Multi-intention recognition and intelligent response method based on large model and domain knowledge base

Abstract

This application relates to the field of intelligent question answering technology, and in particular to a multi-intent recognition and intelligent response method based on a large model and a domain knowledge base. The method includes: segmenting the user-input consultation text to obtain a semantic sequence; inputting the semantic sequence into a feature extractor to obtain attention distribution data for multi-head attention; calculating the significance index of each semantic segment based on the attention distribution data; determining multiple intent boundaries along the significance index curve formed by the semantic sequence to split the semantic sequence into multiple intent segments; retrieving the semantic features of each intent segment, obtaining corresponding knowledge slices from the domain knowledge base, inputting each knowledge slice into a large language model, and outputting a response text containing multiple response sub-blocks. The technical solution of this application can split consultation text into multiple intent segments, significantly improving the precision of knowledge retrieval and achieving segmented responses according to intent segments.

Description

Technical Field

[0001] This application relates to the field of intelligent question answering technology, and in particular to a multi-intent recognition and intelligent response method based on a large model and a domain knowledge base. Background Technology

[0002] With the rapid iteration of Large Language Modeling (LLM) and Retrieval Augmentation (RAG) technologies, intelligent customer service has been widely applied in vertical fields such as government affairs consultation and public resource transactions. In large service platforms, users often expect to quickly obtain professional guidance covering multiple dimensions, including laws and regulations, project processes, and financial settlements, through a single window. This requires intelligent systems to handle highly complex interaction scenarios and accurately extract business needs from unstructured natural language.

[0003] To address these needs, current retrieval methods typically employ keyword matching or single-intent classification. When a user inputs their query text, a pre-trained model vectorizes it, then searches a knowledge base. The retrieved knowledge slices are then concatenated with prompts, and a larger model generates the response. This approach is highly efficient for handling simple, semantically singular queries and is the mainstream practice within the RAG architecture.

[0004] However, since the above methods treat composite text as a single vector for retrieval, the semantic features of different intentions interfere with each other in the vector space. This results in retrieval results that often only cover one intention, or even semantic deviation, leading to irrelevant answers. It is difficult to identify multiple intentions in the consultation text and cannot achieve logically rigorous segmented responses. Summary of the Invention

[0005] To address the technical problem of difficulty in identifying multiple intents in consultation texts and the inability to achieve logically rigorous segmented responses, this application provides a multi-intent recognition and intelligent response method based on a large model and domain knowledge base. This method can pinpoint the critical points of intent switching in consultation texts, split the consultation texts into multiple intent fragments, significantly improve the precision of knowledge retrieval, and simultaneously achieve logically rigorous segmented responses according to intent fragments.

[0006] In a first aspect, this application provides a multi-intent recognition and intelligent response method based on a large model and a domain knowledge base. The response method includes: segmenting user-input consultation text using a business dictionary to obtain a semantic sequence; inputting the semantic sequence into a feature extractor to obtain attention distribution data for multi-head attention; calculating the saliency index of each semantic segment in the semantic sequence based on the attention distribution data, including: calculating the business item density of any semantic segment, wherein the business item density is directly proportional to the number of words in the business dictionary within the neighborhood of the semantic segment and inversely proportional to the total number of words; using the attention variance of multi-head attention on the semantic segment as the attention dispersion; the saliency index is positively correlated with both the business item density and the attention dispersion, and negatively correlated with the attention entropy of the semantic segment; determining multiple intent boundaries along the saliency index curve formed by the semantic sequence to split the semantic sequence into multiple intent segments; retrieving the semantic features of each intent segment, obtaining corresponding knowledge slices from the domain knowledge base, inputting each knowledge slice into a large language model in the order of intent segments, and outputting a response text containing multiple response sub-blocks.

[0007] By utilizing business dictionaries, attention distribution features, and saliency exponential curves to perform multi-dimensional intent determination and text segmentation, the mapping from unstructured text to intent fragments is realized, effectively avoiding the phenomenon of irrelevant answers caused by intent confusion, and improving the retrieval precision of the intelligent question answering system in multi-intent scenarios.

[0008] Preferably, the step of segmenting the user-input consultation text using a business dictionary to obtain a semantic sequence includes: the business dictionary includes legal terms, project terms, and financial terms; and the word segmentation tool is used to determine the boundaries of each semantic segment in the consultation text to obtain a semantic sequence composed of multiple semantic segments.

