs the retrieval results and the generation link as context into the large model through simple splicing, lacking deep analysis and fusion of retrieval knowledge. These problems seriously limit the reliable application of large language models in high-risk fields.

[0027] Therefore, the embodiments of the present disclosure provide an artificial intelligence-based interaction method. By constructing a multi-source heterogeneous knowledge base architecture, various retrieval methods such as vector-based semantic retrieval and keyword-based exact match retrieval are integrated, so that the retrieval results cover both unstructured general content and professional structured content, thereby realizing the deep fusion of general knowledge and domain-specific knowledge. Compared with traditional single retrieval methods, the comprehensive coverage of knowledge is effectively improved, the hallucination problem of large models in professional fields is solved, and reliable knowledge services with domain specificity, factual accuracy and logical coherence are provided for learners.

[0028] Specifically, the embodiments of the present disclosure provide an artificial intelligence-based interaction method, which includes: receiving input content input by a target object in an interaction interface; obtaining a reference knowledge set related to the input content based on a first reference knowledge set determined from a special knowledge base by using a keyword indexing method and a second reference knowledge set determined from a general knowledge base by using a vector indexing method; inputting the input content and the reference knowledge set into a large model to obtain feedback content for feeding back the input content; and displaying the feedback content in the interaction interface.

[0029] Figure 1 An exemplary system architecture to which the artificial intelligence-based interaction method and device according to the embodiments of the present disclosure can be applied is schematically shown.

[0030] It should be noted that, Figure 1The examples shown are merely examples of system architectures to which the embodiments of the present disclosure may be applied, to help those skilled in the art understand the technical content of the present disclosure, but do not imply that the embodiments of the present disclosure may not be applied to other devices, systems, environments, or scenarios. For example, in another embodiment, an exemplary system architecture to which the artificial intelligence-based interaction method and apparatus may be applied may include a terminal device, but the terminal device may implement the artificial intelligence-based interaction method and apparatus provided by the embodiments of the present disclosure without interacting with a server.

[0031] like Figure 1 As shown, the system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is used as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0032] A user may use a first terminal device 101, a second terminal device 102, or a third terminal device 103 to interact with a server 105 via a network 104 to receive or send messages, etc. The first terminal device 101, the second terminal device 102, or the third terminal device 103 may be installed with various communication client applications, such as knowledge reading applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (for example only).

[0033] The first terminal device 101 , the second terminal device 102 , and the third terminal device 103 may be various electronic devices having display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0034] The server 105 may be a server that provides various services, such as a background management server (for example only) that supports content browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103. The background management server may analyze and process received data such as user requests, and feed back processing results (e.g., web pages, information, or data obtained or generated based on user requests) to the terminal devices.

[0035] It should be noted that the content processing method provided in the embodiments of the present disclosure can generally be executed by the first terminal device 101, the second terminal device 102, or the third terminal device 103. Accordingly, the artificial intelligence-based interaction device provided in the embodiments of the present disclosure can also be set in the first terminal device 101, the second terminal device 102, or the third terminal device 103.

[0036] Alternatively, the AI-based interaction method provided by the embodiments of the present disclosure can also be executed by the server 105 in general. Accordingly, the AI-based interaction apparatus provided by the embodiments of the present disclosure can be arranged in the server 105 in general. The AI-based interaction method provided by the embodiments of the present disclosure can also be executed by a server or a server cluster different from the server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105. Accordingly, the content processing apparatus provided by the embodiments of the present disclosure can also be arranged in a server or a server cluster different from the server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105.

[0037] It should be understood that Figure 1 The number of the first terminal device, the second terminal device, the third terminal device, the network and the server in the above description is only illustrative. Any number of terminal devices, networks and servers can be provided according to the implementation needs.

[0038] In the technical solutions of the present disclosure, the collection, storage, use, processing, transmission, provision, disclosure and application of the user personal information involved all comply with the relevant legal regulations, necessary security measures are taken, and the public order and good customs are not violated.

[0039] In the technical solutions of the present disclosure, the authorization or consent of the user is obtained before the user personal information is acquired or collected.

[0040] It should be noted that the serial numbers of the operations in the following method are only used to represent the operations for description, and should not be regarded as representing the execution sequence of the operations. Unless explicitly stated, the method does not need to be executed in the order shown.

[0041] Figure 2 A flowchart of the AI-based interaction method according to the embodiments of the present disclosure is schematically shown.

[0042] As Figure 2 shown, the method includes operations S210-S240.

[0043] In operation S210, the input content input by the target object in the interaction interface is received.

[0044] In operation S220, the reference knowledge set related to the input content is obtained based on the first reference knowledge set determined from the special knowledge base by using the keyword indexing method, and the second reference knowledge set determined from the general knowledge base by using the vector indexing method.

[0045] In operation S230, the input content and the reference knowledge set are input into the large model to obtain feedback content for feeding back the input content.

[0046] In operation S240, the feedback content is displayed on the interactive interface.

[0047] In this embodiment, the target object can represent an operating user. The interactive interface can represent a terminal interface supporting user input, such as an APP, a webpage, a chat robot, and the like. The input form of the interactive interface can include text, voice, pictures, and the like.

[0048] In this embodiment, the input content can represent the query requirement of the user. For example, in an education scenario, the input content can be "how to improve the speed of solving math problems", "changes in the examination syllabus", and the like.

[0049] According to embodiments of the present disclosure, the knowledge stored in the special knowledge base and the knowledge stored in the general knowledge base have different sources.

[0050] In this embodiment, the knowledge in the special knowledge base can come from authoritative or official institutions in the field. For example, in the field of education, the knowledge in the special knowledge base can include "official enrollment brochure of colleges and universities", "professional examination syllabus", and the like. In the field of finance, the knowledge in the special knowledge base can include "policy documents", "product terms", and the like.

[0051] In this embodiment, since the content structure published by the authoritative or official institutions is relatively complete, the knowledge terms have standard uniformity, the update frequency is relatively low, but the content accuracy and universality are relatively high. Therefore, for the special knowledge base, the reference knowledge or standardized terms in a specific field, such as "examination subject code 001", and the like, can be simply and effectively obtained from the special knowledge base based on the keyword indexing method through the inverted index, so as to quickly locate the accurate knowledge fragments.

[0052] In specific embodiments, the first reference knowledge set can represent the knowledge highly matched with the literal meaning or explicit field requirement of the input content extracted from the special knowledge base through the keyword indexing method, such as documents, entries, or structured data directly containing the user query keywords.

[0053] In this embodiment, the knowledge in the general knowledge base can be unstructured or generalized data that is publicly available. The sources of the knowledge in the general knowledge base can include web articles, forum Q&As, popular science video texts, and the like.

[0054] In this embodiment, since the knowledge sources in the general knowledge base are relatively universal, the content structure is diversified and the universality is low, therefore, for the general knowledge base, the vector indexing mode can be used to map the user input and the general knowledge to the same vector space, obtain the generalized knowledge similar in semantics to the user input, and take it as the second reference knowledge set, so as to improve the retrieval accuracy. For example, the user input "math score" can obtain related content such as "problem solving skills" and "mistake analysis", and the user input "thank you" can obtain related content such as "polite language" and "emotional expression" through vector indexing.

