A method and system for intelligent dialogue based on intention recognition based on large model

Abstract

The present invention discloses an intention recognition intelligent dialogue method and system based on a large model, which relates to the technical field of dialogue processing and is used to solve the problem that multiple conflicting intentions in multi-round dialogues cannot be efficiently recognized and processed in the prior art; by using a pre-trained large-scale language model to perform semantic analysis on input data in multi-round dialogues, multiple preliminary intentions of users are extracted; an intention conflict detection mechanism is established based on a predefined intention conflict rule library, the extracted preliminary intentions are analyzed, and logical conflict intention combinations are screened; the degree of opposition of logical conflict intention combinations in logical mutual exclusion and fuzzy overlap is evaluated through semantic graphs and graph neural networks, the dynamic relationship entropy of logical conflict intention combinations is analyzed, and the degree of fluctuation of semantic dependence in multi-round dialogues is evaluated, and finally the processing order of logical conflict intention combinations is dynamically adjusted, thereby improving the accuracy of intention conflict processing and the rationality of priority allocation.

Description

Technical Field

[0001] The present invention relates to the field of dialog processing technology, and more specifically, to an intention recognition intelligent dialog method and system based on a large model. Background Art

[0002] In natural language processing and dialogue systems, intent recognition is one of the core technologies for realizing intelligent dialogue. By recognizing the intent of user input, the system can understand user needs and provide corresponding services. However, in multi-round dialogue scenarios, user expressions are often complex and diverse, and may contain multiple interrelated or conflicting intents, which places higher requirements on the system's parsing and decision-making capabilities.

[0003] Existing technologies usually rely on semantic parsing models to extract user intentions and perform intent conflict detection through predefined rules or simple logical judgments. They cannot efficiently identify and process multiple conflicting intentions of users in multi-round conversations. This will limit the effectiveness of intent parsing and response decision-making, making it difficult to meet the needs of complex conversation scenarios. This will lead to unreasonable priority allocation in the system during intent conflict processing, affecting the continuity of the conversation and the accuracy of the response. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides an intention recognition intelligent dialogue method and system based on a large model to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An intention recognition intelligent dialogue method based on a large model includes the following steps:

[0007] Obtain the user's input data in multiple rounds of conversations, use the pre-trained large-scale language model to perform semantic analysis on the input data, and extract multiple preliminary intentions of the user;

[0008] An intent conflict detection mechanism is established based on a predefined intent conflict rule library, and the extracted preliminary intents are analyzed based on the intent conflict detection mechanism to screen out logical conflict intent combinations;

[0009] The semantic relationship of the logical conflict intention combination is analyzed through semantic graph and graph neural network to evaluate the degree of opposition of the logical conflict intention combination in logical mutual exclusion and fuzzy overlap.

[0010] The dynamic relation entropy of logical conflict intention combinations is analyzed to evaluate the fluctuation degree of semantic dependency of logical conflict intention combinations in multiple rounds of dialogues.

[0011] Based on the degree of opposition of logically conflicting intention combinations in logical mutual exclusion and fuzzy overlap and the degree of fluctuation of semantic dependency in multi-round dialogues, the processing order of logically conflicting intention combinations is dynamically adjusted.

[0012] In a preferred embodiment, the input data of the user in multiple rounds of dialogue is obtained, and the input data is semantically parsed using a pre-trained large-scale language model to extract multiple preliminary intentions of the user, specifically:

[0013] By receiving input data from the user during the conversation;

[0014] Preprocess the received input data, including denoising, formatting and data cleaning;

[0015] Use the pre-trained large-scale language model to perform semantic analysis on the pre-processed input data and extract the key semantic information of the user input, which includes the user's expression purpose, content focus, and potential additional requirements;

[0016] Through the key semantic information extracted after semantic parsing, multiple preliminary intentions of the user in the current conversation turn are identified, and a semantic representation vector corresponding to each intention is generated.

[0017] In a preferred embodiment, an intention conflict detection mechanism is established based on a predefined intention conflict rule library, and the extracted preliminary intentions are analyzed based on the intention conflict detection mechanism to screen out logical conflict intention combinations, specifically:

[0018] Construct an intention conflict rule base, which includes logical mutual exclusion relationships, conditional dependency relationships, and priority relationships;

[0019] Match the extracted preliminary intent with the rules in the intent conflict rule library, and determine whether there is a conflict between the preliminary intents through a matching algorithm;

[0020] Screening the conflicting intentions in the matching results, including extracting the intention combinations that satisfy the logical mutual exclusion relationship, and screening out the logical conflicting intention combinations that satisfy the conflict conditions in combination with the conditional dependency relationship;

[0021] Among them, if two intentions satisfy the logical mutual exclusion relationship and dependency conditions at the same time and meet the priority constraints, they are marked as a logical conflict intention combination.

[0022] In a preferred embodiment, the semantic relationship of the logical conflict intention combination is analyzed by using a semantic graph and a graph neural network to evaluate the degree of opposition of the logical conflict intention combination in logical mutual exclusion and fuzzy overlap, specifically:

[0023] Establish a semantic graph, where the nodes of the semantic graph represent the various intentions in the combination of logically conflicting intentions, and the edges represent the semantic relationships between the intentions. The semantic relationships include logical mutually exclusive relationships and fuzzy overlapping relationships.

[0024] The constructed semantic graph is input into the graph neural network, and features of nodes and edges in the semantic graph are extracted to obtain high-dimensional feature data representing the relationship between various intentions.

[0025] Each pair of intentions in the logically conflicting intention combination is analyzed based on high-dimensional feature data, and the degree of opposition of the logically conflicting intention combination in logical mutual exclusion and fuzzy overlap is evaluated.

