A method and system for mining old-age care intentions based on semantic correlation analysis

Abstract

The application discloses a pension intention mining method and system based on semantic association analysis and belongs to the technical field of semantic analysis. The method comprises the following steps: obtaining to-be-processed interaction data in a pension consultation scene and performing semantic role labeling to obtain a semantic role sequence after labeling; extracting a plurality of candidate intention trigger words and assigning initial semantic weights; performing dependency syntax analysis on the interaction data based on the initial semantic weights, mapping the obtained dependency relationship path to a preset pension semantic relationship graph, and generating an associated semantic subgraph corresponding to each candidate intention trigger word; calculating the semantic association strength between different candidate intention trigger words and clustering to obtain an intention trigger word cluster; for each intention trigger word cluster, a pension intention feature vector is constructed by using a cluster semantic weight, a main dependency relationship type, a main part-of-speech label and a cluster semantic center vector; and the pension intention feature vector is input into an intention classifier to output a pension intention category. The method realizes deep semantic mining of the pension intention.

Description

Technical Field

[0001] This invention relates to the field of semantic analysis technology, specifically to a method and system for mining elderly care intentions based on semantic association analysis. Background Technology

[0002] In elderly care consultation scenarios, automatically mining the potential elderly care service needs of seniors through dialogue interaction data is the foundation for achieving intelligent elderly care recommendations and precise service matching. Existing elderly care intent mining technologies mostly rely on keyword matching or rule template-based methods. They scan text using a pre-set elderly care keyword library and simply categorize sentences containing matching keywords into corresponding intent categories. Some methods introduce text classification techniques, using bag-of-words models or shallow neural networks to classify the intent of the entire conversation text. Other methods, while incorporating semantic role labeling and dependency parsing for sentence component pruning, still ultimately rely on joint encoding of the pruned and original sentences for semantic matching. They fail to map candidate trigger words along dependency paths to a domain semantic relationship graph to form clusterable associative semantic subgraphs, nor do they measure the semantic association strength between trigger words based on subgraph topological overlap. Both of these methods treat interaction data as a flattened or merely pruned sequence of words, lacking an organizational mechanism for the deep semantic associations between words at the domain concept level, resulting in insufficient ability to identify implicit and complex intents. When elderly people express their needs indirectly or implicitly, such as by describing difficulties in life or emotional states, existing methods often only capture surface-level words and cannot aggregate multiple semantically related trigger words scattered in the context into a complete expression of intent. The shortcomings of existing technologies are mainly reflected in two aspects. First, the extraction process of intent trigger words is isolated, lacking a quantification and organization mechanism for the deep semantic relationships between trigger words. Each potentially implied word is processed independently, and even if they all point to the same intention regarding elderly care, they cannot be identified as a semantic community, resulting in fragmented intent and information redundancy. Second, at the level of intent feature representation, the feature vectors constructed by existing methods are usually based only on word frequency or simple word vector averaging, failing to integrate the structural statistical information and syntactic dependency patterns of trigger word clusters in the semantic relationship graph. This makes it difficult to characterize the overall semantic tendency of the collaborative expression of words within an intent cluster, limiting the accuracy of intent classification. Therefore, it is necessary to address how to extract intrinsically related intent trigger words from elderly care interaction data based on semantic roles and dependencies, organize them into semantic clusters expressing the same intent, and generate multi-dimensional semantic features for the formed intent trigger word clusters to achieve high-resolution intent category determination. Summary of the Invention

[0003] This invention provides a method and system for mining elderly care intentions based on semantic association analysis. It aims to deeply mine implicit elderly care intentions from elderly care consultation interaction data. Through semantic role labeling, dependency parsing, and mapping of elderly care semantic relationship graph, it constructs a semantic subgraph of associations between trigger words, and calculates the semantic association strength based on the overlap of subgraphs to achieve trigger word clustering. Then, it constructs a multi-dimensional feature vector that integrates cluster statistical features and semantic centers to achieve highly accurate classification of elderly care intentions.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method and system for mining elderly care intentions based on semantic association analysis, which aims to accurately extract and classify users' implicit elderly care intentions from elderly care consultation interaction data, thereby improving the accuracy and completeness of intention recognition.

[0005] This method acquires interactive data to be processed in elderly care consultation scenarios, performs semantic role labeling on the interactive data, and generates a labeled semantic role sequence. Preferably, the interactive data is segmented into multiple clauses according to punctuation marks and word segmentation is performed to obtain a word sequence; the word sequence is input into a semantic role labeling model jointly trained based on a bidirectional long short-term memory network and a conditional random field, which identifies predicates and labels agent role, patient role, time role, location role, and manner role, and then combines each role according to the original order of words in the clause to form the labeled semantic role sequence of each clause.

[0006] Based on the annotated semantic role sequence, multiple candidate intent trigger words are extracted, and an initial semantic weight is assigned to each candidate intent trigger word. As a technical solution of this invention, the predicate, its agent role, and patient role are extracted from the annotated semantic role sequence corresponding to each clause, forming a candidate intent trigger word group. The candidate intent trigger word groups of all clauses are merged to obtain multiple candidate intent trigger words. For each candidate intent trigger word, its frequency of occurrence in all clauses is counted, and the term frequency-inverse document frequency (TIF) value is obtained. The product of the occurrence frequency and the TIF value is used as the initial semantic weight of the candidate intent trigger word. Furthermore, a synonym normalization process can be performed on the candidate intent trigger words using a thesaurus of intent-related terms: the set of synonyms corresponding to the candidate intent trigger word is searched; if a synonym with a higher TIF value exists, the candidate intent trigger word is replaced with that synonym to obtain updated candidate intent trigger words, thereby enhancing the semantic representativeness of the intent trigger words.

[0007] Based on the initial semantic weight of each candidate intent trigger word, dependency parsing is performed on the interaction data to be processed to obtain multiple dependency relationship paths. These paths are then mapped to a pre-defined elderly care semantic relationship graph, generating an associated semantic subgraph corresponding to each candidate intent trigger word. The elderly care semantic relationship graph contains entity nodes in the elderly care domain and semantic relationship edges between these nodes, including causal, temporal, affiliation, and demand relationships. Specifically, a dependency parser is used to obtain the dependency relationship type and direction between every two words with a dependency relationship. Around each candidate intent trigger word, dependency path branches are obtained by traversing along the dependency direction. Nodes in the elderly care semantic relationship graph are matched and associated with words in the dependency path branches, thereby constructing an associated semantic subgraph and effectively introducing domain knowledge into the intent representation process.

[0008] Based on the associated semantic subgraphs corresponding to different candidate intent trigger words, the semantic association strength is calculated, and all candidate intent trigger words are clustered accordingly to obtain at least one intent trigger word cluster. The node overlap rate and edge overlap rate of the associated semantic subgraphs of any two candidate intent trigger words are calculated, and the weighted sum of these rates is used as the semantic association strength. An intent trigger word relationship graph is constructed using all candidate intent trigger words as nodes and the semantic association strength as the edge weight. This graph is then divided into communities, and the candidate intent trigger words within each community are identified as an intent trigger word cluster. Further quality verification of the intent trigger word clusters is performed: the average semantic association strength within the cluster is calculated. If it is less than a preset cluster cohesion threshold, connected component analysis is performed on the associated semantic subgraph corresponding to the cluster, and the candidate intent trigger words in each connected subgraph are taken as a new intent trigger word cluster, thus ensuring close semantic consistency within the cluster.

[0009] For each intent-triggered word cluster, a feature vector for elderly care intent is constructed by combining candidate intent-triggered words within the cluster and their associated semantic subgraphs. Specifically, the initial semantic weights of all candidate intent-triggered words within the cluster are summed to obtain the cluster semantic weight value; the frequency of dependency relation types and part-of-speech tags are statistically analyzed from the associated semantic subgraphs to determine the primary dependency relation type and primary part-of-speech tag, respectively; the word vectors of all nodes in the subgraph are extracted and average pooled to obtain the cluster semantic center vector; the cluster semantic weight value, the type encoding corresponding to the primary dependency relation type, the tag encoding corresponding to the primary part-of-speech tag, and the cluster semantic center vector are concatenated to form a multi-dimensional feature vector for elderly care intent, so that the feature simultaneously carries semantic weight, structural information, and distribution semantics.