[0055] In specific embodiments, the second reference data set can represent knowledge related to or implicitly associated with the input content but not literally matched, such as generalized background knowledge or experience suggestions, extracted from the general knowledge base through the vector indexing mode.

[0056] According to embodiments of the present disclosure, the reference knowledge set related to the input content can represent a set of knowledge associated with the user input content obtained from the special knowledge base and the general knowledge base through keyword indexing and vector indexing respectively.

[0057] In this embodiment, the reference knowledge set related to the input content and the input content are input into the large model at the same time, the understanding ability of the large model is used to integrate the knowledge, and the feedback content corresponding to the input content is generated based on the reference knowledge set, and the feedback content is displayed to the user on the interaction interface. The feedback content can be in the form of text, voice, picture, etc. matching the input interface.

[0058] Based on this, the embodiments of the present disclosure integrate various retrieval methods such as vector-based semantic retrieval and keyword-based accurate matching retrieval by constructing a multi-source heterogeneous knowledge base architecture, so that the retrieval results cover both unstructured generalized content and professional structured content, thereby realizing the deep integration of general knowledge and domain-specific knowledge. Compared with traditional single retrieval method, the overall knowledge coverage is effectively improved, the illusion problem of large model in professional field application is solved, and reliable knowledge service with domain specificity, factual accuracy and logical coherence is provided for learners.

[0059] According to embodiments of the present disclosure, the first reference knowledge set is determined from the special knowledge base using the keyword indexing method, including: determining an initial first reference knowledge set from the special knowledge base based on the similarity between the index content set and the keyword label set, and removing the same knowledge modality from the reference knowledge in the initial first reference knowledge set to obtain the first reference knowledge set.

[0060] In this embodiment, the keyword tag set can represent a set of special knowledge tags in a special knowledge base. In the education scenario, for example, the keyword tags can include topics, types, sources, knowledge question types, etc.

[0061] In this embodiment, the index content set can represent a set of structured queries generated after the input content of the user is rewritten for intent, for accurate matching of keyword tags in the special knowledge base. In specific embodiments, the index content set can include domain terms, constraint conditions, etc.

[0062] In this embodiment, rewriting the input content of the user for intent can include analyzing the intent of the input content of the user and rewriting it into a set of query statements that are more structured and can more accurately express the intent.

[0063] For example, the input content of the user is “How to review math”, and the intent analysis can obtain that the intent is “get math review methods”. The index content set after rewriting the input content for intent can include: {“math”, “review method”, “knowledge point analysis”}.

[0064] According to embodiments of the present disclosure, the similarity calculation method between the index content set and the keyword tag set can be determined based on text matching similarity calculation such as term frequency-inverse document frequency matching (TF-IDF) and / or semantic matching similarity calculation.

[0065] In specific embodiments, the term frequency-inverse document frequency matching method is to calculate the frequency of each word in the index content set in the keyword tag set of the special knowledge base, and combine the document discrimination degree of the word to generate a keyword weight vector such as a TF-IDF weight. For example, in the index content “math question type analysis”, “question type analysis” is less likely to appear in the keyword tag set of the special knowledge base, so the TF-IDF weight is higher, to preferentially match the knowledge containing the label. Then, the index content set and the keyword tag set are converted into word vectors, for example, each word corresponds to a dimension, and the value is the TF-IDF weight, and the similarity between the index content set and the keyword tag set is calculated by the cosine similarity.

[0066] In another specific embodiment, the semantic matching method is to convert the index content and the keyword tag into a semantic vector by using a model, and calculate the similarity between the index content set and the keyword tag set by the cosine similarity. For example, the index content “how to memorize words” is different from the literal keyword of the label “vocabulary memory strategy”, but the model can capture the semantic association between “memorize words” and “memory strategy”, and the similarity is 0.85, so the related knowledge is matched.

[0067] In yet another specific embodiment, the similarity scores calculated based on the term frequency-inverse document frequency matching method and the similarity scores calculated based on the semantic matching method can be weighted and summed to obtain a fused similarity of the two methods.

[0068] According to an embodiment of the present disclosure, by the similarity calculation described above, all the knowledge related to the semantic and / or literal of the index content of the tags is obtained from the special library knowledge base to form an initial first reference knowledge set. For example, if the input content of the user is “high school physics score”, the first reference knowledge set can include an official table with the tag “high school- physics- score distribution” (similarity 1.0) and / or an analysis text with the tag “high school physics- question type- score proportion” (similarity 0.8).

[0069] According to an embodiment of the present disclosure, the knowledge modal can represent the form of knowledge, and the knowledge modal can include a text modal (such as a text paragraph, a policy interpretation), a structured modal (such as a table, a formula, a list), etc.

[0070] Before deduplicating the reference knowledge in the initial first reference knowledge set according to the knowledge modal, the modal type can be labeled for each reference knowledge, and the reference knowledge in the initial first reference knowledge set is deduplicated based on the modal type to obtain the first reference knowledge set. For example, for the reference knowledge of the same modal type, only the most relevant reference knowledge can be retained to avoid duplication and redundancy, and to ensure that the knowledge input by the large model is concise and diverse.

[0071] Based on this, the embodiments of the present disclosure utilize the keyword indexing method to determine the first reference knowledge set from the special knowledge base based on the similarity between the index content set and the keyword tag set, and through the intent rewriting and similarity matching, the index content is closer to the tags of the special knowledge base, thereby improving the accuracy of the retrieval and avoiding missing the retrieval results due to the inconsistency between the user expression and the knowledge base terminology. In addition, the knowledge entries with the same content but different expressions or repetitions are deduplicated to filter the accurate granularity knowledge, avoid knowledge redundancy, ensure the consistency of the generated results, and reduce the confusion caused by repeated information.

[0072] According to an embodiment of the present disclosure, the human-computer interaction method based on artificial intelligence further includes: rewriting the input content based on the context content of the input content and the intent information of the input content to obtain an initial index content set; and filtering the initial index content set based on the relevance between the initial index content in the initial index content set and the input content to obtain the index content set.

[0073] In this embodiment, the intent information of the input content can be identified and extracted by a natural language processing model such as a long short-term memory network, a coding decoder, etc., such as a knowledge point, a difficulty, a frequency of investigation, etc.

[0074] In this embodiment, a contextual understanding model can be used to analyze the semantic relevance between the user's current input and historical conversation records, and rewrite the input based on the input's intent to obtain initial index content. For example, if the user's current input is "recommended real questions" and the user's input in the historical conversation record is "how to prepare for English reading comprehension exams," the contextual model can use the above context to rewrite the current input to "recommended real questions for English reading comprehension" rather than the literal translation of "recommended real questions."

[0075] In this embodiment, the context understanding model can be combined with the sentiment analysis module to rewrite the input content based on the user's emotional characteristics to obtain the initial index content. For example, based on the punctuation mark "!", the user's current emotional characteristics can be judged as anxious, and the rewriting can include constrained words such as "latest" in the content.

[0076] In this embodiment, a domain-adaptive fine-tuning strategy can also be introduced to fine-tune the large model for specific domains, such as education, so that it can identify potential drift phenomena of user intent to obtain initial index content. For example, in an education scenario, if the user inputs "translate", it may indicate a drift from a general intent such as language translation to a "question type translation" intent, such as translating an English question into Chinese. If the user's current input is "thank you" and the previous input is "please translate", it may indicate a drift from a general intent such as simply expressing gratitude to an intent of "needing further explanation".