[0009] By using a business dictionary covering law, projects, and finance to assist in word segmentation, we ensured that core terms in the vertical domain were not broken down during the word segmentation stage, thus guaranteeing that the smallest processing unit of the semantic sequence contained business information.

[0010] Preferably, the feature extractor includes multiple parallel attention heads; the acquisition of attention distribution data for multi-head attention includes: calculating the attention between all semantic segments in each attention head, wherein the attention distribution data includes the attention of any semantic segment to other semantic segments in each attention head.

[0011] Preferably, the method for obtaining the attention entropy of the semantic segment includes: obtaining the attention weights of any semantic segment in an attention head to other semantic segments, normalizing each attention weight to obtain the attention probability distribution of the semantic segment in the attention head, and calculating the information entropy of the attention probability distribution; and calculating the average value of the information entropy of the attention probability distribution of the semantic segment in each attention head as the attention entropy of the semantic segment.

[0012] By calculating the information entropy of the attention probability distribution and taking the average value, the model's information certainty on specific segments is quantified. This effectively identifies and suppresses colloquial noise in consultation texts that lack business value, thereby enhancing the ability to focus on the core intent area.

[0013] Preferably, the significance index is equal to the sum of the business item density and the first weighted dispersion, minus the second weighted attention entropy; wherein, the first weighted dispersion is the product of the attention dispersion and the first weight parameter, and the second weighted attention entropy is the product of the attention entropy and the second weight parameter.

[0014] By integrating business item density, weighted dispersion, and attention entropy, the resulting significance index can capture attention focus conflicts caused by intent switching, providing data support for accurately locating intent switching points.

[0015] Preferably, determining multiple intent boundaries along the saliency exponential curve formed by the semantic sequence includes: obtaining local minima in the saliency exponential curve and determining the local minima as intent boundaries.

[0016] By identifying the local minimum points of the significant exponential curve as the intent boundary, automated intent segmentation point locking is achieved.

[0017] Preferably, the determination of multiple intent boundaries further includes: starting from the first local minimum, calculating the semantic similarity of adjacent semantic segments before and after any local minimum point; in response to the semantic similarity being less than a preset similarity threshold, determining the local minimum point as an intent boundary, otherwise determining the local minimum point as a non-intent boundary.

[0018] Introducing semantic similarity comparison between adjacent segments can effectively eliminate false segmentation points caused by normal grammatical fluctuations, ensuring accurate and reliable intent boundary recognition and preventing semantic fragmentation caused by overly fine segmentation.

[0019] Preferably, the step of splitting the semantic sequence into multiple intent segments further includes: dividing the semantic sequence into multiple intent segments based on multiple intent boundaries; determining that the intent segment is a transitional expression in response to the business item density of the intent segment being less than a preset lower density limit and the attention dispersion being greater than a preset upper dispersion limit; calculating the cosine similarity between the transitional expression and adjacent semantic segments; and merging the transitional expression into the adjacent semantic segment with the highest similarity in response to the maximum similarity value being greater than a merging threshold, otherwise treating the transitional expression as an independent semantic segment to obtain the verified intent segment.

[0020] By performing transitional judgment and merging verification on segments with low business density and high dispersion, it is ensured that transitional phrases can be reasonably attributed to relevant intent segments, thus guaranteeing the semantic coherence of the split text and avoiding contextual breaks caused by hard segmentation.

[0021] Preferably, the step of retrieving the semantic features of each intent fragment includes: calculating the semantic vector of the intent fragment; calculating the cosine similarity between the semantic vector and the knowledge slice vectors in the domain knowledge base; and obtaining the matching knowledge slices based on the cosine similarity.

[0022] Preferably, the output, which includes response text comprising multiple response sub-blocks, comprises: identifying an identifier of intent boundaries using a large language model during the decoding stage; and automatically constructing multiple response sub-blocks based on the identifier.