[0026] In a preferred embodiment, each pair of intentions in the logical conflict intention combination is analyzed based on high-dimensional feature data to evaluate the degree of opposition of the logical conflict intention combination in logical mutual exclusion and fuzzy overlap, specifically:

[0027] Calculate the similarity of each pair of high-dimensional feature vectors of logical conflicting intentions: ;in, Express intention and intention The semantic similarity of and Intention and intention The characteristic vector of and Intention and intention The norm of the corresponding vector;

[0028] Calculating intent using Gaussian kernel function and intention The fuzzy overlap of: ;in, Express intention and intention The fuzzy overlap of Express intention and intention The Euclidean distance between the eigenvectors of is the adjustment parameter of the Gaussian kernel;

[0029] After the semantic similarity and fuzzy overlap were dimensionlessly processed and assigned weight coefficients respectively, the opposition degree index was calculated.

[0030] In a preferred embodiment, the dynamic relation entropy of the logical conflict intention combination is analyzed to evaluate the fluctuation degree of the semantic dependency of the logical conflict intention combination in multiple rounds of dialogue, specifically:

[0031] In multiple rounds of dialogue, based on the contextual relationships corresponding to the logical conflicting intention combinations, a semantic similarity matrix is ​​generated for the conflicting intention pairs in each round of dialogue. The elements of the semantic similarity matrix represent the semantic association strength between each pair of conflicting intentions.

[0032] Map the change relationship of each round of semantic similarity matrix in multiple rounds of dialogue into information entropy value;

[0033] The information entropy values ​​generated in multi-round conversations are modeled using a temporal adaptive graph neural network, and the change pattern of the information entropy values ​​in the time dimension is analyzed to generate a dynamic relationship fluctuation score to evaluate the degree of fluctuation in the semantic dependency of logical conflict intention combinations.

[0034] In a preferred embodiment, the change relationship of the semantic similarity matrix of each round in multiple rounds of dialogue is mapped to the information entropy value, specifically:

[0035] The difference in the semantic similarity matrix in each round of dialogue is quantified as a change matrix: ;in, Express intention and intention In the and The change in semantic similarity in a turn-based dialogue, and Respectively and Semantic similarity in turn-based dialogues, Represents the time turn index in a multi-turn conversation;

[0036] Calculate the semantic relationship information entropy of each round of dialogue. The formula is as follows: ;in, Indicates The semantic relationship information entropy of the turn-based dialogue, is the normalized probability value of the elements in the change matrix.

[0037] In a preferred embodiment, based on the degree of opposition between logical conflicting intention combinations in logical mutual exclusion and fuzzy overlap and the degree of fluctuation of semantic dependency in multiple rounds of dialogue, the processing order of logical conflicting intention combinations is dynamically adjusted, specifically:

[0038] Set the preset threshold of the opposition index to ; Set the preset threshold of dynamic relationship fluctuation score to ;

[0039] when and , then the logical conflict intention combination is set to the highest priority;

[0040] when and , then the logical conflict intention combination is set to medium priority;

[0041] when and , then the logical conflict intention combination is set to medium priority;

[0042] when and , then the logical conflict intention combination is set to the lowest priority;

[0043] in, is the opposition index, Represents the dynamic relationship fluctuation score.

[0044] On the other hand, the present invention provides an intention recognition intelligent dialogue system based on a large model, including a semantic parsing and extraction module, an intention conflict detection module, an opposition degree assessment module, a semantic dependency analysis module, and a processing order adjustment module;

[0045] Semantic parsing and extraction module: obtains the user's input data in multiple rounds of dialogue, uses the pre-trained large-scale language model to perform semantic parsing on the input data, and extracts multiple preliminary intentions of the user;

[0046] Intent conflict detection module: establishes an intent conflict detection mechanism based on a predefined intent conflict rule base, analyzes the extracted preliminary intents based on the intent conflict detection mechanism, and screens out logical conflict intent combinations;

[0047] Opposition degree assessment module: Analyze the semantic relationship of logical conflict intention combinations through semantic graphs and graph neural networks, and assess the opposition degree of logical conflict intention combinations in logical mutual exclusion and fuzzy overlap;

[0048] Semantic dependency analysis module: Analyzes the dynamic relationship entropy of logical conflict intention combinations and evaluates the degree of fluctuation of semantic dependency of logical conflict intention combinations in multiple rounds of dialogue;

[0049] Adjust processing order module: Dynamically adjust the processing order of logically conflicting intention combinations based on the degree of opposition between logically conflicting intention combinations in logical mutual exclusion and fuzzy overlap and the degree of fluctuation of semantic dependency in multiple rounds of dialogue.

[0050] The technical effects and advantages of the large-model-based intention recognition intelligent dialogue method and system of the present invention are as follows:

[0051] 1. The present invention uses a pre-trained large-scale language model to perform semantic analysis on the user's input data in multiple rounds of conversations, and can efficiently extract multiple preliminary intentions from complex user expressions, and establish an intention conflict detection mechanism based on a predefined intention conflict rule library to analyze and screen the preliminary extracted intentions, so as to accurately obtain logical conflict intention combinations; this method of combining semantic analysis and conflict rule detection not only improves the accuracy and efficiency of intent extraction, but also provides a high-quality data foundation for subsequent conflict analysis and processing, effectively solving the problem of insufficient intent conflict detection capabilities in the prior art.

[0052] 2. The degree of opposition of logically conflicting intention combinations in logical mutual exclusion and fuzzy overlap is evaluated through semantic graphs and graph neural networks, and the degree of fluctuation of semantic dependencies in multi-round conversations is analyzed in combination with dynamic relationship entropy, so as to comprehensively judge the complexity and dynamic characteristics of intention conflicts, dynamically adjust the processing order of logically conflicting intention combinations, and give priority to high-priority intentions. This significantly improves the rationality and adaptability of intention conflict processing, ensures the coherence of multi-round conversations and the accuracy of responses, solves the problem of unreasonable priority allocation in the existing technology, and provides a more efficient solution for complex conversation scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A schematic diagram of an intention recognition intelligent dialogue method based on a large model of the present invention;

[0054] Figure 2 This is a structural diagram of an intention recognition intelligent dialogue system based on a large model of the present invention. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] Embodiment 1: Figure 1 The present invention provides an intention recognition intelligent dialogue method based on a large model, which includes the following steps:

[0057] Obtain the user's input data in multiple rounds of conversations, use the pre-trained large-scale language model to perform semantic analysis on the input data, and extract the user's multiple preliminary intentions.

