Content generation method and device, electronic equipment and computer program product

Abstract

The invention relates to a content generation method and device, electronic equipment and a computer program product, and relates to the technical field of artificial intelligence and natural language processing. The method comprises the steps of obtaining current query content, and performing intention uncertainty quantification processing on a current query intention according to the current query content, historical interaction records and a current retrieval information set generated in a current round to obtain an intention uncertainty value; when the numerical value of the intention uncertainty is smaller than or equal to a preset threshold value, obtaining a multi-modal information set obtained after intention uncertainty quantization processing; performing multi-modal belief fusion processing according to a cross-modal argument atlas constructed based on the multi-modal information set to obtain a fusion belief function; and based on the fusion belief function, generating target reply content corresponding to the current query content. According to the method, the overall intention uncertainty of the current query requirement can be reduced, and the execution efficiency, robustness and adaptability for complex tasks are improved.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence and natural language processing technology, and more specifically, to a content generation method, a content generation device, an electronic device, and a computer program product. Background Technology

[0002] Agentic Retrieval-Augmented Generation (RAG) refers to the introduction of one or more "agents" into the Retrieval-Augmented Generation (RAG) framework. These agents can simulate human thinking, planning, action execution (such as retrieving, using tools, and generating content), and reflection to handle more complex tasks. RAG is an artificial intelligence framework that enhances the accuracy and timeliness of content generated by large language models by retrieving relevant information from external knowledge bases.

[0003] Agentic RAG technology significantly improves the performance of large language models (LLMs) in complex task processing by giving them the ability to plan, iterate, and use external tools. However, existing agentic RAG systems still face challenges as application scenarios expand.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this disclosure is to provide a content generation method, content generation apparatus, electronic device, computer-readable storage medium, and computer program product, thereby overcoming, to at least a certain extent, the limitations of existing Agentic RAG systems in understanding user intent and the inefficiency of complex task planning and resource management.

[0006] Other features and advantages of this disclosure will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0007] According to a first aspect of this disclosure, a content generation method is provided, comprising: obtaining current query content; performing intent uncertainty quantification processing on the current query intent based on the current query content, historical interaction records, and a set of current search information generated in the current round, to obtain an intent uncertainty value; when the intent uncertainty value is less than or equal to a preset threshold, obtaining a multimodal information set obtained after the intent uncertainty quantification processing; performing multimodal belief fusion processing on a cross-modal argument graph constructed based on the multimodal information set, to obtain a fused belief function; and generating target response content corresponding to the current query content based on the fused belief function.

[0008] In one exemplary embodiment of this disclosure, the step of quantifying the intent uncertainty of the current query intent based on the current query content, historical interaction records, and the current search information set generated in the current round to obtain an intent uncertainty value includes: constructing a dynamic intent graph based on the current query content and the current search information set; determining the Shannon entropy, intent ambiguity, and confidence uncertainty index corresponding to the dynamic intent graph; and determining the intent uncertainty value based on the Shannon entropy, the intent ambiguity, and the confidence uncertainty index.

[0009] In one exemplary embodiment of this disclosure, the step of constructing a dynamic intent graph based on the current query content and the current search information set includes: performing information decomposition processing on the current query content and the current search information set to obtain multiple intent nodes, each intent node having a corresponding node confidence level; determining the node relationships between the intent nodes; constructing an initial dynamic intent graph based on the multiple intent nodes and the node relationships; and updating the intent confidence level of the initial dynamic intent graph to obtain the dynamic intent graph.

[0010] In one exemplary embodiment of this disclosure, the method further includes: when the intent uncertainty value is greater than the preset threshold, obtaining an exploration action set, the exploration action set including multiple exploration actions; determining the intent information gain corresponding to each exploration action after performing an operation; selecting a target exploration action from the multiple exploration actions according to the intent information gain, the target exploration action corresponding to the largest intent information gain; generating updated query content based on the target exploration action, the intent uncertainty value corresponding to the updated query intent of the updated query content being less than or equal to the preset threshold; and performing an update operation on the current search information set according to the updated query content to obtain an updated search information set.

[0011] In one exemplary embodiment of this disclosure, the step of performing multimodal belief fusion processing based on the cross-modal argument map constructed based on the multimodal information set to obtain a fused belief function includes: constructing a cross-modal argument map based on the multimodal information set, the cross-modal argument map including multiple cross-modal arguments; determining the argument acceptance probability distributions corresponding to the multiple cross-modal arguments, the argument acceptance probability distributions including a single argument acceptance probability distribution where a single argument is accepted, and a combined argument acceptance probability distribution where a combined argument is accepted, the combined argument consisting of multiple compatible arguments; and performing multimodal belief fusion processing based on the argument acceptance probability distributions to obtain the fused belief function.

[0012] In one exemplary embodiment of this disclosure, constructing a cross-modal argument graph based on the multimodal information set includes: extracting cross-modal arguments based on the multimodal information set, wherein the cross-modal arguments include a core claim and multimodal evidence supporting the core claim; determining the relationship strength weight between any two cross-modal arguments, wherein the relationship strength weight is determined based on one or more of attack support relationship, semantic consistency, and evidence reliability between the two cross-modal arguments; and constructing the cross-modal argument graph based on multiple cross-modal arguments and the relationship strength weight.

[0013] In one exemplary embodiment of this disclosure, determining the argument acceptance probability distribution corresponding to the plurality of cross-modal arguments includes: obtaining the relationship strength weight between each of the cross-modal arguments; determining the acceptance probability distribution of each of the cross-modal arguments under different acceptance semantics based on the relationship strength weight; and determining the individual argument acceptance probability distribution and the combined argument acceptance probability distribution of each of the cross-modal arguments under different acceptability semantics based on the acceptance probability distribution.

[0014] In one exemplary embodiment of this disclosure, the step of performing multimodal belief fusion processing based on the argument acceptance probability distribution to obtain the fused belief function includes: determining the comprehensive arguments corresponding to multiple cross-modal arguments, and unresolved conflict information; and performing multimodal belief fusion processing based on the argument acceptance probability distribution, the comprehensive arguments, and the unresolved conflict information to construct the fused belief function.

[0015] According to a second aspect of this disclosure, a content generation apparatus is provided, comprising: an intent determination module, configured to acquire current query content, and perform intent uncertainty quantification processing on the current query intent based on the current query content, historical interaction records, and a set of current search information generated in the current round, to obtain an intent uncertainty value; an information set determination module, configured to acquire a multimodal information set obtained after the intent uncertainty quantification processing when the intent uncertainty value is less than or equal to a preset threshold; a fusion function determination module, configured to perform multimodal belief fusion processing based on a cross-modal argument graph constructed based on the multimodal information set, to obtain a fusion belief function; and a content generation module, configured to generate target response content corresponding to the current query content based on the fusion belief function.

[0016] In one exemplary embodiment of this disclosure, the intent determination module includes an intent determination unit, configured to: construct a dynamic intent graph based on the current query content and the current retrieval information set; determine the Shannon entropy, intent ambiguity, and confidence uncertainty index corresponding to the dynamic intent graph; and determine the intent uncertainty value based on the Shannon entropy, the intent ambiguity, and the confidence uncertainty index.