[0023] The technical solution of this application has the following beneficial technical effects: The system breaks down consultation texts into semantic sequences comprising multiple semantic segments. It calculates the saliency index of each semantic segment within the consultation text by integrating business item density, attention dispersion, and attention entropy. Based on the saliency index curve, it accurately locates intent segmentation points in complex and composite consultation texts. This allows for precise capture of multiple implicit intents from users across different fields such as law, projects, and finance within a single input. Breaking long consultations down into multiple intent segments for querying significantly improves the precision of domain knowledge retrieval. Furthermore, it enables logically rigorous segmented responses based on intent segments, effectively solving the problem of irrelevant answers. Attached Figure Description

[0024] Figure 1 This is a flowchart of a multi-intent recognition and intelligent response method based on a large model and domain knowledge base according to an embodiment of this application.

[0025] Figure 2 This is a schematic diagram of the significance index curve according to an embodiment of this application. Detailed Implementation

[0026] This application provides a multi-intent recognition and intelligent response method based on a large model and domain knowledge base, applicable to smart customer service, capable of responding to inquiries containing multiple business requests. Taking a bidding scenario as an example, a bidder enters through smart customer service: "Where is the bid opening location for the GZ2026 section? By the way, can I withdraw and modify the bid bond transfer voucher if it was sent incorrectly?" This inquiry spans two independent intents: project information inquiry and financial process intervention. This application can dynamically identify the intent boundaries and accurately match the corresponding answer in a massive domain knowledge base, avoiding irrelevant answers due to semantic confusion. Figure 1 This is a flowchart of a multi-intent recognition and intelligent response method based on a large model and domain knowledge base, according to an embodiment of this application. Figure 1 As shown, the multi-intent recognition and intelligent response method based on a large model and domain knowledge base includes steps S101 to S105, which are described in detail below.

[0027] S101, use the business dictionary to segment the user-input consultation text to obtain a semantic sequence.

[0028] In one embodiment, a business dictionary refers to a pre-built set of professional terms in a specific vertical field; a semantic sequence refers to a set of semantic fragments with independent business meanings arranged in a sequential order.

[0029] To transform user-input consultation text into the smallest processing unit with business attributes and provide a data foundation for subsequent feature calculations, word segmentation is required. This involves segmenting the user-input consultation text using a business dictionary to obtain a semantic sequence. The business dictionary includes legal terms, project terms, and financial terms. A word segmentation tool is used to determine the boundaries of each semantic segment in the consultation text, resulting in a semantic sequence composed of multiple semantic segments.

[0030] For example, legal terms include conditions for bid rejection and determination of bid rigging; project terms include bid section number and qualification review method; and financial terms include bid bond and transfer voucher.

[0031] The business dictionary provides powerful semantic calibration during the word segmentation stage of consultation texts, ensuring that each semantic fragment can carry a clear business intent, guaranteeing the integrity of business terms, and avoiding semantic breaks caused by general word segmentation.

[0032] In this way, by transforming natural language into a semantic sequence with business attributes, the smallest processing unit with business intent is provided for subsequent feature calculation.

[0033] S102, the semantic sequence is input into the feature extractor to obtain the attention distribution data of multi-head attention.

[0034] In one embodiment, the feature extractor is a deep neural network encoder based on a Transformer structure. The Transformer structure includes multiple attention heads, and any attention head can acquire attention between semantic segments through a self-attention mechanism. The attention distribution data refers to the weight matrix assigned to each semantic segment by the multiple attention heads, and the weight matrix includes the attention of any attention head to all semantic segments.

[0035] The feature extractor comprises multiple parallel attention heads; obtaining the attention distribution data for multi-head attention includes calculating the attention between all semantic segments in each attention head, wherein the attention distribution data includes the attention of any semantic segment to other semantic segments in each attention head. Specifically, the feature extractor comprises 12 parallel attention heads, each of which maps the input semantic sequence to the query space, key space, and value space through a linear transformation, and determines the association weights between semantic segments by calculating dot product attention.

[0036] No. In the attention head, the first The semantic fragment for the first attention to semantic fragments Satisfying the relation:

[0037] In the formula, Indicates the first The first one in the attention. The semantic fragment for the first Attention probability values ​​for each semantic segment; and For the first The first one in the attention. The query vector of the semantic fragment and the first semantic fragment The key vector of each semantic segment; In order to query the feature dimension of the vector and the key vector, in this embodiment of the application, The value is 64; This represents the total number of semantic segments in the semantic sequence.