[0058] An intention conflict detection mechanism is established based on a predefined intention conflict rule library. The extracted preliminary intentions are analyzed based on the intention conflict detection mechanism to screen out logical conflicting intention combinations.

[0059] The semantic relations of logically conflicting intention combinations are analyzed through semantic graphs and graph neural networks to evaluate the degree of opposition between logically conflicting intention combinations in logical mutual exclusion and fuzzy overlap.

[0060] The dynamic relation entropy of logical conflict intention combinations is analyzed to evaluate the fluctuation degree of semantic dependency of logical conflict intention combinations in multiple rounds of dialogue.

[0061] Based on the degree of opposition of logically conflicting intention combinations in logical mutual exclusion and fuzzy overlap and the degree of fluctuation of semantic dependency in multi-round dialogues, the processing order of logically conflicting intention combinations is dynamically adjusted.

[0062] Obtain the user's input data in multiple rounds of conversations, use the pre-trained large-scale language model to perform semantic analysis on the input data, and extract multiple preliminary intentions of the user, specifically:

[0063] By receiving input data from the user during the conversation, the input data includes input text and voice, or multimodal data, wherein the multimodal data includes a combination of voice, image and text:

[0064] Among them, text data is the natural language text information entered by the user through the text input box; voice data is the audio data entered by the user through voice, with a sampling rate of 16kHz and an encoding format of PCM; multimodal data is a combination of text, voice and image information provided by the user, for example, the user describes the needs through voice and attaches pictures as supplementary explanations.

[0065] Preprocess the received input data, including denoising, formatting, and data cleaning, to ensure that the input data meets the requirements of semantic parsing:

[0066] Preprocessing of text data:

[0067] Character cleaning: remove redundant symbols, garbled characters, and meaningless spaces in the text.

[0068] Spelling correction: Use word vector-based spelling correction algorithms (such as edit distance-based optimization methods) to correct spelling errors in text.

[0069] Word segmentation: For continuously input text, the Chinese word segmenter based on the BERT word vector model is used to divide the sentences into independent word fragments and mark the parts of speech.

[0070] Preprocessing of speech data:

[0071] Noise reduction: The spectral subtraction method is used to remove background noise in speech while retaining the integrity of the speech signal.

[0072] Speech-to-text: Use end-to-end speech recognition models (such as Transformer-based ASR systems) to convert speech data into text in real time.

[0073] Preprocessing of multimodal data:

[0074] Image processing: Convert the format of images uploaded by users (such as JPG to PNG) and adjust the size to ensure that the images are suitable for subsequent semantic parsing model input.

[0075] Modality alignment: Synchronize the relevance of text, speech, and images based on timestamps or input order to ensure temporal and semantic consistency of modality data.

[0076] Use the pre-trained large-scale language model to perform semantic analysis on the pre-processed input data and extract the key semantic information of the user input. The key semantic information includes the user's expression purpose, content focus, and potential additional requirements:

[0077] After completing data preprocessing, use a pre-trained large-scale language model (such as Wenxinyiyan) to perform semantic analysis on the preprocessed data. The specific steps of semantic analysis are as follows:

[0078] Concatenate text and speech-to-text data, as well as image description text in multimodal data, into a standardized input sequence.

[0079] The input sequence is encoded through a large-scale language model to extract the semantic features of each paragraph of content. For example, the attention mechanism in the Transformer model is used to capture the contextual information in the input data and the logical relationship between sentences.

[0080] Based on specific task objectives (such as intent recognition), the following key semantic information is filtered out from the semantic features output by the model:

[0081] Express purpose: The overall goal or task of the user input, such as "book a flight" or "check the weather."

[0082] Content focus: The core information that needs to be processed first in the input data, such as "time: tomorrow morning" and "location: Beijing".

[0083] Potential additional needs: Additional intents extracted from the implicit semantics of the context, such as “hope the price is cheap” and “recommend comfortable seats”.

[0084] Through the key semantic information extracted after semantic parsing, multiple preliminary intentions of the user in the current conversation round are identified, and the semantic representation vector corresponding to each intention is generated:

[0085] The extracted key semantic information is input into a classification model (e.g., a deep neural network model based on multi-task learning) to identify possible preliminary intent categories. For example, the main intent is "book a flight ticket" and the additional intents are "recommend comfortable seats" and "select direct flights."

[0086] The semantic vector is extracted through the output layer of the large-scale language model. Each intent corresponds to a high-dimensional vector, and the vector dimension is usually 768 dimensions (BERT-base model). The vector expresses the semantic features of the intent, including contextual relevance and potential semantic information.

[0087] An intent conflict detection mechanism is established based on the predefined intent conflict rule library. The extracted preliminary intent is analyzed based on the intent conflict detection mechanism to screen out logical conflict intent combinations, specifically:

[0088] Build an intention conflict rule base, which includes logical mutual exclusion relationships, conditional dependencies, and priority relationships between common intentions. Logical mutual exclusion relationships are used to describe the contradictions between intentions, conditional dependencies are used to describe the sequential execution conditions of intentions, and priority relationships are used to describe the execution priority order of intentions:

[0089] The intent conflict rule base includes the logical relationships between common intents, which are mainly divided into the following three categories:

[0090] Logical mutual exclusion: defines the contradiction when two intentions cannot be satisfied at the same time. For example, "refund" and "change" are mutually exclusive because they cannot be executed at the same time in the same order.

[0091] Conditional dependencies: define the order constraints of intents, for example, "select seat" must be effective after "change ticket".

[0092] Priority relationship: Define the priority of multiple intents in the same context. For example, "ticket refund" takes precedence over "seat selection" to ensure that the user's core needs are met.

[0093] Use a hierarchical structure to build an intent conflict rule base:

[0094] First level: intent categories, such as “book”, “modify and cancel”.

[0095] The second layer: intention relationships, including mutual exclusion, dependency, and priority.

[0096] The third layer: specific rule details, for example, "refund" and "change" are mutually exclusive.