[0010] Each elderly care intention feature vector is input into the intention classifier to obtain the corresponding elderly care intention category and output it. Preferably, standard intention feature vectors for multiple preset elderly care intention categories are obtained, the cosine similarity between the elderly care intention feature vector and each standard intention feature vector is calculated, the elderly care intention category corresponding to the largest cosine similarity is taken as the classification result, and each category is sorted and output according to the original occurrence order of its corresponding intention trigger word cluster in the interaction data, ensuring that the classification result can be traced back to the logical order of the original dialogue context.

[0011] After outputting the elderly care intent category, an intent summary can also be generated. Candidate intent trigger words belonging to the same elderly care intent category are identified as the intent evidence word set. The degree centrality value of each intent evidence word in the associated semantic subgraph is calculated, and the word with the highest degree centrality value is selected as the core evidence word for that elderly care intent category. The core evidence words are arranged in the order of category output to form an intent summary sequence, which intuitively displays the user's core intent points during the consultation process.

[0012] A system for mining elderly care intentions based on semantic association analysis includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the aforementioned method for mining elderly care intentions based on semantic association analysis. This system, through semantic role labeling, semantic association strength calculation, and subgraph clustering, can automatically mine and differentiate multi-layered elderly care intentions from complex and fragmented elderly care consultation dialogues. While reducing manual intervention, it effectively improves the precision and recall rate of intention capture, providing a reliable basis for intelligent processing and personalized service recommendations in elderly care consultation services.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: Semantic role labeling extracts predicates and their associated agent and patient roles from the interaction data to be processed as candidate intent trigger words. An initial semantic weight is assigned to each trigger word based on its frequency of occurrence and word frequency-inverse document frequency (IF-IVF) values. Based on dependency parsing, the parsed dependency relationship paths are mapped to a predefined semantic relationship graph, generating an associated semantic subgraph for each candidate intent trigger word. The weighted sum of node overlap rate and edge overlap rate for the associated semantic subgraphs of any two candidate intent trigger words is calculated as the semantic association strength. Based on this, an intent trigger word relationship graph is constructed and communities are divided to form intent trigger word clusters. This approach uses the degree of topological overlap of trigger words in the semantic relationship graph as the basis for measuring their intent coreference, aggregating previously scattered and independent trigger words into semantic clusters pointing to the same intent. This overcomes the shortcomings of traditional methods that only process keywords in isolation, thus severing semantic associations. Compared to solutions that simply perform semantic role and syntactic pruning on sentences and then directly perform semantic matching, this solution uses a pre-defined domain topological constraint on the elderly care semantic relationship graph. This allows scattered trigger words to be clustered according to the domain concept relevance rather than simply aggregating them based on the surface syntactic structure, providing a targeted adaptation for the roundabout and implicit expressions of the elderly in elderly care consultation scenarios. Words within an intent trigger word cluster share nodes and edges in the elderly care semantic relationship graph, indicating that they are in a closely related local structure within the elderly care domain concept network. This allows indirect and implicit intent expressions to be explicitly aggregated into a single, processable intent unit through the graph structure. For each intent trigger word cluster, the initial semantic weights of all candidate intent trigger words within the cluster are summed to obtain the cluster semantic weight value. The most frequent dependency relationship type is selected as the primary dependency relationship type from the corresponding association semantic subgraph of the cluster, and the most frequent part-of-speech tag is selected as the primary part-of-speech tag. The word vectors of all nodes in the subgraph are averaged and pooled to obtain the cluster semantic center vector. These four parts are then concatenated to form the elderly care intent feature vector. This feature vector simultaneously encompasses the overall saliency of the intent cluster, its core syntactic dependency pattern, dominant part-of-speech distribution, and its positional representation in the continuous semantic space, jointly representing the intent cluster from four dimensions: statistical strength, structural pattern, grammatical tendency, and semantic vector. Inputting such a feature vector into an intent classifier enables the classifier to perceive the collaborative expression patterns of words within the intent cluster, thereby achieving stronger discriminative power among different categories of elderly care intents, especially demonstrating stable classification performance for complex elderly care intents with varied expressions and flexible word combinations. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of a method for mining elderly care intentions based on semantic association analysis; Figure 2 This is a flowchart of a method for mining elderly care intentions based on semantic association analysis; Figure 3 It is a scatter plot of the overlap rate between the intended trigger word cluster nodes and edges; Figure 4 This is a principal component analysis distribution diagram of the feature vectors of elderly care intentions. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See Figure 1 This invention provides a method for mining elderly care intentions based on semantic association analysis, comprising: acquiring interaction data to be processed in an elderly care consultation scenario, and performing semantic role labeling on the interaction data to be processed to obtain a labeled semantic role sequence; extracting multiple candidate intention trigger words from the interaction data to be processed according to the labeled semantic role sequence, and assigning a corresponding initial semantic weight to each candidate intention trigger word; performing dependency parsing on the interaction data to be processed based on the initial semantic weight corresponding to each candidate intention trigger word to obtain multiple dependency relationship paths, and mapping the multiple dependency relationship paths to a preset elderly care semantic relationship graph to generate each candidate intention trigger word. Select the associated semantic subgraph corresponding to the intent trigger word; calculate the semantic association strength between different candidate intent trigger words based on the associated semantic subgraph corresponding to each candidate intent trigger word, and cluster all candidate intent trigger words according to the semantic association strength to obtain at least one intent trigger word cluster; for each intent trigger word cluster, construct the elderly care intent feature vector corresponding to the intent trigger word cluster based on the candidate intent trigger words contained in the intent trigger word cluster and the associated semantic subgraph corresponding to the intent trigger word cluster; input each elderly care intent feature vector into the intent classifier to obtain the elderly care intent category corresponding to the elderly care intent feature vector, and output the elderly care intent category.

[0018] Example 1: In specific implementation, please refer to Figure 2This paper implements a semantic association analysis-based method for mining elderly care intentions. This method acquires unprocessed interaction data from elderly care consultation scenarios. This unprocessed interaction data can be dialogue text between elderly users and elderly care service personnel, or text consultation content entered by elderly users through an interactive terminal. After acquiring the unprocessed interaction data, semantic role labeling is performed on the data to obtain a labeled semantic role sequence.

[0019] The process of semantic role labeling (SRP) of the interaction data to be processed to obtain a labeled SRP sequence includes the following operations: The interaction data to be processed is divided into multiple clauses according to punctuation marks, including periods, question marks, exclamation marks, semicolons, and other symbols indicating the end of a sentence. Each clause is segmented to obtain a corresponding word sequence. Word segmentation can be performed using dictionary-based matching or statistical methods. The word sequence corresponding to each clause is input into a SRP labeling model. The model identifies the predicates in each word sequence and labels the agent role, patient role, time role, place role, and manner role of each identified predicate from its corresponding word sequence. Finally, the agent role, patient role, time role, place role, and manner role corresponding to all predicates in each clause are combined according to the original order of the words in the clause to obtain the labeled SRP sequence for that clause. The labeled SRP sequence for each clause can be represented as a sequence composed of SRP labels and corresponding words.

[0020] The semantic role labeling model is a deep learning model jointly trained using a bidirectional long short-term memory (BSSM) network and a conditional random field (CRF). The core architecture of this model includes an input layer, an embedding layer, a BSSM network layer, a CRF layer, and an output layer. The input layer receives the word sequence corresponding to a clause and converts each word into a corresponding word index. The embedding layer maps the word indices to word vectors of a preset dimension, which can be set to 200 dimensions. The embedding layer can load pre-trained word vectors or be updated during model training. The BSSM network layer contains a forward BSSM network and a backward BSSM network. The forward BSSM network processes the word vector sequence from front to back, while the backward BSSM network processes the word vector sequence from back to front. The hidden states output by the forward and backward networks are concatenated to obtain the context feature vector corresponding to each word. The CRF layer receives the context feature vector sequence output by the BSSM network layer, calculates the conditional probability of the label sequence corresponding to the entire word sequence, and outputs the optimal semantic role label sequence. The output layer maps the label sequence output by the conditional random field layer to the semantic role annotation result of each word. The annotation result includes predicate, agent role, patient role, time role, location role, manner role, and non-role label.