[0017] In one exemplary embodiment of this disclosure, the intent determination unit includes an intent graph construction subunit, configured to perform information decomposition processing on the current query content and the current search information set to obtain multiple intent nodes, each intent node having a corresponding node confidence level; determine the node relationships between the intent nodes; construct an initial dynamic intent graph based on the multiple intent nodes and the node relationships; and update the intent confidence level of the initial dynamic intent graph to obtain the dynamic intent graph.

[0018] In one exemplary embodiment of this disclosure, the intent determination module further includes an information set update unit, configured to: when the intent uncertainty value is greater than the preset threshold, obtain an exploration action set, the exploration action set including multiple exploration actions; determine the intent information gain corresponding to each exploration action after performing an operation; select a target exploration action from the multiple exploration actions according to the intent information gain, the target exploration action having the largest intent information gain; generate updated query content based on the target exploration action, the intent uncertainty value corresponding to the updated query intent of the updated query content being less than or equal to the preset threshold; and update the current search information set according to the updated query content to obtain an updated search information set.

[0019] In one exemplary embodiment of this disclosure, the fusion function determination module includes a fusion function determination unit, configured to: construct a cross-modal argument map based on the multimodal information set, the cross-modal argument map including multiple cross-modal arguments; determine the argument acceptance probability distributions corresponding to the multiple cross-modal arguments, the argument acceptance probability distributions including a single argument acceptance probability distribution where a single argument is accepted, and a combined argument acceptance probability distribution where a combined argument is accepted, the combined argument consisting of multiple compatible arguments; and perform multimodal belief fusion processing based on the argument acceptance probability distributions to obtain the fused belief function.

[0020] In one exemplary embodiment of this disclosure, the fusion function determination unit includes a cross-modal graph construction subunit, configured to: extract cross-modal arguments based on the multimodal information set, wherein the cross-modal arguments include a core claim and multimodal evidence supporting the core claim; determine the relationship strength weight between any two cross-modal arguments, wherein the relationship strength weight is determined based on one or more of attack support relationship, semantic consistency, and evidence reliability between the two cross-modal arguments; and construct the cross-modal argument graph based on multiple cross-modal arguments and the relationship strength weight.

[0021] In one exemplary embodiment of this disclosure, the fusion function determination unit includes a probability distribution determination subunit, configured to: obtain the relationship strength weights between the cross-modal arguments; determine the acceptability probability distribution of each cross-modal argument under different acceptability semantics based on the relationship strength weights; and determine the individual argument acceptability probability distribution and the combined argument acceptability probability distribution of each cross-modal argument under different acceptability semantics based on the acceptability probability distribution.

[0022] In one exemplary embodiment of this disclosure, the fusion function determination unit includes a fusion function determination subunit, configured to: determine the comprehensive arguments corresponding to the multiple cross-modal arguments, and unresolved conflict information; and perform multimodal belief fusion processing based on the argument acceptance probability distribution, the comprehensive arguments, and the unresolved conflict information to construct the fused belief function.

[0023] According to a third aspect of this disclosure, an electronic device is provided, comprising: a processor; and a memory storing computer-readable instructions that, when executed by the processor, implement the content generation method according to any one of the preceding claims.

[0024] According to a fourth aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the content generation method according to any one of the preceding claims.

[0025] According to a fifth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the content generation method described in any of the preceding claims.

[0026] The technical solution provided in this disclosure may include the following beneficial effects:

[0027] The content generation method in the exemplary embodiments of this disclosure, on the one hand, accurately assesses the confidence level of the current query intent to determine the true intent demand, thereby reducing overall intent uncertainty. On the other hand, by deeply analyzing the consistency between multimodal information, a more comprehensive and reliable cross-modal cognitive foundation is formed to guide subsequent content generation or decision-making, which can improve the execution efficiency, robustness, and adaptability for complex tasks.

[0028] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0029] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0030] Figure 1 A flowchart illustrating an exemplary embodiment of the content generation method according to this disclosure is shown schematically;

[0031] Figure 2 The diagram illustrates an overall flowchart of information processing and decision-making centered on an intelligent agent according to an exemplary embodiment of the present disclosure.

[0032] Figure 3 A flowchart illustrating an exemplary embodiment of the HII-IUQ module according to this disclosure is shown schematically;

[0033] Figure 4 A flowchart illustrating an exemplary embodiment of the MMCD-IF module according to the present disclosure is shown schematically;

[0034] Figure 5 A block diagram of a content generation apparatus according to an exemplary embodiment of the present disclosure is shown schematically;

[0035] Figure 6 A block diagram of an electronic device according to an exemplary embodiment of the present disclosure is shown schematically;

[0036] Figure 7The illustration shows a schematic diagram of a computer-readable storage medium according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0037] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0038] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced without one or more of the specific details described, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known structures, methods, apparatuses, implementations, materials, or operations are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0039] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, or in one or more software-hardened modules, or in different network and / or processor devices and / or microcontroller devices.

[0040] With the expansion of application scenarios, existing Agentic RAG systems still face challenges in the following aspects: (1) Limitations in understanding user intent. Current Agentic RAG systems, when understanding complex or ambiguous user queries, usually remain at the level of surface keyword matching or single intent recognition, making it difficult to capture the user's deep, implicit task goals or multiple intents. When the user intent is unclear, ambiguous, or changes dynamically with the task progress, the agent has difficulty effectively adjusting its exploration strategy and generated content, which can easily lead to information bias, inefficient retrieval, or the generation of answers that do not meet the user's real needs.

[0041] (2) Insufficient efficiency in complex task planning and resource management. When dealing with complex tasks involving multiple steps and stages, the existing Agentic RAG system's task planning capabilities are usually limited by preset rules or simple chained calls, lacking the ability to dynamically adapt to complex environments and efficiently manage diverse external tools. When the task path is uncertain, resources are limited, or unexpected situations occur during execution, the system struggles to flexibly adjust its planning, easily falling into suboptimal loops or failing to complete the task.

[0042] Based on this, according to embodiments of the present disclosure, a content generation method, a content generation apparatus, a computer-readable storage medium, an electronic device, and a computer program product are proposed.

[0043] In this article, it is important to understand the terminology involved. Hierarchical Intent Inference (HII) refers to the process by which an agent performs a multi-granular and multi-layered deep understanding of its user queries or task objectives, gradually inferring from surface intent (such as keyword matching) to deeper intent (such as implicit user needs and the ultimate goal of the task).

[0044] Intent Uncertainty Quantification (IUQ) is a method for quantifying an agent’s understanding of a user’s intent, including the assessment of confidence, ambiguity, and potential intent space at different levels of intent.

[0045] A Dynamic Intent Graph (DIG) is a graph structure used to dynamically represent and update an agent’s understanding of user intent, including intent nodes, relationships between intents (such as inclusion, juxtaposition, and conflict), and the activation probability or confidence of each intent node.

[0046] Dynamic Task Planning (DTP) refers to the process by which an agent, based on its current understanding of the task, available resources, and environmental feedback, breaks down the task, generates action sequences, and adjusts the plan in real time.

[0047] Adaptive Tool Orchestration (ATO) refers to an agent intelligently selecting, invoking, and coordinating one or more external tools (such as search engines, APIs, calculators, etc.) to perform tasks, and dynamically adjusting its usage strategy based on the tool's execution results.

