Intention recognition method and device based on domain knowledge seed words and computer device

Abstract

This application relates to an intent recognition method, apparatus, and computer device based on domain knowledge seed words. The method includes: acquiring and preprocessing a planning document corpus; calculating term importance weights to filter initial seed words; dividing the corpus into hierarchical seed words using a large language model; expanding the corpus using a dual-dimensional approach of general vocabulary similarity and domain context embedding similarity; filtering the corpus with a domain dictionary to obtain a hierarchical expanded seed word dictionary; increasing the Dirichlet prior weight of the corresponding level's expanded seed words in the topic model based on hierarchical matching relationships according to the expanded word dictionary corresponding to each level; performing topic modeling on the planning document corpus; outputting the vocabulary distribution of each topic and the topic distribution of each document; constructing a clustering tree based on topic semantic distance; filtering the optimal partition; and generating hierarchical intent results by combining the hierarchical structure. This method can effectively achieve automated and hierarchical intent recognition, ensuring that the recognition results align with domain business logic.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to an intent recognition method, apparatus, and computer device based on domain knowledge seed words. Background Technology

[0002] With the rapid growth of text data, automatically identifying latent intents from massive amounts of unstructured text has become a crucial requirement in the field of information processing. Traditional intent recognition methods mainly fall into two categories: one relies on manual qualitative analysis, where professionals interpret the text to extract the core intent. This approach is not only inefficient and difficult to adapt to batch processing scenarios of large-scale text, but it is also easily influenced by the subjective experience of analysts, leading to insufficient consistency and reliability of the recognition results. The other category employs purely data-driven topic modeling techniques, generating latent topics by mining the co-occurrence patterns of words in the text, and then deriving the intent.

[0003] However, these methods lack domain-specific guidance, resulting in generated topics that are often weakly related to actual business scenarios. They struggle to establish effective connections between underlying textual features and high-level intent concepts, leading to semantic gaps and poor interpretability of intent recognition results, failing to meet the demands for precise intent mining. Therefore, there is an urgent need for a text intent recognition technology that integrates domain knowledge while balancing recognition efficiency and accuracy to address the aforementioned problems of traditional methods. Summary of the Invention

[0004] Therefore, it is necessary to provide an intent recognition method, apparatus, and computer device based on domain knowledge seed words to address the aforementioned technical problems.

[0005] An intent recognition method based on domain knowledge seed words, the method comprising:

[0006] Obtain a corpus of planning documents containing target behavior planning, preprocess the planning document corpus, and calculate the importance weight of each term in the preprocessed planning document corpus;

[0007] An initial seed word set is obtained by filtering based on importance weights. The initial seed word set is then divided into seed words at each level according to the concept granularity of a preset hierarchical structure using a large language model.

[0008] Semantic expansion is performed on seed words at each level. First candidate words and second candidate word sets are generated based on lexical similarity from a general vocabulary semantic database and cosine similarity based on domain context embedding. The first candidate words and second candidate word sets are merged and filtered by a domain-specific dictionary to obtain the extended seed word dictionary corresponding to each level.

[0009] Based on the extended word dictionary corresponding to each level, the Dirichlet prior weight of the corresponding level extended seed words in the topic model is increased according to the hierarchical matching relationship. To the topic modeling of the planned document corpus, the vocabulary distribution of each topic and the topic distribution of each document are output.

[0010] The semantic distance between topics is calculated based on the word distribution of each topic, and a topic distance matrix is ​​constructed. All topics are then hierarchically organized according to the topic distance matrix to construct a topic clustering tree. The topic clustering tree is then subjected to clustering partitioning and filtering to obtain the optimal clustering partitioning result.

[0011] Based on the optimal clustering results and the preset hierarchical structure, a hierarchical intent recognition result is generated.

[0012] An intent recognition device based on domain knowledge seed words, the device comprising:

[0013] The document acquisition module is used to acquire a corpus of planning documents containing target behavior planning, preprocess the planning document corpus, and calculate the importance weight of each term in the preprocessed planning document corpus.

[0014] The seed word segmentation module is used to filter and obtain an initial seed word set based on importance weights. The initial seed word set is then segmented into seed words at each level according to the concept granularity of a preset hierarchical structure using a large language model.

[0015] The semantic expansion module is used to semantically expand the seed words at each level. It calculates the first candidate word and the second candidate word set based on the lexical similarity of the general vocabulary semantic database and the cosine similarity based on the domain context embedding. The first candidate word and the second candidate word set are merged and filtered by the domain professional dictionary to obtain the expanded seed word dictionary corresponding to each level.

[0016] The topic modeling module is used to increase the Dirichlet prior weight of the corresponding level extended seed words in the topic model according to the hierarchical matching relationship based on the extended word dictionary corresponding to each level, to perform topic modeling on the planned document corpus, and output the lexical distribution of each topic and the topic distribution of each document.

[0017] The clustering module is used to calculate the semantic distance between topics based on the word distribution of each topic and construct a topic distance matrix. Based on the topic distance matrix, all topics are hierarchically organized to construct a topic clustering tree. The topic clustering tree is then subjected to clustering and filtering to obtain the optimal clustering result.

[0018] The result output module is used to generate hierarchical intent recognition results based on the optimal clustering partitioning results and the preset hierarchical structure.

[0019] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0020] Obtain a corpus of planning documents containing target behavior planning, preprocess the planning document corpus, and calculate the importance weight of each term in the preprocessed planning document corpus;

[0021] An initial seed word set is obtained by filtering based on importance weights. The initial seed word set is then divided into seed words at each level according to the concept granularity of a preset hierarchical structure using a large language model.

[0022] Semantic expansion is performed on seed words at each level. First candidate words and second candidate word sets are generated based on lexical similarity from a general vocabulary semantic database and cosine similarity based on domain context embedding. The first candidate words and second candidate word sets are merged and filtered by a domain-specific dictionary to obtain the extended seed word dictionary corresponding to each level.

[0023] Based on the extended word dictionary corresponding to each level, the Dirichlet prior weight of the corresponding level extended seed words in the topic model is increased according to the hierarchical matching relationship. To the topic modeling of the planned document corpus, the vocabulary distribution of each topic and the topic distribution of each document are output.

[0024] The semantic distance between topics is calculated based on the word distribution of each topic, and a topic distance matrix is ​​constructed. All topics are then hierarchically organized according to the topic distance matrix to construct a topic clustering tree. The topic clustering tree is then subjected to clustering partitioning and filtering to obtain the optimal clustering partitioning result.

[0025] Based on the optimal clustering results and the preset hierarchical structure, a hierarchical intent recognition result is generated.