[0021] The training process of the semantic role labeling model includes: constructing a training dataset containing multiple clause samples, each with manually labeled predicates and corresponding semantic role labels; inputting the clause samples from the training dataset into the semantic role labeling model to obtain the model's predicted label sequence; calculating the loss function based on the predicted label sequence and the manually labeled label sequence, using the negative log-likelihood loss of a conditional random field; updating the model parameters using a stochastic gradient descent optimization algorithm, with 50 training epochs, a batch size of 32, an initial learning rate of 0.001, and a learning rate decay strategy, decreasing to 0.5 times the original value every 10 epochs. After training, the semantic role labeling model is used to label new clauses with semantic roles.

[0022] After obtaining the annotated semantic role sequence, multiple candidate intent trigger words are extracted from the interaction data to be processed based on the annotated semantic role sequence, and a corresponding initial semantic weight is assigned to each candidate intent trigger word. The specific process is as follows: The agent role and patient role corresponding to each predicate in each clause are extracted from the annotated semantic role sequence of each clause. Each predicate, its corresponding agent role, and its corresponding patient role are collectively identified as the candidate intent trigger word group for that clause. It can be understood that multiple predicates may be identified in a clause, therefore a clause may generate multiple candidate intent trigger word groups. Each candidate intent trigger word group contains a predicate and its corresponding agent role and patient role words. The candidate intent trigger word groups of all clauses are merged to obtain multiple candidate intent trigger words in the interaction data to be processed. The merging operation involves treating all predicates, agent role words, and patient role words appearing in all candidate intent trigger word groups as independent candidate intent trigger words, removing duplicate words, and forming a set of candidate intent trigger words.

[0023] For each candidate intent trigger word, the frequency of its occurrence in all clauses is counted, and the term frequency-inverse document frequency (TFM) value of that candidate intent trigger word in a pre-built corpus of elderly care-related documents is obtained. The elderly care-related corpus is a pre-built corpus that collects documents related to the elderly care field, and the TFM value of each word within it is calculated. The formula for calculating the TFM value is: in, This represents the term frequency-inverse document frequency value of candidate intent-triggered words in the basic elderly care thesaurus. This indicates the total number of times candidate intent trigger words appear in the basic elderly care vocabulary. This indicates the total number of documents contained in the basic thesaurus for elderly care. This represents the number of documents in the basic lexicon for elderly care where the candidate intent trigger word has appeared. A constant of 1 is used to prevent the denominator from being zero. The product of the occurrence frequency and the term frequency minus the inverse document frequency is determined as the initial semantic weight corresponding to the candidate intent trigger word.

[0024] After merging the candidate intent trigger words from all clauses to obtain multiple candidate intent trigger words for the interaction data to be processed, synonym replacement updates can be performed on the candidate intent trigger words to enhance the consistency of their expression. Specifically, a pre-defined thesaurus for elderly care intent is used. This thesaurus stores multiple sets of synonyms, each set consisting of different words expressing the same elderly care intent. Each word in each set has a pre-calculated term frequency-inverse document frequency (TNF) value, calculated in the same way as the basic elderly care thesaurus. For each candidate intent trigger word, the pre-defined thesaurus searches for its corresponding set of synonyms. If a set of synonyms containing the candidate intent trigger word exists in the thesaurus, all other words in that set are considered synonyms of the candidate intent trigger word. The TNF value of each synonym in the set is then compared with the TNF value of the candidate intent trigger word. If a target synonym with a higher term frequency-inverse document frequency (TNF) value exists in the synonym set, the candidate intent trigger word is replaced with the target synonym to obtain an updated candidate intent trigger word; otherwise, the original candidate intent trigger word is retained. All updated candidate intent trigger words are then re-determined as candidate intent trigger words for the interaction data to be processed, and subsequent dependency parsing and other steps are performed based on the updated candidate intent trigger words. In this way, by replacing synonyms, the candidate intent trigger words are selected with more representative words and higher TNF values ​​while maintaining semantic integrity, thus improving the accuracy of subsequent semantic analysis.

[0025] Example 2: In practice, based on the initial semantic weight corresponding to each candidate intent trigger word, dependency parsing is performed on the interaction data to be processed to obtain multiple dependency relationship paths. These multiple dependency relationship paths are then mapped to a preset elderly care semantic relationship graph to generate an associated semantic subgraph corresponding to each candidate intent trigger word.

[0026] The interaction data to be processed is input into the dependency parser. The dependency parser is a pre-trained deep learning model whose core architecture includes an input layer, a word embedding layer, a bidirectional long short-term memory (LSTM) network layer, a dual affine attention layer, and an output layer. The input layer receives the raw text of the interaction data to be processed. The word embedding layer maps each word in the text to a word vector of a preset dimension, set to 300 dimensions. The bidirectional LSM network layer contains a forward LSM network and a backward LSM network. The forward LSM network processes the word vector sequence from front to back, while the backward LSM network processes the word vector sequence from back to front. The hidden states output by the forward and backward networks are concatenated to obtain the context feature vector corresponding to each word, with the hidden state dimension set to 400 dimensions. The dual affine attention layer receives the context feature vector of each word from the bidirectional long short-term memory network layer. It then uses four different feedforward neural networks to convert each word's context feature vector into a dependency arc score vector and a dependency relation type score vector, respectively. The dimension of the dependency arc score vector is the same as the number of words in the interaction data to be processed, and the dimension of the dependency relation type score vector is equal to the preset number of dependency relation types. The output layer of the dependency parser determines whether a dependency relation exists between any two words by calculating the dependency arc score between each pair of words, and determines the dependency direction. For word pairs with a dependency relation, it outputs the dependency relation type and dependency direction between the two words. The dependency direction includes the direction from the core word to the dependent word. The dependency parser is pre-trained using a general domain dependency treebank and fine-tuned using a dependency treebank annotated in the elderly care domain. The learning rate during the fine-tuning phase is set to 0.0001, the training epochs are set to 20, and the batch size is set to 16.

[0027] The dependency parser outputs the dependency relationship type and direction between every two dependent words in the interaction data to be processed, and defines each pair of dependent words and their corresponding dependency relationship type and direction as a dependency relationship path. Starting from each candidate intent trigger word in all dependency relationship paths, the process traverses according to the dependency direction of each dependency relationship path, obtaining at least one dependency path branch for each candidate intent trigger word. Each dependency path branch contains the words and dependency relationship types traversed from that candidate intent trigger word. The traversal process stops when a punctuation mark is encountered or a sentence boundary is reached.

[0028] The pre-built semantic relationship graph for elderly care is a directed graph structure containing entity nodes in the elderly care domain and semantic relationship edges between these nodes. Entity nodes in the elderly care domain include nodes for elderly care service types, elderly person needs types, health status, assistive devices, housing styles, personnel roles, and institution names. Semantic relationship edges include at least causal relationships, temporal relationships, affiliation relationships, and demand relationships. A causal relationship indicates a cause-and-effect connection between two entity nodes; a temporal relationship indicates a chronological order between two entity nodes; an affiliation relationship indicates that one entity node is part of or belongs to another entity node; and a demand relationship indicates that one entity node has a need for or depends on another entity node.

[0029] Each node in the pre-defined semantic relationship graph for elderly care is matched with a word in at least one dependency path branch corresponding to each candidate intent trigger word. The matching process uses a string-complete match method; if the node name in the elderly care semantic relationship graph is exactly the same as a word in a dependency path branch, the match is considered successful. The successfully matched nodes in the elderly care semantic relationship graph and the words in the dependency path branches are then associated. The existing connections between successfully matched nodes in the elderly care semantic relationship graph are then defined as edges in the associated semantic subgraph corresponding to that candidate intent trigger word. The associated semantic subgraph is a subgraph with the successfully matched nodes as the node set and the edges between successfully matched nodes in the elderly care semantic relationship graph as the edge set; this subgraph preserves the topological structure of the elderly care semantic relationship graph.