[0048] In this example embodiment, a content generation method is first provided. The content generation method of this disclosure can be implemented using a server or using a terminal device. The terminal described in this disclosure can include mobile terminals such as mobile phones, tablets, laptops, handheld computers, and personal digital assistants (PDAs), as well as fixed terminals such as desktop computers. Figure 1 The illustration shows a schematic diagram of a content generation method flow according to some embodiments of the present disclosure. Reference Figure 1 The content generation method may include the following steps:

[0049] Step S110: Obtain the current query content; based on the current query content, historical interaction records, and the current search information set generated in the current round, perform intent uncertainty quantification on the current query intent to obtain the intent uncertainty value.

[0050] Step S120: When the intention uncertainty value is less than or equal to a preset threshold, obtain the multimodal information set obtained after the intention uncertainty quantification process;

[0051] Step S130: Based on the cross-modal argument graph constructed based on the multimodal information set, perform multimodal belief fusion processing to obtain the fused belief function;

[0052] Step S140: Based on the fusion belief function, generate the target response content corresponding to the current query content.

[0053] According to the content generation method in this example embodiment, on the one hand, by accurately assessing the confidence level of the current query intent to determine the true intent demand, the overall intent uncertainty can be reduced. On the other hand, by deeply analyzing the consistency between multimodal information, a more comprehensive and reliable cross-modal cognitive foundation is formed to guide subsequent content generation or decision-making, which can improve the execution efficiency, robustness, and adaptability for complex tasks.

[0054] The content generation method in this example embodiment will be further explained below.

[0055] refer to Figure 2 , Figure 2 This illustration schematically shows an overall flowchart of information processing and decision-making centered on an intelligent agent according to an exemplary embodiment of this disclosure. The content generation method of this disclosure can be implemented based on the Agentic RAG system. Figure 2 The overall architecture of the Agentic RAG system proposed in this disclosure and the interaction flow between its main modules are shown. The Agentic RAG system mainly includes a hierarchical intent reasoning and intent uncertainty quantification module (HII-IUQ module) and a multimodal collaborative dialectical-information fusion (MMCD-IF) module.

[0056] When a user enters their query, the HII-IUQ module first attempts to understand and clarify the user's intent until the uncertainty is reduced to an acceptable level. Then, the MMCD-IF module performs dynamic planning and adaptive tool orchestration to form the final execution strategy, which guides answer generation or action execution.

[0057] In one exemplary embodiment of this disclosure, for step S110, the intent uncertainty of the current query intent is quantified based on the current query content, historical interaction records, and the current search information set generated in the current round to obtain the intent uncertainty value. This includes: constructing a dynamic intent graph based on the current query content and the current search information set; determining the Shannon entropy, intent ambiguity, and confidence uncertainty index corresponding to the dynamic intent graph; and determining the intent uncertainty value based on the Shannon entropy, intent ambiguity, and confidence uncertainty index.

[0058] Here, "current query content" refers to the query content corresponding to the current round. "Historical interaction records" can be query interaction records generated based on historical query operations. "Current retrieval information set" can be the information set generated by information retrieval operations up to the current round. Historical interaction records can serve as the data foundation for constructing the current retrieval information set. "Current query intent" can be the true intent corresponding to the current query content. "Intent uncertainty value" can be an indicator value determined after quantifying the intent uncertainty of the current query intent. "Dynamic intent graph" can be a dynamically changing intent graph.

[0059] refer to Figure 3 , Figure 3 A flowchart illustrating an exemplary embodiment of the HII-IUQ module according to this disclosure is shown. The current query content is obtained, which can be a query entered by the user or a query selected by the user based on multiple candidate queries pre-provided by the question-and-answer system. For example, if the user currently enters the question "How to make a cake in an oven, and provide some recipes," then this sentence is the current query content.

[0060] Since the user may have already interacted with the Agentic RAG system before this query, the historical interaction records generated before the current round and the current search information generated in the current round will be used as input to the HII-IUQ module. The HII-IUQ module will then perform hierarchical intent reasoning and intent uncertainty quantification to obtain the intent uncertainty value corresponding to the current query intent. If this query is the first interaction, there will be no historical interaction records. Figure 3 The HII-IUQ module in the system decomposes the input content to construct a dynamic intent graph; then, it performs intent uncertainty quantification based on the dynamic intent graph to obtain the intent uncertainty value.

[0061] In one exemplary embodiment of this disclosure, a dynamic intent graph is constructed based on the current query content and the current search information set, including: performing information decomposition processing on the current query content and the current search information set to obtain multiple intent nodes, each intent node having a corresponding node confidence level; determining the node relationships between the intent nodes; constructing an initial dynamic intent graph based on the multiple intent nodes and node relationships; and updating the intent confidence level of the initial dynamic intent graph to obtain the dynamic intent graph.

[0062] In this context, an intent node can be a node reflecting a specific intent or sub-intent; for example, an intent node can be represented using keywords, phrases, or other abstract concepts. Node confidence can be the confidence associated with each intent node, representing the intelligent agent's degree of belief in the accuracy of that intent. Node relationships can be the mutual relationships between two intent nodes, including but not limited to inclusion, parallel, and conflict relationships. The initial dynamic intent graph can be a structured graph constructed based on multiple intent nodes and their relationships. The dynamic intent graph can be an intent graph obtained by updating the intent confidence in the initial dynamic intent graph based on new evidence using an evidence-based belief update mechanism.

[0063] The input to the HII-IUQ module includes the current query content Q. user Historical interaction records H hist and the information set D retrieved in the current round curr ={d1,d2,…,d k The HII-IUQ module first performs intent decomposition and dynamic intent graph construction. First, intent decomposition is performed, including intent node identification and hierarchical processing. The current query content Q is then... user and the current search information set D curr The information in the data is decomposed into a series of intent nodes I = {I1, I2, ..., I...} k Each intent node I j Representing a specific intention or sub-intention, it can be expressed using keywords, phrases, or more abstract concepts. Through information decomposition, intentions can be organized into a hierarchical structure, from macro-intentions to micro-intentions.

[0064] Secondly, dynamic intent graph construction. The Agentic RAG system can construct an initial dynamic intent graph G based on multiple intent nodes and the relationships between them. intent =(V intent E intent ), where V intent It is a collection of intent nodes, containing multiple intent nodes, E inten It is a set of relationships between intent nodes, such as Eintent This can include, but is not limited to, inclusion relationships (A includes B), parallel relationships (A and B are parallel), and conflict relationships (A and B conflict). And each intent node I... j ∈V intent Associate a confidence level C(I) j ), Node confidence C(I) j This can represent the degree of belief that the Agentic RAG system has in the accuracy of the intent.

[0065] Next, the intent confidence update operation. An evidence-based belief update mechanism is used to update the intent confidence in the intent graph based on new evidence D_curr. For the intent update operation, the HII-IUQ module uses this new evidence to correct its previous understanding of the user's intent, i.e., updating the confidence of the initial dynamic intent graph. The intent update operation is a cyclical process; whenever new information (i.e., new evidence) comes in, the HII-IUQ module can re-evaluate its understanding of the user's intent and obtain the updated intent confidence. After the intent confidence update operation is complete, the updated intent is used as the dynamic intent graph. The updated confidence C new (I j ) and the confidence level C before the update o(d (I j The relationship between the new evidence and the strength of support for the intention is directly proportional to the strength of the new evidence, as shown in Formula 1.