[0097] In the intention conflict rule, the relationship between each two intentions can be expressed as mutually exclusive or non-mutually exclusive. If two intentions cannot be satisfied at the same time, it is defined as a mutually exclusive relationship; if two intentions can coexist, it is defined as a non-mutually exclusive relationship. For example, refund and change belong to a mutually exclusive relationship because they are logically mutually exclusive and cannot be executed at the same time. On the contrary, seat selection and adding luggage may belong to a non-mutually exclusive relationship because they can be satisfied at the same time. Through the definition of the rule base, the mutually exclusive and non-mutually exclusive relationships between all intentions are clearly listed to guide subsequent conflict detection and processing.

[0098] The rule base is stored as an extensible graph structure, where nodes represent intents and edges represent logical relationships between intents.

[0099] Use dynamic update strategies to continuously expand and improve the rule base based on user feedback or new intents.

[0100] The extracted preliminary intent is matched with the rules in the intent conflict rule base, and the matching algorithm is used to determine whether the preliminary intent conflicts. The matching algorithm includes a layer-by-layer matching algorithm based on the rule tree structure and a conflict judgment algorithm based on graph relationship reasoning:

[0101] Layer-by-layer matching algorithm based on rule tree: The preliminary intentions are input into the rule tree in sequence, starting from the root node, and the categories and relationships in the rule base are matched layer by layer to quickly screen out potential conflicting intentions.

[0102] Conflict judgment algorithm based on graph relationship reasoning:

[0103] Construct a semantic graph, where each node represents a preliminarily extracted intent, and the edges between nodes represent the regular relationships between intents. These regular relationships can include logical mutual exclusion, conditional dependency, or priority relationships.

[0104] In the semantic graph, the conflict score between every two intent nodes is calculated through the message passing mechanism. Message passing refers to the gradual propagation and integration of information between nodes and edges in the semantic graph in combination with rule relationships, and finally the score between nodes is obtained.

[0105] The conflict score between each intent node is determined by the combined impact of multiple rules. The influence of the rules is measured by their weights, and the degree of adaptability of each rule to a specific intent combination also needs to be considered.

[0106] The calculation method of the conflict score can be understood as: multiplying the weights of all rules by their adaptability one by one, then summing up the results of all rules, and finally normalizing them to generate a conflict score between intents.

[0107] A higher conflict score value indicates a higher degree of conflict between the two intents.

[0108] All rules are considered simultaneously, and the impact of each rule on the score is determined by its weight. Therefore, the number of rules and the distribution of their weights will directly affect the final conflict score.

[0109] The input of the preliminary intent includes the category label of the intent and the corresponding context information. For example, the "refund" intent contains the user's order number and the reason for the refund.

[0110] A semantic parser is used to convert the textual representation of preliminary intent into standardized tags that align with the node format in the rule base.

[0111] According to the matching algorithm, the preliminary intention is compared with the rules in the rule base one by one, and all successfully matched rules and corresponding conflict relationships are recorded.

[0112] Filter the conflicting intents in the matching results. The filtering includes extracting the intent combinations that satisfy the logical mutual exclusion relationship, and filtering out the logical conflicting intent combinations that satisfy the conflict conditions in combination with the conditional dependency relationship:

[0113] Filter based on logical mutual exclusion relationships: Extract all intent combinations with logical mutual exclusion relationships, and mark these intent combinations as conflicting intent combinations based on the mutual exclusion conditions defined in the rule base.

[0114] Filter based on conditional dependencies: Eliminate intent combinations that do not meet the dependency conditions. For example, if the completion of one intent depends on another intent, and this order is not met, the combination is considered to not meet the filtering conditions.

[0115] Filter based on priority relationship: For multiple conflicting intent combinations, the intent combinations with higher priorities defined in the rule base are retained first, and the intents with lower priorities are marked as minor conflicts or deferred.

[0116] The filtering results between intents are represented by a filtering matrix. Each element in the filtering matrix indicates whether the two intents meet the filtering conditions.

[0117] If two intentions satisfy both the logical mutual exclusion relationship and dependency conditions and meet the priority constraints, they are marked as a logical conflicting intention combination. Otherwise, they are marked as not meeting the filter conditions.

[0118] The conflicting intentions that meet the conditions are combined and stored and used as input data for subsequent steps (such as dynamically adjusting the processing order).

[0119] Through this screening process, it can be ensured that the screened logical conflicting intention combinations comply with the logical mutual exclusion relationships, dependency conditions and priority constraints defined in the rule base, which facilitates subsequent intention processing and priority adjustment.

[0120] The semantic relationship of the logical conflict intention combination is analyzed through semantic graph and graph neural network to evaluate the degree of opposition of the logical conflict intention combination in logical mutual exclusion and fuzzy overlap, specifically:

[0121] A semantic graph is established. The nodes of the semantic graph represent the intentions in the combination of logically conflicting intentions, and the edges represent the semantic relationships between the intentions. The semantic relationships include logical mutually exclusive relationships and fuzzy overlapping relationships:

[0122] Each node of the semantic graph represents a specific intention in the logical conflict intention combination. For example, in the logical conflict intention combination "refund" and "change", "refund" and "change" are used as nodes respectively.

[0123] Each node not only contains the semantic information of the intent, but also carries relevant contextual information, such as the conversation turn to which the intent belongs, related user input features, etc.

[0124] Edges represent semantic relationships between intents, and relationships include logical mutual exclusion and fuzzy overlap.

[0125] Logical mutual exclusion relationship: indicates that there is a clear contradiction between two intents, such as "refund" and "change" cannot be executed at the same time. Fuzzy overlapping relationship: indicates that two intents have a certain similarity or overlap in semantics, such as "recommend flight" and "search flight".

[0126] Edge weights are used to quantify the strength of relationships. The weight values ​​can be set according to the relationship strength defined in the rule base or the matching degree of semantic analysis.

[0127] The semantic graph is stored as a weighted graph structure, where the node set represents all logical conflicting intents, and the edge set represents the relationship weights between intents. For example, the weighted graph structure can be represented as: ;in, represents a weighted graph structure, Represents a node set (corresponding to all logical conflicting intentions), Represents a set of edges (corresponding to logical mutually exclusive relations and fuzzy overlapping relations).