[0030] Based on the associated semantic subgraph corresponding to each candidate intent trigger word, the semantic association strength between different candidate intent trigger words is calculated. For any two candidate intent trigger words, the first associated semantic subgraph corresponding to the first candidate intent trigger word and the second associated semantic subgraph corresponding to the second candidate intent trigger word are obtained. The first and second associated semantic subgraphs are subjected to node intersection operation to obtain a common node set, which is the set of nodes that exist in both the first and second associated semantic subgraphs. The first and second associated semantic subgraphs are subjected to edge intersection operation to obtain a common edge set, which is the set of edges that exist in both the first and second associated semantic subgraphs. The ratio of the number of nodes in the common node set to the number of nodes in the union of the first and second associated semantic subgraphs is determined as the node overlap rate. The number of nodes in the union is the total number of nodes in the union of the node sets of the first and second associated semantic subgraphs. The ratio of the number of edges in the common edge set to the number of edges in the union of the first and second associated semantic subgraphs is determined as the edge overlap rate. The number of edges in the union is the total number of edges in the union of the edge sets of the first and second associated semantic subgraphs. The weighted sum of node overlap rate and edge overlap rate is determined as the semantic association strength between the first and second candidate intent trigger words. The semantic association strength is denoted as... It can be expressed by the following formula: in, This indicates the semantic association strength between the first candidate intent trigger word and the second candidate intent trigger word; The weighting coefficients representing the node overlap rate. The range of values ​​is , The value is set to 0.4. This setting is based on the premise that the degree of node overlap has a moderate to low impact on the semantic association strength, in order to highlight the importance of edge overlap. This indicates the number of nodes in the common node set; This represents the number of nodes in the union of the first and second associated semantic subgraphs. The weighting coefficients representing the edge overlap rate. The range of values ​​is , The setting is 0.6. This setting is based on the fact that the degree of edge overlap has a medium to high impact on the semantic association strength, because edges carry information about the type of semantic relationship, making the determination of semantic association more critical. This indicates the number of edges in the common edge set; This represents the number of edges in the union of the first and second associated semantic subgraphs.

[0031] Using all candidate intent trigger words as nodes and the semantic association strength between any two candidate intent trigger words as the edge weight, an intent trigger word relationship graph is constructed. The intent trigger word relationship graph is a weighted undirected graph where each node represents a candidate intent trigger word, and each edge connects two different candidate intent trigger words, with the edge weight being the semantic association strength between the two connected candidate intent trigger words. The intent trigger word relationship graph is then divided into communities, and all candidate intent trigger words in each community are identified as an intent trigger word cluster. The Louvain community detection algorithm is used for community division. The Louvain community discovery algorithm executes as follows: During initialization, each node in the intent trigger word graph is treated as an independent community. For each node, the modularity gain gained by moving it to the community of each of its neighboring nodes is calculated. The modularity gain is calculated by subtracting the modularity value of the entire graph before the move from the modularity value of the entire graph after the move. The node is then moved to the neighboring community that yields the maximum modularity gain. If the modularity gain of all neighboring communities is not greater than zero, the node's community remains unchanged. This process is repeated until the community affiliation of all nodes no longer changes. Each community in the current community structure is abstracted as a supernode, with the edge weight between supernodes being the sum of the edge weights between the two connected communities, constructing a new network graph. The community partitioning process is repeated on the new network graph until the modularity of the entire network no longer increases. The modularity calculation formula adopts the standard modularity calculation formula of the Louvain algorithm. After community partitioning, the candidate intent trigger words represented by all nodes in each community constitute an intent trigger word cluster, thus obtaining at least one intent trigger word cluster.

[0032] In some implementation scenarios, after obtaining the intent trigger word clusters, cohesion verification and splitting can be performed on them. For each intent trigger word cluster, the average semantic association strength among all candidate intent trigger words in the cluster is calculated. This calculation iterates through all candidate intent trigger word pairs in the cluster, obtaining the semantic association strength between each pair. The sum of the semantic association strengths of all candidate intent trigger word pairs is then divided by the total number of candidate intent trigger word pairs in the cluster to obtain the average semantic association strength. This average semantic association strength is compared with a preset intra-cluster cohesion threshold. The intra-cluster cohesion threshold is preset to a fixed value of 0.5, based on the theoretical median value of semantic association strength. When the average semantic association strength is lower than the median value, it indicates insufficient intra-cluster cohesion. If the average semantic association strength is less than the intra-cluster cohesion threshold, the corresponding intent trigger word cluster is identified as an intent trigger word cluster to be split. Connectivity component analysis is then performed on the associated semantic subgraph corresponding to the intent trigger word cluster to be split. The process of connected component analysis is as follows: The associated semantic subgraph corresponding to the intent triggering word cluster to be split is treated as an undirected graph. All nodes in the undirected graph are traversed, and all connected components are found using depth-first search or breadth-first search. Each connected component is a maximal connected subgraph of the original undirected graph, meaning that any two nodes in a connected component are connected by a path, while different connected components are not connected by edges. All candidate intent triggering words in each connected subgraph are redefined as a new intent triggering word cluster, and the intent triggering word cluster to be split is replaced with this new cluster. After replacement, the subsequent step of constructing the elderly care intent feature vector is performed for each new intent triggering word cluster.

[0033] It is understandable that, through the above process, based on the initial semantic weights and dependency parsing results, combined with the preset semantic relationship graph of elderly care, an associated semantic subgraph of each candidate intent trigger word is generated. The semantic association strength is quantified by the overlap of nodes and edges between subgraphs. Then, the community detection algorithm is used to cluster the candidate intent trigger words, and adaptive splitting is performed when the cohesion within the cluster is insufficient, so that the obtained intent trigger word clusters can more accurately reflect the semantic structure in the interaction data.

[0034] See Figure 3 In the figure, the horizontal axis represents the node overlap rate, and the vertical axis represents the edge overlap rate, both ranging from 0 to 1. The legend distinguishes between two types of candidate intent trigger word pairs: "intra-cluster candidate intent trigger word pairs" (represented by dark dots) and "inter-cluster candidate intent trigger word pairs" (represented by light squares).

[0035] The figure shows that the node overlap rate of candidate intent trigger word pairs within the cluster is mainly concentrated in the range of 0.4 to 1.0, with the node overlap rate of the vast majority of points exceeding 0.6; the corresponding edge overlap rate is also generally high, with most points distributed in the range of 0.5 to 1.0, and the edge overlap rate of some points approaching 1.0. This indicates that the associated semantic subgraphs corresponding to candidate intent trigger words within the cluster have a high degree of overlap in both nodes and edges, reflecting the strong semantic association of words within the cluster.

[0036] In contrast, the node overlap rate of inter-cluster candidate intent trigger word pairs is mainly concentrated between 0 and 0.4, and the edge overlap rate of most points is less than 0.6, with a relatively scattered distribution. This indicates that the overlap of nodes and edges in the associated semantic subgraphs corresponding to inter-cluster candidate intent trigger words is low, and the semantic association strength is significantly weaker than that of intra-cluster word pairs.

[0037] This data distribution reflects the associated semantic subgraph constructed based on dependency parsing and semantic relation graph mapping for elderly care. When used to calculate the semantic association strength between candidate intent trigger words, it successfully distinguishes between intra-cluster and inter-cluster semantic relationships. Intra-cluster word pairs exhibit strong semantic tightness due to the high overlap rate of nodes and edges, which is beneficial for accurate segmentation in the subsequent intent trigger word clustering process based on the Louvain community discovery algorithm. Meanwhile, the low overlap rate of inter-cluster word pairs avoids confusion between different intent clusters, improving the purity and accuracy of clustering.

[0038] Example 3: In practice, for each intent-triggered word cluster, an elderly care intent feature vector is constructed based on the candidate intent-triggered words contained in the intent-triggered word cluster and the associated semantic subgraph corresponding to the intent-triggered word cluster.

[0039] The process of constructing the elderly care intent feature vector specifically includes the following operations. For each intent trigger word cluster, the initial semantic weights of all candidate intent trigger words in the intent trigger word cluster are accumulated to obtain the cluster semantic weight value of the intent trigger word cluster. The initial semantic weight is the weight value assigned to each candidate intent trigger word in the previous step. This weight value is obtained by multiplying the frequency of the candidate intent trigger word in all clauses by the word frequency-inverse document frequency value of the candidate intent trigger word in the preset elderly care basic lexicon. The accumulation operation is to add the initial semantic weights of all candidate intent trigger words contained in the intent trigger word cluster one by one.

