Intelligent interaction system and method based on dynamic intention recognition

Abstract

The invention discloses an intelligent interaction system and method based on dynamic intention recognition, and relates to the technical field of intelligent interaction.The intelligent interaction method comprises the steps that firstly, current query of a user and a context set including historical dialogues, browsing records and other information are obtained; a parallel embedding model is introduced to extract semantic features of user query and semantic features of each piece of information in a context set, and dynamic correlation weighting and screening are performed on context features through a gating network with current query as guidance, so that self-adaptive deep fusion of key context information and user query is realized; and performing intention classification inference on the basis of the semantic representation after fusion enhancement to obtain a final intention recognition result. Therefore, the problems of insufficient utilization of the context information and serious noise interference can be solved, so that the intention understanding accuracy and robustness of the system when the system processes the reference unknown query are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent interaction, and more particularly, to an intelligent interaction system and method based on dynamic intent recognition. BACKGROUND

[0002] With the rapid development of artificial intelligence and natural language processing technology, intelligent interaction systems represented by intelligent customer service, voice assistants and chat robots have been deeply integrated into many fields such as e-commerce, online services and smart homes. The core value of these systems lies in their ability to accurately and efficiently understand user intent and provide corresponding services or feedback, thereby improving user experience and interaction efficiency. However, in real multi-turn interaction scenarios, user intent expression is often dynamic and highly dependent on context, and single, isolated analysis of the current query cannot meet the needs of complex interactions. Each input from the user may be closely related to previous dialogue content, browsing behavior or other environmental information, which requires intelligent interaction systems to have the ability to dynamically understand and utilize context.

[0003] When performing intent recognition, existing technologies attempt to incorporate context information such as dialogue history. The common approach is to simply concatenate historical dialogue text with the current user query or to perform sequential encoding through recurrent neural networks and other models in order to capture context associations. However, such methods generally treat context information indiscriminately and fail to effectively distinguish the relevance of different context information to the current query, resulting in key information being overwhelmed by a large amount of irrelevant or weakly related historical data noise, which affects the accuracy of intent recognition. Especially when dealing with user queries containing ambiguous references or omitted components, existing technologies face the difficult problem of intent completion. For example, when a user issues non-explicit instructions such as "how about this" or "change one" after browsing a specific product page, if the system cannot dynamically and accurately identify the specific entity implicitly associated with "this" from the complex context set (such as browsing history, historical dialogue), it is likely to cause a deviation in understanding the user's true intent or even complete failure.

[0004] Therefore, there is an urgent need for an optimized intelligent interaction system and method based on dynamic intent recognition. SUMMARY

[0005] To solve the above technical problems, the present application is proposed.

[0006] According to an aspect of the present application, an intelligent interaction method based on dynamic intent recognition is provided, which includes:

[0007] obtaining a user query and a context set;

[0008] parallelly embedding the user query and the context set to obtain a user query embedding encoding vector and a context embedding encoding vector set;

[0009] performing gating network-based relevance calculation and screening on the context embedding encoding vector set based on the user query embedding encoding vector to obtain a gating context representation vector;

[0010] performing feature fusion on the gating context representation vector and the user query embedding encoding vector to obtain a user query semantic enhancement encoding vector with fused context information;

[0011] performing intent inference based on the user query semantic enhancement encoding vector with fused context information to obtain a final intent.

[0012] According to another aspect of the present application, an intelligent interaction system based on dynamic intent recognition is provided, which comprises:

[0013] a user query context acquisition module, configured to acquire a user query and a context set;

[0014] a parallel embedding encoding module, configured to parallelly embed the user query and the context set to obtain a user query embedding encoding vector and a context embedding encoding vector set;

[0015] a relevance calculation and screening module, configured to perform gating network-based relevance calculation and screening on the context embedding encoding vector set based on the user query embedding encoding vector to obtain a gating context representation vector;

[0016] a feature fusion module, configured to perform feature fusion on the gating context representation vector and the user query embedding encoding vector to obtain a user query semantic enhancement encoding vector with fused context information;

[0017] an intent inference module, configured to perform intent inference based on the user query semantic enhancement encoding vector with fused context information to obtain a final intent.

[0018] Compared with the prior art, the intelligent interaction system and method based on dynamic intent recognition provided by the present application first acquires a user's current query and a context set containing information such as historical dialogues and browsing records, and introduces a parallel embedding model to extract semantic features of the user query and semantic features of each piece of information in the context set, respectively. A gating network oriented to the current query is used to dynamically weight and screen the context features, so as to realize adaptive deep fusion of key context information and the user query, and then perform intent classification inference based on the fused and enhanced semantic representation, to obtain a final intent recognition result. In this way, the problem of insufficient utilization of context information and serious noise interference can be solved, thereby significantly improving the intent understanding accuracy and robustness of the system when processing ambiguous queries. Attached Figure Description

[0019] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0020] Figure 1 This is a flowchart of an intelligent interaction method based on dynamic intent recognition according to an embodiment of this application.

[0021] Figure 2 This is a data flow diagram of an intelligent interaction method based on dynamic intent recognition according to an embodiment of this application.

[0022] Figure 3 This is a flowchart of sub-step S2 of the intelligent interaction method based on dynamic intent recognition according to an embodiment of this application.

[0023] Figure 4 This is a flowchart of sub-step S3 of the intelligent interaction method based on dynamic intent recognition according to an embodiment of this application.

[0024] Figure 5 This is a flowchart of sub-step S4 of the intelligent interaction method based on dynamic intent recognition according to an embodiment of this application.

[0025] Figure 6 This is a flowchart of sub-step S5 of the intelligent interaction method based on dynamic intent recognition according to an embodiment of this application.