[0038] In this embodiment, the method for obtaining the attention entropy of the semantic segment includes: obtaining the attention weight of any semantic segment in an attention head to other semantic segments, normalizing each attention weight to obtain the attention probability distribution of the semantic segment in the attention head, and calculating the information entropy of the attention probability distribution; and calculating the average value of the information entropy of the attention probability distribution of the semantic segment in each attention head as the attention entropy of the semantic segment.

[0039] No. Attention entropy of semantic segments Satisfying the relation:

[0040] In the formula, Indicates the first The first one in the attention. The semantic fragment for the first Attention probability values ​​for each semantic segment; The total number of attention heads is preferably 12 in this embodiment.

[0041] Understandably, the first The attention entropy of the semantic segment reflects the feature extractor's attention entropy in the first semantic segment. The certainty of information in semantic fragments varies. High-entropy semantic fragments are typically transitional, connective, and colloquial, while low-entropy semantic fragments usually correspond to core business intent. Therefore, when A high value indicates that the feature extractor's focus at that semantic segment is extremely scattered, making it impossible to form an effective semantic focus. In the subsequent construction of the saliency index, by subtracting the weighted attention entropy, the interference of colloquial noise on the determination of intent boundaries can be effectively suppressed.

[0042] Thus, by utilizing the self-attention mechanism, multi-dimensional weight matrices and entropy features were extracted, providing data support for the subsequent construction of a saliency index to locate intent switching points.

[0043] S103, calculate the saliency index of each semantic segment in the semantic sequence based on attention distribution data.

[0044] In one embodiment, service item density is used to quantify the proportion of the strength of the real service signal carried by the semantic segment; attention dispersion refers to the quantitative value used to measure semantic focus conflict by calculating the attention difference of the same semantic segment at different attention heads.

[0045] Understandably, attention entropy measures the average level of determinism in the attention distribution within each attention head, while attention dispersion measures the difference in attention to the same semantic segment across different attention heads, used to distinguish whether there is a semantic focus conflict between attention heads for the same semantic segment.

[0046] To quantify which segments carry genuine business signals and effectively eliminate colloquial background noise, a saliency index is constructed based on attention distribution data by coupling business attributes with semantic focus conflicts within the model. The calculation of the saliency index for each semantic segment in the semantic sequence based on attention distribution data includes: calculating the business item density of any semantic segment, where the business item density is directly proportional to the number of words in the business dictionary within the semantic segment's neighborhood and inversely proportional to the total number of words; using the variance of multi-head attention on the semantic segment as the attention dispersion; the saliency index is positively correlated with both business item density and attention dispersion, and negatively correlated with the attention entropy of the semantic segment.

[0047] The purpose of building a business item density is to filter out high-value signal areas related to the business, in order to distinguish between core business needs and routine pleasantries. Business item density of semantic fragments Satisfying the relation:

[0048] In the formula, Indicates the first The neighborhood range of the semantic fragment, including the first The semantic fragment and the first The semantic segments on the left and right sides of each semantic segment; It is a collection of business dictionaries containing legal terms, project terms, and financial terms; For the first Within the semantic segment, the first The tag value of the nth word, if the nth word If a word belongs to the business dictionary set, then The value is 1, otherwise The value is 0; For the first The total number of words within the neighborhood of a semantic segment.

[0049] The reason for constructing attention discreteness is to use the conflict performance of multi-head attention to characterize the differences of a semantic segment among various attention heads, and to deduce the critical point of intent switching in reverse.

[0050] No. The first attention point in the Attention weights on semantic segments Satisfying the relation:

[0051] In the formula, Indicates the first In the attention head, the first The semantic fragment for the first Attention probability values ​​for each semantic segment; Let be the total number of semantic segments in the semantic sequence. At this point, for the ... Given the _th semantic segment, we can obtain the attention weight of each attention head for that semantic segment. The attention discreteness of each semantic segment satisfies the following relation:

[0052] In the formula, This represents the attention dispersion, i.e., the variance of the weights of each attention head; Indicates the total number of heads of attention; Indicates the first The first attention point in the Attention weights on each semantic segment; For all attention head in the first The average attention value of each semantic segment.