[0040] Extract the dependency relation types of all edges from the semantic subgraph corresponding to the intent-triggered word cluster. Each edge in the semantic subgraph corresponds to a dependency relation type, including subject-verb, verb-object, indirect object, pre-object, pivot, attributive-head, adverbial-head structure, verb-complement structure, parallel, prepositional, left-attached, right-attached, independent, and core relation. Count the frequency of each dependency relation type in the semantic subgraph corresponding to the intent-triggered word cluster. Type frequency refers to the total number of times the same dependency relation type appears in all edges of the semantic subgraph. The dependency relation type with the highest frequency is determined as the primary dependency relation type of the intent-triggered word cluster. If multiple dependency relation types occur with the highest frequency, the dependency relation type with the highest priority is selected as the primary dependency relation type according to the preset priority order of dependency relation types. The preset priority order of dependency relation types is arranged from high to low according to the coreness of the dependency relation type in the syntactic structure, in the following order: core relation, subject-verb relation, verb-object relation, prepositional object relation, attributive-head relation, adverbial-head structure, verb-complement structure, coordinate relation, pivot relation, prepositional object relation, indirect object relation, left-attached relation, right-attached relation, and independent structure.

[0041] Extract part-of-speech tags (POS) from all nodes in the semantic subgraph corresponding to the intent-triggered word cluster. Each node in the semantic subgraph corresponds to a word, and each word has a POS tag, including nouns, verbs, adjectives, adverbs, pronouns, numerals, classifiers, prepositions, conjunctions, auxiliary words, interjections, and onomatopoeia. Count the frequency of each POS tag in the semantic subgraph corresponding to the intent-triggered word cluster. Tag frequency refers to the total number of times the same POS tag appears in all nodes of the semantic subgraph. The POS tag with the highest frequency is determined as the primary POS tag of the intent-triggered word cluster. If multiple POS tags have the same highest frequency, the POS tag with the highest priority is selected as the primary POS tag according to a preset priority order, arranged from highest to lowest information carrying capacity in the expression of the intent for elderly care: nouns, verbs, adjectives, adverbs, pronouns, numerals, classifiers, prepositions, conjunctions, auxiliary words, interjections, and onomatopoeia.

[0042] Word vectors are extracted from the semantic subgraph corresponding to the intent-triggered word cluster. Each node's word vector is obtained using a pre-trained word vector model trained on a Word2Vec Skip-gram model on a corpus related to the elderly care domain, with a word vector dimension of 200. Average pooling is performed on the word vectors of all nodes to obtain the cluster semantic center vector of the intent-triggered word cluster. Average pooling involves summing the word vectors of all nodes in the semantic subgraph along the same dimension and then dividing by the total number of nodes. The formula for calculating the cluster semantic center vector is as follows: in, The cluster semantic center vector representing the intent-triggered word cluster has a dimension of 200. This represents the total number of nodes in the associated semantic subgraph corresponding to the intent-triggered word cluster; Represents the node index variable. The value range is from 1 to Integers; The first word in the semantic subgraph corresponding to the intention-triggered word cluster is... Each node corresponds to a word vector, and the word vector dimension is 200.

[0043] The cluster semantic weight value, the type encoding corresponding to the primary dependency relation type, the label encoding corresponding to the primary part-of-speech tag, and the cluster semantic center vector are concatenated to obtain the elderly care intent feature vector corresponding to the intent triggering word cluster. The concatenation uses a vector concatenation method to form a fixed-dimensional multi-dimensional feature vector. The cluster semantic weight value is a scalar value, represented as a single-dimensional scalar value. The type encoding corresponding to the primary dependency relation type is generated using a one-hot encoding method, with the encoding dimension set to 14 dimensions. Each dimension corresponds to a dependency relation type, including subject-verb relation, verb-object relation, indirect object relation, prepositional object relation, pivotal relation, attributive-head relation, adverbial-head structure, verb-complement structure, parallel relation, prepositional object relation, left-attached relation, right-attached relation, independent structure, and core relation. In the type encoding, the value of the dimension corresponding to the primary dependency relation type is 1, and the value of the other dimensions is 0. The tag encoding corresponding to the main part-of-speech tag is generated using a one-hot encoding method, with a 12-dimensional encoding dimension. Each dimension corresponds to a part-of-speech tag, including nouns, verbs, adjectives, adverbs, pronouns, numerals, quantifiers, prepositions, conjunctions, auxiliary words, interjections, and onomatopoeia. In the tag encoding, the value of the dimension corresponding to the main part-of-speech tag is 1, and the values ​​of the other dimensions are 0. The cluster semantic center vector is a 200-dimensional real number vector. By concatenating the single-dimensional cluster semantic weight value, the 14-dimensional type encoding, the 12-dimensional tag encoding, and the 200-dimensional cluster semantic center vector in sequence, a multi-dimensional feature vector with a fixed dimension of 227 is obtained, which is the elderly care intention feature vector.

[0044] Each elderly care intention feature vector is input into the intention classifier to obtain the elderly care intention category corresponding to the feature vector, and the elderly care intention category is output. The intention classifier is a pre-trained classification model. The model construction process of the intention classifier includes: constructing a training dataset, which contains multiple elderly care intention feature vector samples, each with a manually labeled elderly care intention category. The elderly care intention categories are pre-defined as having multiple types; the elderly care intention categories are divided according to the dimension of elderly care needs, into health management intention, rehabilitation and nursing intention, daily care intention, spiritual comfort intention, cultural and entertainment intention, social interaction intention, legal consultation intention, financial and insurance intention, age-friendly modification intention, and emergency assistance intention. The number of elderly care intention feature vector samples corresponding to each category in the training dataset is no less than 500. The core architecture of the intention classifier adopts a multilayer perceptron model, which includes one input layer, two hidden layers, and one output layer. The input layer receives a 227-dimensional elderly care intention feature vector, with 227 nodes, each node corresponding to one dimension of the elderly care intention feature vector. The first hidden layer contains 256 neurons and uses the ReLU activation function. The second hidden layer contains 128 neurons and uses the ReLU activation function. The output layer contains the same number of neurons as the number of elderly care intention categories. There are 10 elderly care intention categories, and the output layer has 10 nodes. The output layer uses the Softmax activation function and outputs the predicted probability for each elderly care intention category. The training process of the intention classifier includes: inputting elderly care intention feature vector samples from the training dataset into the multilayer perceptron model to obtain the predicted probability distribution output by the model; calculating the cross-entropy loss between the predicted probability distribution and the manually labeled elderly care intention category labels; updating the model parameters using the Adam optimization algorithm, with the learning rate set to 0.0005, the training epochs set to 80, the batch size set to 64, and multiplying the learning rate by a decay factor of 0.8 after every 15 training epochs.

[0045] The intent classifier, when in use, also includes the process of acquiring and matching standard intent feature vectors. Multiple standard intent feature vectors corresponding to various preset elderly care intent categories are acquired, with one standard intent feature vector for each category. The standard intent feature vectors are acquired as follows: for each elderly care intent category, typical standard intent description texts are collected. These texts are written by experts in the elderly care field, and at least 20 standard intent description texts are collected for each category. Each standard intent description text undergoes the same processing flow as the interaction data to be processed, including semantic role labeling, candidate intent trigger word extraction, initial semantic weight allocation, dependency parsing, generation of related semantic subgraphs, generation of intent trigger word clusters, and construction of elderly care intent feature vectors. The elderly care intent feature vectors generated from all standard intent description texts under the same elderly care intent category are averaged and pooled. The average pooling result is used as the standard intent feature vector for that category. The dimension of the standard intent feature vector for each category is the same as the dimension of the elderly care intent feature vector generated from the interaction data to be processed, both being 227 dimensions.

[0046] For each elderly care intent feature vector generated from the interaction data to be processed, the cosine similarity between the elderly care intent feature vector and each standard intent feature vector is calculated. The cosine similarity is calculated by taking the inner product of two 227-dimensional vectors and then dividing by the product of their Euclidean norms. This results in a similarity sequence, the length of which is equal to the number of elderly care intent categories, which is 10. The elderly care intent category corresponding to the standard intent feature vector with the highest cosine similarity in the similarity sequence is determined as the elderly care intent category corresponding to the feature vector. All elderly care intent categories corresponding to the feature vectors are sorted according to the original order of occurrence of the intent trigger word clusters corresponding to each feature vector in the interaction data to be processed. The original order of occurrence of the intent trigger word clusters in the interaction data to be processed is determined by the position of the first candidate intent trigger word in that cluster, with earlier positions ranked higher. The sorted elderly care intent categories are output as a list of elderly care intent category labels arranged in order.