[0077] In this embodiment, the initial index content set is filtered and screened for intent-based consistency by calculating the semantic similarity between the initial index content and the input content. This eliminates semantically deviating rewrites, ensuring that the returned knowledge fragments are highly consistent with the user's actual needs, resulting in an indexed content set. For example, if the input content is "thank you," in the translation context, only rewrites related to the "translation" intent, such as "English translation of thank you," are retained, while irrelevant rewrites such as "ways to express gratitude" are eliminated.

[0078] Figure 3 The following schematically illustrates a method for constructing an index content set in a specific embodiment of the present disclosure.

[0079] like Figure 3 As shown, based on the previous content of the input content 310, the input content 310 is mined for intent to obtain intent information 320 of the input content 310, the input content 310 is rewritten based on the intent information 320 to obtain an initial index content set 330, and based on the correlation between the initial index content in the initial index content set 330 and the input content 310, the initial index content set 330 is filtered to obtain an index content set 340.

[0080] Based on this, the embodiments of the present disclosure rewrite the input content based on the context of the input content and the intention information of the input content, to improve the richness of the index content set and ensure that the intention of each index content is consistent with the real intention of the user after context analysis Figure 1 , solve the "polysemy" and "intention drift" problems through context and emotion analysis, so that the generated index content is more in line with the potential needs of the user. In addition, based on the relevance between the initial index content in the initial index content set and the input content, the initial index content set is filtered to retain the rewritten content with high relevance, thereby improving the retrieval efficiency and accuracy.

[0081] According to an embodiment of the present disclosure, the second reference knowledge set is determined from the general knowledge base by using a vector index method, including: determining an initial second reference knowledge set from the general knowledge base based on the similarity between the index vector set and the general knowledge vector, wherein the index vector set is obtained by rewriting the input content based on the intention information of the input content and vector conversion, and the general knowledge vector is obtained by vector conversion on the general knowledge in the general knowledge base; and performing the same knowledge modality deduplication on the reference knowledge in the initial second reference knowledge set to obtain the second reference knowledge set.

[0082] In this embodiment, the index vector set can be used to represent a set of numerical vectors obtained by converting the input content of the user after intention rewriting in the vector space, for semantic matching with the knowledge in the knowledge base.

[0083] In specific embodiments, the operation of vector conversion may, for example, include text vectorization using a domain-adapted pre-trained model. Such a model, which has been pre-trained on data in a specific domain, can better understand the semantic information in that domain. The rewritten query statement is input into the model, which converts it into a vector representation. These vectors can constitute the index vector set.

[0084] In this embodiment, the general knowledge vector can be used to represent the vector representation of the unstructured knowledge in the general knowledge base. In specific embodiments, a domain-adapted pre-trained model can also be used to convert each document segment in the general knowledge base into a vector, so that the knowledge in the general knowledge base is stored in the same semantic space in the form of a vector.

[0085] In this embodiment, the index vector set is mapped into the same semantic space as the general knowledge vector, and then the cosine similarity between the index vector set and the general knowledge vector is calculated. The closer the cosine similarity is to 1, the more similar the directions of the two vectors are, and the more similar the corresponding text semantics are.

[0086] In this embodiment, a dynamic threshold strategy can also be included to extract the candidate knowledge with the highest semantic relevance according to the weight of different types of knowledge. The dynamic threshold strategy can represent that different similarity thresholds are set for different types of knowledge. For example, for some key and authoritative knowledge, a higher threshold can be set, and only the knowledge with a similarity higher than the threshold can be extracted. For some auxiliary and less relevant knowledge, the threshold can be appropriately reduced. In this way, the knowledge with high semantic relevance to the user query can be filtered from the general knowledge base to form the initial second reference knowledge set.

[0087] In this embodiment, the initial second reference knowledge set is de-duplicated with the same knowledge modalities to obtain the second reference knowledge set. Specifically, in the de-duplication process, for knowledge of the same modality type, only the most relevant knowledge can be retained, thereby avoiding redundancy of knowledge in the second reference knowledge set.

[0088] Based on this, the embodiments of the present disclosure utilize a vector index method to determine the second reference knowledge set from the general knowledge base based on the similarity between the index vector set and the general knowledge vector. Even if the terms in the user query and the knowledge base are different, by rewriting the input content, clarifying the intention, and then performing vector conversion, the relevant general knowledge can be more accurately matched, and the potential relevant information between the index vector set and the general knowledge vector can be found to improve the semantic generalization capability. In addition, the knowledge entries with repeated or different expressions but the same content are de-duplicated to avoid knowledge redundancy, ensure the consistency of the generated results, and reduce the confusion caused by repeated information.

[0089] According to the embodiments of the present disclosure, the artificial intelligence-based interaction method further includes: rewriting the input content based on the context content of the input content and the intention information of the input content to obtain an initial index content set; performing vector encoding on the initial index content in the initial index content set to obtain an initial index vector set; and adding identification information representing a reference degree to the initial index vector in the initial index vector set to obtain the index vector set.

[0090] In this embodiment, a context understanding model can be used to analyze the semantic relevance of the current input content of the user and the historical dialogue record, and the input content can be rewritten based on the intention information of the input content to obtain the initial index content.

[0091] In a specific embodiment, for the scenario of semantic ambiguity and information missing of the input content of the user, the input content of the user can also be rewritten by a multi-hypothesis document generation and semantic expansion method to realize the alignment of the input content of the user and the semantic space of the knowledge base.

[0092] Specifically, the context understanding model can be used to generate a plurality of initial index contents with similar semantics after rewriting according to the historical dialogue record of the user and the scene label such as "high school English" and "postgraduate examination" in the current input content or historical dialogue content, and form an initial index content set. Each index content in the initial index content set realizes semantic enrichment by adding domain terminology, scene label, constraint condition and the like on the premise of retaining the core intent of the input content, so as to solve the problems of implied ambiguity and information loss of the input content.

[0093] For example, the current input content of the user is "recommend real questions", and the previous dialogue is "postgraduate examination English review", and the initial index content set can be: ["postgraduate examination English real question recommendation", "postgraduate examination English reading real question analysis", "postgraduate examination English real question type distribution"].

[0094] In this embodiment, the initial index content can also be obtained by rewriting the input content by the context understanding model combined with the sentiment analysis module through the emotional characteristics of the user.

[0095] In this embodiment, a fine-tuning strategy of domain adaptation can also be introduced. The large model is fine-tuned for a specific domain such as the education domain, so that it can identify the potential drift phenomenon of the user's intent, to obtain the initial index content.

[0096] In this embodiment, the initial index content can be vector encoded to convert the index content set of natural language description obtained after rewriting into an initial index vector set that can be provided to a computer for efficient processing.