[0026] Figure 7 This is a block diagram of an intelligent interaction system based on dynamic intent recognition according to an embodiment of this application. Detailed Implementation

[0027] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0028] To address the problems mentioned above in the background technology, this application proposes an intelligent interaction method based on dynamic intent recognition. Figure 1 This is a flowchart of an intelligent interaction method based on dynamic intent recognition according to an embodiment of this application. Figure 2This is a data flow diagram of an intelligent interaction method based on dynamic intent recognition according to an embodiment of this application. For example... Figure 1 and Figure 2 As shown, the intelligent interaction method based on dynamic intent recognition includes the following steps: S1, obtaining a user query and context set; S2, performing parallel embedding on the user query and context set to obtain a user query embedding encoding vector and a context embedding encoding vector set; S3, based on the user query embedding encoding vector, performing relevance calculation and filtering on the context embedding encoding vector set based on a gating network to obtain a gated context representation vector; S4, performing feature fusion on the gated context representation vector and the user query embedding encoding vector to obtain a user query semantic enhancement encoding vector with fused context information; S5, performing intent inference based on the user query semantic enhancement encoding vector with fused context information to obtain the final intent.

[0029] In the aforementioned intelligent interaction method based on dynamic intent recognition, step S1 involves acquiring the user query and context set. It should be understood that user queries often contain ambiguous references or incomplete information, such as using pronouns like "this" or "that" or omitting key entities. Relying solely on the query text cannot accurately infer the user's true intent. Furthermore, the user's intent expression is closely related to the surrounding context; contextual information such as the currently viewed page and historical dialogue content are crucial for understanding the intent. Therefore, this application acquires user queries and related contextual information in real time, forming a user query and context set. This provides complete raw data support for subsequent intent inference, ensuring that the system can analyze ambiguous references and implicit information in the query in conjunction with contextual information. This avoids misjudgments of intent due to missing information, lays a data foundation for subsequent normalization, embedding, and fusion processing steps, and ensures the accuracy of the entire intent inference process.

[0030] Specifically, in one possible embodiment, step S1 is implemented as follows: First, a listener deployed in the front-end application captures in real time the text entered by the user in the input box of the interactive interface. For example, when a user is browsing the product details page of a specific model of refrigerator, they enter the text "Is this in stock?" in the input box of the interactive interface as a user query. At the same time, the system synchronously collects context data related to the current user session through the interface with the front-end application and the back-end service. This data is constructed into a set containing multiple context units. One context unit records the UI interface information that the user is currently browsing, including fields such as page type, entity ID, and entity type; another context unit records the historical dialogue content within a preset time window. Finally, the user query text and the collected context set are integrated and used as input for subsequent processing.

[0031] In the aforementioned intelligent interaction method based on dynamic intent recognition, step S2 involves parallel embedding of the user query and the context set to obtain a user query embedding encoding vector and a context embedding encoding vector set. It should be understood that the user query and the context set are unstructured information, such as text sequences. This form cannot be directly used by subsequent intent recognition models for semantic association calculation. Furthermore, to avoid semantic information confusion or loss, this application further performs parallel embedding processing on the user query and the context set, transforming the unstructured query and context information into structured vector representations, while ensuring that their respective semantic features are fully preserved during the transformation process. This provides a directly computable numerical basis for calculating the semantic association between the user query and the context in subsequent steps, enabling the model to accurately capture the inherent connection between the two, thereby improving the accuracy of dynamic intent recognition.

[0032] In particular, in one specific embodiment, Figure 3 This is a flowchart of sub-step S2 of the intelligent interaction method based on dynamic intent recognition according to an embodiment of this application. Figure 3 As shown, step S2 includes: S21, inputting the user query into a text encoder containing a text segmenter to obtain the user query embedding encoding vector; S22, extracting a first context from the context set; S23, reading the data source field value of the first context; S24, in response to the data source field value being UI, performing entity embedding encoding on the entity ID and entity type of the first context to obtain a context embedding encoding vector; S25, in response to the data source field value being dialogue history, inputting the text field of the first context into a text encoder containing a text segmenter to obtain a context embedding encoding vector.

[0033] Specifically, in step S21, the user query is input into a text encoder containing a text segmenter to obtain the user query embedding encoding vector. It should be understood that user queries exist in the form of natural language text, which cannot be directly used by subsequent intent recognition models for semantic calculation and association analysis. Therefore, this application inputs the user query into a text encoder containing a text segmenter. The text encoder uses a Transformer-based pre-trained language model to transform the unstructured text query into a vector representation with fixed dimensions. During the transformation process, the segmenter parses the word structure and word order relationships of the text, and then the encoder deeply extracts semantic features. In this way, a user query embedding encoding vector that retains the core semantics of the user query and can be directly computed by the model can be obtained, providing a numerical basis for subsequent semantic matching with context vectors and ensuring accurate capture of query semantics during intent recognition.

[0034] Specifically, in one possible embodiment, step S21 is implemented as follows: First, the text segmenter receives the user query text, such as "Does this have stock?", preprocesses the text including removing punctuation and unifying character case, and then uses a sub-word segmentation algorithm to split the text into a series of tokens, such as "this", "have", "stock", and "does". Special classification and delimiter tags are added to the beginning and end of the sequence. Subsequently, the text encoder receives the token sequence and maps each token to an initial word vector through its internal embedding layer. These word vectors are then input into a network structure composed of multiple stacked Transformer encoders. After multiple layers of self-attention calculation and feedforward network transformation, the hidden state vector corresponding to the classification tag in the output sequence of the last layer of the encoder is taken as the user query embedding encoding vector representing the entire user query semantics.