[0047] See Figure 4The figure shows the distribution of the feature vectors of different categories of elderly care intentions on a two-dimensional plane after dimensionality reduction by principal component analysis (PCA) based on the feature vectors constructed in Example 3. The horizontal axis represents the first principal component, with values ​​ranging from approximately -7 to 7, and the vertical axis represents the second principal component, with values ​​ranging from approximately -6 to 7. Both principal components are real number axes. The legend lists 10 categories of elderly care intentions: health management intention, rehabilitation and nursing intention, daily care intention, spiritual comfort intention, cultural and recreational intention, social interaction intention, legal consultation intention, financial and insurance intention, age-friendly renovation intention, and emergency assistance intention. Each category of intention corresponds to a different symbol.

[0048] The feature points of various elderly care intentions exhibit a clear clustered distribution trend in the two-dimensional principal component coordinate system, with the clusters remaining largely separated, demonstrating that the elderly care intention feature vector constructed in Example 3 has good discriminative ability. Among them: Health management intentions are represented by circular dots, mainly concentrated in the [0,4] range of the positive interval of the first principal component and the [0,5] range of the second principal component, forming relatively dense clusters; The rehabilitation nursing intentions are represented by square points, distributed in the [-4,0] interval of the first principal component and the [-1,4] interval of the second principal component, showing a moderate density; The intention to provide daily care is represented by the upward points of the triangle, and the distribution is relatively scattered, mainly in the [1,4] interval of the first principal component and the [-5,0] interval of the second principal component; The intention to provide emotional comfort is represented by the downward point of a triangle, concentrated in the [-4,1] interval of the first principal component and the [-5,0] interval of the second principal component, forming a clear lower cluster; Cultural and entertainment intentions are represented by diamond-shaped points, mainly distributed in the [1,7] interval of the first principal component and the [-2,2] interval of the second principal component, showing a right-side clustering trend; Social interaction intentions are represented by pentagonal points, which are relatively dispersed but concentrated in the [-6,-2] interval of the first principal component and the [-3,2] interval of the second principal component; The intent of legal consultation is represented by a pentagram, which is located in the interval [-3,3] of the first principal component and the interval [2,7] of the second principal component, forming a significant cluster above. Financial and insurance intentions are represented by cross points, distributed in the [-2,2] interval of the first principal component and the [-6,-2] interval of the second principal component, located in the lower middle region; The intention to make modifications suitable for the elderly is represented by plus signs, mainly concentrated in the [1,7] interval of the first principal component and the [-5,-1] interval of the second principal component; The intent to provide emergency assistance is represented by hexagonal points, which are relatively dispersed and are located in the interval [-6, -1] of the first principal component and the interval [1, 5] of the second principal component.

[0049] Example 4: In practice, after inputting each elderly care intention feature vector into the intention classifier to obtain the elderly care intention category corresponding to the elderly care intention feature vector, core evidence words can be extracted from the interaction data to be processed based on the elderly care intention category, and an intention summary sequence can be generated.

[0050] The set of intent evidence words for each elderly care intent category is determined by identifying all candidate intent trigger words in the interaction data to be processed, corresponding to each category. Candidate intent trigger words are those extracted and merged from the labeled semantic role sequence of the interaction data in the previous steps. The specific method for determining the intent evidence word set is as follows: traverse all candidate intent trigger words in the interaction data. For each candidate intent trigger word, determine the final intent trigger word cluster to which it belongs. Then, determine the elderly care intent feature vector corresponding to the intent trigger word cluster, and further determine the elderly care intent category corresponding to the feature vector. The elderly care intent category is the label obtained after classifying the feature vector using an intent classifier. Associate the candidate intent trigger words with their corresponding elderly care intent category labels. After traversal, group all candidate intent trigger words associated with the same elderly care intent category together to form the intent evidence word set for that category. One intent evidence word set is obtained for each elderly care intent category.

[0051] For each elderly care intent category, calculate the degree centrality value of each intent evidence word in the intent evidence word set of that category within the associated semantic subgraph corresponding to that category. The associated semantic subgraph corresponding to the elderly care intent category refers to the associated semantic subgraph corresponding to the intent trigger word cluster of that category. The degree centrality value represents the tightness of the connection between the intent evidence word and other nodes in the associated semantic subgraph, and is obtained by calculating the degree of the node corresponding to the intent evidence word in the associated semantic subgraph. For each intent evidence word in the intent evidence word set, locate the node corresponding to that intent evidence word in the associated semantic subgraph, count the number of edges directly connected to that node, and use the counted number of edges as the degree of the node corresponding to that intent evidence word. The formula for calculating the degree centrality value is expressed as: in, The degree centrality value of the intention evidence word in the semantic subgraph corresponding to the elderly care intention category. The range of values ​​is ; This indicates the node corresponding to the intent evidence word in the associated semantic subgraph; It is a node identifier index. The value range is an integer from 1 to the total number of nodes in the associated semantic subgraph; Represents a node The degree, that is, the degree with the node The number of directly connected edges; This represents the total number of nodes in the associated semantic subgraph. It is an integer greater than or equal to 1; In the semantic subgraph representing the association, excluding nodes The number of nodes other than itself is used to normalize the degree. When the association semantic subgraph contains only one node... The value is 1, and the denominator is zero. For this special case, the degree centrality value is directly... Set to 1. For any intent evidence word in the intent evidence word set, if the intent evidence word does not appear in the associated semantic subgraph, then the degree centrality value of the intent evidence word is set to 0.

[0052] The intent evidence word with the highest degree centrality value in the intent evidence word set is identified as the core evidence word for the elderly care intent category. If there are multiple intent evidence words with the highest degree centrality value, the intent evidence word with the earliest first occurrence in the interaction data to be processed is selected as the core evidence word. The method for determining the earliest first occurrence in the interaction data to be processed is as follows: the interaction data to be processed is indexed sequentially by word order, and the index position of the first occurrence of each candidate intent trigger word in the interaction data to be processed is obtained. The smaller the index position value, the earlier the occurrence position.

[0053] Arrange the core evidence words corresponding to all elderly care intent categories according to the output order of the elderly care intent categories to generate an intent summary sequence of the interaction data to be processed. The output order of the elderly care intent categories is the order obtained by sorting all elderly care intent feature vectors corresponding to the elderly care intent categories according to the original occurrence order of intent trigger word clusters in the interaction data to be processed in the previous step. Extract the core evidence words corresponding to each elderly care intent category in sequence according to this output order, and arrange the extracted core evidence words into a sequence, which is the intent summary sequence of the interaction data to be processed. The intent summary sequence is presented in the form of a word list, where each element is a core evidence word, and the order of the core evidence words is consistent with the output order of the elderly care intent categories.

[0054] Example 5: In a specific implementation, a system for mining elderly care intentions based on semantic association analysis is provided. This system includes a memory, a processor, and a computer program stored in the memory and running on the processor. The memory stores the computer program, intermediate data, and final result data generated during system operation. The memory type includes at least one of random access memory, read-only memory, and flash memory. The processor includes at least one of a central processing unit, a graphics processing unit, or a neural network processor. When the processor executes the computer program, it implements the complete steps of a method for mining elderly care intentions based on semantic association analysis.

[0055] The processor reads and loads computer program instructions from memory, and sequentially completes the elderly care intent mining process according to the execution order of the computer program instructions. The processor first executes the program segment corresponding to the data acquisition and semantic role labeling module to acquire the interaction data to be processed in the elderly care consultation scenario, and then performs semantic role labeling on the acquired interaction data to obtain a labeled semantic role sequence. The interaction data to be processed can be acquired through methods including receiving it from external input devices, reading it from a pre-defined elderly care interaction data storage area in memory, or obtaining it from a remote server via a communication interface.