[0128] The constructed semantic graph is input into the graph neural network, and features of nodes and edges in the semantic graph are extracted to obtain high-dimensional feature data representing the relationship between various intents:

[0129] The node and edge data in the semantic graph are digitized: the initial feature vector of the node represents the semantic information of the intent, which can be generated by a pre-trained language model (such as BERT); the weight of the edge represents the strength of logical mutual exclusion or fuzzy overlap, and the specific value is set according to the definition in the rule base during initialization.

[0130] Use graph neural networks based on message passing mechanism (such as Graph Convolutional Network, GCN) to perform joint feature extraction on nodes and edges.

[0131] In the graph neural network, each layer updates the node features through the following formula: ;in, Indicates Node in layer The eigenvector of Indicates Node in layer The characteristic vector of Is a node Neighbor nodes of It is The weight matrix of the layer; It is The bias term of the layer; is a nonlinear activation function (such as ReLU); Representation Node The set of neighbor nodes of Represents the layer index of the graph neural network, which is used to distinguish the calculation process of different layers in the network; Represents the index of the current computation node, which is used to identify a specific node in the graph; Indicates the current node The neighbor node index of is used to collect its associated information.

[0132] Among them, the bias term is a parameter in the neural network, which adjusts the eigenvalue after weighted sum, provides additional flexibility, and facilitates fitting nonlinear data; the nonlinear activation function is a function that introduces nonlinear capabilities (such as ReLU), which is used to enhance the network's ability to express complex data.

[0133] Each node generates a high-dimensional feature vector, which represents the relationship between the node and other nodes in the semantic graph, including information about logical mutual exclusivity and fuzzy overlap.

[0134] Based on high-dimensional feature data, each pair of intentions in the logical conflict intention combination is analyzed to evaluate the degree of opposition between the logical conflict intention combination in logical mutual exclusion and fuzzy overlap:

[0135] Similarity is calculated for each pair of high-dimensional feature vectors of logically conflicting intents to evaluate their semantic similarity: ;in, Express intention and intention The semantic similarity of and Intention and intention The characteristic vector of and Intention and intention The norm of the corresponding vector.

[0136] Calculating intent using Gaussian kernel function and intention The fuzzy overlap of: ;in, Express intention and intention The fuzzy overlap of Express intention and intention The Euclidean distance between the eigenvectors of is the adjustment parameter of the Gaussian kernel.

[0137] in, It is the first intention in a combination of logically conflicting intentions, such as "refund"; It is the second intention in the combination of logical conflicting intentions, such as "change the ticket". The adjustment parameter of the Gaussian kernel controls the sensitivity of the semantic space distance in the fuzzy overlap calculation. The larger the adjustment parameter value of the Gaussian kernel, the weaker the influence of the distant relationship between the feature vectors, and the smaller the value, the more sensitive it is.

[0138] After the semantic similarity and fuzzy overlap are dimensionlessly processed and weighted, the opposition index is calculated and expressed as: ;in, Is an index of opposition, used to indicate intention and intention the degree of opposition in logical mutual exclusion and ambiguous overlap; and are the weight coefficients of semantic similarity and fuzzy overlap, respectively, and and Both are greater than 0.

[0139] From the expression of the opposition degree index, it can be seen that the larger the opposition degree index is, the higher the opposition degree of the logical conflict intention combination in logical mutual exclusion and fuzzy overlap is.

[0140] The dynamic relation entropy of logical conflict intention combinations is analyzed to evaluate the fluctuation degree of semantic dependency of logical conflict intention combinations in multiple rounds of dialogues. Specifically:

[0141] In multi-round dialogues, based on the contextual relationships corresponding to logical conflicting intention combinations, a semantic similarity matrix is ​​generated for the conflicting intention pairs in each round of dialogue. The elements of the semantic similarity matrix represent the semantic association strength between each pair of conflicting intentions:

[0142] From the context of multiple rounds of dialogue, each pair of conflicting intents of logical conflicting intent combinations is extracted and their contextual relationships are annotated. Each pair of conflicting intents is used as the basic unit for similarity calculation. For example, "refund" and "change" constitute a pair of conflicting intents.

[0143] The semantic information of each intent is parsed through a pre-trained large-scale language model to generate a high-dimensional feature representation of the intent. For each conflicting intent pair, the two intents are respectively generated their feature vectors, and each dimension of the feature vector represents a specific feature in the semantic space.

[0144] Based on the feature vectors of conflicting intentions, the semantic similarity between each pair of conflicting intentions is calculated to generate a semantic similarity matrix in each round of dialogue. The formula for calculating the elements of the semantic similarity matrix is ​​as follows: ;in, Express intention and intention The semantic similarity of and Intention and intention The high-dimensional feature vector of and Intention and intention The norm of the corresponding eigenvector.

[0145] It is the first intention in the logical conflict intention combination in the scenario of analyzing the dynamic relationship entropy of the logical conflict intention combination. The second intention in the logical conflict intention combination in the scenario of analyzing the dynamic relationship entropy of the logical conflict intention combination and Represents a pair of intents in a logically conflicting intent combination, used to generate a cell in the semantic similarity matrix.

[0146] The change relationship of each round of semantic similarity matrix in multiple rounds of dialogue is mapped to the information entropy value, which is used to quantify the relationship complexity of conflicting intention combinations and the fluctuation of semantic dependency relationships:

[0147] For each pair of conflicting intentions, the semantic similarity in multiple rounds of dialogue may change as the dialogue progresses. The difference in the semantic similarity matrix in each round of dialogue is quantified as a change matrix, and the calculation formula is as follows: ;in, Express intention and intention In the and The change in semantic similarity in a turn-based dialogue, and Respectively and Semantic similarity in turn-based dialogues, Represents the time turn index in a multi-turn conversation.

[0148] According to the similarity change in the change matrix, the semantic relationship information entropy of each round of dialogue is calculated. The formula is as follows: ;in, Indicates The semantic relationship information entropy of the turn-based dialogue, is the normalized probability value of the elements in the change matrix, where .