[0097] Specifically, the initial index content in the initial index content set can be vector encoded by using a double encoder architecture. The double encoder architecture can include a main encoder and an auxiliary encoder. The main encoder can use a domain-adapted pre-training model to map each initial index content to a dense vector to capture the context semantics. For example, the semantics of "real questions" is equivalent to "test questions". The auxiliary encoder can be used to calculate the similarity or semantic deviation degree of the generated initial index vector and the input content vector. If the similarity or semantic deviation degree exceeds a threshold, it can be determined that the initial index vector is noise and filtered. Through the double encoding architecture, effective initial index vectors can be screened to form an initial index vector set, ensuring the purity of the vector space and avoiding irrelevant index content interference with the retrieval result.

[0098] In this embodiment, the identification information can be represented as a weight representing the degree of reference. The identification information can be used to quantify the semantic association degree of each initial index vector in the initial index vector set and the original intent of the input content. The identification information is added to each initial index vector to obtain an index vector set.

[0099] In specific embodiments, the weight can be configured as a dynamic weight. Among them, a weighting fusion algorithm based on an attention mechanism can be used to assign a dynamic weight to each initial index vector to generate a fusion vector containing both the original query semantics and the potential answer mode features. Specifically, the dynamic weight is calculated as follows:

[0100] (1);

[0101] wherein, denotes the dynamic weight of the i-th initial index vector, denotes the i-th initial index vector, denotes the query vector corresponding to the input content, denotes a weight coefficient for adjusting the weight distribution, sim is a cosine similarity, and softmax is an activation function.

[0102] In this specific embodiment, based on each initial index vector and the corresponding dynamic weight, a fusion vector can be generated to form an index vector set. Specifically, the expression of the fusion vector is as follows:

[0103] (2);

[0104] wherein, denotes the fusion vector.

[0105] In this embodiment, the identification information representing the reference degree can be used for priority sorting in subsequent retrieval, such as the vector with high reference degree having a higher matching priority in retrieval to ensure that the knowledge related to the core intent is acquired first. The identification information can also be used as an input of the attention mechanism to guide the large model to pay more attention to the knowledge fragments corresponding to the high-correlation index vector when generating answers.

[0106] Figure 4 An illustrative diagram of a method for constructing an index content set in specific embodiments of the present disclosure is shown.

[0107] As shown in Figure 4 , based on the content of the input content 410, the intent of the input content 410 is mined to obtain the intent information 420 of the input content, the input content 410 is rewritten based on the intent information 420 to obtain the initial index content set 430, the initial index content in the initial index content set 430 is vector encoded to obtain the initial index vector set 440, and the initial index vector in the initial index vector set 440 is filtered and identification information is added to obtain the index vector set 450.

[0108] For example, the initial index vector is compared with the vector encoding of the input content in terms of vector similarity, and a vector similarity is obtained. Based on the vector similarity, it is determined whether the initial index vector is retained. If the vector similarity is higher than a threshold, the initial index vector is added with identification information and divided into the index vector set. If not, the initial index vector is removed.

[0109] Based on this, the embodiments of the present disclosure rewrite the input content by analyzing the semantic relevance between the historical dialogue and the input content intention, so as to ensure that the rewritten initial index content set strictly aligns with the real intention of the user rather than the literal meaning. The initial index content is vector encoded to convert the index content set of natural language description obtained after rewriting into an initial index vector set that can be provided to a computer for efficient processing, and the index vector across the fields and with low relevance is automatically filtered based on the degree of semantic deviation, so as to improve the semantic matching accuracy of the index vector. In addition, the initial index vector is weighted and fused by using the identification information representing the reference degree, so as to ensure that the search result covers the core intention and contains high-value expansion information, thereby improving the search accuracy.

[0110] According to the embodiments of the present disclosure, the reference knowledge set is obtained based on the first reference knowledge set and the second reference knowledge set, including: filtering the first reference knowledge set and the second reference knowledge set based on the attribute information of the reference knowledge in the first reference knowledge set and the reference knowledge in the second reference knowledge set, respectively, to obtain the reference knowledge set.

[0111] In this embodiment, the attribute information of the reference knowledge can include the relevance between the reference knowledge and the input content, the professionalism of the reference knowledge, and the timeliness of the reference knowledge.

[0112] The attribute information of the reference knowledge can further include intention information, for example, whether it meets the user's intention based on the scene label and category information.

[0113] The relevance between the reference knowledge and the input content can represent the matching degree of the knowledge content and the user input content in terms of semantics, theme, and intention. The professionalism of the reference knowledge can represent the authority and field professionalism of the reference knowledge source. The timeliness of the reference knowledge can represent the update time or effective time range of the reference knowledge.

[0114] In this embodiment, the attribute information extraction and quantification can be performed on the reference knowledge in the first reference knowledge set and the second reference knowledge set, respectively.

[0115] Specifically, for the special knowledge in the first reference knowledge set, the relevance between the special knowledge and the input content can be determined based on the similarity between the keyword label and the index content set. For example, the input content is "exam score line", and the matching degree with the "xx university admission score" label in the special knowledge base is 1.0. Secondly, the professionalism of the special knowledge can be determined based on the knowledge source annotation. For example, the professional authority score of "Ministry of Education official website" is 100, and the professional score of "educational institution" is 80. In addition, the timeliness of the special knowledge can be determined based on the official notice release update time, effective time, etc. For example, the current time is January 2025, and the timeliness score of the score line released in June 2024 is 0.95.

[0116] For the general knowledge in the second reference knowledge set, the relevance between the general knowledge and the input content can be determined based on the cosine similarity of vector retrieval, for example, if the user input content is "English reading skills", the similarity between the article vector in the general knowledge base is 0.85. Secondly, the professionalism of the general knowledge can be determined based on the knowledge source weight, for example, the weight of the authoritative education website is 0.9, and the weight of the personal blog is 0.6. In addition, the timeliness of the general knowledge can be determined based on the article release time, etc. Specifically, a timestamp can be added to each reference knowledge, and a decay rule can be set according to the field to generate a timeliness score. For example, the timeliness score of the general knowledge in the last 3 months is 0.9, and the timeliness score of the general knowledge more than 1 year is 0.5.

[0117] In this embodiment, based on the attribute information of each reference knowledge in the first reference knowledge set and the attribute information of each reference knowledge in the second reference knowledge set, the first reference knowledge set and the second reference knowledge set are respectively screened to obtain the reference knowledge set.

[0118] Specifically, the invalid knowledge can be filtered and removed based on the relevance quantization value, the professionalism quantization value, and the timeliness quantization value, etc. to screen the valid reference knowledge and constitute the reference knowledge set. For example, reference knowledge with a knowledge matching degree <0.7 in the special knowledge base, or reference knowledge with a vector similarity <0.6 in the general knowledge base can be removed. For another example, invalid and outdated reference knowledge can also be removed.

[0119] In this specific embodiment, the reference knowledge in the first reference knowledge set and the second reference knowledge set can also be attribute priority sorted based on the relevance quantization value, the professionalism quantization value, and the timeliness quantization value, and then screened to preferentially retain high-value knowledge such as authoritative data and recent notifications.

[0120] The ranking method can include rule-based weighted summation, normalized discount cumulative gain, etc. For example, the normalized discount cumulative gain method is used to measure the attribute priority of high-value knowledge in the ranking result. For example, if high-authority and high-relevance knowledge is ranked first, the normalized discount cumulative gain score is 1.0; if it is ranked third due to fixed weight, the normalized discount cumulative gain score decreases to 0.6.