[0056] The process of the processor performing semantic role annotation on the interaction data to be processed to obtain an annotated semantic role sequence has been described in detail in previous embodiments. In general, it includes: dividing the interaction data to be processed into multiple clauses according to punctuation marks, and performing word segmentation on each clause to obtain a word sequence corresponding to each clause; inputting the word sequence corresponding to each clause into a semantic role annotation model, identifying the predicates in each word sequence through the semantic role annotation model, and for each identified predicate, annotating the agent role, patient role, time role, place role, and manner role of the predicate from the word sequence corresponding to the predicate; combining the agent role, patient role, time role, place role, and manner role corresponding to all predicates in each clause according to the original order of the words in the clause to obtain the annotated semantic role sequence corresponding to the clause. The semantic role labeling model is a deep learning model jointly trained by a bidirectional long short-term memory network and a conditional random field. The structure and training method of the semantic role labeling model have been described in detail in previous embodiments. When the processor is running, it loads the trained semantic role labeling model parameter file, loads the model parameters in the semantic role labeling model parameter file into memory, constructs the computation graph of the semantic role labeling model, performs forward computation on the input data, and outputs the labeling results.

[0057] After generating the annotated semantic role sequence, the processor extracts multiple candidate intent trigger words from the interaction data to be processed based on the annotated semantic role sequence, and assigns a corresponding initial semantic weight to each candidate intent trigger word. The extraction process includes: extracting the agent role and patient role corresponding to each predicate in each clause from the annotated semantic role sequence of each clause, and determining each predicate, its corresponding agent role, and its corresponding patient role in the clause as the candidate intent trigger word group for that clause; merging the candidate intent trigger word groups of all clauses to obtain multiple candidate intent trigger words in the interaction data to be processed. The process of assigning initial semantic weights includes: for each candidate intent trigger word, counting the frequency of occurrence of the candidate intent trigger word in all clauses, obtaining the word frequency-inverse document frequency value of the candidate intent trigger word in the preset basic lexicon, and determining the product of the occurrence frequency and the word frequency-inverse document frequency value as the initial semantic weight corresponding to the candidate intent trigger word. The pre-defined basic vocabulary for elderly care is stored in the memory. The term frequency-inverse document frequency value of each word in the basic vocabulary for elderly care has been pre-calculated and stored in key-value pairs before the candidate intent trigger word extraction is performed. The processor obtains the corresponding term frequency-inverse document frequency value by querying the basic vocabulary for elderly care through the word string.

[0058] In some optional implementations, after merging the candidate intent trigger words of all clauses to obtain multiple candidate intent trigger words, the processor also performs a synonym replacement update step for the candidate intent trigger words: for each candidate intent trigger word, the processor searches for the corresponding set of synonyms in a preset elderly care intent synonym library; it compares the term frequency-inverse document frequency (TNF) value of each synonym in the TNF set with the TNF value of the candidate intent trigger word; if there is a target synonym in the TNF set whose TNF value is greater than that of the candidate intent trigger word, the candidate intent trigger word is replaced with the target synonym, resulting in an updated candidate intent trigger word; the updated candidate intent trigger word is then re-determined as the candidate intent trigger word for the interaction data to be processed. The preset elderly care intent synonym library is also stored in memory, and the elderly care intent synonym library is organized as a synonym index table, with each entry word corresponding to a set of synonyms and their TNF values.

[0059] After obtaining candidate intent trigger words and their corresponding initial semantic weights, the processor performs dependency parsing on the interaction data to be processed based on the initial semantic weights of each candidate intent trigger word. This yields multiple dependency relationship paths, which are then mapped to a pre-defined semantic relationship graph, generating an associated semantic subgraph corresponding to each candidate intent trigger word. The dependency parsing is performed by the processor calling a dependency parser, which is a pre-trained deep learning model whose structure has been described in previous embodiments. The dependency parser outputs the dependency relationship type and direction between every two dependent words in the interaction data to be processed, and defines each pair of dependent words and their corresponding dependency relationship type and direction as a dependency relationship path. The processor uses each candidate intent trigger word in all dependency relationship paths as a starting point and traverses according to the dependency direction of each dependency relationship path to obtain at least one dependency path branch corresponding to each candidate intent trigger word. A pre-defined semantic relationship graph for elderly care is stored in memory. This graph uses a graph data structure, containing a list of nodes and edges. Each node includes a name and attributes, and each edge includes a start node, an end node, and a semantic relationship type. The processor matches each node in the pre-defined semantic relationship graph with words in the dependency path branches. Successfully matched nodes are associated with the words in the dependency path branches, and the connections between successfully matched nodes are defined as edges in the associated semantic subgraph corresponding to the candidate intent trigger word, thus obtaining the associated semantic subgraph.

[0060] After obtaining the associated semantic subgraph, the processor calculates the semantic association strength between different candidate intent trigger words based on the associated semantic subgraph corresponding to each candidate intent trigger word, and clusters all candidate intent trigger words according to the semantic association strength to obtain at least one intent trigger word cluster. The process of calculating the semantic association strength includes: for any two candidate intent trigger words, obtaining the first associated semantic subgraph corresponding to the first candidate intent trigger word and the second associated semantic subgraph corresponding to the second candidate intent trigger word; performing node intersection operation on the first and second associated semantic subgraphs to obtain a common node set, and performing edge intersection operation to obtain a common edge set; determining the node overlap rate as the ratio between the number of nodes in the common node set and the number of nodes in the union of the first and second associated semantic subgraphs, and determining the edge overlap rate as the ratio between the number of edges in the common edge set and the number of edges in the union of the first and second associated semantic subgraphs; and determining the semantic association strength between the first and second candidate intent trigger words by the weighted sum of the node overlap rate and the edge overlap rate. The clustering process includes: constructing an intent trigger word relationship graph using all candidate intent trigger words as nodes and the semantic association strength between any two candidate intent trigger words as edge weights; dividing the intent trigger word relationship graph into communities; and identifying all candidate intent trigger words in each community as an intent trigger word cluster. The community division is implemented using the Louvain community detection algorithm, whose program code is stored in memory. The processor calls this algorithm to iteratively divide the intent trigger word relationship graph.

[0061] In some optional implementations, after obtaining the intent trigger word cluster, the processor also performs an intra-cluster cohesion check and splitting step: for each intent trigger word cluster, the average semantic association strength among all candidate intent trigger words in the intent trigger word cluster is calculated, and the average semantic association strength is compared with a preset intra-cluster cohesion threshold; if the average semantic association strength is less than the intra-cluster cohesion threshold, the intent trigger word cluster is determined as the intent trigger word cluster to be split, and connected component analysis is performed on the associated semantic subgraph corresponding to the intent trigger word cluster to be split to obtain multiple connected subgraphs; all candidate intent trigger words in each connected subgraph are re-determined as a new intent trigger word cluster, and the intent trigger word cluster to be split is replaced with the new intent trigger word cluster.

[0062] After determining the intent trigger word cluster, the processor constructs the elderly care intent feature vector corresponding to each intent trigger word cluster based on the candidate intent trigger words contained in the intent trigger word cluster and the associated semantic subgraph corresponding to the intent trigger word cluster. The construction process includes: accumulating the initial semantic weights of all candidate intent trigger words in the intent trigger word cluster to obtain the cluster semantic weight value; extracting the dependency relationship types of all edges from the associated semantic subgraph corresponding to the intent trigger word cluster, counting the frequency of each dependency relationship type, and determining the dependency relationship type with the highest frequency as the primary dependency relationship type of the intent trigger word cluster; extracting the part-of-speech tags of all nodes from the associated semantic subgraph corresponding to the intent trigger word cluster, counting the tag frequency of each part-of-speech tag, and determining the part-of-speech tag with the highest frequency as the primary part-of-speech tag of the intent trigger word cluster; extracting the word vector of each node from the associated semantic subgraph corresponding to the intent trigger word cluster, performing average pooling on the word vectors of all nodes to obtain the cluster semantic center vector of the intent trigger word cluster; concatenating the cluster semantic weight value, the type encoding corresponding to the primary dependency relationship type, the tag encoding corresponding to the primary part-of-speech tag, and the cluster semantic center vector, and forming a fixed-dimensional multi-dimensional feature vector using a vector concatenation method. The word vectors are obtained by the processor calling a pre-trained word vector model, and the word vector model file is stored in memory.