[0035] Specifically, step S22 involves extracting a first context from the context set. It should be understood that the context set contains information from multiple sources and categories. Processing it simultaneously would lead to information redundancy and semantic confusion. Furthermore, different contexts have varying degrees of relevance to the user query, necessitating priority processing of the most relevant information to improve efficiency. Therefore, this application extracts a first context from the context set to focus on the context information most closely related to the current user query, reducing interference from irrelevant information in subsequent processing, providing targeted information support for accurate intent recognition, and simplifying the processing flow to improve system response speed.

[0036] Specifically, in one possible embodiment, step S22 is implemented as follows: The context set contains multiple context units, each with a timestamp, information type label, and content data. For example, it may contain multiple units, such as a unit with an earlier timestamp, type "UI," and content of product page information, or a unit with a later timestamp, type "dialogue history," and content of the user's previous question. When extracting the first context, the units are first sorted in reverse chronological order by timestamp, and valid units with timestamps within a preset time period before the current query time are selected. Units with low relevance are then excluded, and finally, the unit ranked first is selected as the first context. During the extraction process, it is necessary to simultaneously verify whether the content format of the unit is complete, ensuring that it contains key fields such as entity information and interaction time to meet subsequent processing requirements.

[0037] Specifically, in step S23, the data source field value of the first context is read. It should be understood that the sources of the first context differ, such as UI interfaces and dialogue history, resulting in significant differences in their data structures and semantic features. Using a uniform processing method would lead to the loss or distortion of semantic information. Therefore, this application clarifies the information source of the first context by reading the data source field value, providing a basis for subsequently selecting the corresponding embedding encoding method. This ensures that the most suitable processing strategy is adopted for contexts from different sources, fully preserving their semantic features and avoiding information distortion caused by improper processing methods, thus laying the foundation for accurate intent recognition.

[0038] Specifically, in step S24, in response to the data source field value being UI, entity embedding encoding is performed on the entity ID and entity type of the first context to obtain a context embedding encoding vector. It should be understood that the first context, where the data source is UI, often contains structured information such as entity IDs (e.g., control identifiers, product numbers) and entity types (e.g., buttons, products). This type of information cannot be effectively converted through text encoding. Therefore, this application further performs entity embedding encoding on the entity ID and entity type of the first context to transform the entity information from the UI source into a context embedding encoding vector containing category semantics and individual characteristics. This accurately preserves the structured features of UI entities, enabling semantic association calculations with the user query embedding encoding vector, thus improving the accuracy of utilizing UI interaction information in intent recognition.

[0039] Specifically, in one possible embodiment, step S24 is implemented as follows: The first context is product page information from the UI source. First, entity ID and entity type are extracted from its content data, where entity ID is obtained from the corresponding field, and entity type is obtained from the corresponding field. Then, a pre-built entity embedding table is called, which contains pre-trained embedding vectors of all product entities. The corresponding embedding vector is obtained by matching entity ID and entity type. Then, the pre-trained embedding vector and the embedding vector are concatenated column-wise into a fusion vector. This fusion vector is then input into a multilayer perceptron composed of multiple fully connected layers. The first layer uses the ReLU activation function, and the second layer uses the linear activation function. The output vector is the corresponding vector. Finally, the vector is subjected to L2 normalization to obtain the context embedding encoding vector, which retains the entity features and type attributes of the UI context.

[0040] Specifically, in step S25, in response to the data source field value being dialogue history, the text field of the first context is input into a text encoder containing a text segmenter to obtain a context embedding encoding vector. It should be understood that the first context, whose data source is dialogue history, exists in the form of natural language text, containing the semantics of user-system interaction. Contextual dependencies and semantic features can be extracted through text encoding, forming an effective association with the semantics of the user's query text. Based on this, this application inputs the text field of the first context into a text encoder containing a text segmenter to transform the text information of the dialogue history into a context embedding encoding vector with the same dimension as the user's query embedding encoding vector, while preserving the semantic associations in the text. This ensures that the contextual semantics in the dialogue history are accurately captured, providing complete historical interaction semantic support for dynamic intent recognition.

[0041] Specifically, in one possible embodiment, step S25 is implemented as follows: First, the text field content of the first context is extracted, which is natural language interactive text. Second, a text segmenter splits the text into word sequences and converts them into index identifiers. Subsequently, a text encoder processes the index sequences through an embedding layer and a neural network layer, extracting semantic features and generating a fixed-dimensional vector. The final output context embedding encoding vector maintains the same dimension as the user query embedding encoding vector.

[0042] In the aforementioned intelligent interaction method based on dynamic intent recognition, step S3 involves performing relevance calculation and filtering on the context embedding encoding vector set using a gating network based on the user query embedding encoding vector to obtain a gated context representation vector. It should be understood that since the context embedding encoding vector set contains multiple vectors from different sources, the relevance of these vectors to the user's current query varies significantly. To avoid irrelevant or weakly related vectors introducing noise and interfering with the accurate understanding of the user's true intent, this application uses a gating network to perform relevance calculation and filtering on the context embedding encoding vector set based on the user query embedding encoding vector. This quantifies the degree of association between each context vector and the user query and extracts the most relevant contextual information. In this way, the obtained gated context representation vector effectively filters out irrelevant contextual noise and highlights the semantic association between key contexts and the user query, providing a high-quality contextual foundation for subsequent feature fusion, thereby improving the system's accuracy in recognizing user intent, especially query intents containing fuzzy references.

[0043] In particular, in one specific embodiment, Figure 4 This is a flowchart of sub-step S3 of the intelligent interaction method based on dynamic intent recognition according to an embodiment of this application. Figure 4As shown, step S3 includes: S31, inputting the user query embedding encoding vector and each context embedding encoding vector in the context embedding encoding vector set into the gating network to obtain a relevance score set; S32, performing normalization processing on the relevance score set based on the Softmax function to obtain a context relevance weight set; S33, selecting the context embedding encoding vector corresponding to the maximum context relevance weight from the context relevance weight set as the gating context representation vector.