[0053] In this embodiment, the significance index is equal to the sum of the business item density and the first weighted dispersion, minus the second weighted attention entropy. Significance index of semantic segments Satisfying the relation:

[0054] In the formula, For the first Business item density of each semantic fragment; Attention dispersion; For the first Attention entropy of a semantic segment; and These are the first weighting parameter and the second weighting parameter, used to balance the impact of traffic density, multi-head dispersion, and attention entropy on intent recognition. For example, the first weighting parameter... The value is 0.3, the second weight parameter. The value is 0.2.

[0055] Understandably, The first weighted dispersion, The second weighted attention entropy is used; the saliency index is used to comprehensively evaluate the consistency between the business value of semantic fragments and the model's focus. When a user's consultation text shifts from one intent to another (e.g., from "bid opening location" to "deposit withdrawal"), the transition words cause a decrease in business item density. Simultaneously, multiple attention heads in the feature extractor may experience feature acquisition conflicts due to the switching of semantic goals, leading to drastic fluctuations in the weight distribution of different attention heads in the transition region. This increases the attention dispersion of semantic fragments within the transition region, resulting in significant numerical fluctuations in the saliency index sequence at the boundaries of each intent. Furthermore, the attention entropy of semantic fragments is introduced into the saliency index to reduce the saliency index of transitional, connective, and colloquial content.

[0056] Thus, a reliable quantitative indicator is provided for the critical point at which the intention to lock on will switch.

[0057] S104, determine multiple intent boundaries along the significant exponential curve formed by the semantic sequence to split the semantic sequence into multiple intent segments.

[0058] In one embodiment, arranging the semantic segments according to their order in the semantic sequence and using the significance index as the ordinate value yields, as shown below. Figure 2 The significance index curve shown reflects the trajectory of the significance index. The intent boundary refers to the critical position where different business intents switch within the consultation text; an intent fragment refers to an independent text block that, after being broken down, possesses a single and clear intent.

[0059] Determining multiple intent boundaries along the saliency exponential curve formed by the semantic sequence includes: obtaining local minima in the saliency exponential curve and determining the local minima as intent boundaries.

[0060] Specifically, if the The saliency index of each semantic segment satisfies and Then determine the first Each semantic segment is a potential segmentation point. The determination of multiple intent boundaries further includes: starting from the first local minimum, calculating the semantic similarity between adjacent semantic segments before and after any local minimum point; in response to a semantic similarity less than a preset similarity threshold, determining the local minimum point as an intent boundary; otherwise, determining the local minimum point as a non-intent boundary.

[0061] Understandably, the saliency index can sensitively capture the dynamic evolution of consultation intent. When transitioning from one intent to another, the combined effect of a drop in business item density and a surge in attention dispersion causes significant troughs, i.e., local minima, to appear at the boundaries between intents in the saliency index sequence. At this point, a secondary verification using a preset similarity threshold can effectively eliminate false segmentation points caused by normal word order fluctuations, ensuring the accuracy of intent segmentation. The preferred preset similarity threshold is 0.6.

[0062] Furthermore, the step of splitting the semantic sequence into multiple intent segments also includes: dividing the semantic sequence into multiple intent segments based on multiple intent boundaries; determining that the intent segment is a transitional expression in response to the business item density of the intent segment being less than a preset lower density limit and the attention dispersion being greater than a preset upper dispersion limit; calculating the cosine similarity between the transitional expression and adjacent semantic segments; and merging the transitional expression into the adjacent semantic segment with the highest similarity in response to the maximum similarity value being greater than a merging threshold, otherwise treating the transitional expression as an independent semantic segment to obtain the verified intent segment, with a merging threshold of 0.6.

[0063] The lower density limit and upper dispersion limit are both preset. For example, the lower density limit is set to 0.2, and the upper dispersion limit is set to 0.5. If a semantic fragment, such as "I would like to ask," has a business item density... And attention dispersion If the expression is transitional, it is then determined to be a transitional expression. Subsequently, the cosine similarity between the semantic segment and its left and right adjacent intent segments is calculated. If the cosine similarity with the right intent segment is the largest and greater than [value missing], then [the expression is considered transitional]. If the cosine similarity with both right-side intention segments is less than 1, then merge it into the right-side intention segment; if the cosine similarity with both right-side intention segments is less than 1, then merge it into the right-side intention segment. If the semantic fragment is retained as an independent semantic fragment, the semantic coherence of the intent fragment is verified, thus avoiding the semantic breakage problem caused by hard merging.