[0026] The aforementioned intent recognition method, apparatus, and computer equipment based on domain knowledge seed words can purify corpus information and objectively quantify the coreness of terms by acquiring and preprocessing a planning document corpus and calculating the importance weights of terms, providing a precise basis for subsequent seed word selection. By selecting an initial seed word set through weighting and dividing it into preset levels using a large language model, core domain terms can be identified, allowing seed words to form a hierarchical system that fits the natural structure of intent, thus solidifying domain knowledge anchors for subsequent analysis. By generating an expanded seed word dictionary through dual-dimensional semantic expansion and fusion filtering, the semantic coverage and domain context of seed words can be balanced, improving their semantic coverage and domain adaptability. By enhancing Dirichlet prior weights through hierarchical matching and conducting topic modeling, domain knowledge can be integrated into the modeling process, generating a topic distribution that fits the domain logic and improving the domain relevance of topics. By calculating topic semantic distance, constructing cluster trees, and selecting the optimal partitioning results, topic associations can be objectively quantified, forming a structurally reasonable hierarchical topic system. By generating hierarchical intent recognition results through the optimal partitioning results and preset levels, the intent can present a clear hierarchical structure, improving the structure and interpretability of intent recognition. The embodiments of the present invention can effectively realize automated and hierarchical identification of intent from planning documents, so that the identification results are consistent with the domain business logic and provide accurate and systematic intent reference for relevant decisions. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating an intent recognition method based on domain knowledge seed words in one embodiment;

[0028] Figure 2 This is a schematic diagram of the mathematical process of a hierarchical planning intent recognition framework in one embodiment;

[0029] Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0031] In one embodiment, such as Figure 1 As shown, an intent recognition method based on domain knowledge seed words is provided, including the following steps:

[0032] Step 102: Obtain the planning document corpus containing the target behavior planning, preprocess the planning document corpus, and calculate the importance weight of each term in the preprocessed planning document corpus.

[0033] Target behavior planning refers to the planned arrangements formulated by a specific entity around business operations, resource allocation, scenario-based operations, and long-term development. In this embodiment of the invention, taking the field of low-altitude operation of civilian unmanned aerial vehicles (UAVs) as an example, target behavior planning refers to various planned arrangements formulated by the operating entity around UAV low-altitude flight activities, including low-altitude flight route planning, airspace usage planning, operational task execution planning, site and take-off / landing point layout planning, equipment scheduling and maintenance planning, safety control and compliant flight planning, as well as related implementation plans and deployment schemes for routine operation, scenario-based operations, and medium- and long-term business expansion. The planning document corpus refers to a collection of documents containing relevant content such as the medium- and long-term behavior plans and layout arrangements of a specific entity (such as an enterprise or institution), which can be in the form of reports, announcements, application materials, etc. Preprocessing is a standardized process of extracting text from the corpus, segmenting sentences, uniformly converting characters, removing common stop words, and restoring word forms. The purpose is to remove irrelevant and interfering information and standardize the text format. The importance weight of terms is to evaluate the contribution of each professional term in the corpus to the core meaning of the text through quantitative indicators. The calculation method in this embodiment of the invention is the word frequency-inverse document frequency weighting algorithm, that is, to combine the frequency of occurrence of the term in a single document with the distribution frequency in the entire corpus to comprehensively determine the importance of the term.

[0034] It is understandable that standardized preprocessing can purify the quality of the corpus and reduce the interference of noise information on subsequent analysis, while accurate calculation of term importance weights can quickly identify the core key terms in the corpus, providing an objective quantitative basis for the subsequent selection of seed words and effectively improving the accuracy and efficiency of seed word selection.

[0035] Step 104: Filter the initial seed word set according to importance weights, and divide the initial seed word set into concept granularity according to the preset hierarchical structure using a large language model to obtain seed words at each level.

[0036] The initial seed word set is a collection of terms that represent core domain concepts, selected from the preprocessed corpus based on importance weights. The large language model is an AI model with powerful natural language understanding and classification capabilities, which can be guided to complete specific classification tasks through prompting engineering. The pre-defined hierarchical structure refers to a hierarchical system divided according to concept granularity from coarse to fine (e.g., macro, meso, micro), where concept granularity reflects the breadth and specificity of the terminology's scope. Seed words at each level are core terms belonging to different levels within the pre-defined hierarchical structure after being divided by the large language model, and they satisfy the distribution characteristic that the number of higher-level seed words is less than the number of lower-level seed words.

[0037] It is understandable that selecting the initial seed word set based on importance weights can ensure the domain relevance and coreness of the seed words. Dividing the seed words into levels according to the concept granularity through a large language model can enable the seed words to form a hierarchical system that conforms to the natural structure of the intent. This is beneficial for subsequent targeted guidance of topic modeling, making the topics generated by the model more in line with the domain logic and intent expression habits.

[0038] Step 106: Semantically expand the seed words at each level. Calculate the first candidate word and the second candidate word set based on the lexical similarity of the general vocabulary semantic database and the cosine similarity based on the domain context embedding. Merge the first candidate word and the second candidate word set and filter them through a domain-specific dictionary to obtain the extended seed word dictionary corresponding to each level.

[0039] Semantic expansion is the process of mining synonyms, near-synonyms, and semantically related terms based on existing seed words, aiming to enrich the semantic coverage of seed words. A general-purpose lexical semantic database stores semantic relationships between words, and its lexical similarity quantifies the degree of semantic association between words. Domain context embedding places terms in a specific domain context, extracting deep semantic vectors through a pre-trained context encoding model; cosine similarity is used to measure the similarity between two semantic vectors. A domain-specific dictionary contains core terms and standardized expressions for a specific domain, used to filter out non-domain-related terms.

[0040] It is understandable that semantic expansion through two complementary similarity calculation methods can take into account both the surface semantic association of words and the deep contextual association of the domain, effectively improving the semantic coverage and domain adaptability of seed words. Meanwhile, filtering by domain-specific dictionaries can ensure the domain specificity of expanded words, avoid interference from irrelevant terms, and provide more comprehensive and accurate domain knowledge anchors for subsequent topic modeling.

[0041] Step 108: Based on the extended word dictionary corresponding to each level, increase the Dirichlet prior weight of the corresponding level extended seed words in the topic model according to the level matching relationship, perform topic modeling on the planned document corpus, and output the vocabulary distribution of each topic and the topic distribution of each document.

[0042] Hierarchical matching refers to the correspondence between the hierarchical level of a topic and the hierarchical level of its extended vocabulary. For example, macro-level topics match macro-level extended vocabulary, meso-level topics match meso-level extended vocabulary, and micro-level topics match micro-level extended vocabulary. Dirichlet prior weights are parameters used in topic models to preset the initial probabilities of topic-vocabulary distributions. Increasing the Dirichlet prior weights of corresponding hierarchical extended seed words increases the probability of these words appearing in the corresponding hierarchical topic during model initialization. Topic models are probabilistic models used to mine potential topics from text sets (such as latent Dirichlet assignment models). The vocabulary distribution of a topic refers to the probability distribution of each word appearing under each topic, while the topic distribution of a document refers to the proportion of each topic in each document.

[0043] It is understandable that increasing the Dirichlet prior weights of extended seed words according to hierarchical matching relationships can effectively integrate domain knowledge into the topic modeling process, guide the model to generate topics that are highly consistent with the domain logic, significantly improve the domain relevance and interpretability of the topics, and the output topic vocabulary distribution and document topic distribution are conducive to accurately capturing the potential intent structure in the corpus.

[0044] Step 110: Calculate the semantic distance between topics based on the word distribution of each topic and construct a topic distance matrix. Organize all topics hierarchically based on the topic distance matrix, construct a topic clustering tree, and perform clustering partitioning and selection on the topic clustering tree to obtain the optimal clustering partitioning result.