[0044] Specifically, in step S31, the user query embedding vector and each context embedding vector in the context embedding vector set are input into a gating network to obtain a relevance score set. It should be understood that since the semantic relevance between the user query and each vector in the context set differs, and this difference cannot be directly quantified through intuitive comparison, it is difficult to accurately determine which contextual information is more valuable to the current query without quantitative indicators. Therefore, this application further inputs the user query embedding vector and each context embedding vector into a gating network to calculate and output a relevance score reflecting the closeness of their association. This provides a quantifiable basis for subsequent selection of key contexts, accurately distinguishes the relevance strength between different contexts and the user query, and avoids the omission of key information or interference from irrelevant information due to subjective judgment.

[0045] Specifically, in one possible embodiment, step S31 is implemented as follows: The gated network may consist of one or more fully connected layers. Specifically, the user query embedding vector and a context embedding vector from the context embedding vector set are concatenated dimensionally to form a fused feature vector. Subsequently, this fused feature vector is input into a feedforward neural network, which may contain one or more hidden layers, for example, using ReLU or GELU as activation functions. Finally, the high-dimensional feature is mapped to a scalar value through an output layer without an activation function; this value is the relevance score between the user query and the current context. This process is repeated for each vector in the context embedding vector set, ultimately forming a relevance score set containing all calculated results.

[0046] Specifically, step S32 involves normalizing the relevance score set using the Softmax function to obtain a context-related weight set. It should be understood that since the values ​​in the relevance score set are at different magnitudes, the raw scores alone cannot directly reflect the relative importance of each context, nor can they serve as a clear basis for subsequent selection. Therefore, this application further employs the Softmax function to normalize the relevance score set, thereby converting the scores into weight values ​​that sum to 1. In this way, the resulting context-related weight set allows for a direct representation of the relative importance of each context, ensuring that the weight values ​​retain both the magnitude relationship of the original scores and a comparable, unified scale, providing a clear quantitative standard for subsequently selecting the most relevant context.

[0047] Specifically, in one possible embodiment, step S32 is implemented as follows: The relevance score set contains two scores, corresponding to the refrigerator entity vector from the UI source and the dialogue history vector, respectively. After inputting these scores into the Softmax function, the function amplifies the difference between high and low scores through exponential operations, and divides the result by the sum of all exponential results, so that the sum of the two output weight values ​​is 1. Among them, the weight value corresponding to the UI source vector is close to 0.8 due to the higher original score, and the weight value corresponding to the dialogue history vector is close to 0.2. The resulting context-related weight set clearly reflects the higher relative importance of the former.

[0048] Specifically, in step S33, the context embedding encoding vector corresponding to the largest context relevance weight is selected from the context relevance weight set as the gated context representation vector. It should be understood that since the context relevance weight set clearly defines the relative importance of each vector, and the understanding of the user's current query intent often depends on the most relevant context information, to avoid redundancy or interference that may be introduced by multiple contexts, this application further selects the context embedding encoding vector corresponding to the largest weight from the context relevance weight set as the gated context representation vector, thereby focusing on the key context most closely related to the query. This ensures that the subsequent feature fusion process is based only on the most valuable context information, effectively filtering noise and improving the accuracy of understanding the user's intent for fuzzy queries (such as queries containing referential meanings).

[0049] Specifically, in one possible embodiment, step S33 is implemented as follows: The context-related weight set includes the weight (0.8) corresponding to the UI source context embedding encoding vector and the weight (0.2) corresponding to the dialogue history context embedding encoding vector. The system compares the weight values ​​and determines that the weight of the UI source context embedding encoding vector is the maximum value. Therefore, the UI source context embedding encoding vector is selected as the gating context representation vector. This vector contains key information such as the entity ID and entity type of the currently viewed product, and has the highest relevance to the user's query "Is this in stock?".

[0050] In the aforementioned intelligent interaction method based on dynamic intent recognition, step S4 involves feature fusion of the gated context representation vector and the user query embedding encoding vector to obtain a user query semantic enhancement encoding vector with fused contextual information. It should be understood that since the gated context representation vector contains key contextual information most relevant to the user query, while the user query embedding encoding vector only reflects the semantics of the query itself, their individual existence cannot fully reflect the intrinsic relationship between the query and the context. This is especially true for queries with ambiguous referents, where relying solely on the query vector leads to semantic loss. Therefore, this application further fuses the gated context representation vector and the user query embedding encoding vector to organically combine key contextual information with query semantics, forming a unified representation that incorporates the core features of both. This allows the obtained user query semantic enhancement encoding vector with fused contextual information to retain the original intent of the query while incorporating supplementary contextual information, effectively solving the semantic ambiguity problem caused by unclear referents and providing a more comprehensive and accurate semantic foundation for subsequent intent inference.

[0051] In particular, in one specific embodiment, Figure 5 This is a flowchart of sub-step S4 of the intelligent interaction method based on dynamic intent recognition according to an embodiment of this application. Figure 5 As shown, step S4 includes: S41, using the user query embedding encoding vector as the query vector and the gating context representation vector as the key vector and value vector; S42, inputting the query vector, key vector and value vector into the cross-attention layer to obtain the user query semantic enhancement encoding vector that integrates the context information.

[0052] Specifically, in step S41, the user query embedding encoding vector is used as the query vector, and the gated context representation vector is used as the key and value vectors. In particular, to accurately focus on the user query's need for contextual information, this application uses the user query embedding encoding vector as the query vector and the gated context representation vector as the key and value vectors. This establishes an attention computation orientation centered on the user query, ensuring that the extraction of contextual information is closely related to the user's intent. In this way, the attention mechanism can revolve around the semantic needs of the user query, selectively mining key features related to the query from the gated context representation vector, avoiding interference from irrelevant information, and laying the foundation for subsequent efficient feature fusion.