[0121] The ranking method can also include machine learning models, knowledge graphs, sliding window evaluation, etc. to accurately rank attribute priorities to ensure that high-value knowledge is ranked first. For example, attribute information can be converted into input features that the model can understand, and the machine learning model can learn text understanding capabilities in general scenarios. Through instruction templates such as "sort by authority from high to low" and "preferentially display relevant knowledge from the past year", the general sorting capability can be transferred to the education vertical field without relying on labeled data for specific tasks.

[0122] Figure 5 An example of a method for constructing a reference knowledge set in the embodiments of the present disclosure is shown.

[0123] As shown in Figure 5 , the input content 501 is mined for intent based on the context of the input content, and intent information 502 of the input content is obtained. The input content 501 is rewritten based on the intent information 502, and an initial index content set 503 is obtained. The initial index content set 503 is filtered based on the relevance between the initial index content in the initial index content set 503 and the input content 501, and an index content set 504 is obtained. The first reference knowledge set 506 is determined from the special knowledge base 505 based on the similarity between the index content set 504 and the keyword label set. The initial index content in the initial index content set 503 is vector encoded to obtain an initial index vector set 507. The initial index vector set 507 is added with identification information to obtain an index vector set 508. The second reference knowledge set 510 is determined from the general knowledge base 509 based on the similarity between the index vector set 508 and the general knowledge vector. The attribute information of the reference knowledge in the first reference knowledge set 506 and the reference knowledge in the second reference knowledge set 510 is aligned and prioritized for intent, and the first reference knowledge set 506 and the second reference knowledge set 510 are filtered to obtain a reference knowledge set 511. The reference knowledge set 511 is input into the large model LLM 512, and the feedback content 513 is obtained based on the input content 501.

[0124] Based on this, the embodiments of the disclosure establish a multi-dimensional reference knowledge screening system based on the semantic relevance, professionalism, timeliness, and other multi-dimensional attribute characteristics of the reference knowledge, and propose low-relevance and low-value knowledge through fine screening of the reference knowledge, thereby improving the refinement and comprehensiveness of the reference knowledge set.

[0125] According to the embodiments of the disclosure, the artificial intelligence-based interaction method further includes: rewriting the initial knowledge to obtain target knowledge; extracting key information of the target knowledge to obtain keyword labels for representing a topic of the target knowledge; and obtaining a general knowledge base based on the target knowledge, the keyword labels, and attribute information of the target knowledge.

[0126] In this embodiment, the initial knowledge can include text knowledge, video knowledge, and picture knowledge, etc. Rewriting and polishing the initial knowledge can solve the problems of diversity of expression, noise interference, and inconsistent format of the initial knowledge, thereby generating standardized target knowledge.

[0127] In specific embodiments, a multi-stage knowledge distillation method can be used to rewrite and polish the initial knowledge. For example, a large model is used to perform term unification and fluency polishing on the initial knowledge to perform semantic standardization processing; an unstructured document is converted into a standardized format with semantic annotations to perform rich text structured conversion processing; and a text generation model is applied to ensure content fluency and context consistency to achieve knowledge coherence enhancement effect.

[0128] Specifically, for text knowledge, the rewriting method can include eliminating redundant information, rewriting knowledge with standardized terms, and unifying formats, etc. The redundant information, for example, includes advertisements, repeated paragraphs, etc. The standardized term rewriting, for example, can rewrite "brushing questions" to "practice problem training". The unified format, for example, can convert free text into structured paragraphs or list format.

[0129] For video knowledge, the rewriting method can include visual feature extraction, speech recognition, and content summary generation, etc. The visual feature extraction can detect key frames through a deep neural network to extract visual features such as scenes and objects. The speech recognition can convert video speech into subtitles through automatic speech recognition technology, and parse the subtitle text using spatiotemporal alignment technology. The content summary generation can generate a text summary of the video content through a large model to serve as the core expression of the target knowledge.

[0130] In this embodiment, the keyword labels can include topics, knowledge types, sources, difficulty coefficients, and examination frequencies, etc. The topics, for example, are "research English", "linear algebra", etc. The knowledge types, for example, are text, video, table, etc. The sources, for example, are official releases, personal blogs, etc.

[0131] In specific embodiments, for video knowledge, composite keyword labels can also be generated based on visual features such as graphs in key frames, high-frequency words in speech text, such as "math formula derivation video - derivative - problem solving skills".

[0132] In this embodiment, for the extracted keyword labels, the accuracy of the keyword labels can be ensured through domain expert review and / or automatic verification based on the similarity between the keyword labels and the knowledge content.

[0133] In this embodiment, the rewritten target knowledge, the extracted keyword labels, and the attribute information of the target knowledge are fused to construct a general knowledge base.

[0134] Figure 6 The construction method of the general knowledge base in the specific embodiments of the present disclosure is schematically shown.

[0135] As shown in Figure 6 The initial knowledge 610 can include professional content, search interface data, and question banks, videos, etc. The initial content 610 is preprocessed by using a multi-stage knowledge distillation method to rewrite and polish, to obtain target knowledge 620. Such as video key frame information, speech high-frequency word information, web page text information, etc. The key information 630 of the target knowledge is extracted to obtain keyword labels 640. Based on the target knowledge 620, the attribute information of the target knowledge 620, and the keyword labels 640, a general knowledge base 650 is obtained.

[0136] Based on this, the embodiments of the present disclosure rewrite and polish the initial knowledge, so that the knowledge in the general knowledge base is consistent in form and unified in terminology, to improve the accuracy and retrieval efficiency of the general knowledge. In addition, by extracting the key information of the target knowledge to form keyword labels, users can not only search according to basic theme labels when searching for knowledge, but also can make more accurate queries in combination with semantic labels and multi-modal labels. At the same time, through the processing and fusion of multi-modal knowledge, the general knowledge base can cover a wider range of knowledge types to meet the diverse needs of users. For example, in the field of education, users can not only obtain knowledge points explained in text, but also can learn more intuitively by watching related teaching videos, enriching the presentation form of knowledge and improving the learning experience of users.

[0137] According to the embodiments of the present disclosure, the artificial intelligence-based interaction method further includes: verifying a plurality of initial knowledge obtained from different sources to obtain target knowledge; performing synonym expansion on keyword labels of the target knowledge based on a special dictionary to obtain a keyword label set for representing a theme of the target knowledge; and obtaining a special knowledge base based on the keyword label set, the target knowledge, and a set of identification information.

[0138] In this embodiment, the initial knowledge can be directly obtained from authoritative data sources in the field, such as enrollment websites and university website data interfaces in the education field, and central bank data interfaces in the financial field. Structured data such as tables, clauses, and question banks from official websites can also be synchronized in real time.

[0139] In another embodiment, the source of initial knowledge can also be obtained by collecting data from multiple institutions, such as "knowledge point analysis" of different educational institutions.

[0140] In this embodiment, a verification method such as a cross-validation mechanism or a redundancy verification mechanism may be used to verify multiple initial knowledge from different sources to obtain target knowledge.