[0045] Semantic distance between topics is an indicator that quantifies the degree of semantic similarity between two topics. A topic distance matrix is ​​a data structure that organizes the semantic distances between all topics in matrix form, intuitively presenting the degree of semantic association between topics. Hierarchical organization is the process of clustering and merging scattered topics according to their semantic association based on topic semantic distance, forming a hierarchical structure. A topic clustering tree is a carrier that presents the hierarchical relationship between topics in a tree structure. Nodes in the tree represent individual topics or merged topic clusters, and the connections between nodes represent the semantic association and merging logic between topics. In this embodiment of the invention, the optimal clustering partitioning result is obtained by calculating the silhouette coefficient (an indicator of cluster compactness and separation) under different numbers of clusters, selecting the clustering partitioning method with the largest average silhouette coefficient. This result ensures optimal cohesion and separation of the clusters. It can be understood that accurately calculating the semantic distance between topics and constructing a distance matrix can provide an objective quantitative basis for hierarchical topic organization. Constructing a topic clustering tree based on the distance matrix can clearly present the semantic association and hierarchical relationship between topics, which is conducive to forming a clear and semantically coherent intent framework.

[0046] Step 112: Generate hierarchical intent recognition results based on the optimal clustering partitioning results and the preset hierarchical structure.

[0047] The preset hierarchical structure is the hierarchical system determined in step 104, divided from coarse to fine in terms of concept granularity. The hierarchical intent recognition result is based on the preset hierarchical structure, assigning the optimally clustered topics to the corresponding levels, forming an intent framework that combines hierarchical relationships and semantic associations.

[0048] It is understandable that assigning topic levels based on the optimal clustering results and the preset hierarchical structure can enable the identified intents to form a hierarchical system that conforms to human cognitive habits, clearly presenting the complete intent chain from top-level planning to bottom-level implementation, significantly improving the interpretability and structuring of intent recognition results, making it easier for users to quickly grasp the core intents and detailed implementation actions, and providing intuitive and systematic intent references for subsequent decisions.

[0049] In the aforementioned intent recognition method based on domain knowledge seed words, by acquiring and preprocessing the planning document corpus and calculating the importance weights of terms, the corpus information can be purified, and the core degree of terms can be objectively quantified, providing a precise basis for subsequent seed word selection. By selecting the initial seed word set through weights and dividing it into preset levels by a large language model, core domain terms can be identified, allowing the seed words to form a hierarchical system that fits the natural structure of intent, thus solidifying the domain knowledge anchor for subsequent analysis. By expanding the seed word dictionary through dual-dimensional semantic expansion and fusion filtering, the semantic coverage and domain context of the words can be taken into account, improving the semantic coverage and domain adaptability of the seed words. By improving the Dirichlet prior weights through hierarchical matching and conducting topic modeling, domain knowledge can be integrated into the modeling process, generating a topic distribution that fits the domain logic and improving the domain relevance of the topics. By calculating the semantic distance of topics, constructing a cluster tree, and selecting the optimal partitioning results, the topic association can be objectively quantified, forming a reasonably structured hierarchical topic system. By generating hierarchical intent recognition results through the optimal partitioning results and preset levels, the intent can present a clear hierarchical structure, improving the structure and interpretability of intent recognition. The embodiments of the present invention can effectively realize automated and hierarchical identification of intent from planning documents, so that the identification results are consistent with the domain business logic and provide accurate and systematic intent reference for relevant decisions.

[0050] In one embodiment, the preset hierarchical structure includes a three-level structure with concept granularity ranging from coarse to fine; each level of seed words includes macro seed words, meso seed words, and micro seed words, and satisfies the condition that the number of macro seed words < the number of meso seed words < the number of micro seed words.

[0051] In this embodiment, by setting the preset hierarchical structure as a multi-level structure with concept granularity from coarse to fine, the hierarchical division of seed words can match the natural granularity characteristics of the planning intent from the top-level goal to the bottom-level implementation. This ensures that high-level seed words focus on the core top-level concept and low-level seed words cover specific implementation details, providing clear and stable hierarchical anchors for subsequent hierarchical semantic expansion and topic modeling, and improving the hierarchical matching degree and logical coherence of subsequent intent recognition.

[0052] Specifically, in the field of low-altitude operation supervision of civilian drones, it is often necessary to identify the full-chain planning intent of the operator's top-level planning layout, core capability building, and specific implementation actions from various documents. Taking the business development plan and airspace use application materials of a low-altitude logistics drone operator in a certain region as an example, macro-level seed keywords are set as top-level concepts such as "low-altitude logistics network construction" and "regional airspace layout," meso-level seed keywords are set as capability-related terms such as "long-endurance drone deployment" and "night flight capability building," and micro-level seed keywords are set as implementation-related terms such as "Area A route planning" and "logistics drone equipment procurement." The number of macro-level seed keywords is strictly controlled to be less than that of meso-level keywords, and the number of meso-level keywords is less than that of micro-level keywords, so as to achieve a precise match between seed keywords and the level of operation planning.

[0053] In one embodiment, generating a first candidate word based on the word similarity calculation of a general vocabulary semantic database includes: obtaining a set of synonyms for seed words and candidate words in the general vocabulary semantic database, calculating the depth of the lowest common ancestor node of each set of synonyms in the vocabulary hierarchy tree; obtaining the word similarity between seed words and candidate words based on the proportion of the lowest common ancestor node depth; and determining candidate words with word similarity greater than or equal to a first preset threshold and appearing in a domain-specific dictionary as the first candidate word.

[0054] In this embodiment, the general semantic associations of seed words can be accurately captured by word similarity calculation, ensuring the basic semantic consistency between extended words and seed words. At the same time, non-domain-related words are filtered out by domain-specific dictionaries, taking into account the semantic relevance and domain adaptability of extended words, effectively enriching the basic semantic coverage of seed words and providing more comprehensive lexical anchors for subsequent topic modeling.

[0055] In one embodiment, generating a second candidate word set based on the cosine similarity calculation of domain context embedding includes: adding domain-specific context descriptions to seed words and candidate words and inputting them into a pre-trained context encoding model; extracting the output vectors of the last four layers of the context encoding model and taking the average to obtain the context embedding vectors of seed words and candidate words; calculating the cosine similarity between the context embedding vectors; and determining the candidate words with a cosine similarity greater than or equal to a second preset threshold as second candidate words and forming a second candidate word set.

[0056] In this embodiment, by using cosine similarity calculation based on domain context embedding, the deep semantic features of seed words in specific domain contexts can be captured, avoiding contextual bias caused by general semantic extension, accurately mining professional expressions that are synonyms or near-synonyms within the domain, further supplementing the domain context semantic coverage of seed words, complementing general vocabulary semantic extension, and improving the domain fit and semantic integrity of extended seed words.