[0053] Specifically, in step S42, the query vector, key vector, and value vector are input into a cross-attention layer to obtain the user query semantic enhancement encoding vector that integrates the context information. It should be understood that, in order to achieve deep semantic fusion of the user query embedding encoding vector and the gated context representation vector, this application further inputs the query vector, key vector, and value vector into a cross-attention layer. This allows for the calculation of the association weights between the query and the context through an attention mechanism, and the selective integration of context features based on these weights, achieving semantic enhancement fusion of the two. In this way, the fused vector retains both the original intent of the user query and deeply integrates the key information in the gated context representation vector, effectively solving the semantic completion problem of ambiguous referencing in the query and significantly improving the vector's ability to represent the user's true intent.

[0054] Specifically, in one possible embodiment, step S42 is implemented as follows: The cross-attention layer first receives the query vector, key vector, and value vector. Within the layer, matrix operations are used to calculate the similarity between the query vector and the key vector, resulting in a weight distribution reflecting the degree of attention the query pays to various contextual features. Features related to model identification and inventory status have significantly higher weights than other features. Subsequently, based on this weight distribution, the features in the value vector are weighted and aggregated to extract the information most relevant to the user's query from the gate control context representation vector. Finally, the aggregation result is integrated with the original features of the query vector to generate a user query semantic enhancement encoding vector that incorporates contextual information. This vector clearly associates the specific product referred to by "this model" and incorporates the product's inventory-related features, fully representing the user's true intent to "query the inventory of a specific model product."

[0055] In the aforementioned intelligent interaction method based on dynamic intent recognition, step S5 involves inferring intent based on the user query semantic augmented encoding vector with integrated contextual information to obtain the final intent. It should be understood that although the user query semantic augmented encoding vector with integrated contextual information has integrated the semantic features of the query and key context, this vector is a low-dimensional, dense numerical representation and cannot be directly used as a structured intent result that can be executed by the system. Therefore, this application further infers intent based on the user query semantic augmented encoding vector with integrated contextual information to transform the semantic representation into a structured final intent containing intent type and key information. This allows abstract semantic information to be transformed into a concrete and clear description of user needs, providing a direct basis for the subsequent intelligent interaction engine to generate accurate responses, ensuring that the system can accurately respond to the user's true intent, especially when dealing with queries containing ambiguous references or omitted information, significantly improving the completeness and accuracy of intent understanding.

[0056] In particular, in one specific embodiment, Figure 6 This is a flowchart of sub-step S5 of the intelligent interaction method based on dynamic intent recognition according to an embodiment of this application. Figure 6 As shown, step S5 includes: S51, inputting the user query semantic augmented encoding vector with fused context information into the intent classifier to obtain the predicted intent label; S52, inputting the user query semantic augmented encoding vector with fused context information into the slot filler to obtain the extracted slot map; S53, performing structured intent assembly on the predicted intent label and the extracted slot map to obtain the final intent.

[0057] Specifically, in step S51, the user query semantic augmentation encoding vector with fused context information is input into the intent classifier to obtain a predicted intent label. It should be understood that although the user query semantic augmentation encoding vector with fused context information contains complete semantic features, it does not clearly define the specific type of user intent, making it impossible to directly determine whether the user is querying information, requesting a service, or performing other operations. Without a clear definition of the intent type, subsequent processing will lack a clear direction. Therefore, this application inputs the user query semantic augmentation encoding vector with fused context information into the intent classifier, which contains intent prototype vectors corresponding to multiple predefined intents, to identify and output predicted intent labels for user intents. This clarifies the core purpose of user interaction, providing a basic framework for subsequent extraction of specific information and generation of responses, ensuring that the system's understanding of user needs transitions from the semantic level to the functional level, and improving the targeting of intent recognition.

[0058] Specifically, in one possible embodiment, step S51 is implemented as follows: The intent classifier comprises a multi-layer fully connected network and an output layer. The output layer corresponds to a preset set of intent categories, such as checking inventory, inquiring about prices, and applying for after-sales service. After the enhanced encoding vector of the user query semantics, which integrates the query semantics of "Is this model in stock?" with the corresponding refrigerator entity information, is input into the classifier, the multi-layer fully connected network performs deep extraction of features from the vector, focusing on capturing intent features related to "in stock" and "model". The output layer maps the extracted features to preset categories through an activation function, ultimately outputting a predicted intent label for "checking product inventory," clearly indicating that the user's current core need is to understand the inventory status of a specific product.

[0059] Specifically, in step S52, the user query semantic enhancement encoding vector with fused context information is input into the slot filler to obtain the extracted slot map. It should be understood that since the predicted intent label only defines the type of user intent and lacks specific entity information supporting that intent (such as the product model, attributes, etc.), the specific target of the user intent cannot be clearly defined if only the label is present, leading to incomplete information and difficulty in accurate execution of subsequent processing. Therefore, this application further inputs the enhancement encoding vector into the slot filler, which generates entity embedding vectors corresponding to the slot entities. This extracts key entity and attribute information related to the intent from the vectors, forming structured slot data. This supplements the intent label with specific details, transforming the abstract intent type into an information set containing specific targets, ensuring the system's understanding of the completeness and accuracy of the user intent, and providing data support for subsequent assembly of the complete intent.

[0060] Specifically, in one possible embodiment, step S52 is implemented as follows: The slot filler adopts a sequence labeling-based model structure, and the preset slot types include product model, query attribute, etc. After the user query semantic enhancement encoding vector with fused contextual information is input into the filler, the model analyzes the fused semantic features in the vector to identify the specific product identifier corresponding to "this model", such as the model code of a refrigerator, and determines that the query attribute pointed to by "Is it in stock?" is inventory status. Subsequently, these identification results are organized according to slot type to form an extracted slot map containing "Product Model: [Specific Code]" and "Query Attribute: Inventory", clearly presenting the key information supporting the user's intent.