[0141] Among them, if the initial knowledge is structured data, such as score lines, formulas, etc., it can be retained as target knowledge through field-level comparison. For example, if the field expressions such as numerical values, units, and effective time are consistent,

[0142] Among them, if the initial knowledge is unstructured data, such as knowledge point analysis, the knowledge similarity of multiple sources can be calculated through the text similarity model, and the expressions with greater similarity can be retained as target knowledge. For example, the "derivative definition" descriptions of multiple institutions can be compared, and the versions with consistent or similar expressions can be retained.

[0143] The initial knowledge that has passed the verification is standardized, such as unifying the units and formats, to form the target knowledge.

[0144] In this embodiment, the keyword tags for the target knowledge can be obtained by constructing a domain term mapping table using a domain-specific dictionary. For example, in the education field, the keyword tags for the target knowledge may include {"practice":"exercise training","outline":"exam outline","knowledge point":"knowledge test point"}.

[0145] Furthermore, keyword tags can be dynamically expanded by combining synonyms of domain terms to form a keyword tag set. For example, based on term similarity, "question practice skills" can be expanded to "exercise training skills" and "real test practice methods" and added to the keyword tag set.

[0146] In this embodiment, the identification information in the identification information set is used to indicate the importance of the keyword tags in the keyword tag set. For example, the identification information can be the weight of each keyword tag, which is used for priority sorting during retrieval. The identification information can also be determined by the relevance of each keyword tag to the knowledge topic calculated using the weight distribution model of the attention mechanism.

[0147] In this embodiment, the keyword tag set, target knowledge and identification information set are fused to construct a dedicated knowledge base.

[0148] Figure 7 A method for constructing a special knowledge base in the embodiments of the present disclosure is schematically shown.

[0149] As shown in Figure 7 The initial knowledge 710 can include the score line information, the enrollment directory, the college admission data, the search interface data, and the like. The initial knowledge 710 is verified to obtain the target knowledge 720. The keyword label 730 can be extracted from the target knowledge 720. The keyword label set 740 is obtained by synonym expansion of the keyword label 730. The keyword label set 740, the target knowledge 720, and the identification information set 750 are fused to obtain the special knowledge base 760.

[0150] Based on this, the embodiments of the present disclosure verify the initial knowledge from multiple sources to ensure that the special knowledge in the special knowledge base is consistent with the official data, to exclude the self-compiled error data of the institution, and to improve the knowledge accuracy and authority of the special knowledge base. In addition, the synonym expansion of the keyword label by the special dictionary can solve the problem of user expression diversity, can improve the retrieval efficiency of the user input content, and thus construct the special knowledge base with field specialization and retrieval precision.

[0151] According to the embodiments of the present disclosure, the interaction method based on artificial intelligence further includes: determining an attribution recognition result of the initial knowledge based on a source type of the initial knowledge, wherein the attribution recognition result represents whether the initial knowledge belongs to a general knowledge base or a special knowledge base.

[0152] In this embodiment, the source type of the knowledge can include the authority, the structured degree, the field specialization, and the field relevance of the knowledge source.

[0153] In this embodiment, since the special knowledge base can accommodate the authoritative, structured, and field-specific knowledge such as policy provisions, standard question banks, and official data, and the general knowledge base can accommodate the generalized and multi-modal user-generated knowledge such as experience sharing, popular science articles, and video tutorials, the attribution of the initial knowledge to the knowledge base can be determined by automatically judging the source type of the initial knowledge.

[0154] For example, the attribution recognition result of the initial knowledge can be determined by establishing a white list of authoritative sources and constructing a multi-level decision. The white list of authoritative sources can include official data, authoritative platforms in the field, and professional institutions. The multi-level decision can include a knowledge structured degree decision and a field specialization decision.

[0155] Figure 8 A multi-level decision method for the attribution recognition result of the embodiments of the present disclosure is schematically shown.

[0156] As Figure 8 shown, the multi-level decision method of identifying the result can include operations S801-S805.

[0157] In operation S801, the source type of the initial knowledge is acquired.

[0158] In operation S802, it is determined whether the knowledge source type belongs to the whitelist of authoritative sources. If yes, operation S803 is performed, and if no, operation S804 is performed.

[0159] In operation S803, it is determined whether the knowledge belongs to structured. If no, operation S804 is performed. If yes, operation S805 is performed.

[0160] In operation S804, it is classified into a general knowledge base.

[0161] In operation S805, it is classified into a special knowledge base.

[0162] Based on this, the embodiments of the present disclosure automatically determine whether the initial knowledge should be classified into a special knowledge base or a general knowledge base by analyzing the source characteristics of the initial knowledge, quickly classify the initial knowledge, effectively update the knowledge base, and improve the retrieval efficiency.

[0163] Based on this, the present disclosure proposes an interactive method based on artificial intelligence, which builds a multi-source heterogeneous knowledge base architecture, adopts a multi-modal mixed retrieval mechanism, integrates various retrieval methods such as vector-based semantic retrieval, keyword-based exact match retrieval, and cross-modal feature retrieval, and realizes the deep integration of general knowledge and domain professional knowledge. In addition, at the retrieval optimization level, a double-layer optimization framework is constructed through semantic expansion and dynamic reordering technology, the user's intention is accurately analyzed through query rewriting based on large models, and the reordering algorithm is used to perform semantic alignment and logical consistency verification on the recall results. Compared with traditional single retrieval methods, the illusion problem of large models in professional fields is effectively solved, and reliable knowledge services with domain professionalism, factual accuracy and logical coherence are provided for learners.

[0164] In addition, the method can be widely applied in educational scenarios and meet the following user needs: in the scene where the user queries the examination time, the examination score line and other highly factual requirements, the method can ensure that the generated content is strictly based on authoritative knowledge base, and eliminate illusion errors; through the mixed retrieval path to trace the knowledge source, it can provide the user with a reliable evidence chain; for different needs such as fact query and creative generation, the retrieval breadth and generation freedom can be dynamically balanced, the collaborative optimization of mixed retrieval and knowledge enhancement provides reliability guarantee for the landing of large models, and promotes the large-scale application of large models in educational scenarios.

[0165] Figure 9 A block diagram of an artificial intelligence-based interaction device is shown schematically.

[0166] As shown in Figure 9 The artificial intelligence-based interaction device 900 includes a receiving module 910, a retrieval enhancement module 920, a model processing module 930, and a display module 940.

[0167] The receiving module 910 is configured to receive input content input by a target object on an interaction interface.

[0168] The retrieval enhancement module 920 is configured to obtain a reference knowledge set related to the input content based on a first reference knowledge set determined from a special knowledge base using a keyword indexing method and a second reference knowledge set determined from a general knowledge base using a vector indexing method.

[0169] The model processing module 930 is configured to input the input content and the reference knowledge set into a large model to obtain feedback content for feeding back the input content.

[0170] The display module 940 is configured to display the feedback content on the interaction interface.

[0171] According to an embodiment of the present disclosure, the retrieval enhancement module 920 includes a first determination submodule and a first deduplication submodule.

[0172] The first determination submodule is configured to determine an initial first reference knowledge set from the special knowledge base based on similarity between the index content set and the keyword tag set.

[0173] The first deduplication submodule is configured to perform same-knowledge modality deduplication on the reference knowledge in the initial first reference knowledge set to obtain the first reference knowledge set.