[0057] In one embodiment, based on the extended word dictionary corresponding to each level, the Dirichlet prior weights of the corresponding level extended seed words in the topic model are increased according to the level matching relationship. Topic modeling is performed on the planned document corpus, and the vocabulary distribution of each topic and the topic distribution of each document are output. This includes: initializing the basic Dirichlet prior matrix of the topic model; each element in the basic Dirichlet prior matrix is ​​the basic Dirichlet prior value of the topic model; matching the corresponding level extended seed word dictionary for each level topic, adding a preset enhancement coefficient to the prior value of the corresponding level extended seed word in the basic Dirichlet prior matrix, and keeping the basic Dirichlet prior value unchanged for words that do not match extended seed words, thus obtaining the improved Dirichlet prior matrix; performing topic modeling on the planned document corpus based on the improved Dirichlet prior matrix, updating the document topic distribution and topic vocabulary distribution through iterative sampling, and outputting the vocabulary distribution of each topic and the topic distribution of each document after iterative convergence.

[0058] In this embodiment, hierarchical domain knowledge can be deeply integrated into the entire topic modeling process, guiding topic generation towards core domain concepts at the corresponding level. This effectively avoids the drawback of pure data-driven models that easily generate irrelevant and redundant topics, bridges the semantic gap between low-level text features and high-level intent concepts, and significantly improves the domain relevance, distinguishability, and interpretability of generated topics. At the same time, the output topic distribution and document distribution data accurately match the domain hierarchy logic, providing stable and reliable core data support for subsequent hierarchical topic organization and intent recognition.

[0059] In one embodiment, calculating the semantic distance between topics and constructing a topic distance matrix based on the vocabulary distribution of each topic includes: for any two topics, calculating the Kolb-Leibler divergence from the vocabulary distribution to the mean distribution, taking the average of the two Kolb-Leibler divergences to obtain the Jensen-Shannon divergence; taking the square root of the Jensen-Shannon divergence as the semantic distance between the two topics, and constructing a topic distance matrix based on the semantic distance between all topics.

[0060] In this embodiment, the semantic differences between any two topics can be quantified symmetrically and without bias, avoiding the problem of unbalanced calculation results caused by asymmetric distance indicators. The constructed topic distance matrix can objectively and stably present the semantic associations and differences between all topics, providing accurate and quantifiable calculation basis for subsequent hierarchical clustering of topics, and ensuring that the subsequent topic clustering results fully match the real semantic associations between topics.

[0061] In one embodiment, hierarchically organizing all topics according to the topic distance matrix and constructing a topic clustering tree includes: initializing each topic as an independent single-sample cluster; calculating the intra-cluster variance increment when merging any two clusters based on the topic distance matrix; iteratively merging the cluster pairs with the smallest intra-cluster variance increments and generating a new cluster; using the new cluster as the parent node and the two merged clusters as child nodes, and updating the cluster hierarchy; continuing to iterate until all clusters are merged into a root cluster, forming a topic clustering tree with clusters as nodes, semantic distance as the connection basis between nodes, and parent-child node relationships as the hierarchical division basis; the cluster corresponding to a node is a single topic or a set of multiple topics merged together.

[0062] In this embodiment, the merging criterion of minimizing the intra-class variance increment can ensure that the generated clustering structure has good compactness and homogeneity, achieving accurate aggregation of semantically highly related topics. At the same time, the tree structure presents the hierarchical relationship and aggregation link between topics in a complete and clear manner, providing a structurally complete and traceable carrier for the subsequent hierarchical division of intent and the establishment of association relationships, greatly improving the rationality and logical coherence of the hierarchical organization of topics.

[0063] In one embodiment, the optimal clustering result is obtained by clustering and filtering the topic clustering tree. Based on the optimal clustering result and a preset hierarchical structure, a hierarchical intent recognition result is generated. This includes: calculating the silhouette coefficients of all samples when the topic clustering tree is divided into different numbers of clusters; selecting the clustering method corresponding to the largest average silhouette coefficient as the optimal clustering result; assigning the root node of the topic clustering tree to the highest level of the preset hierarchical structure, the intermediate nodes to the intermediate levels, and the leaf nodes to the lowest level; establishing a hierarchical and support relationship between clusters at different levels based on the parent-child connection relationship of the clustering tree nodes; clustering into a single topic or a collection of multiple topics; and generating a hierarchical intent framework with a tree structure as its carrier based on the hierarchical correspondence and the hierarchical and support relationship between clusters, thereby obtaining the hierarchical intent recognition result.

[0064] In this embodiment, the optimal clustering method with cohesion and separation can be selected by using the contour coefficient, ensuring the rationality and accuracy of the topic clustering results. At the same time, through the precise correspondence between nodes and preset levels, and the establishment of the attribution and support relationships between parent and child nodes, the scattered topics can form a logically coherent and hierarchically clear complete intent framework, clearly presenting the entire chain of intent from top-level planning to bottom-level implementation, greatly improving the interpretability and structuring of intent recognition results. The generated tree-like intent framework can be directly adapted to the analysis and decision-making needs of different levels, and has strong practical applicability.

[0065] This invention combines Large Language Models (LLMs) with structured domain knowledge through a mathematically formalized process. Using budget and planning documents from the field of low-altitude operation supervision of civil unmanned aerial vehicles (UAVs) as data sources, this method introduces a domain knowledge-guided seed word enhancement mechanism. The processing flow includes: (1) hierarchical seed word classification using LLMs; (2) semantic expansion using WordNet-based lexical similarity measurement and BERT-based contextual embedding; (3) seeded LDA topic modeling (by adjusting the Dirichlet prior distribution to favor seed words); and (4) constructing an intent hierarchy using Ward's join method for agglomerated hierarchical clustering. This method achieves guidance by increasing the prior probability of seed words at specific topic levels and uses Jensen-Shannon divergence to measure topic similarity. Experimental evaluation based on data from the field of low-altitude operation supervision of civil UAVs shows that this framework significantly outperforms baseline methods in terms of topic consistency (42.6% improvement) and human evaluation (19.8% improvement), while enhancing the interpretability of the results through a hierarchical structure. This research contributes to the field of intelligent optimization by providing a formalized, scalable, and interpretable framework for analyzing planning intent. The mathematical foundation of this framework ensures transparency and repeatability, while its algorithmic implementation supports practical applications in large-scale budget analysis. The mathematical foundation guarantees the transparency and reproducibility of the analysis process, while the engineering implementation of the algorithm enables its practical application in scenarios such as large-scale defense budget analysis.

[0066] In one specific embodiment, such as Figure 2The proposed framework transforms semi-structured budget text into a hierarchical intent representation through four mathematically defined stages. The diagram illustrates the process of automatically constructing a hierarchical planning intent framework from data. The process begins with seed word construction, defining conceptual anchors by integrating domain knowledge. These seed words guide the dataset construction and preprocessing stages. Subsequently, an LDA-based seed topic modeling method is employed, using seed words to guide the topic discovery process and ensure domain relevance. The generated topics are organized through hierarchical clustering based on semantic similarity, as shown in the tree diagram on the right. The final output is a multi-level generative planning intent framework. The bar chart on the left shows a comparison of the seed word hierarchical expansion results, used to evaluate the effectiveness of seed propagation. The specific process is as follows: 2.1 Seed Word Extraction and Classification. Given a corpus of planning documents... First, candidate seed words are extracted using the TF-IDF weighted method for each document in the corpus. The formula for calculating term weights is:

[0067] ;

[0068] in, For the term In the document word frequency in For the term Document frequency, This represents the total number of documents in the corpus. For the term In the document The TF-IDF weights in the equation.