[0061] Specifically, step S53 involves assembling the predicted intent tags and extracted slot maps into a structured intent to obtain the final intent. Specifically, the predicted intent tags and extracted slot maps describe user intent from the perspectives of type and specific information, respectively. This application performs a structured assembly of the predicted intent tags and extracted slot maps, associating and integrating intent types with specific information to form a unified final intent. This transforms intent information from scattered types and details into a structured and complete description of requirements, ensuring that the system's understanding of user intent is both clear and specific, providing a direct and complete basis for the intelligent interaction engine to generate targeted responses.

[0062] Specifically, in one possible embodiment, step S53 is implemented as follows: First, the system reads the predicted intent label "Query Product Inventory" to determine the core framework of the assembly; then, it extracts information such as "Product Model: [Specific Code]" and "Query Attribute: Inventory" from the extraction slot map; finally, according to the template "[Intent Type]: [Query Attribute] for [Product Model]", the two are integrated into the final intent "Query Product Inventory: Inventory Status of Refrigerator for [Specific Code]". This final intent clearly defines the user's interaction purpose and includes the specific product and attribute, fully reflecting the user's true needs.

[0063] In particular, in another possible preferred embodiment, intent inference is performed based on the user query semantic augmented encoding vector with fused contextual information to obtain the final intent. This includes: before performing intent inference, performing semantic space correction on the user query semantic augmented encoding vector to obtain a corrected user query semantic augmented encoding vector; calculating the semantic distance between the corrected user query semantic augmented encoding vector and each intent prototype vector, and determining the intent corresponding to the closest intent prototype vector as the predicted intent label; and adjusting the entity embedding vector through contrastive learning to enhance its semantic similarity with the intent prototype vector corresponding to the predicted intent label; wherein the intent prototype vector corresponding to the predicted intent label and the adjusted entity embedding vector together constitute the final intent and serve as contextual information for the next round of interaction to ensure the homogeneity and continuity of contextual semantic representation in multiple rounds of interaction.

[0064] Specifically, here, where the intent is predicted as a discrete label (string) and the entity is extracted as a text fragment (string), if it continues to be used as context input in the next round of dialogue, it still needs to go through the text-to-embedded conversion process, which will cause the following problems.

[0065] First, the lossy transformation path from vector to label / string to text and back to vector results in rich semantic representations being compressed into string labels. When used as context in the next round, the semantics need to be re-inferred, leading to semantic discontinuity. Furthermore, the original UI context is directly embedded, the dialogue history is semantically encoded, and the intent / entity context from the previous round is encoded by newly generated descriptive text. This means that the generation methods and semantic spaces of context vectors from these three sources may be inconsistent, making it difficult to establish precise cross-modal associations during fusion—that is, there is an encoding consistency problem under representational heterogeneity.

[0066] Therefore, when predicting discrete labels, the intent classifier is expected to generate or match semantic vectors to directly and losslessly serve as context in subsequent dialogues to ensure encoding consistency. Specifically, prototype vectors can be pre-learned or defined for each predefined intent, making the intent classification task a matter of calculating which intent prototype the query's fusion embedding is closest to. Furthermore, for entities extracted from text, not only should their entity types be identified, but contrastive learning should also be used to ensure that the generated embedding vectors are close to the corresponding intent prototype representations.

[0067] First, assume there are K possible intent labels, each associated with a learnable embedding vector, i.e., an intent prototype vector, denoted as P. i For each i = 1 to K, the intent classification process becomes calculating the distance between the user query semantic augmentation encoded vector that incorporates contextual information and all intent prototype vectors, and selecting the closest distance as the prediction result.

[0068] Here, considering the potential offset between the user query semantic enhancement encoding vector formed by the fused context information and the intent semantic space constituted by the intent prototype vector, for the user query semantic enhancement encoding vector fused with context information, for example denoted as V, firstly, a P-query embedding is performed on the intent prototype matrix, for example denoted as a two-dimensional arrangement of all intent prototype vectors. Simultaneously, before performing intent inference, the user query semantic enhancement encoding vector is semantically space-corrected to obtain the corrected user query semantic enhancement encoding vector, i.e., spatial class constraint boundaryization is performed:

[0069]

[0070] Where exp(·) denotes an exponential function with base e, V represents the user query semantic enhancement encoding vector, and P represents the intent prototype matrix. ||P|| represents the square of the L2 norm of the semantic enhancement encoding vector for user queries. F The F-norm of the prototype matrix represents the intention. ⊙ represents matrix multiplication, ⊙ represents element-wise multiplication, and in the above formula, it represents element-wise scaling of scalars and vectors. V' represents the corrected user query semantic enhancement encoding vector.

[0071] In other words, semantic space correction is applied to the semantically enhanced encoding vector of the user query to eliminate the offset between the vector and the intent semantic space. This ensures that the corrected semantically enhanced encoding vector of the user query can be accurately mapped to the semantic space where the intent prototype vector is located. This ensures that the semantic distance calculated subsequently truly reflects the correlation strength between the query vector and each intent prototype vector. For example, when a user queries the inventory status of a specific product model, the corrected query vector can accurately fall into the semantic space region corresponding to the intent to query the product inventory, laying the foundation for accurate matching of intent prototype vectors, avoiding misjudgment of intent due to spatial offset, and ensuring that the subsequent intent classification process can be carried out based on accurate vector correlations.

[0072] Then, the semantic distance between the corrected user query semantic enhancement encoding vector and each intent prototype vector is calculated, that is, the Euclidean distance between vectors in different semantic spaces is calculated. Here, if the spatial offset of the overall semantic space is not corrected, it may affect the distribution characteristics of the Euclidean distance. For example, when the distance between vectors is in the same direction as the spatial offset, the distance representation may increase, while when the distance between vectors is in the opposite direction to the spatial offset, the distance representation may decrease.