[0066]

[0067] Among them, C new (I j ) is the updated intent node I j confidence level, C o(d (I j ) is the intention node I before the update. j The confidence level, P(e|I) j ) indicates that evidence e is intended for I j The probability of the following occurring is given by R(e), which represents the reliability of evidence e.

[0068] By decomposing user queries and retrieval information into structured intent nodes and constructing a dynamic intent graph, detailed intent modeling is achieved. At the same time, an intent confidence update mechanism is introduced, enabling the agent to reasonably handle the support of different information sources for intent.

[0069] After constructing the dynamic intent graph, multiple complementary metrics are used to quantify its intent uncertainty. Specifically, the intent entropy corresponding to the dynamic intent graph is calculated. For the discrete intent space, its Shannon entropy can be calculated, as shown in Formula 2.

[0070]

[0071] Where H[P(i)] can be the Shannon entropy corresponding to the dynamic intent graph; P(i) represents the probability of a specific intent i, and the specific intent i can be a characteristic intent node I. j The corresponding intent; P(i) can be obtained by normalizing its confidence level C(i); I is the set of all possible intents. Specifically, the higher the entropy value, the greater the uncertainty of the intent.

[0072] The intent ambiguity of a dynamic intent graph can be defined as the degree of ambiguity between multiple conflicting intents with high confidence. If multiple conflicting intents exist, their degree of ambiguity can be assessed. This can be quantified by calculating a similarity metric (e.g., cosine similarity or semantic distance) between the conflicting intent pairs and combining it with their confidence levels. One such quantification method is shown in Equation 3.

[0073]

[0074] Ambiguity can be the degree of intent ambiguity; Conflicting Pairs refers to all conflicting intent pairs identified in the dynamic intent map; C(I a ) can represent intention I a Confidence level; C(I) b ) can represent intention I b Confidence level; Similarity(I) a ,I b Intent I can be computed based on semantic embedding. a and Intention I b The degree of ambiguity is higher when there is a high degree of similarity but also a conflicting relationship.

[0075] Confidence uncertainty metrics can be used to measure the confidence level corresponding to key intent nodes. For example, confidence uncertainty metrics can be quantified using key intent uncertainty and key intent coverage. Identify the current query content Q. user Which key intent nodes still have high uncertainty or low coverage in terms of confidence level? This can be quantified using Formula 4.

[0076]

[0077] Coverage can be an indicator of the uncertainty of the confidence level of key intentions; key It is the set of key intents identified in the user query; C(I) j ) is the intention node I jThe confidence level. The confidence level uncertainty index can be obtained based on the coverage and uncertainty of key intents.

[0078] Low coverage can mean that the agent's identification and understanding of all the key intents in a user's query is incomplete or insufficient, meaning that the agent may have missed the core intents in the user's query. High uncertainty means that although the agent has identified a key intent, the accuracy or confidence level of that intent is very low.

[0079] Taking the query "How to make a cake in an oven, and provide some recipes" as an example, low coverage means that if the agent only identifies the two key intents "make a cake" and "oven," but completely ignores the key intent "recipe," then the agent's coverage of the entire query is low. High uncertainty means that if the agent identifies the key intent "make a cake," but finds many recipes for "baking bread" or "making pizza" while searching for information, this information conflicts with the intent "make a cake," leading to low confidence in the intent "make a cake," i.e., high uncertainty.

[0080] After determining the above-mentioned uncertainty quantification indicators, the individual uncertainty can be measured by combining the confidence variance or standard deviation of these key intent nodes, as shown in Formula 5.

[0081]

[0082] in, It is the numerical value of intent uncertainty, that is, the specific value of the comprehensive intent uncertainty index, which can be abbreviated as U. total H[P(i)] is the Shannon entropy corresponding to the dynamic intent graph; Ambiguity can be the degree of intent ambiguity; Coverage can be the confidence uncertainty index of key intents; w1, w2, and w3 can be the weight values ​​corresponding to the three indices. By using a hierarchical intent graph to dynamically and quantitatively represent its intent state for user queries, the belief and uncertainty of intents at each level can be clearly reflected.

[0083] In one exemplary embodiment of this disclosure, when the intent uncertainty value is greater than or equal to a preset threshold, an exploration action set is obtained, which includes multiple exploration actions; the intent information gain corresponding to each exploration action after execution is determined; based on the intent information gain, a target exploration action is selected from the multiple exploration actions, and the intent information gain corresponding to the target exploration action is the largest; based on the target exploration action, updated query content is generated, and the intent uncertainty value corresponding to the updated query intent is less than or equal to a preset threshold; the current search information set is updated according to the updated query content to obtain the updated search information set.

[0084] The set of exploration actions can be a collection of multiple exploration actions. An exploration action can be an action used to generate new query content. Intent information gain can be the information gain resulting from employing a particular exploration action. The target exploration action can be the exploration action in the set that maximizes the intent information gain. Updated query content can be the query content obtained by updating the current query content based on the target exploration action. Updated search information set can be the search information set obtained by updating the current search information set based on the updated query content.

[0085] After calculating the intent uncertainty value through the above steps, if the intent uncertainty value is greater than or equal to a preset threshold, then the proactive intent clarification strategy planning is executed. The agent's goal is to select a target exploration action A from the set of exploration actions, which maximizes the expected Intent Information Gain (EIIG) or minimizes the expected posterior intent uncertainty. The target exploration action A is determined as shown in Equation 6.

[0086]

[0087] in, The optimal clarification action is the identified target exploration action A; target exploration action A can be generating a new clarification query Qclarify′, asking the user a question Qask, or calling an external tool Tcall, etc. actions It is the set of all possible exploration actions; EIG(A) represents the expected intention information gain after performing the target exploration action A. The intention information gain is roughly represented by the sum of the current intention uncertainty and the expected new observation O after the target exploration action A. new The difference in uncertainty of the expected intention after the event is shown in Formula 7.

[0088] EIG = Uncertainty(Q) user )-ExpectedUncertainty(After actionA andobservationO new )(Formula 7)

[0089] Where EIG is the intent information gain; Uncertainty(Q) user ) represents the uncertainty of the current user's intent, i.e., the intent uncertainty value corresponding to the current query intent; ExpectedUncertainty(AfteractionAandobservationOnew (A) is the execution of target exploration action A and the acquisition of new observations O. new The uncertainty of the expected outcome, i.e., the uncertainty value of the intent corresponding to the updated query content, is generated through the goal exploration action A to generate a new query. By planning a proactive intent clarification strategy based on information theory principles (such as expected intent information gain), the agent can be guided to choose the optimal information acquisition action in a goal-oriented manner, which can effectively reduce the uncertainty of key intents.

[0090] After performing the above proactive intent clarification strategy planning operation, if the updated intent uncertainty value U tota(_intent Below the preset threshold U threshold_intent Then the HII-IUQ module outputs the currently constructed dynamic intent graph G. intent This information is then passed to the MMCD-IF module. The goal of the HII-IUQ module is to enable the Agent to perform a quantitative assessment of its current intent state regarding the user query (including explicit intent, implicit intent, and degree of uncertainty), just like a human, and to determine the next exploratory action based on this assessment in order to maximize the approximation of the user's true needs or clarify key intents.