[0064] In this way, by utilizing the changing trend of the significant exponential curve, the problem of ambiguous intent boundaries in complex consultation scenarios is solved. Through dynamic segmentation and semantic coherence verification, adaptive decoupling of consultation text is achieved, laying the foundation for subsequent accurate retrieval.

[0065] S105: Retrieve the semantic features of each intent fragment, obtain the corresponding knowledge slices from the domain knowledge base, input each knowledge slice into the large language model in the order of intent fragments, and output the response text containing multiple response sub-blocks.

[0066] In one embodiment, a knowledge slice refers to the smallest unit of information retrieved from a domain knowledge base that is highly semantically matched to a specific intent fragment.

[0067] To achieve intelligent responses in complex consultation scenarios, the semantic feature retrieval of each intent fragment includes: calculating the semantic vector of the intent fragment; calculating the cosine similarity between the semantic vector and the knowledge slice vectors in the domain knowledge base; and obtaining the matching knowledge slice based on the cosine similarity. Specifically, a pre-trained encoder is used to transform the decoupled intent fragments into numerical representations in a high-dimensional feature space. Retrieval in the domain knowledge base ensures that each retrieval request has a single and clear semantic focus.

[0068] Furthermore, the output, comprising response text with multiple response sub-blocks, includes: identifying identifiers of intent boundaries using a large language model during the decoding phase; and automatically constructing multiple response sub-blocks based on these identifiers. It should be noted that inputting each knowledge slice into the large language model in the order of intent fragments is to ensure that the logical order of the responses remains consistent with the user's initial inquiry.

[0069] Understandably, after receiving knowledge slices assembled in atomic intent order, the large language model will guide text generation by recognizing preset intent boundary markers, which can be special placeholders or line break markers.

[0070] For example, the semantic sequence of the consultation text “Where is the bid opening location for the GZ2026 section? By the way, can the bid security deposit transfer voucher be withdrawn and modified if it is sent incorrectly?” includes a total of 5 semantic segments, as shown in Table 1, Semantic Segment Index Table.

[0071] Table 1 Semantic Fragment Index Table Index 1 Location of bid opening for section GZ2026 Index 2 Where Index 3 By the way, could you please tell me? Index 4 Margin transfer voucher Index 5 Can the changes be reversed? along Figure 2 The significant exponential curve shown is traversed, and index 3 is identified as a local minimum point of the curve, satisfying the condition... and The judgment criteria are used to further calculate the semantic similarity before and after index 3, since the cosine similarity of the semantic vectors of index 2 and index 4 is... Therefore, index 3 was confirmed as the intent boundary point. Subsequent splitting and validation were performed, due to the high density of business items in index 3. And attention dispersion It was determined to be a transitional expression. By calculating the similarity between index 3 and its adjacent segments, it was found that its cosine similarity to the subsequent intent "Can the wrong transfer voucher for the margin deposit be withdrawn and modified?" was higher. Therefore, index 3 was merged into the subsequent intent. Finally, the original semantic sequence was precisely split into two intent segments: Intent segment A (composed of indices 1-2): semantic focus is "bid opening location query"; Intent segment B (composed of indices 3-5): semantic focus is "margin deposit process intervention".

[0072] The final output response text will contain two distinct sub-blocks: the first part, based on the first set of knowledge slices, informs the user of the specific bidding location; the second part, based on the second set of knowledge slices, provides detailed guidance on the withdrawal and modification process after an error in transmitting the deposit transfer voucher. This segmented generation method not only meets the multiple needs of users in a single inquiry text but also enhances the interactive experience of intelligent customer service through structured expression.

[0073] It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this patent application shall be determined by the appended claims.