[0073] Then, by activating the softmax function with a negative distance, the intent prototype vector corresponding to the highest probability value is obtained to determine the predicted intent label. That is, the intent corresponding to the intent prototype vector with the closest distance is determined as the predicted intent label and used as the contextual semantic representation for the next round of dialogue. For example, when the semantic distance between the corrected user query semantic enhancement encoding vector and the intent prototype vector for querying product inventory is the closest, it can be clearly determined that the user's current intent is to query the inventory status of a specific product, rather than inquiring about the price or requesting after-sales service. This provides a clear intent framework for the subsequent extraction of key entity information such as product model and inventory attributes, ensuring a smooth transition of intent recognition from the semantic level to the functional level and avoiding deviations in subsequent processing direction due to unclear intent categories.

[0074] Simultaneously, for entity fragments identified from the query text, when obtaining their embedded representations, this application adjusts the entity embedding vectors through contrastive learning to ensure that these embedding representations possess clear type semantics. This enhances the semantic similarity between the entity embedding vectors and the corresponding intent prototype vectors of the predicted intent labels, enabling the entity vectors to possess semantic features matching the intent type. This allows the entity embedding vectors to form a close semantic association with the corresponding intent prototype vectors. For example, after adjustment, the entity vector for a specific product model will exhibit stronger similarity in the semantic space to the intent prototype vector for querying product inventory. This ensures that during subsequent intent assembly, the entity can accurately serve as supporting information for the intent to query product inventory, avoiding user misunderstandings caused by entity-intent misalignment and guaranteeing a high degree of consistency between the entity and intent in the final intent.

[0075] That is, for an entity embedding vector, for example denoted as E, a contrastive loss function is introduced:

[0076]

[0077] Where E represents the entity embedding vector, Sim represents the similarity function (e.g., cosine similarity), and P is the entity embedding vector. i It is the intent prototype vector corresponding to the predicted intent label, P j τ is the prototype vector of intent corresponding to other intent labels in the intent label set. τ is a hyperparameter used to enable the model to better distinguish between consistent positive samples and inconsistent negative samples, forcing the model to learn features with more consistent discriminative power. That is, τ can be used to form a cluster representation with good discriminative power in the embedding space. Loss represents the contrastive loss function.

[0078] In this way, both intent and entity representations are transformed from a discrete, semantically unrelated label space to a continuous, semantically rich vector space compatible with other contexts (such as UI entities). This ensures that all contexts (UI, history, previous intent / entity) exist in a homogeneous semantic vector form, guaranteeing encoding consistency. Specifically, the intent prototype vector corresponding to the predicted intent label and the adjusted entity embedding vector together constitute the final intent and serve as context information for the next round of interaction, ensuring the homogeneity and continuity of contextual semantic representations across multiple rounds of interaction. For example, when a user, after checking the inventory of a specific product, further inquires "Is there a discount?", the system can directly utilize the intent prototype vector and product model entity vector from the previous round as context to accurately identify the specific product referred to by "it," quickly positioning the user's current intent as a query for a specific product discount. This eliminates the need to re-parse the contextual semantics, improving the efficiency and accuracy of intent understanding in multi-round interactions and ensuring that user needs are responded to consistently and accurately throughout continuous interactions.

[0079] Specifically, in another specific embodiment of this application, the aforementioned intelligent interaction method based on dynamic intent recognition further includes: inputting the final intent and user query into an intelligent interaction engine based on a large language model to obtain interactive response text. It should be understood that although the final intent clearly defines the user's structured needs, it lacks the natural language expression features of the user's original query. While the user query retains its original expression, it does not integrate key information supplemented by context. If a response is generated based solely on one of these, it may sound stiff due to a lack of natural language, or it may be inaccurate due to incomplete information. Therefore, this application further inputs the final intent and user query into an intelligent interaction engine based on a large language model to combine the precise structured needs with the expression style of the original query, generating interactive response text that meets the user's expectations. This ensures that the response accurately carries the information needed by the user and conforms to the user's language habits, improving the naturalness of the interaction and user satisfaction, while avoiding response deviations caused by information fragmentation.

[0080] Specifically, in one possible embodiment, the intelligent interaction engine first receives the final intent and the user query, where the final intent is "to check the inventory status of refrigerator model XXX," and the user query is "Is this model in stock?". The engine first parses the final intent, extracting its core information, namely, the need to provide feedback on the inventory status of a specific refrigerator model. It then analyzes the language style of the user query, recognizing that the user is using a concise and direct approach. Subsequently, the large language model, based on preset response generation rules, combines the key information from the final intent with the expressive features of the user query. It first confirms the specific refrigerator model corresponding to "this model," and then, based on real-time inventory data obtained by the system, generates a response text such as "The refrigerator model XXX you inquired about is currently in stock and can be ordered directly." This text accurately responds to the user's inquiry about inventory status while maintaining the user's concise and direct expression style, ensuring a natural and accurate interaction.

[0081] In summary, the intelligent interaction method based on dynamic intent recognition, as described in the embodiments of this application, is explained. It first acquires the user's current query and a context set including historical dialogues and browsing history. A parallel embedding model is then introduced to extract the semantic features of the user's query and the semantic features of each piece of information in the context set. A gating network guided by the current query dynamically weights and filters the context features based on relevance, achieving adaptive deep fusion of key context information and the user query. Intent classification and inference are then performed based on this enhanced semantic representation to obtain the final intent recognition result. This approach solves the problems of insufficient utilization of context information and severe noise interference, thereby significantly improving the accuracy and robustness of intent understanding when handling queries with unclear referential meanings.