[0091] refer to Figure 4 , Figure 4 A flowchart of an MMCD-IF module according to an exemplary embodiment of the present disclosure is illustrated. The MMCD-IF module receives processed data and intent graphs (including a set of multimodal information) transmitted from the HII-IUQ module, and through argument extraction, argument graph construction, cross-modal conflict assessment, and dynamic conflict resolution combined with dialectical reasoning, it ultimately forms a final belief to guide subsequent multimodal answer generation or decision-making.

[0092] In one exemplary embodiment of this disclosure, step S130, which involves performing multimodal belief fusion processing based on a cross-modal argument map constructed from a multimodal information set to obtain a fused belief function, includes: constructing a cross-modal argument map based on the multimodal information set, the cross-modal argument map including multiple cross-modal arguments; determining the argument acceptance probability distributions corresponding to the multiple cross-modal arguments, the argument acceptance probability distributions including a single argument acceptance probability distribution where a single argument is accepted, and a combined argument acceptance probability distribution where a combined argument is accepted, the combined argument consisting of multiple compatible arguments; and performing multimodal belief fusion processing based on the argument acceptance probability distributions to obtain the fused belief function.

[0093] The multimodal information set can be a collection of information composed of multiple different modalities. The cross-modal argument map can be a map constructed by structuring the multimodal information set. A cross-modal argument can be an argument composed of multiple different modalities. The argument acceptability probability distribution can be the probability that one or more cross-modal arguments are accepted under different acceptable semantics, and can be represented by a probability distribution. A combined argument can be an argument composed of multiple compatible arguments. The combined argument acceptance probability distribution can be the probability distribution of the combined argument being accepted. Compatible arguments can be multiple mutually compatible arguments. The fusion belief function can be a fusion function constructed by analyzing the consistency between multimodal information and performing belief fusion processing according to cross-modal conflict resolution strategies.

[0094] The MMCD-IF module enables intelligent agents to deeply integrate and dialectically analyze multimodal information (including text, images, audio, etc.), dynamically construct cross-modal cognitive models, and generate reliable arguments and beliefs. Specifically, it extracts structured cross-modal arguments from processed data (such as multimodal information sets) and constructs a weighted cross-modal argument graph, clearly representing the support and attack relationships between different modalities; it introduces a multimodal computational argumentation framework, combining dialectical reasoning to evaluate the acceptability of arguments and resolve cross-modal conflicts; and it designs a dynamic fusion mechanism based on conflict scoring, allowing the agent to choose to directly fuse beliefs or generate qualified answers based on the degree of conflict, thereby significantly improving the agent's accuracy and robustness in handling multimodal uncertainty and conflict tasks.

[0095] In one exemplary embodiment of this disclosure, constructing a cross-modal argument graph based on a multimodal information set includes: extracting cross-modal arguments based on the multimodal information set, wherein the cross-modal arguments include core claims and multimodal evidence supporting the core claims; determining the relationship strength weight between any two cross-modal arguments, wherein the relationship strength weight is determined based on one or more of the attack-support relationship, semantic consistency, and evidence reliability between the two cross-modal arguments; and constructing a cross-modal argument graph based on multiple cross-modal arguments and relationship strength weights.

[0096] The core claim can be the insights presented in a cross-modal argument. Multimodal evidence can be evidence used to support the core claim, and can consist of evidence from multiple modalities (text, images, audio, etc.). Relationship strength weight can be a weight reflecting the degree of attack or support relationship between two cross-modal arguments; it measures the credibility of the relationship between the arguments. Attack-support relationship can be an indicator reflecting the attack and support relationships between two cross-modal arguments. Semantic consistency can be an indicator reflecting whether the semantics of two cross-modal arguments are consistent. Evidence reliability can be an indicator reflecting whether the evidence used to support the cross-modal argument is reliable.

[0097] When the HII-IUQ module determines that the uncertainty of the current user's intent has been reduced to an acceptable level (i.e., the updated intent uncertainty value is less than or equal to a preset threshold), the MMCD-IF module can deeply analyze the consistency between multimodal information during the process of deeply integrating the existing information that the uncertainty of the current user's intent has been reduced to an acceptable level. It can also handle potential conflicts by simulating the multimodal dialectical reasoning process, and finally form a more comprehensive and reliable cross-modal cognitive foundation to guide subsequent answer generation or decision-making.

[0098] Specifically, the output of the HII-IUQ module is used as the input of the MMCD-IF module, such as the multimodal information set D′ after preliminary intent clarification. processed Multimodal information set D′ processed It includes text fragments, images, audio, and other information collected from various sources, as well as related dynamic intent graphs G. intent .

[0099] Perform cross-modal argument formalization and graph construction steps. First, extract cross-modal arguments from the multimodal information set D′. processed The system automatically identifies and extracts structured cross-modal arguments. Each cross-modal argument argi∈Narg; where Narg is a set of cross-modal arguments. Each cross-modal argument argi typically contains a core claim (ci) and evidence (ei) supporting that claim. The evidence ei can be fragments of different modalities such as text, images, and audio.

[0100] Then, the cross-modal relationship identification and weighting steps are performed. This is done from the multimodal information set D′. processed The system automatically identifies and extracts structured cross-modal argument argi, where each argument contains a core claim (ci) and supporting evidence (ei) from different modalities (text, image, audio, etc.). This constitutes multimodal evidence. Simultaneously, it identifies attack and support relationships between any two cross-modal arguments and assigns a strength weight W(rel) to the relationship between them, considering factors such as semantic consistency and evidence reliability. These cross-modal arguments and their relationships are then integrated into a weighted directed graph, namely the Cross-Modal Argument Graph GAMM = (N... arg E re( ); where N arg It can be a set of multiple cross-modal arguments, E re( This can be the relationship between cross-modal arguments. By extracting structured multimodal arguments from the processed data and constructing a weighted cross-modal argument graph, the support and attack relationships between different modal arguments can be clearly represented.

[0101] In one exemplary embodiment of this disclosure, determining the argument acceptance probability distribution corresponding to multiple cross-modal arguments includes: obtaining the relationship strength weight between each cross-modal argument; determining the acceptance probability distribution of each cross-modal argument under different acceptance semantics based on the relationship strength weight; and determining the individual argument acceptance probability distribution and the combined argument acceptance probability distribution of each cross-modal argument under different acceptability semantics based on the acceptance probability distribution.

[0102] Here, acceptability semantics can refer to the acceptability of a certain cross-modal argument, and acceptability semantics can include, but are not limited to, basic semantics, complete semantics, and priority semantics. The acceptance probability distribution of a single argument can be the probability distribution of a single cross-modal argument being accepted under different acceptability semantics, which can be understood as the reliability of that single argument. The acceptance probability distribution of a combined argument can be the probability distribution of a combined cross-modal argument being accepted, which can be understood as the reliability of the combined argument.

[0103] The system performs probabilistic multimodal argumentation and acceptability assessment steps. Based on the constructed cross-modal argument graph (GAMM), the system applies a probabilistic multimodal argumentation framework, which evaluates the probability distribution of the arguments themselves and their possible relationships based on the relation strength weight W(rel).