[0069] Select the K terms with the highest average TF-IDF scores to form the initial seed word set. Subsequently, a specific prompt template was constructed to call the large language model API for automated classification. This prompt template requires the model to act as a domain expert, performing term selection and three-level classification tasks. The template explicitly defines and provides examples for the macro (planning concepts), meso (capability system), and micro (technology / platform) layers, and instructs the model to output results in a specified JSON format, ensuring that the output meets the constraint that the number of macro-level terms < the number of meso-level terms < the number of micro-level terms. This template is then used to initialize the seed word set. The data is processed and divided into three levels: macro, meso, and micro. This hierarchical structure reflects the natural granularity of the planning concept, from coarse to fine.

[0070] 2.2 Semantic expansion based on hybrid similarity. Seed words for each level. We employ two complementary similarity measurement methods to generate an extended vocabulary set that combines semantic relevance and domain adaptability. The specific similarity measurement methods include WordNet-based lexical similarity and BERT-based contextual similarity.

[0071] 2.2.1 WordNet-based lexical similarity calculation. For seed words... and candidate words The word similarity based on WordNet is defined as:

[0072] ;

[0073] in, Seed word With candidate words WordNet word similarity, Representation of terms In WordNet, the LCS (Lowest Common Ancestor) node represents the lowest common ancestor node of two synonym sets in the WordNet hierarchy. For nodes Depth in the WordNet hierarchy tree.

[0074] To ensure the domain-specific adaptability of the expanded terms, a dictionary of low-altitude operation supervision of civilian unmanned aerial vehicles is incorporated. Domain-specific filtering is performed, and the formula for the filtered candidate word set is:

[0075] ;

[0076] in, Seed word The effective candidate word set after WordNet expansion and filtering The preset threshold for WordNet similarity. A professional dictionary for the field of low-altitude operation supervision of civilian drones. For logical AND operator.

[0077] 2.2.2 Contextual Similarity Calculation Based on BERT. To capture the deep semantic features of terms in the context of civil UAV low-altitude operation supervision, specific contexts related to civil UAV low-altitude operation supervision (e.g., "In the context of civil UAV low-altitude operation supervision: [terminology]") are added to seed words and candidate words. These contexts are then input into a pre-trained BERT model, and the outputs of the last four layers of the model are averaged to obtain the context embedding vectors of the seed words and candidate words. The contextual similarity between the seed words and candidate words is obtained by calculating the cosine similarity of the embedding vectors, using the following formula:

[0078] ;

[0079] in, Seed word With candidate words BERT context similarity, and Seed words and candidate words The context embedding vector, The dot product operation is performed on two vectors. Let L2 be the norm (modulus) of the vector. Each vector is obtained by inputting its corresponding prefix text into the BERT model and performing average pooling on the token representations of the last four layers. These embeddings capture the semantic meaning of each term in its specific domain context, and their cosine similarity measures the seed words. and candidate words The semantic relationship between them.

[0080] 2.2.3 Hybrid Expansion Algorithm. Algorithm 1 formalizes the semantic expansion process based on hybrid similarity described above. By fusing the expansion results from WordNet and BERT, it achieves efficient and accurate expansion of seed words. The specific steps are as follows:

[0081] Algorithm 1 Hybrid Semantic Extension:

[0082] Input: Seed word WordNet dictionary BERT model Dictionary of the Supervision of Low-Altitude Operations of Civil Unmanned Aerial Vehicles WordNet similarity threshold BERT similarity threshold ;

[0083] Output: Hierarchical expanded vocabulary dictionary ;

[0084] 1: ;

[0085] 2: For each seed word Execution loop:

[0086] 3: ;

[0087] 4: ;

[0088] 5: ;

[0089] 6: Mixed candidate word set according to Sort in descending order;

[0090] 7: ;

[0091] 8: End the loop;

[0092] 9: Return to the hierarchical expanded dictionary .

[0093] 2.3 Improved Dirichlet Prior Seed LDA Topic Modeling. After generating a hierarchical extended seed word dictionary, this embodiment adopts an improved Dirichlet Prior Seed LDA topic modeling method, which integrates hierarchical domain knowledge into the topic modeling process, guiding the direction of topic generation and avoiding the generation of irrelevant and redundant topics by purely data-driven models.

[0094] 2.3.1 Standard LDA Model Definition. The core assumptions of the standard LDA model are: each document corresponds to a multidimensional topic distribution, and each topic corresponds to a multidimensional vocabulary distribution, both of which follow a Dirichlet prior distribution; the generation process of each word in the document is as follows: first, a topic is randomly sampled from the topic distribution of the document, and then a word is randomly sampled from the vocabulary distribution of the sampled topic.

[0095] 2.3.2 Seed-word-guided prior improvement. To integrate domain knowledge into the topic modeling process, the Dirichlet prior of the LDA model is improved. (In the Gensim library, it is usually written as) To incorporate seed word information, set For a hierarchical extended seed word dictionary, belonging to the hierarchy Theme The improved Dirichlet prior is defined as follows:

[0096] ;

[0097] in, For the revised theme The following terms Dirichlet priors, To revise the previous topic The following terms The basic Dirichlet prior values, The seed word prior enhancement coefficient, This is an indicator function (it takes the value 1 if the condition in parentheses is true, and 0 otherwise). Theme Belonging level The corresponding extended seed word dictionary, Theme The corresponding levels (macro / meso / micro). This improvement tilts the topic-vocabulary distribution towards seed words, enabling domain knowledge to provide targeted guidance for topic modeling.

[0098] Algorithm 2 formalizes the improved Dirichlet prior seeded-LDA model and completes the iterative update of model parameters through Gibbs sampling. The specific steps are as follows:

[0099] Algorithm 2 improves the seeded LDA model based on Dirichlet priors:

[0100] Input: Corpus Preset number of topics Layered seed word dictionary Basic Dirichlet Priors Seed word enhancement coefficient ;

[0101] Output: Lexical distribution for each topic Thematic distribution of each document ;

[0102] 1: Initialization for Dimensional matrix, all elements are initially assigned values ​​based on prior knowledge. , For the size of the vocabulary;

[0103] 2: For each topic arrive Execution loop:

[0104] 3: Determine the topic hierarchy (Macro / Meso / Micro)

[0105] 4: For each term Execution loop:

[0106] 5: ;

[0107] 6: Finish;

[0108] 7: End;

[0109] 8: Initialize the vocabulary distribution for each topic , ;

[0110] 9: Initialize the topic distribution of each document , , Number of documents;

[0111] 10: Number of Gibbs sampling iterations arrive Execution loop:

[0112] 11: For each document arrive Execution loop:

[0113] 12: For the document Each word position in Execution loop:

[0114] 13: Sampling topic ;

[0115] 14: Update count: , ;

[0116] 15: Finish;

[0117] 16: Finish;

[0118] 17: Update document topic distribution and topic vocabulary distribution: , ;

[0119] 18: End;

[0120] 19: Return to the distribution of vocabulary by topic Thematic distribution of each document ;

[0121] 2.4 Hierarchical Clustering for Intent Mapping. After obtaining the latent topics of the text through the seeded-LDA model, a hierarchical clustering method is used to organize the topics hierarchically and construct a hierarchical structure of planning intent. The core includes two steps: topic similarity calculation and cohesive hierarchical clustering based on Ward connections.