[0174] According to an embodiment of the present disclosure, the retrieval enhancement module 920 further includes a second determination submodule and a second deduplication submodule.

[0175] The second determination submodule is configured to determine an initial second reference knowledge set from the general knowledge base based on similarity between the index vector set and the general knowledge vector.

[0176] The second deduplication submodule is configured to perform same-knowledge modality deduplication on the reference knowledge in the initial second reference knowledge set to obtain the second reference knowledge set.

[0177] According to an embodiment of the present disclosure, the retrieval enhancement module 920 further includes a first screening submodule.

[0178] The first screening submodule is configured to screen the first reference knowledge set and the second reference knowledge set based on attribute information of the reference knowledge in the first reference knowledge set and the reference knowledge in the second reference knowledge set, respectively, to obtain the reference knowledge set.

[0179] According to an embodiment of the present disclosure, the retrieval enhancement module 920 further includes a first rewriting submodule, a second screening submodule.

[0180] The first rewriting submodule is configured to rewrite the input content based on the context of the input content and the intention information of the input content, to obtain an initial index content set.

[0181] The second screening submodule is configured to screen the initial index content set based on a correlation between the initial index content in the initial index content set and the input content, to obtain an index content set.

[0182] According to an embodiment of the present disclosure, the retrieval enhancement module 920 further includes a second rewriting submodule, a vector encoding submodule, and an identification submodule.

[0183] The second rewriting submodule is configured to rewrite the input content based on the context of the input content and the intention information of the input content, to obtain an initial index content set.

[0184] The vector encoding submodule is configured to vector encode the initial index content in the initial index content set, to obtain an initial index vector set.

[0185] The identification submodule is configured to add identification information representing a reference degree to the initial index vector in the initial index vector set, to obtain an index vector set.

[0186] According to an embodiment of the present disclosure, the human-computer interaction device 900 based on artificial intelligence further includes a rewriting module, an extraction module, and a general knowledge base determination module.

[0187] The rewriting module is configured to rewrite the initial knowledge, to obtain target knowledge.

[0188] The extraction module is configured to extract key information of the target knowledge, to obtain a keyword label representing a theme of the target knowledge.

[0189] The general knowledge base determination module is configured to obtain a general knowledge base based on the target knowledge, the keyword label, and attribute information of the target knowledge.

[0190] According to an embodiment of the present disclosure, the human-computer interaction device 900 based on artificial intelligence further includes a verification module, an expansion module, and a special knowledge base determination module.

[0191] The verification module is configured to verify a plurality of initial knowledge obtained from different sources, to obtain target knowledge.

[0192] The expansion module is configured to perform synonym expansion on the keyword label of the target knowledge based on a special dictionary, to obtain a keyword label set for representing a topic of the target knowledge.

[0193] The special knowledge base determination module is configured to determine a special knowledge base based on the keyword label set, the target knowledge, and the set of identification information.

[0194] According to an embodiment of the present disclosure, the artificial intelligence based interaction device 900 further includes a belonging determination module.

[0195] The belonging determination module is configured to determine a belonging recognition result of the initial knowledge based on a source type of the initial knowledge.

[0196] Any one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present disclosure, or at least part of the functions of any one or more of the modules, sub-modules, units, sub-units can be implemented in one module. Any one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present disclosure can be split into multiple modules for implementation. Any one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present disclosure can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on chip, a system on substrate, a system on package, an application specific integrated circuit (ASIC), or any other reasonable way of hardware or firmware through integration or packaging of circuits, or in any one of software, hardware and firmware or in an appropriate combination of any of the above. Alternatively, one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present disclosure can be at least partially implemented as computer program modules that can perform corresponding functions when the computer program modules are run.

[0197] For example, any of the receiving module 910, the retrieval enhancement module 920, the model processing module 930, and the presentation module 940 can be combined in one module / unit / subunit, or any of the modules / units / subunits can be split into multiple modules / units / subunits. Alternatively, at least part of the function of one or more of the modules / units / subunits can be combined with at least part of the function of other modules / units / subunits and implemented in one module / unit / subunit. According to embodiments of the present application, at least one of the receiving module 910, the retrieval enhancement module 920, the model processing module 930, and the presentation module 940 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on chip, a system on board, a system on package, an application specific integrated circuit (ASIC), or any other reasonable manner of integrating or packaging a circuit, etc. hardware or firmware, or in any one of software, hardware, and firmware implementations or in a suitable combination of any of them. Alternatively, at least one of the receiving module 910, the retrieval enhancement module 920, the model processing module 930, and the presentation module 940 can be at least partially implemented as a computer program module that can perform the corresponding function when executed.

[0198] It should be noted that the part of the interaction device based on artificial intelligence in the embodiments of the present application corresponds to the part of the interaction method based on artificial intelligence in the embodiments of the present application, and the description of the part of the interaction device based on artificial intelligence is specifically referred to the part of the interaction method based on artificial intelligence, which will not be repeated here.

[0199] Figure 10 An illustrative structural block diagram of an artificial intelligence agent is shown according to an embodiment of the present disclosure.

[0200] In embodiments of the present disclosure, as shown in Figure 10 The AI agent 1000 can include an input module 1010, a processing module 1020, and an output module 1030.

[0201] The input module 1010 is configured to receive input information;

[0202] The processing module 1020 is configured to determine a target task based on the input information received by the input module, determine a large model based on the target task, execute the interaction method based on artificial intelligence according to an embodiment of the present disclosure by calling the large model, and obtain output information;

[0203] The output module 1030 is configured to output the output information obtained by the processing module.

[0204] According to an embodiment of the present disclosure, the input module 1010 is responsible for receiving or perceiving information such as queries, requests, instructions, signals or data from the outside world (e.g., a user or an external environment), and converting them into a format that the AI agent 1000 can understand and process. The input module 1010 is the first step for the AI agent 1000 to interact with the outside world, and it enables the AI agent 1000 to efficiently and accurately obtain necessary "sensory" information from the outside world and respond to it.

[0205] In an example, the input module 1010 can input the input content described above.

[0206] In an example, the processing module 1020 is the core support for the AI agent 1000 to process complex tasks. The processing module 1020 can execute the artificial intelligence-based interaction method described above.

[0207] In an example, the performance of the processing module 1020 can be closely related to the large model on which the AI agent 1000 is based. In order to fully exert the capabilities of the large model, the internal structure of the processing module 1020 can be designed to be highly configurable and scalable in order to cope with various different types of tasks and requirements in real scenarios.

[0208] In an example, after the AI agent 1000 obtains the input content, the processing module 1020 can process the special knowledge base and the general knowledge base to obtain a reference knowledge set by using a text recognition large model, process the reference knowledge set and the input content by using the large model, obtain feedback content, and pass the feedback content to the output module 1030.

[0209] It can be understood that, although the large model has excellent language understanding and generation capabilities, it, like a human, can only solve a limited number of tasks without the aid of any tools. When the AI agent 1000 is endowed with the ability to call tools, it can achieve tasks such as completing mathematical operations with a calculator, completing data analysis with Python, and completing weather forecasts with a search engine.

[0210] The AI agent 1000 according to an embodiment of the present disclosure can simply and effectively improve the degree of intelligence, and improve flexibility and versatility.