[0149] The information entropy values ​​generated in multiple rounds of dialogues are modeled using a time-series adaptive graph neural network, and the change pattern of the information entropy values ​​in the time dimension is analyzed to generate a dynamic relationship fluctuation score to evaluate the degree of fluctuation of the semantic dependency of the logical conflict intention combination:

[0150] Arrange the information entropy values ​​generated by each round of calculation in multiple rounds of dialogue into a time series and construct an entropy value sequence matrix: ;in, represents the entropy value sequence matrix, Indicates the total number of conversation turns.

[0151] Each node is defined to represent a round of dialogue and its corresponding information entropy value, and the edge weights between nodes represent the temporal dependency between adjacent dialogue rounds.

[0152] The initial edge weight is calculated based on the time interval and entropy difference between dialogue turns. Specifically, the edge weight reflects the similarity between the information entropy values ​​of adjacent dialogue turns. When the difference in information entropy values ​​between two dialogue turns is small, the edge weight is large, indicating that the semantic relationship between the two turns is close; conversely, when the difference in information entropy values ​​is large, the edge weight is small. The adjustment parameter is used to control the sensitivity of the edge weight to the difference in information entropy values. The larger the value, the weaker the impact of the difference on the edge weight; the smaller the value, the more significant the impact of the difference on the edge weight.

[0153] Initialize the entropy value sequence matrix as the feature vector of each node.

[0154] In the time-series adaptive graph neural network, information between nodes is aggregated through the message passing mechanism. Each node aggregates feature information through the relationship between its neighbor nodes, thereby updating its own features. The specific process is as follows:

[0155] Each node in the current layer receives information from its neighboring nodes and combines this information with its own features.

[0156] The contribution of each neighbor node is weighted according to the strength of its relationship with the central node (attention weight).

[0157] The aggregated information is linearly transformed through a weight matrix and a bias term is added to represent the weighted sum of the features of the current layer.

[0158] In order to introduce nonlinear expression capabilities, the weighted results are finally processed by a nonlinear activation function (such as ReLU) to generate the feature vectors of the nodes in the next layer.

[0159] The node feature vectors in each layer are calculated by combining the features of neighbor nodes in the previous layer and its own features. The feature calculation will go deeper layer by layer to capture richer graph structure and time series information.

[0160] In order to aggregate information more accurately, the time-series adaptive graph neural network assigns an attention weight to each neighbor node to measure the importance of the neighbor node to the current node. The specific process is as follows:

[0161] The attention weight of each neighbor node is calculated based on the feature correlation between the central node and the neighbor node.

[0162] The relevance is modeled through an attention mechanism, which combines the feature vector of the central node with the feature vectors of the neighboring nodes to generate a relevance score.

[0163] The relevance scores are normalized into a probability distribution to represent the importance of each neighbor node.

[0164] To ensure that the sum of the attention weights of all neighbor nodes is 1, the relevance score is processed through a normalization function (such as the softmax function).

[0165] The attention mechanism contains a parameter vector that maps the features of the center node and neighbor nodes into a similarity space. In addition, there are two feature weight matrices that act on the features of the center node and neighbor nodes respectively to adjust the feature representation to adapt to the attention mechanism.

[0166] After multiple layers of message transmission, the final node feature vector represents the temporal characteristics of each round of entropy value in multiple rounds of dialogue. The dynamic relationship fluctuation score is generated by weighted aggregation of node feature vectors. The calculation formula is: ;in, represents the dynamic relationship fluctuation score, Indicates The node weight of the turn dialogue, Indicates The feature vector of the turn dialogue node in the last layer of the temporal adaptive graph neural network, Indicates The norm of the feature vector of the turn dialogue node, Represents the total number of layers of the temporal adaptive graph neural network.

[0167] The larger the dynamic relationship fluctuation score is, the more significant the fluctuation of the semantic dependency of the logical conflict intention combination is, indicating that the semantic dependency of the logical conflict intention combination in multiple rounds of dialogue fluctuates more dramatically in the time dimension, and the complexity and instability of the dependency are higher.

[0168] in, Generated according to the changing trend of information entropy value, used to measure the The importance of turn-taking in rating dynamic relationship fluctuations; Generated through a multi-layer message passing mechanism, including the The temporal dependencies and relational characteristics of wheel nodes; It reflects the strength or size of the feature vector and is used to quantify the contribution of this round of dialogue nodes to the dynamic relationship fluctuation score. In practical applications, The value of needs to be adjusted according to the complexity of the task and the relationship characteristics of the nodes. Generally speaking, shallower networks (smaller values) are suitable for capturing local relationships, while deeper networks (larger values) can extract more global relationship patterns.

[0169] Based on the degree of opposition between logically conflicting intention combinations in logical mutual exclusion and fuzzy overlap and the degree of fluctuation of semantic dependency in multi-round dialogues, the processing order of logically conflicting intention combinations is dynamically adjusted, specifically:

[0170] Set the preset threshold of the opposition index to , It is used to judge whether the degree of opposition between the combination of logical conflicting intentions in logical mutual exclusion and fuzzy overlap is significant. The size of the threshold should be set according to the complexity of the conflicting intentions in the task scenario and the influence of the mutually exclusive logic.

[0171] Set the preset threshold of dynamic relationship fluctuation score to , It is used to evaluate the significance of fluctuations in semantic dependencies of logical conflict intention combinations. The threshold can be set by analyzing the average fluctuation amplitude of semantic dependencies in historical conversations.

[0172] like , it is determined that the logical conflict intention combination has a high degree of opposition in logical mutual exclusion and fuzzy overlap, indicating that the degree of opposition of the logical conflict intention combination is significant, and there may be strong mutual exclusion or fuzzy overlap. The system needs to give priority to these intentions to resolve potential conflicts as soon as possible and avoid user confusion.

[0173] like , then it is determined that the degree of opposition between the logical conflict intention combination in logical mutual exclusion and fuzzy overlap is normal, indicating that the degree of opposition is low, the logical mutual exclusion and fuzzy overlap relationship between the intentions are not strong, and the conflict risk is small. The system can lower its processing priority and save resources for more urgent tasks.