[0104] Based on the relation strength weight and in conjunction with the selected acceptability semantics (such as basic semantics, complete semantics, and priority semantics), the arg for each argument is calculated. i Belongs to a certain acceptable extension E xttype The probability P(arg) i ∈E xttype This aims to identify a set of arguments that can coexist under uncertainty and quantify their likelihood of occurrence. It yields the probability distribution of acceptance for each cross-modal argument under different acceptable semantics, or the probability distribution of acceptance for multiple combined arguments.

[0105] Taking the query "How to make a cake in an oven and provide some recipes" as an example, suppose the Agent finds three arguments: Argument 1: Recipe A says the cake needs 200 grams of flour; Argument 2: Recipe B says the cake needs 180 grams of flour; Argument 3: An ingredient list image shows that the cake needs 150 grams of flour.

[0106] After constructing the Cross-Modal Argument Graph (GAMM), it was found that Argument 1 and Argument 2 are in an attack relationship because of the different amounts of flour used. The Agent will assign a weight to the attack relationship between the two recipes based on their source (e.g., one from a well-known baking website, the other from a regular blog). For example, if the information from a well-known baking website is more reliable, the attack weight of the argument corresponding to that source will be lower, and vice versa.

[0107] Argument 2 and Argument 3 are also in an attack relationship, but the amount of flour used is different. The system will use a probabilistic multimodal argumentation framework, which calculates the probability of each argument being "accepted" based on these weighted relationships between arguments and in conjunction with acceptability semantics.

[0108] This framework analyzes all arguments and their relationships, considering which arguments can coexist without contradiction (forming an "acceptable extension"), and calculates the probability of this combination occurring. Higher-weighted attack relationships reduce the probability that related arguments are simultaneously accepted. The system ultimately obtains a probability distribution. For example, a probability distribution with P(argument 1 accepted) = 0.3; P(argument 2 accepted) = 0.6; P(argument 3 accepted) = 0.1, and finally calculates that each argument belongs to a certain acceptable extension (E). xttype The framework can also output the acceptability probability distribution corresponding to combined arguments, i.e., a set of multiple probabilistic acceptability extensions. The above steps, combined with dialectical reasoning, assess the acceptability of arguments and resolve cross-modal conflicts.

[0109] In one exemplary embodiment of this disclosure, multimodal belief fusion processing is performed based on the argument acceptance probability distribution to obtain a fused belief function, including: determining the comprehensive argument corresponding to multiple cross-modal arguments and unresolved conflict information; and constructing the fused belief function by performing multimodal belief fusion processing based on the argument acceptance probability distribution, the comprehensive argument, and the unresolved conflict information.

[0110] Among these, a comprehensive argument can be a multimodal understanding generated from multiple cross-modal arguments when facing complex conflicts, capable of explaining the source of the conflict and reconciling different modal perspectives. Unresolved conflict information can be information that clearly points out the points of contention, the arguments of different modal viewpoints, and the strength of evidence when the conflict cannot be completely resolved, in order to generate a final answer.

[0111] Through multi-probabilistic multimodal argumentation and acceptability assessment steps, the calculated argument acceptability probability distribution can include a set of probabilistically acceptable extensions, as well as the probability distribution of each argument being accepted. After obtaining the above argument acceptability probability distribution, this disclosure will integrate beliefs by designing cross-modal conflict resolution strategies. Specifically, these include: modal advantage adoption, multimodal dialectical synthesis, and restricted presentation.

[0112] Modal advantage adoption refers to prioritizing the adoption of argument expansions supported by modal information that is more reliable or relevant. Modal advantage adoption can be based on the probability distribution of argument acceptability, prioritizing the adoption of corresponding arguments.

[0113] Multimodal dialectical synthesis generates comprehensive cross-modal arguments. For complex conflicts, the multimodal understanding and generation capabilities of Large Language Models (LLMs) are utilized to synthesize a comprehensive argument that can explain the source of the conflict and reconcile different modal perspectives.

[0114] A regulated presentation. If the conflict cannot be completely resolved, the points of contention, the arguments of different modalities, and the strength of evidence should be clearly indicated when generating the final answer.

[0115] Finally, the processing results are integrated into a structured fusion belief index, Bel, through a fusion belief function. final_MM Specifically, as shown in Formula 8, the Agent will invoke its multimodal generation module to produce the final multimodal answer or execute the corresponding decision.

[0116] Bel final_MM =F fuse_MM ({P(arg i ∈E xttype SynthesisArguments MM UnresolvedConflicts MM ) (Formula 8)

[0117] Among them, F {use_MM It is a fusion belief function, a multimodal fusion function that comprehensively considers the acceptance probability of each argument, the synthesized reconciled argument, and unresolved conflict information; Bel final_MM It can be a fusion belief index; P(arg) i ∈E xttype ) can be each argument's arg i Belongs to a certain acceptable extension E xttype The probability of; Synthesis Arguments MM It can be a comprehensive argument, Unresolved Conflicts MM This can be unresolved conflict information. By designing a dynamic fusion mechanism based on conflict scores, the agent can choose to directly fuse beliefs or generate limited answers according to the degree of conflict, thereby significantly improving the agent's accuracy and robustness in handling multimodal uncertainty and conflict tasks.

[0118] As can be seen from the above processing procedure, after receiving the current query content, the HII-IUQ module of this disclosure sequentially performs intent identification and hierarchicalization, intent graph construction and updating, and intent uncertainty quantification. Finally, based on the uncertainty level, it decides whether to proactively clarify the intent or pass the current intent state to the MMCD-IF module. The HII-IUQ module enables the Agent to build and maintain a dynamic intent graph of the user query, inferring user intent from multiple granularities and levels, accurately assessing the confidence level of intent at different levels, identifying intent ambiguities, and formulating an optimized information exploration strategy based on this. This clarifies the user's true needs in the most efficient way, reducing overall intent uncertainty.

[0119] The MMCD-IF module introduces dynamic task planning and adaptive tool orchestration, enabling the agent to dynamically decompose and plan complex tasks in multiple steps based on its continuously updated intent understanding and task status. This module intelligently evaluates the applicability and expected effects of available tools, optimizes tool call sequences, and performs real-time replanning based on feedback during execution, thereby significantly improving the agent's execution efficiency, robustness, and adaptability in complex and open tasks.

[0120] This disclosure aims to significantly improve the robustness, accuracy, and intelligence of Agentic RAG systems when handling real-world tasks fraught with ambiguity, uncertainty, and multi-step execution by endowing agents with the ability to perform multi-layered intent understanding and dynamic planning for complex tasks. This will provide solid technical support for cutting-edge AI applications such as advanced question answering systems, automated research assistants, and complex decision support systems, and drive them towards deeper levels of intelligent interaction and knowledge discovery.