[0122] 2.4.1 Topic Similarity Calculation. For two topics with similar word distributions... and The semantic similarity between topics is measured using the Jensen-Shannon divergence (JSD), a symmetric improvement of the Körbek-Leibler divergence that addresses the asymmetry issue. JSD is defined as the average of the Körbek-Leibler divergences of two topic distributions to their mean distributions. To better reflect distance characteristics, the square root of the Jensen-Shannon divergence is used as the final inter-topic distance, as shown in the formula:

[0123] ;

[0124] in, Theme With the topic Jensen-Shannon divergence, The mean distribution of the two topic vocabulary distributions. The Korbeck-Leibler divergence is defined as follows:

[0125] ;

[0126] For distribution Relative to distribution The Kulbeck-Leibler divergence, , Distributions , Chinese terminology The probability of its occurrence, The total size of the vocabulary of the corpus.

[0127] Construct a topic distance matrix based on the distance between topics. Defined as:

[0128] ;

[0129] in, For the topic in the distance matrix With the topic The distance between them.

[0130] 2.4.2 Ward Connectivity in Agglomerative Hierarchical Clustering. The core of the Ward Connectivity criterion is to minimize the intra-cluster variance increment after cluster merging. By merging clusters using this criterion, the resulting cluster structure can be guaranteed to be compact and homogeneous. For clusters with centroids of... , The sample size is , Two clusters and The formula for the intra-class variance increment during merging is:

[0131] ;

[0132] in, For clustering Clustering Increment of intra-class variance during merging , Clustering , The centroid vector, , Clustering , Number of samples within, It represents the Euclidean distance on the probability simplex.

[0133] Algorithm 3 formalizes the agglomerative hierarchical clustering process based on Ward connections. By iteratively merging cluster pairs with the smallest variance increment, it constructs a hierarchical clustering tree of topics, ultimately mapping it to a hierarchical structure of planning intent. The specific steps are as follows:

[0134] Algorithm 3: Agglomerative Hierarchical Clustering

[0135] Input: Topic distribution output by the seeded-LDA model Number of preset planning intention levels ;

[0136] Output: Topic hierarchical clustering tree The hierarchical allocation results for each topic;

[0137] 1: Calculate the topic distance matrix ,in

[0138] 2: Initialize each topic as an independent single-sample cluster: , ;

[0139] 3: Initialize the clustering tree (Leaf nodes are initialized independent clusters with no parent node).

[0140] 4: Number of clusters Decrease to 2 and execute the loop;

[0141] 5: Iterate through all cluster pairs to find the one that increases the variance. Minimal clustering pairs ;

[0142] 6: Merge cluster pairs to generate new clusters: ;

[0143] 7: Update the clustering tree Add to , New clusters for child nodes ;

[0144] 8: Remove merged clusters from the active cluster set. , ;

[0145] 9: The newly generated clusters Join the active cluster set;

[0146] 10: End the loop;

[0147] 11: Based on the clustering tree, calculate the partitioning into The silhouette coefficients of all samples in a cluster;

[0148] 12: Select the clustering method with the highest average silhouette coefficient as the optimal hierarchical partitioning;

[0149] 13: Assign hierarchical labels to each cluster: macro (root node of the cluster tree), meso (middle node of the cluster tree), and micro (leaf node of the cluster tree);

[0150] 14: Return to the topic hierarchical clustering tree The hierarchical allocation results for each topic.

[0151] In one specific embodiment, using relevant documents in the field of low-altitude operation supervision of civil unmanned aerial vehicles (UAVs) as experimental corpus, the hierarchical planning intent recognition framework was implemented from four dimensions: seed word extraction and classification, comparison of semantic expansion methods, topic modeling performance, and hierarchical structure analysis. Multiple sets of quantitative indicators were used to evaluate the implementation effect of each stage, ensuring the stability and effectiveness of the framework's implementation. The experimental corpus consists of publicly available authoritative documents in the field of low-altitude operation supervision of civil UAVs, ensuring the validity and representativeness of the experimental data.

[0152] First, text extraction and preprocessing are performed on the original document: the text content is extracted from the PDF document to obtain the original corpus. The original corpus is then subjected to standardized preprocessing, including sentence segmentation based on NLTK Punkt tokenizer, full-text character lowercase conversion, general stop word removal, and WordNet-based lemmatization. Finally, about 1200 effective sentences are obtained, which are used for subsequent candidate seed word extraction.

[0153] Candidate seed words were identified from the preprocessed corpus using the TF-IDF method. To filter out low-frequency meaningless terms, bigrams and trigrams with a minimum document frequency of 2 were selected as candidates, resulting in approximately 1,000 candidate seed words. After manual screening and deduplication of the candidate seed words, they were divided according to macroscopic, mesoscopic, and microscopic hierarchical features to ensure the hierarchical nature and domain coverage of the seed words.

[0154] As shown in Table 1, 45 seed words were finally selected and hierarchically divided, which strictly met the hierarchical feature of increasing number of terms from macro to micro, and conformed to the granular change law of planning concepts from coarse to fine. This can comprehensively represent the core planning concepts in the field of low-altitude operation supervision of civil drones.

[0155] Table 1. Examples of Seed Word Hierarchical Classification

[0156]

[0157] In multi-source heterogeneous planning documents, the expression forms of the same planning concept differ. To solve the problem of the single expression form of seed words, it is necessary to semantically expand the seed words to improve their semantic coverage. This invention compares three different semantic expansion methods and uses three core quantitative indicators to evaluate the performance of different expansion methods, comprehensively measuring the expansion effect and computational efficiency: (1) Total number of expanded words: the total number of unique terms successfully expanded by all seed words, reflecting the coverage of the expansion; (2) Average number of expansions: the average number of expanded terms for a single seed word, reflecting the efficiency of the expansion; (3) Processing time: the total time spent completing the expansion of all 45 seed words, reflecting the computational efficiency of the method. It is confirmed that the hybrid similarity method has comprehensive advantages in balancing semantic depth and processing efficiency, and can provide comprehensive and accurate domain vocabulary support for subsequent topic modeling. The three methods are as follows:

[0158] Enhanced WordNet: It generates candidate words using only the lexical relationships (synonyms, hyponyms) of WordNet, and filters them using a dictionary for the supervision of low-altitude operations of civilian drones to ensure the domain relevance of the expanded words;

[0159] Professional BERT: Only BERT models with specific hints added for the field of low-altitude operation supervision of civilian drones are used to generate contextual embedding vectors, and candidate words are selected by cosine similarity to capture the deep contextual semantics of terms;

[0160] Hybrid approach: Combining the advantages of the first two methods, it first generates initial candidate words through WordNet, and then reorders and filters the candidate words based on BERT contextual similarity, taking into account both lexical relevance and contextual semantics.