[0082] Figure 7 This is a block diagram of an intelligent interaction system based on dynamic intent recognition according to an embodiment of this application. Figure 7 As shown, the intelligent interaction system 100 based on dynamic intent recognition according to an embodiment of this application includes: a user query context acquisition module 110, used to acquire a user query and a context set; a parallel embedding encoding module 120, used to perform parallel embedding on the user query and the context set to obtain a user query embedding encoding vector and a context embedding encoding vector set; a relevance calculation and filtering module 130, used to perform relevance calculation and filtering on the context embedding encoding vector set based on the user query embedding encoding vector to obtain a gated context representation vector; a feature fusion module 140, used to perform feature fusion on the gated context representation vector and the user query embedding encoding vector to obtain a user query semantic enhancement encoding vector with fused context information; and an intent inference module 150, used to perform intent inference based on the user query semantic enhancement encoding vector with fused context information to obtain the final intent.

[0083] As described above, the intelligent interaction system 100 based on dynamic intent recognition according to the embodiments of this application can be implemented in various wireless terminals, such as servers with intelligent interaction algorithms based on dynamic intent recognition. In one possible implementation, the intelligent interaction system 100 based on dynamic intent recognition according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the intelligent interaction system 100 based on dynamic intent recognition can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the intelligent interaction system 100 based on dynamic intent recognition can also be one of many hardware modules of the wireless terminal.

[0084] Alternatively, in another example, the intelligent interaction system 100 based on dynamic intent recognition and the wireless terminal can also be separate devices, and the intelligent interaction system 100 based on dynamic intent recognition can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0085] Here, those skilled in the art will understand that the specific operations of each step in the above-described intelligent interaction system based on dynamic intent recognition have been referenced above. Figures 1 to 6 The description of the intelligent interaction method based on dynamic intent recognition is detailed here, and therefore, its repeated description will be omitted.

Claims

1. An intelligent interaction method based on dynamic intent recognition, characterized in that, include: Retrieve user query and context set; Parallel embedding is performed on the user query and the context set to obtain the user query embedding encoding vector and the context embedding encoding vector set; Based on the user query embedding encoding vector, the context embedding encoding vector set is subjected to relevance calculation and filtering based on a gated network to obtain the gated context representation vector; Feature fusion is performed on the gating context representation vector and the user query embedding encoding vector to obtain a user query semantic enhancement encoding vector that incorporates contextual information; Intent inference is performed on the semantically enhanced encoded vector of the user query based on fused contextual information to obtain the final intent.

2. The intelligent interaction method based on dynamic intent recognition according to claim 1, characterized in that, Parallel embedding of user queries and context sets to obtain user query embedding encoding vectors and context embedding encoding vector sets includes: inputting user queries into a text encoder containing a text segmenter to obtain the user query embedding encoding vectors.

3. The intelligent interaction method based on dynamic intent recognition according to claim 2, characterized in that, Parallel embedding is performed on the user query and the context set to obtain a set of user query embedding encoding vectors and a set of context embedding encoding vectors, including: Extract the first context from the context set; Read the data source field value of the first context; In response to the data source field value being UI, entity embedding encoding is performed on the entity ID and entity type of the first context to obtain a context embedding encoding vector; In response to the data source field value being dialogue history, the text field of the first context is input into a text encoder containing a text segmenter to obtain a context embedding encoding vector.

4. The intelligent interaction method based on dynamic intent recognition according to claim 1, characterized in that, Based on the user query embedding encoding vector, the context embedding encoding vector set is subjected to relevance calculation and filtering based on a gating network to obtain the gating context representation vector, including: The user query embedding encoding vector and each context embedding encoding vector in the context embedding encoding vector set are input into a gating network to obtain a set of relevance scores; The set of relevance scores is normalized using the Softmax function to obtain the set of context-related weights; The context embedding encoding vector corresponding to the largest context-related weight is selected from the set of context-related weights as the gated context representation vector.

5. The intelligent interaction method based on dynamic intent recognition according to claim 4, characterized in that, Feature fusion is performed on the gating context representation vector and the user query embedding encoding vector to obtain a user query semantically enhanced encoding vector with fused context information, including: The user query embedding encoding vector is used as the query vector, and the gating context representation vector is used as the key vector and value vector; The query vector, key vector, and value vector are input into the cross-attention layer to obtain the user query semantic enhancement encoding vector that incorporates the contextual information.

6. The intelligent interaction method based on dynamic intent recognition according to claim 5, characterized in that, Intent inference is performed based on the semantically augmented encoded vector of the user query, which incorporates contextual information, to obtain the final intent, including: The user query semantic augmentation encoding vector with fused context information is input into the intent classifier to obtain the predicted intent label; The user query semantic enhancement encoding vector with fused context information is input into the slot filler to obtain the extracted slot map; The predicted intent labels and extracted slot maps are used to perform structured intent assembly to obtain the final intent.

7. The intelligent interaction method based on dynamic intent recognition according to claim 1, characterized in that, Also includes: The final intent and user query are input into an intelligent interaction engine based on a large language model to obtain interactive response text.

8. An intelligent interaction system based on dynamic intent recognition, characterized in that, include: The user query context acquisition module is used to obtain user queries and context sets. The parallel embedding encoding module is used to perform parallel embedding on the user query and the context set to obtain the user query embedding encoding vector and the context embedding encoding vector set; The relevance calculation and filtering module is used to perform relevance calculation and filtering on the context embedding encoding vector set based on the user query embedding encoding vector to obtain the gated context representation vector. The feature fusion module is used to fuse the gating context representation vector and the user query embedding encoding vector to obtain a user query semantic enhancement encoding vector with fused context information. The intent inference module is used to infer intent based on the user query semantic augmented encoding vector with fused contextual information to obtain the final intent.

Images