[0174] like , it is determined that the semantic dependency relationship of the logical conflict intention combination fluctuates significantly, indicating that the semantic dependency relationship of the logical conflict intention combination fluctuates violently, which may cause new conflicts due to changes in user context or multiple rounds of dialogue. We should respond as soon as possible to avoid the spread of conflicts.

[0175] like , it is determined that the fluctuation of the semantic dependency relationship of the logical conflict intention combination is not significant, indicating that the fluctuation change is not significant, the semantic dependency relationship is relatively stable, the conflict risk is low, and the adjustment or processing between intentions can be postponed, which will not have a significant impact on the overall dialogue logic.

[0176] when and , then the logical conflict intention combination is set to the highest priority, indicating that the intention combination has both high opposition and high volatility, and is the core object of conflict handling. It should be immediately prioritized to stabilize the dialogue logic and avoid affecting the user experience.

[0177] when and , then the logical conflict intention combination is set to medium priority, indicating that although the degree of opposition is high, the semantic dependency fluctuation is small, and the conflict of the intention combination is relatively stable. It can be set to medium priority and resolved in the next round of processing.

[0178] when and , then the logical conflict intention combination is set to medium priority, indicating that although the degree of opposition is low, the semantic dependency fluctuates significantly, which may bring conflict risks due to context changes and needs to be dynamically monitored and set to medium priority for processing.

[0179] when and , then the logical conflict intention combination is set to the lowest priority, indicating that the degree of opposition and the fluctuation of semantic dependence are not significant, the impact of intention conflict can be ignored, the priority is the lowest, and it can be postponed or even merged into a secondary task.

[0180] Priority is used to determine the order in which logically conflicting intent combinations are processed in the dialogue system. Through dynamic adjustment of priority, the system can prioritize the conflicts that have the greatest impact on user experience, improving dialogue coherence and user satisfaction. By regularly recalculating the opposition index and dynamic relationship fluctuation score, the priority of logically conflicting intent combinations is updated in real time in each round of dialogue. Intent combinations with high priority will be parsed and responded to first, and intentions with low priority can be deferred or merged into secondary tasks, thereby achieving dynamic adaptation and efficient processing of complex dialogues.

[0181] Example 2: The difference between Example 2 of the present invention and Example 1 is that this example introduces an intention recognition intelligent dialogue system based on a large model.

[0182] Figure 2 A structural schematic diagram of an intention recognition intelligent dialogue system based on a large model of the present invention is given, an intention recognition intelligent dialogue system based on a large model, including a semantic parsing and extraction module, an intention conflict detection module, an opposition degree assessment module, a semantic dependency analysis module and a processing order adjustment module.

[0183] Semantic parsing and extraction module: obtains the user's input data in multiple rounds of conversations, uses a pre-trained large-scale language model to perform semantic parsing on the input data, and extracts multiple preliminary intentions of the user.

[0184] Intent conflict detection module: establishes an intent conflict detection mechanism based on a predefined intent conflict rule base, analyzes the extracted preliminary intents based on the intent conflict detection mechanism, and screens out logical conflicting intent combinations.

[0185] Opposition degree assessment module: Analyze the semantic relationship of logical conflict intention combinations through semantic graphs and graph neural networks, and evaluate the opposition degree of logical conflict intention combinations in logical mutual exclusion and fuzzy overlap.

[0186] Semantic dependency analysis module: Analyzes the dynamic relationship entropy of logical conflict intention combinations and evaluates the degree of fluctuation in the semantic dependency of logical conflict intention combinations in multiple rounds of dialogue.

[0187] Adjust processing order module: Dynamically adjust the processing order of logically conflicting intention combinations based on the degree of opposition between logically conflicting intention combinations in logical mutual exclusion and fuzzy overlap and the degree of fluctuation of semantic dependency in multiple rounds of dialogue.

[0188] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0189] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or may be transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium may be a solid-state hard disk.

[0190] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0191] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0192] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0193] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0194] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0195] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0196] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0197] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A large model-based intention recognition intelligent dialogue method, characterized in that: The steps include: Obtain the user's input data in multiple rounds of conversations, use the pre-trained large-scale language model to perform semantic analysis on the input data, and extract multiple preliminary intentions of the user; An intent conflict detection mechanism is established based on a predefined intent conflict rule library, and the extracted preliminary intents are analyzed based on the intent conflict detection mechanism to screen out logical conflict intent combinations; The semantic relationship of the logical conflict intention combination is analyzed through semantic graph and graph neural network to evaluate the degree of opposition of the logical conflict intention combination in logical mutual exclusion and fuzzy overlap. The dynamic relation entropy of logical conflict intention combinations is analyzed to evaluate the fluctuation degree of semantic dependency of logical conflict intention combinations in multiple rounds of dialogues. Dynamically adjust the processing order of logically conflicting intention combinations based on the degree of opposition between logically conflicting intention combinations in logical mutual exclusion and fuzzy overlap and the degree of fluctuation of semantic dependence in multi-round dialogues; An intent conflict detection mechanism is established based on the predefined intent conflict rule library. The extracted preliminary intent is analyzed based on the intent conflict detection mechanism to screen out logical conflict intent combinations, specifically: Construct an intention conflict rule base, which includes logical mutual exclusion relationships, conditional dependency relationships, and priority relationships; Match the extracted preliminary intent with the rules in the intent conflict rule library, and determine whether there is a conflict between the preliminary intents through a matching algorithm; Screening the conflicting intentions in the matching results, including extracting the intention combinations that satisfy the logical mutual exclusion relationship, and screening out the logical conflicting intention combinations that satisfy the conflict conditions in combination with the conditional dependency relationship; Among them, if two intentions satisfy the logical mutual exclusion relationship and dependency conditions at the same time and meet the priority constraints, they are marked as a logical conflict intention combination.