[0161] This embodiment employs multiple intrinsic quantitative metrics to comprehensively evaluate the trained seeded-LDA model, balancing the quality of topics generated by the model and its generalization ability. Specifically, these metrics include topic consistency, log perplexity, and Jaccard similarity between topics. The calculation methods and evaluation significance of each metric are as follows:

[0162] Thematic consistency Topic consistency measures the semantic relevance between high-weight words within a single topic. A higher score indicates stronger word association within the topic and better topic interpretability. For topics containing... The topic of high-weight terms The formula for calculating topic consistency is:

[0163] ;

[0164] in, Score for thematic consistency. Selected top within a single topic Number of high-weight terms For the term and Point-to-point information, For the term and Number of document co-occurrences in the corpus.

[0165] Log-perplexity: Log-perplexity quantifies a model's predictive ability on unseen test documents, reflecting the model's generalization performance. A lower value indicates more accurate predictions and better generalization for new documents. The formula is:

[0166] ;

[0167] in, The logarithmic perplexity of the model, The total number of documents in the test set. The total number of words in all documents of the test set. The total size of the corpus's vocabulary. For the term In the document Frequency of occurrence in For a given document Theme distribution At that time, terminology The generation probability, For document The topic distribution probability vector.

[0168] Jaccard Similarity: Jaccard similarity measures the degree of overlap between the high-weight vocabulary sets of two topics, assessing topic diversity and discriminability. A lower value indicates less overlap between the two topics' vocabulary sets and higher topic discriminability. The formula is as follows:

[0169] ;

[0170] in, Theme With the topic The similarity of Jaccard, and Themes and theme The former A set of high-weight terms, The number of elements in the intersection of the two thesaurus sets. The number of elements in the union of the two thesaurus sets.

[0171] To evaluate the overall diversity of all topics generated by the model, the average Jaccard similarity of all topic pairs is calculated using the following formula:

[0172] ;

[0173] in, The average Jaccard similarity for all topic pairs. The total number of topics generated by the model.

[0174] The proposed seeded-LDA model, tailored to the specific scenarios of professional terminology in the field of low-altitude operation supervision of civilian unmanned aerial vehicles (UAVs), effectively integrates domain knowledge into the topic modeling process, guiding the model to generate topics that fit the semantics of domain planning and significantly improving the interpretability of topics. Simultaneously, it effectively enhances the model's generalization ability to domain corpora, adapting to unseen domain-related documents; and significantly reduces the similarity between topics, giving the generated topics excellent distinguishability. Each topic can capture a unique dimension within the domain, avoiding topic overlap, and ultimately extracting potential topics with clear planning meaning from domain planning documents.

[0175] Building upon this foundation, this embodiment employs a condensed hierarchical clustering method based on Jensen-Shannon divergence and Ward connectivity to hierarchically structure the topics generated by the model, constructing a hierarchical structure of planning intent from macro to micro levels. This embodiment maximizes the silhouette coefficient to select the optimal number of clusters, using this as the basis for classifying macro-level planning intent categories. Each macro category corresponds to a core high-level planning theme. For example, in this embodiment, the optimal silhouette coefficient corresponds to eight macro-level clusters, representing core high-level planning themes in the field of low-altitude operation supervision of civil unmanned aerial vehicles (UAVs), including "low-altitude logistics network construction" and "regional airspace layout." These macro clusters can be further decomposed into 30 meso- and micro-level themes based on the hierarchical structure of the clustering tree, achieving a refined depiction of planning priorities and implementation details in the field of low-altitude operation supervision of civil UAVs. This provides a complete perspective from the overall to the local for planning intent analysis, ultimately forming a clearly structured and logically coherent hierarchical planning intent framework.

[0176] This invention addresses the issues of strong subjectivity and lack of domain relevance in traditional planning intent recognition methods and purely data-driven models. It proposes a mathematically formalized hierarchical framework for planning intent recognition, deeply integrating domain knowledge of low-altitude operation supervision of civilian unmanned aerial vehicles (UAVs) with natural language processing technology to achieve automated and hierarchical recognition of operational planning intent. The invention proposes a hybrid semantic expansion mechanism that combines WordNet lexical similarity calculation and BERT contextual embedding, solving the problem of limited seed word expression and improving the semantic coverage and domain adaptability of seed words. It constructs an improved Dirichlet prior-seeded-LDA model, using Gibbs sampling to achieve hierarchical enhancement of seed word probabilities, effectively integrating domain knowledge into the topic modeling process and solving the semantic gap problem of purely data-driven models. Finally, it designs a hierarchical clustering method based on Jensen-Shannon divergence and Ward connectivity to achieve hierarchical organization of potential topics, constructing a highly interpretable hierarchical structure of operational planning intent and providing a systematic perspective for operational planning intent analysis. Experimental verification results based on low-altitude operation planning documents for civilian drones show that the framework proposed in this invention can effectively extract potential topics related to operation planning. The generated hierarchical structure of operation planning intent has good interpretability and distinguishability, and can provide data support and analytical reference for low-altitude operation supervision and industry decision-making.

[0177] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated in this invention, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Furthermore, Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0178] In one embodiment, an intent recognition device based on domain knowledge seed words is provided, comprising:

[0179] The document acquisition module is used to acquire a corpus of planning documents containing target behavior planning, preprocess the planning document corpus, and calculate the importance weight of each term in the preprocessed planning document corpus.

[0180] The seed word segmentation module is used to filter and obtain an initial seed word set based on importance weights. The initial seed word set is then segmented into seed words at each level according to the concept granularity of a preset hierarchical structure using a large language model.

[0181] The semantic expansion module is used to semantically expand the seed words at each level. It calculates the first candidate word and the second candidate word set based on the lexical similarity of the general vocabulary semantic database and the cosine similarity based on the domain context embedding. The first candidate word and the second candidate word set are merged and filtered by the domain professional dictionary to obtain the expanded seed word dictionary corresponding to each level.

[0182] The topic modeling module is used to increase the Dirichlet prior weight of the corresponding level extended seed words in the topic model according to the hierarchical matching relationship based on the extended word dictionary corresponding to each level, to perform topic modeling on the planned document corpus, and output the lexical distribution of each topic and the topic distribution of each document.

[0183] The clustering module is used to calculate the semantic distance between topics based on the word distribution of each topic and construct a topic distance matrix. Based on the topic distance matrix, all topics are hierarchically organized to construct a topic clustering tree. The topic clustering tree is then subjected to clustering and filtering to obtain the optimal clustering result.

[0184] The result output module is used to generate hierarchical intent recognition results based on the optimal clustering partitioning results and the preset hierarchical structure.