[0121] In summary, the content generation method disclosed herein obtains the current query content, and based on the current query content, historical interaction records, and the set of current retrieval information generated in the current round, performs intent uncertainty quantification processing on the current query intent to obtain an intent uncertainty value. When the intent uncertainty value is less than or equal to a preset threshold, it obtains a multimodal information set obtained after intent uncertainty quantification processing. Based on the cross-modal argument graph constructed based on the multimodal information set, it performs multimodal belief fusion processing to obtain a fused belief function. Based on the fused belief function, it generates the target response content corresponding to the current query content. On the one hand, by accurately assessing the confidence level of the current query intent to determine the true intent demand, the overall intent uncertainty can be reduced. On the other hand, by deeply analyzing the consistency between multimodal information, a more comprehensive and reliable cross-modal cognitive foundation is formed to guide subsequent content generation or decision-making, which can improve the execution efficiency, robustness, and adaptability for complex tasks.

[0122] It should be noted that although the steps of the method in this invention are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0123] Furthermore, in this example embodiment, a content generation apparatus is also provided. (See reference) Figure 5 The content generation device 500 may include: an intent determination module 510, an information set determination module 520, a fusion function determination module 530, and a content generation module 540.

[0124] Specifically, the intent determination module 510 is used to obtain the current query content, and based on the current query content, historical interaction records, and the current search information set generated in the current round, to perform intent uncertainty quantification processing on the current query intent to obtain an intent uncertainty value; the information set determination module 520 is used to obtain the multimodal information set obtained after intent uncertainty quantification processing when the intent uncertainty value is less than or equal to a preset threshold; the fusion function determination module 530 is used to perform multimodal belief fusion processing based on the cross-modal argument graph constructed based on the multimodal information set to obtain a fused belief function; and the content generation module 540 is used to generate the target response content corresponding to the current query content based on the fused belief function.

[0125] In one exemplary embodiment of this disclosure, the intent determination module 510 includes an intent determination unit, configured to: construct a dynamic intent graph based on the current query content and the current search information set; determine the Shannon entropy, intent ambiguity, and confidence uncertainty index corresponding to the dynamic intent graph; and determine the intent uncertainty value based on the Shannon entropy, intent ambiguity, and confidence uncertainty index.

[0126] In one exemplary embodiment of this disclosure, the intent determination unit includes an intent graph construction subunit, which is used to perform information decomposition processing on the current query content and the current search information set to obtain multiple intent nodes, each intent node having a corresponding node confidence level; determine the node relationship between each intent node; construct an initial dynamic intent graph based on the multiple intent nodes and node relationships; and update the intent confidence level of the initial dynamic intent graph to obtain the dynamic intent graph.

[0127] In one exemplary embodiment of this disclosure, the intent determination module 510 further includes an information set update unit, configured to: when the intent uncertainty value is greater than or equal to a preset threshold, acquire an exploration action set, the exploration action set including multiple exploration actions; determine the intent information gain corresponding to each exploration action after performing an operation; select a target exploration action from the multiple exploration actions according to the intent information gain, the target exploration action having the largest intent information gain; generate updated query content based on the target exploration action, the intent uncertainty value corresponding to the updated query intent of the updated query content being less than or equal to the preset threshold; and perform an update operation on the current search information set according to the updated query content to obtain an updated search information set.

[0128] In one exemplary embodiment of this disclosure, the fusion function determination module 530 includes a fusion function determination unit, configured to: construct a cross-modal argument map based on a multimodal information set, the cross-modal argument map including multiple cross-modal arguments; determine the argument acceptance probability distribution corresponding to the multiple cross-modal arguments, the argument acceptance probability distribution including a single argument acceptance probability distribution where a single argument is accepted, and a combined argument acceptance probability distribution where a combined argument is accepted, the combined argument consisting of multiple compatible arguments; and perform multimodal belief fusion processing based on the argument acceptance probability distribution to obtain a fused belief function.

[0129] In one exemplary embodiment of this disclosure, the fusion function determination unit includes a cross-modal graph construction subunit, configured to: extract cross-modal arguments based on a multimodal information set, wherein the cross-modal arguments include a core claim and multimodal evidence supporting the core claim; determine the relationship strength weight between any two cross-modal arguments, wherein the relationship strength weight is determined based on one or more of the attack-support relationship, semantic consistency, and evidence reliability between the two cross-modal arguments; and construct a cross-modal argument graph based on multiple cross-modal arguments and relationship strength weights.

[0130] In one exemplary embodiment of this disclosure, the fusion function determination unit includes a probability distribution determination subunit, configured to: obtain the relationship strength weights between each cross-modal argument; determine the acceptability probability distribution of each cross-modal argument under different acceptability semantics based on the relationship strength weights; and determine the individual argument acceptability probability distribution and the combined argument acceptability probability distribution of each cross-modal argument under different acceptability semantics based on the acceptability probability distribution.

[0131] In one exemplary embodiment of this disclosure, the fusion function determination unit includes a fusion function determination subunit, which is used to: determine the comprehensive argument corresponding to multiple cross-modal arguments and unresolved conflict information; and perform multimodal belief fusion processing based on the argument acceptance probability distribution, the comprehensive argument and the unresolved conflict information to construct a fusion belief function.

[0132] The specific details of the virtual modules of each content generation device mentioned above have been described in detail in the corresponding content generation methods. For any undisclosed details, please refer to the implementation methods in the method section, and therefore will not be repeated here.

[0133] It should be noted that although several modules or units of the content generation apparatus have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0134] Exemplary embodiments of this disclosure also provide a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the above-described content generation method.

[0135] In one implementation, the computer program product may be a tangible product containing a computer program, such as a computer-readable storage medium storing the computer program. (See reference...) Figure 6 , Figure 6 The schematic diagram illustrates a computer-readable storage medium 600 according to an exemplary embodiment of the present disclosure. The computer-readable storage medium 600 can be a storage medium based on electrical, magnetic, optical, electromagnetic, infrared, or other signals, including but not limited to: random access memory (RAM), read-only memory (ROM), magnetic tape, floppy disk, flash memory, hard disk drive (HDD), solid-state drive (SSD), etc. Exemplarily, a computer program product can be implemented as a non-volatile storage medium storing a computer program, such as read-only memory, NAND flash memory, etc.

[0136] In one implementation, the computer program product can be an intangible product containing a computer program. For example, the computer program product can be implemented as a virtual digital product, such as an executable file, installation package, or other digital file storing the computer program.

[0137] Computer program code can be written in one or more programming languages. Examples of programming languages ​​include C, Java, and C++. Program code can execute entirely on the user's computing device, partially on the user's computing device, or as a standalone software package. It can also execute partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, such as a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via an internet connection provided by a mobile network operator).

[0138] Computer programs can be carried or transmitted via signals such as electrical, magnetic, optical, electromagnetic, and infrared rays. Electronic devices can convert the signals carrying computer programs into digital signals, thereby running the computer programs. When a computer program runs on an electronic device, its code is used to cause the electronic device to execute (more specifically, to execute) the method steps of various exemplary embodiments of this disclosure, such as the content generation method described above.

[0139] Exemplary embodiments of this disclosure also provide an electronic device, which may include a processor and a memory. The memory stores executable instructions of the processor, such as a computer program. The processor executes the executable instructions to perform the method steps of various exemplary embodiments of this disclosure. Furthermore, the electronic device may also include a display for displaying a graphical user interface.

[0140] The following is for reference. Figure 7 The electronic device is illustrated by way of a general-purpose computing device. It should be understood that... Figure 7 The electronic device 700 shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments disclosed herein.