2. According to claim 1, the method for intelligent dialogue based on intention recognition based on a large model is characterized in that: Obtain the user's input data in multiple rounds of conversations, use the pre-trained large-scale language model to perform semantic analysis on the input data, and extract multiple preliminary intentions of the user, specifically: By receiving input data from the user during the conversation; Preprocess the received input data, including denoising, formatting and data cleaning; Use the pre-trained large-scale language model to perform semantic analysis on the pre-processed input data and extract the key semantic information of the user input, which includes the user's expression purpose, content focus, and potential additional requirements; Through the key semantic information extracted after semantic parsing, multiple preliminary intentions of the user in the current conversation turn are identified, and a semantic representation vector corresponding to each intention is generated.

3. The method for intelligent dialogue based on intention recognition based on a large model according to claim 2, characterized in that: The semantic relationship of the logical conflict intention combination is analyzed through semantic graph and graph neural network to evaluate the degree of opposition of the logical conflict intention combination in logical mutual exclusion and fuzzy overlap, specifically: Establish a semantic graph, where the nodes of the semantic graph represent the various intentions in the combination of logically conflicting intentions, and the edges represent the semantic relationships between the intentions. The semantic relationships include logical mutually exclusive relationships and fuzzy overlapping relationships. The constructed semantic graph is input into the graph neural network, and features of nodes and edges in the semantic graph are extracted to obtain high-dimensional feature data representing the relationship between various intentions. Each pair of intentions in the logically conflicting intention combination is analyzed based on high-dimensional feature data, and the degree of opposition of the logically conflicting intention combination in logical mutual exclusion and fuzzy overlap is evaluated.

4. The method for intelligent dialogue based on intention recognition based on a large model according to claim 3 is characterized in that: Based on high-dimensional feature data, each pair of intentions in the logical conflict intention combination is analyzed to evaluate the degree of opposition between the logical conflict intention combination in logical mutual exclusion and fuzzy overlap, specifically: Calculate the similarity of each pair of high-dimensional feature vectors of logical conflicting intentions: ;in, Express intention and intention The semantic similarity of and Intention and intention The characteristic vector of and Intention and intention The norm of the corresponding vector; Calculating intent using Gaussian kernel function and intention The fuzzy overlap of: ;in, Express intention and intention The fuzzy overlap of Express intention and intention The Euclidean distance between the eigenvectors of is the adjustment parameter of the Gaussian kernel; After the semantic similarity and fuzzy overlap were dimensionlessly processed and assigned weight coefficients respectively, the opposition degree index was calculated.

5. The method for intelligent dialogue based on intention recognition based on a large model according to claim 4, characterized in that: The dynamic relation entropy of logical conflict intention combinations is analyzed to evaluate the fluctuation degree of semantic dependency of logical conflict intention combinations in multiple rounds of dialogues. Specifically: In multiple rounds of dialogue, based on the contextual relationships corresponding to the logical conflicting intention combinations, a semantic similarity matrix is ​​generated for the conflicting intention pairs in each round of dialogue. The elements of the semantic similarity matrix represent the semantic association strength between each pair of conflicting intentions. Map the change relationship of each round of semantic similarity matrix in multiple rounds of dialogue into information entropy value; The information entropy values ​​generated in multi-round conversations are modeled using a temporal adaptive graph neural network, and the change pattern of the information entropy values ​​in the time dimension is analyzed to generate a dynamic relationship fluctuation score to evaluate the degree of fluctuation in the semantic dependency of logical conflict intention combinations.

6. The method for intelligent dialogue based on intention recognition based on a large model according to claim 5, characterized in that: The change relationship of each round of semantic similarity matrix in multiple rounds of dialogue is mapped to the information entropy value, specifically: The difference in the semantic similarity matrix in each round of dialogue is quantified as a change matrix: ;in, Express intention and intention In the and The change in semantic similarity in a turn-based dialogue, and Respectively and Semantic similarity in turn-based dialogues, Represents the time turn index in a multi-turn conversation; Calculate the semantic relationship information entropy of each round of dialogue. The formula is as follows: ;in, Indicates The semantic relationship information entropy of the turn-based dialogue, is the normalized probability value of the elements in the change matrix.

7. The method for intelligent dialogue based on intention recognition based on a large model according to claim 6, characterized in that: Based on the degree of opposition between logically conflicting intention combinations in logical mutual exclusion and fuzzy overlap and the degree of fluctuation of semantic dependency in multi-round dialogues, the processing order of logically conflicting intention combinations is dynamically adjusted, specifically: Set the preset threshold of the opposition index to ; Set the preset threshold of dynamic relationship fluctuation score to ; when and , then the logical conflict intention combination is set to the highest priority; when and , then the logical conflict intention combination is set to medium priority; when and , then the logical conflict intention combination is set to medium priority; when and , then the logical conflict intention combination is set to the lowest priority; in, is the opposition index, Represents the dynamic relationship fluctuation score.

8. An intention recognition intelligent dialogue system based on a large model, used to implement an intention recognition intelligent dialogue method based on a large model as described in any one of claims 1 to 7, characterized in that: It includes a semantic parsing and extraction module, an intention conflict detection module, an opposition degree assessment module, a semantic dependency analysis module, and a processing order adjustment module; Semantic parsing and extraction module: obtains the user's input data in multiple rounds of dialogue, uses the pre-trained large-scale language model to perform semantic parsing on the input data, and extracts multiple preliminary intentions of the user; Intent conflict detection module: establishes an intent conflict detection mechanism based on a predefined intent conflict rule base, analyzes the extracted preliminary intents based on the intent conflict detection mechanism, and screens out logical conflict intent combinations; Opposition degree assessment module: Analyze the semantic relationship of logical conflict intention combinations through semantic graphs and graph neural networks, and assess the opposition degree of logical conflict intention combinations in logical mutual exclusion and fuzzy overlap; Semantic dependency analysis module: Analyzes the dynamic relationship entropy of logical conflict intention combinations and evaluates the degree of fluctuation of semantic dependency of logical conflict intention combinations in multiple rounds of dialogue; Adjust processing order module: Dynamically adjust the processing order of logically conflicting intention combinations based on the degree of opposition between logically conflicting intention combinations in logical mutual exclusion and fuzzy overlap and the degree of fluctuation of semantic dependency in multiple rounds of dialogue.

Images