[0185] Specific limitations regarding the intent recognition device based on domain knowledge seed words can be found in the limitations of the intent recognition method based on domain knowledge seed words mentioned above, and will not be repeated here. Each module in the aforementioned intent recognition device based on domain knowledge seed words can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0186] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 3As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements an intent recognition method based on domain knowledge seed words. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0187] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0188] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0189] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0190] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. 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 protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An intent recognition method based on domain knowledge seed words, characterized in that, The method includes: Obtain a corpus of planning documents containing target behavior planning, preprocess the planning document corpus, and calculate the importance weight of each term in the preprocessed planning document corpus; An initial seed word set is obtained by filtering based on importance weights. The initial seed word set is then divided into seed words at each level according to the concept granularity of a preset hierarchical structure using a large language model. Semantic expansion is performed on seed words at each level. First candidate words and second candidate word sets are generated based on lexical similarity from a general vocabulary semantic database and cosine similarity based on domain context embedding. The first candidate words and second candidate word sets are merged and filtered by a domain-specific dictionary to obtain the extended seed word dictionary corresponding to each level. Based on the extended word dictionary corresponding to each level, the Dirichlet prior weight of the corresponding level extended seed words in the topic model is increased according to the hierarchical matching relationship. To the topic modeling of the planned document corpus, the vocabulary distribution of each topic and the topic distribution of each document are output. The semantic distance between topics is calculated based on the word distribution of each topic, and a topic distance matrix is ​​constructed. All topics are then hierarchically organized according to the topic distance matrix to construct a topic clustering tree. The topic clustering tree is then subjected to clustering partitioning and filtering to obtain the optimal clustering partitioning result. Based on the optimal clustering results and the preset hierarchical structure, a hierarchical intent recognition result is generated.

2. The method of claim 1, wherein, The preset hierarchical structure includes a three-level structure with concept granularity ranging from coarse to fine; The seed words at each level include macro seed words, meso seed words, and micro seed words, and satisfy the condition that the number of macro seed words < the number of meso seed words < the number of micro seed words.

3. The method according to claim 1, characterized in that, The first candidate words generated based on lexical similarity calculation using a general vocabulary semantic database include: Obtain the synonym sets of seed words and candidate words from the general lexical semantic database, and calculate the depth of the lowest common ancestor node of each synonym set in the lexical hierarchy tree; The lexical similarity between seed words and candidate words is obtained based on the percentage of the depth of the lowest common ancestor node. Candidate words with a similarity greater than or equal to the first preset threshold and that appear in the domain-specific dictionary are identified as the first candidate words.

4. The method according to claim 1, characterized in that, The second candidate word set generated based on cosine similarity calculation using domain context embedding includes: After adding domain-specific context descriptions to seed words and candidate words, input them into a pre-trained context encoding model, extract the output vectors of the last four layers of the context encoding model and take the average to obtain the context embedding vectors of seed words and candidate words; Calculate the cosine similarity between the context embedding vectors, and determine the candidate words with a cosine similarity greater than or equal to a second preset threshold as the second candidate words and form the second candidate word set.

5. The method according to claim 1, characterized in that, The process involves increasing the Dirichlet prior weights of corresponding level-specific extended seed words in the topic model based on the hierarchical matching relationship, performing topic modeling on the planned document corpus, and outputting the lexical distribution of each topic and the topic distribution of each document, including: Initialize the basic Dirichlet prior matrix of the topic model; each element in the basic Dirichlet prior matrix is ​​the basic Dirichlet prior value of the topic model. For each level of topic, an extended seed word dictionary corresponding to the level is matched. A preset enhancement coefficient is added to the prior value of the extended seed word corresponding to the level in the basic Dirichlet prior matrix. The basic Dirichlet prior value of words that do not match extended seed words remains unchanged, resulting in an improved Dirichlet prior matrix. Based on the improved Dirichlet prior matrix, topic modeling is performed on the planned document corpus. The document topic distribution and topic vocabulary distribution are updated through iterative sampling. After the iteration converges, the vocabulary distribution of each topic and the topic distribution of each document are output.

6. The method according to claim 1, characterized in that, The step of calculating the semantic distance between topics and constructing a topic distance matrix based on the vocabulary distribution of each topic includes: For any two topics, calculate the Körbek-Leibler divergence from the word distribution to the mean distribution, and take the average of the two Körbek-Leibler divergences to obtain the Jensen-Shannon divergence. The square root of the Jensen-Shannon divergence is used as the semantic distance between two topics, and a topic distance matrix is ​​constructed based on the semantic distance between all topics.

7. The method according to claim 1, characterized in that, Based on the topic distance matrix, all topics are hierarchically organized, and a topic clustering tree is constructed, including: Each topic is initialized as an independent single-sample cluster, and the intra-cluster variance increment when merging any two clusters is calculated based on the topic distance matrix. Iteratively merge the cluster pairs with the smallest intra-class variance increment and generate a new cluster. Use the new cluster as the parent node and the two merged clusters as child nodes to update the cluster hierarchy. The process continues iterating until all clusters are merged into a single root cluster, forming a topic clustering tree with clusters as nodes, semantic distance as the basis for connecting nodes, and parent-child node relationships as the basis for hierarchical division; the clusters corresponding to the nodes are single topics or sets of multiple topics merged together.

8. The method according to claim 1, characterized in that, The step of performing clustering partitioning and filtering on the topic clustering tree to obtain the optimal clustering partitioning result, and generating hierarchical intent recognition results based on the optimal clustering partitioning result and the preset hierarchical structure, includes: Calculate the silhouette coefficient of all samples when the topic clustering tree is divided into different numbers of clusters, and select the clustering method corresponding to the largest average silhouette coefficient as the optimal clustering result; The root node of the topic clustering tree corresponds to the highest level of the preset hierarchical structure, the intermediate nodes correspond to the intermediate levels, and the leaf nodes correspond to the lowest level. The affiliation and support relationships between different levels of clusters are established based on the parent-child connection relationship of the nodes in the clustering tree; the cluster is a single topic or a collection of multiple topics. Based on the hierarchical correspondence and the attribution and support relationships between clusters, a hierarchical intent framework with a tree structure is generated to obtain the hierarchical intent recognition results.

9. An intent recognition device based on domain knowledge seed words, characterized in that, The device includes: The document acquisition module is used to acquire a corpus of planning documents containing target behavior planning, preprocess the planning document corpus, and calculate the importance weight of each term in the preprocessed planning document corpus. The seed word segmentation module is used to filter and obtain an initial seed word set based on importance weights. The initial seed word set is then segmented into seed words at each level according to the concept granularity of a preset hierarchical structure using a large language model. The semantic expansion module is used to semantically expand the seed words at each level. It calculates the first candidate word and the second candidate word set based on the lexical similarity of the general vocabulary semantic database and the cosine similarity based on the domain context embedding. The first candidate word and the second candidate word set are merged and filtered by the domain professional dictionary to obtain the expanded seed word dictionary corresponding to each level. The topic modeling module is used to increase the Dirichlet prior weight of the corresponding level extended seed words in the topic model according to the hierarchical matching relationship based on the extended word dictionary corresponding to each level, to perform topic modeling on the planned document corpus, and output the lexical distribution of each topic and the topic distribution of each document. The clustering module is used to calculate the semantic distance between topics based on the word distribution of each topic and construct a topic distance matrix. Based on the topic distance matrix, all topics are hierarchically organized to construct a topic clustering tree. The topic clustering tree is then subjected to clustering and filtering to obtain the optimal clustering result. The result output module is used to generate hierarchical intent recognition results based on the optimal clustering partitioning results and the preset hierarchical structure.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images