[0141] like Figure 7 As shown, the electronic device 700 may include: a processor 710, a memory 720, a bus 730, an I / O (input / output) interface 740, a network adapter 750, and a display 760.

[0142] The memory 720 may include volatile memory, such as RAM 721 and cache unit 722, and may also include non-volatile memory, such as ROM 723. The memory 720 may also include one or more program modules 724, including but not limited to: an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. For example, program module 724 may include the modules described above.

[0143] The processor 710 may include one or more processing units, such as an AP (Application Processor), a modem processor, a GPU (Graphics Processing Unit), an ISP (Image Signal Processor), a controller, an encoder, a decoder, a DSP (Digital Signal Processor), a baseband processor, and / or an NPU (Neural-Network Processing Unit).

[0144] The processor 710 can be used to execute executable instructions stored in the memory 720, such as the above-mentioned content generation method.

[0145] Bus 730 is used to connect different components of electronic device 700 and may include a data bus, an address bus and a control bus.

[0146] Electronic device 700 can communicate with one or more external devices 800 (such as keyboard, mouse, external controller, etc.) through I / O interface 740.

[0147] Electronic device 700 can communicate with one or more networks via network adapter 750. For example, network adapter 750 can provide mobile communication solutions such as 3G / 4G / 5G, or wireless communication solutions such as wireless LAN, Bluetooth, and near-field communication. Network adapter 750 can communicate with other modules of electronic device 700 via bus 730.

[0148] Electronic device 700 can display a graphical user interface via display 760, such as an interface that displays the answer information to a user's question.

[0149] although Figure 7 As not shown in the diagram, other hardware and / or software modules may also be configured in the electronic device 700, including but not limited to: a display (if the electronic device already includes a display, then "display" is removed here), microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0150] As can be seen from the above, the technical solutions disclosed herein can be implemented as methods, apparatus, systems, computer program products, storage media, electronic devices, etc. Those skilled in the art will understand that various aspects of this disclosure can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, which may be referred to as "circuit," "module," or "system," respectively.

[0151] It should be understood that this disclosure is not limited to the specific methods, steps, or structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. Those skilled in the art will readily conceive of other embodiments based on the specific implementations provided in this disclosure. Therefore, the specific implementations provided in this disclosure are merely exemplary, and the scope and spirit of this disclosure are indicated by the claims, and should cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary technical means in the art not disclosed in this disclosure.

Claims

1. A content generation method, characterized in that, include: Obtain the current query content, and based on the current query content, historical interaction records, and the set of current search information generated in the current round, perform intent uncertainty quantification processing on the current query intent to obtain the intent uncertainty value; When the intention uncertainty value is less than or equal to a preset threshold, a multimodal information set obtained after the intention uncertainty quantification process is acquired. Based on the cross-modal argument graph constructed from the multimodal information set, multimodal belief fusion processing is performed to obtain the fused belief function; Based on the fusion belief function, the target response content corresponding to the current query content is generated.

2. The method according to claim 1, characterized in that, The step of quantifying the intent uncertainty of the current query intent based on the current query content, historical interaction records, and the set of current retrieval information generated in the current round, to obtain an intent uncertainty value, includes: Based on the current query content and the current search information set, a dynamic intent graph is constructed; Determine the Shannon entropy, intent ambiguity, and confidence uncertainty index corresponding to the dynamic intent graph; The intention uncertainty value is determined based on the Shannon entropy, the intention ambiguity, and the confidence uncertainty index.

3. The method according to claim 2, characterized in that, The step of constructing a dynamic intent graph based on the current query content and the current retrieval information set includes: The current query content and the current search information set are decomposed to obtain multiple intent nodes, each of which has a corresponding node confidence level. Determine the node relationships between each of the stated intent nodes; Based on the relationships between multiple intent nodes and the nodes, an initial dynamic intent graph is constructed; The initial dynamic intent graph is updated with intent confidence to obtain the dynamic intent graph.

4. The method according to claim 1 or 2, characterized in that, The method further includes: When the uncertainty value of the intent is greater than the preset threshold, an exploration action set is obtained, which includes multiple exploration actions; Determine the intent information gain corresponding to each of the aforementioned exploration actions after the operation is performed; Based on the intent information gain, a target exploration action is selected from the plurality of exploration actions, and the target exploration action corresponds to the largest intent information gain. Based on the target exploration action, an updated query content is generated, and the intent uncertainty value corresponding to the updated query intent is less than or equal to the preset threshold. The current search information set is updated based on the updated query content to obtain the updated search information set.

5. The method according to claim 1, characterized in that, The process of performing multimodal belief fusion processing based on the cross-modal argument graph constructed based on the multimodal information set to obtain a fused belief function includes: A cross-modal argument map is constructed based on the multimodal information set, and the cross-modal argument map includes multiple cross-modal arguments; Determine the argument acceptance probability distribution corresponding to multiple cross-modal arguments, the argument acceptance probability distribution including the single argument acceptance probability distribution of a single argument being accepted, and the combined argument acceptance probability distribution of a combined argument being accepted, the combined argument being composed of multiple compatible arguments; Based on the probability distribution of the acceptance of the argument, a multimodal belief fusion process is performed to obtain the fused belief function.

6. The method according to claim 5, characterized in that, The construction of a cross-modal argument graph based on the multimodal information set includes: Cross-modal arguments are extracted based on the multimodal information set, and the cross-modal arguments include core claims and multimodal evidence supporting the core claims; Determine the strength weight of the relationship between any two of the cross-modal arguments, the strength weight being determined based on one or more of the attack support relationship, semantic consistency, and evidence reliability between the two cross-modal arguments; The cross-modal argument graph is constructed based on multiple cross-modal arguments and the relationship strength weights.

7. The method according to claim 5, characterized in that, Determining the argument acceptance probability distribution corresponding to the multiple cross-modal arguments includes: Obtain the strength weights of the relationships between the various cross-modal arguments; Based on the relation strength weights, determine the acceptability probability distribution of each of the cross-modal arguments under different acceptability semantics; Based on the acceptance probability distribution, determine the acceptance probability distribution of individual arguments and the acceptance probability distribution of combined arguments under different acceptable semantics for each of the cross-modal arguments.

8. The method according to claim 5, characterized in that, The multimodal belief fusion process based on the argument acceptance probability distribution, to obtain the fused belief function, includes: Identify the comprehensive arguments corresponding to the multiple cross-modal arguments, as well as information on unresolved conflicts; Based on the probability distribution of the acceptance of the argument, the comprehensive argument, and the unresolved conflict information, multimodal belief fusion processing is performed to construct the fused belief function.

9. A content generation device, characterized in that, include: The intent determination module is used to obtain the current query content, and perform intent uncertainty quantification processing on the current query intent based on the current query content, historical interaction records and the current search information set generated in the current round, to obtain the intent uncertainty value; The information set determination module is used to obtain the multimodal information set after the intent uncertainty quantification process when the intent uncertainty value is less than or equal to a preset threshold. The fusion function determination module is used to perform multimodal belief fusion processing based on the cross-modal argument map constructed based on the multimodal information set to obtain the fused belief function; The content generation module is used to generate target response content corresponding to the current query content based on the fusion belief function.

10. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the content generation method as described in any one of claims 1 to 8.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the content generation method according to any one of claims 1 to 8.

Images