[0211] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0212] According to an embodiment of the present disclosure, an electronic device comprises at least one processor and a memory connected in communication with the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described above.

[0213] According to an embodiment of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute the method described above.

[0214] According to an embodiment of the present disclosure, a computer program product includes a computer program, and when the computer program is executed by a processor, the computer program implements the method described above.

[0215] Figure 11 A schematic block diagram of an example electronic device 1100 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are provided as examples only and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0216] like Figure 11 As shown, device 1100 includes a computing unit 1101, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1102 or a computer program loaded from a storage unit 1108 into a random access memory (RAM) 1103. RAM 1103 may also store various programs and data required for the operation of device 1100. Computing unit 1101, ROM 1102, and RAM 1103 are connected to each other via a bus 1104. An input / output (I / O) interface 1105 is also connected to bus 1104.

[0217] Various components in device 1100 are connected to an input / output (I / O) interface 1105, including an input unit 1106, such as a keyboard and mouse; an output unit 1107, such as various types of displays and speakers; a storage unit 1108, such as a magnetic disk and optical disk; and a communication unit 1109, such as a network card, a modem, a wireless communication transceiver, etc. Communication unit 1109 allows device 1100 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0218] The computing unit 1101 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1101 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1101 performs various methods and processes described above, such as the artificial intelligence-based interaction method. For example, in some embodiments, the artificial intelligence-based interaction method can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 1108. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device 1100 via the ROM 1102 and / or the communication unit 1109. When the computer program is loaded onto the RAM 1103 and executed by the computing unit 1101, one or more steps of the artificial intelligence-based interaction method described above can be performed. Alternatively, in other embodiments, the computing unit 1101 can be configured to perform the artificial intelligence-based interaction method by any other suitable means, such as by means of firmware.

[0219] Various implementations of the systems and techniques described above can be realized in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a complex programmable logic device (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0220] Program code for carrying out methods of the present disclosure can be written in any combination of one or more programming languages. The program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, produces a means for implementing the functions / acts specified in the flowcharts and / or block diagrams. The program code can be executed entirely on a machine, partially on a machine, partially on a machine and partially on a remote machine or entirely on a remote machine or server.

[0221] In the context of this disclosure, a machine-readable medium can be a tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include but is not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0222] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0223] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0224] The computer system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server can be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0225] It should be understood that the various forms of flow shown above can be used to reorder, add, or remove steps. For example, the steps described in the present disclosure can be performed in parallel, in series, or in a different order, without limitation herein, so long as the desired results of the technology described in the present disclosure are achieved.

[0226] The specific implementation described above does not constitute a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. An artificial intelligence-based interaction method, comprising: Receive input content entered by the target object in the interactive interface; Obtaining a reference knowledge set related to the input content based on a first reference knowledge set determined from a dedicated knowledge base using a keyword indexing method and a second reference knowledge set determined from a general knowledge base using a vector indexing method; Inputting the input content and the reference knowledge set into the large model to obtain feedback content for feeding back the input content; as well as Displaying the feedback content on the interactive interface; The knowledge stored in the dedicated knowledge base and the knowledge stored in the general knowledge base have different sources.

2. The method according to claim 1, wherein Determining a first reference knowledge set from a dedicated knowledge base by using a keyword index method includes: determining an initial first reference knowledge set from the dedicated knowledge base based on a similarity between an index content set and a keyword tag set, wherein the keyword tag set is used to represent the dedicated knowledge in the dedicated knowledge base, and the index content set is obtained by rewriting the input content based on the intent information of the input content; and The reference knowledge in the initial first reference knowledge set is deduplicated based on the same knowledge modality to obtain the first reference knowledge set.

3. The method according to claim 1 or 2, wherein: Determining a second reference knowledge set from a general knowledge base using a vector index method includes: Determining an initial second reference knowledge set from the general knowledge base based on a similarity between an index vector set and general knowledge vectors, wherein the index vector set is obtained by rewriting and vector-converting the input content based on intent information of the input content, and the general knowledge vectors are obtained by vector-converting general knowledge in the general knowledge base; and The reference knowledge in the initial second reference knowledge set is deduplicated based on the same knowledge modality to obtain the second reference knowledge set.

4. The method according to any one of claims 1 to 3, wherein The reference knowledge set is obtained based on the first reference knowledge set and the second reference knowledge set, including: Filtering the first reference knowledge set and the second reference knowledge set based on attribute information of each of the reference knowledge in the first reference knowledge set and the reference knowledge in the second reference knowledge set to obtain the reference knowledge set; The attribute information includes at least one of the following: The relevance between the reference knowledge and the input content, the professionalism of the reference knowledge, and the timeliness of the reference knowledge.

5. The method according to claim 2, further comprising: rewriting the input content based on the previous content of the input content and the intention information of the input content to obtain an initial index content set; as well as The initial index content set is screened based on the correlation between the initial index content in the initial index content set and the input content to obtain the index content set.

6. The method according to claim 3, further comprising: rewriting the input content based on the previous content of the input content and the intention information of the input content to obtain an initial index content set; Performing vector encoding on the initial index contents in the initial index content set to obtain an initial index vector set; as well as Identification information representing a reference degree is added to the initial index vectors in the initial index vector set to obtain the index vector set.

7. The method according to any one of claims 1 to 6, further comprising: Rewrite the initial knowledge to obtain the target knowledge; Extracting key information of the target knowledge to obtain keyword tags for characterizing the target knowledge topic; as well as The general knowledge base is obtained based on the target knowledge, the keyword tags and the attribute information of the target knowledge.

8. The method according to any one of claims 1 to 7, further comprising: Use multiple initial knowledge obtained from different sources to verify and obtain target knowledge; Based on a dedicated dictionary, performing synonymous expansion on the keyword tags of the target knowledge to obtain a keyword tag set for representing the target knowledge topic; and The dedicated knowledge base is obtained based on the keyword tag set, the target knowledge and the identification information set, wherein the identification information in the identification information set is used to represent the importance of the keyword tags in the keyword tag set.

9. The method according to claim 1, further comprising: Based on the source type of the initial knowledge, an attribution identification result of the initial knowledge is determined, wherein the attribution identification result indicates whether the initial knowledge belongs to the general knowledge base or the dedicated knowledge base.

10. An artificial intelligence-based interactive device, comprising: A receiving module, configured to receive input from a target object on an interactive interface; a retrieval enhancement module, configured to obtain a reference knowledge set related to the input content based on a first reference knowledge set determined from a dedicated knowledge base using a keyword indexing method and a second reference knowledge set determined from a general knowledge base using a vector indexing method; A model processing module, configured to input the input content and the reference knowledge set into a large model, and obtain feedback content for feeding back the input content; as well as A display module, configured to display the feedback content on the interactive interface; The knowledge stored in the dedicated knowledge base and the knowledge stored in the general knowledge base have different sources.

11. An artificial intelligence agent, comprising: An input module, used for receiving input information; a processing module, configured to determine a target task based on the input information received by the input module, determine a large model based on the target task, and execute the method according to any one of claims 1 to 9 by calling the large model to obtain output information; An output module is used to output the output information obtained by the processing module.

12. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 9.

13. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 9.

14. A computer program product comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 9.

Images