Claims

1. A multi-intent recognition and intelligent response method based on a large model and domain knowledge base, characterized in that, The response method includes: The user-input consultation text is segmented using a business dictionary to obtain a semantic sequence; The semantic sequence is input into a feature extractor to obtain attention distribution data for multi-head attention; The significance index of each semantic segment in the semantic sequence is calculated based on attention distribution data, including: calculating the business item density of any semantic segment, wherein the business item density is directly proportional to the number of words in the business dictionary within the neighborhood of the semantic segment and inversely proportional to the total number of words; using the attention variance of multi-head attention on the semantic segment as the attention dispersion; the significance index is positively correlated with both business item density and attention dispersion, and negatively correlated with the attention entropy of the semantic segment. Multiple intent boundaries are determined along the significant exponential curve formed by the semantic sequence to split the semantic sequence into multiple intent segments; The semantic features of each intent fragment are retrieved, the corresponding knowledge slices are obtained from the domain knowledge base, and the knowledge slices are input into the large language model in the order of the intent fragments. The output is a response text containing multiple response sub-blocks.

2. The multi-intent recognition and intelligent response method based on a large model and domain knowledge base according to claim 1, characterized in that, The step of segmenting the user-input consultation text using a business dictionary to obtain a semantic sequence includes: The business dictionary includes legal terms, project terms, and financial terms; By using word segmentation tools to determine the boundaries of each semantic segment in the consultation text, a semantic sequence composed of multiple semantic segments is obtained.

3. The multi-intent recognition and intelligent response method based on a large model and domain knowledge base according to claim 1, characterized in that, The feature extractor comprises multiple parallel attention heads; The attention distribution data for acquiring multi-head attention includes: Calculate the attention among all semantic segments in each attention head, wherein the attention distribution data includes the attention of any semantic segment to other semantic segments in each attention head.

4. The multi-intent recognition and intelligent response method based on a large model and domain knowledge base according to claim 1, characterized in that, The method for obtaining the attention entropy of the semantic segment includes: obtaining the attention weight of any semantic segment in an attention head to other semantic segments, normalizing each attention weight to obtain the attention probability distribution of the semantic segment in the attention head, and calculating the information entropy of the attention probability distribution; The average value of the information entropy of the attention probability distribution of the semantic segment in each attention head is calculated as the attention entropy of the semantic segment.

5. The multi-intent recognition and intelligent response method based on a large model and domain knowledge base according to claim 1, characterized in that, The significance index is equal to the sum of the business item density and the first weighted dispersion, minus the second weighted attention entropy; Wherein, the first weighted dispersion is the product of the attention dispersion and the first weight parameter, and the second weighted attention entropy is the product of the attention entropy and the second weight parameter.

6. The multi-intent recognition and intelligent response method based on a large model and domain knowledge base according to claim 1, characterized in that, Determining multiple intent boundaries along the saliency exponential curve formed by the semantic sequence includes: Obtain the local minimum points in the significant exponential curve, and determine the local minimum points as the intention boundary.

7. The multi-intent recognition and intelligent response method based on a large model and domain knowledge base according to claim 6, characterized in that, The determination of multiple intent boundaries also includes: Starting from the first local minimum, calculate the semantic similarity between adjacent semantic segments before and after any local minimum point; If the semantic similarity is less than a preset similarity threshold, the local minimum point is determined as the intention boundary; otherwise, the local minimum point is determined as the non-intention boundary.

8. The multi-intent recognition and intelligent response method based on a large model and domain knowledge base according to claim 1, characterized in that, The step of splitting the semantic sequence into multiple intent fragments also includes: The semantic sequence is segmented into multiple intent segments based on multiple intent boundaries; If the business item density of an intent fragment is less than a preset lower density limit and the attention dispersion is greater than a preset upper dispersion limit, the intent fragment is determined to be a transitional expression. Calculate the cosine similarity between the transitional expression and adjacent semantic segments; In response to a maximum similarity value greater than the merging threshold, the transitional expression is merged into the adjacent semantic segment with the highest similarity; otherwise, the transitional expression is treated as an independent semantic segment to obtain the verified intent segment.

9. The multi-intent recognition and intelligent response method based on a large model and domain knowledge base according to claim 1, characterized in that, The retrieval of semantic features for each intent fragment includes: Calculate the semantic vector of the intent segment; Calculate the cosine similarity between the semantic vector and each knowledge slice vector in the domain knowledge base, and obtain the matching knowledge slice based on the cosine similarity.

10. The multi-intent recognition and intelligent response method based on a large model and domain knowledge base according to claim 1, characterized in that, The output includes response text comprising multiple response sub-blocks, including: Large language models are used to identify intent boundaries during the decoding stage; Multiple response sub-blocks are automatically constructed based on the identifier.

Images