[0063] After constructing the elderly care intention feature vectors, the processor inputs each feature vector into the intention classifier to obtain the corresponding elderly care intention category and outputs the category. The intention classifier is a pre-trained classification model stored in memory. The processor loads the model structure and parameters of the intention classifier and performs forward inference computation. The classification process includes: obtaining multiple standard intention feature vectors corresponding to various preset elderly care intention categories (the standard intention feature vectors are pre-calculated and stored in memory); calculating the cosine similarity between each elderly care intention feature vector and each standard intention feature vector to obtain a similarity sequence; determining the elderly care intention category corresponding to the standard intention feature vector with the highest cosine similarity in the similarity sequence; and sorting all the elderly care intention categories corresponding to the feature vectors according to the original order of occurrence of the intent trigger word clusters corresponding to each feature vector in the interaction data to be processed, and outputting the sorted categories. The output of the elderly care intention categories can be achieved by writing the sorted category label list into memory, visualizing it through a display device, or sending it to an external system via a communication interface.

[0064] In some optional implementations, after outputting the elderly care intent category, the processor further performs an intent summary sequence generation step: determining all candidate intent trigger words in the interaction data to be processed that correspond to each elderly care intent category as the intent evidence word set for the elderly care intent category; for each elderly care intent category, calculating the degree centrality value of each intent evidence word in the intent evidence word set of the elderly care intent category in the associated semantic subgraph corresponding to the elderly care intent category, and determining the intent evidence word with the highest degree centrality value as the core evidence word of the elderly care intent category; arranging all the core evidence words corresponding to the elderly care intent categories according to the output order of their corresponding elderly care intent categories to generate the intent summary sequence of the interaction data to be processed, and outputting the intent summary sequence.

[0065] It is understandable that when the processor executes a computer program, it sequentially calls the program segments corresponding to each functional module in the above order. Data is transferred between functional modules through a shared data area in memory, with the output data of the previous module serving as the input data for the next. The system's basic elderly care lexicon, elderly care intent thesaurus, elderly care semantic relationship graph, semantic role labeling model parameters, dependency parser model parameters, word vector model parameters, intent classifier model parameters, and standard intent feature vectors are all stored in memory as persistent files. The processor loads the required files into memory when the system starts or is used for the first time. The entire elderly care intent mining process is fully automated, requiring no manual intervention. The processor completes all processing steps from semantic role labeling to elderly care intent category output and intent summary sequence generation based on the input interaction data.

[0066] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for mining elderly care intentions based on semantic association analysis, characterized in that, The method includes: Acquire the interaction data to be processed in the elderly care consultation scenario, and perform semantic role labeling on the interaction data to be processed to obtain the labeled semantic role sequence; Based on the labeled semantic role sequence, multiple candidate intent trigger words are extracted from the interaction data to be processed, and a corresponding initial semantic weight is assigned to each candidate intent trigger word; Based on the initial semantic weight corresponding to each candidate intent trigger word, the interaction data to be processed is parsed using dependency parsing to obtain multiple dependency relationship paths. These multiple dependency relationship paths are then mapped to a preset semantic relationship graph for elderly care, generating an associated semantic subgraph corresponding to each candidate intent trigger word. Based on the associated semantic subgraph corresponding to each candidate intent trigger word, the semantic association strength between different candidate intent trigger words is calculated, and all candidate intent trigger words are clustered according to the semantic association strength to obtain at least one intent trigger word cluster; For each intent trigger word cluster, construct the elderly care intent feature vector corresponding to the intent trigger word cluster based on the candidate intent trigger words contained in the intent trigger word cluster and the associated semantic subgraph corresponding to the intent trigger word cluster; Each elderly care intention feature vector is input into the intention classifier to obtain the elderly care intention category corresponding to the feature vector, and the elderly care intention category is output.

2. The method for mining elderly care intentions based on semantic association analysis according to claim 1, characterized in that, The interaction data to be processed is semantically labeled to obtain a labeled semantic role sequence, including: The interactive data to be processed is divided into multiple clauses according to punctuation marks, and each clause is segmented into words to obtain the word sequence corresponding to each clause; The word sequence corresponding to each clause is input into the semantic role labeling model. The semantic role labeling model identifies the predicate in each word sequence and, for each identified predicate, labels the agent role, patient role, time role, place role, and manner role of the predicate from the word sequence corresponding to the predicate. The agent, patient, time, place, and manner roles corresponding to all predicates in each clause are combined according to the original order of the words in the clause to obtain the labeled semantic role sequence corresponding to the clause.

3. The method for mining elderly care intentions based on semantic association analysis according to claim 2, characterized in that, The semantic role labeling model is a deep learning model jointly trained on a bidirectional long short-term memory network and a conditional random field.

4. The method for mining elderly care intentions based on semantic association analysis according to claim 2, characterized in that, Based on the labeled semantic role sequence, multiple candidate intent trigger words are extracted from the interaction data to be processed, and each candidate intent trigger word is assigned a corresponding initial semantic weight, including: Extract the agent role and patient role corresponding to each predicate in each clause from the semantic role sequence after annotation, and determine each predicate, the agent role corresponding to the predicate, and the patient role corresponding to the predicate as the candidate intent trigger words for the clause. The candidate intent trigger words of all clauses are merged to obtain multiple candidate intent trigger words of the interaction data to be processed; For each candidate intent trigger word, the frequency of its occurrence in all clauses is counted, and the word frequency-inverse document frequency value of the candidate intent trigger word in the preset elderly care basic lexicon is obtained. The product of the occurrence frequency and the word frequency-inverse document frequency value is determined as the initial semantic weight corresponding to the candidate intent trigger word.

5. The method for mining elderly care intentions based on semantic association analysis according to claim 4, characterized in that, The process involves performing dependency parsing on the interaction data to be processed based on the initial semantic weight corresponding to each candidate intent trigger word, obtaining multiple dependency relationship paths, and mapping these multiple dependency relationship paths to a preset semantic relationship graph for elderly care, thereby generating an associated semantic subgraph corresponding to each candidate intent trigger word, including: The interaction data to be processed is input into the dependency parser. The dependency parser outputs the dependency relationship type and dependency direction between every two words with a dependency relationship in the interaction data to be processed, and determines every two words with a dependency relationship and their corresponding dependency relationship type and dependency direction as a dependency relationship path. Starting from each candidate intent trigger word in all dependency paths, traverse according to the dependency direction of each dependency path to obtain at least one dependency path branch corresponding to each candidate intent trigger word. Each dependency path branch contains the words and dependency relationship types traversed from the candidate intent trigger word. Each node in the preset semantic relationship graph of elderly care is matched with the words in at least one dependency path branch corresponding to each candidate intent trigger word. The successfully matched nodes and the words in the dependency path branches are associated. The connection between the successfully matched nodes is determined as the edge of the associated semantic subgraph corresponding to the candidate intent trigger word, thus obtaining the associated semantic subgraph corresponding to the candidate intent trigger word.

6. The method for mining elderly care intentions based on semantic association analysis according to claim 5, characterized in that, The step involves calculating the semantic association strength between different candidate intent trigger words based on the associated semantic subgraph corresponding to each candidate intent trigger word, and clustering all candidate intent trigger words according to the semantic association strength to obtain at least one intent trigger word cluster, including: For any two candidate intent trigger words, obtain the first associated semantic subgraph corresponding to the first candidate intent trigger word and the second associated semantic subgraph corresponding to the second candidate intent trigger word; Perform node intersection operation on the first associated semantic subgraph and the second associated semantic subgraph to obtain a common node set, and perform edge intersection operation on the first associated semantic subgraph and the second associated semantic subgraph to obtain a common edge set; The ratio between the number of nodes in the common node set and the number of nodes in the union of the first associated semantic subgraph and the second associated semantic subgraph is determined as the node overlap rate, and the ratio between the number of edges in the common edge set and the number of edges in the union of the first associated semantic subgraph and the second associated semantic subgraph is determined as the edge overlap rate. The weighted sum of the node overlap rate and the edge overlap rate is determined as the semantic association strength between the first candidate intent trigger word and the second candidate intent trigger word; Using all candidate intent trigger words as nodes and the semantic association strength between every two candidate intent trigger words as the edge weight, an intent trigger word relationship graph is constructed. The intent trigger word relationship graph is then divided into communities, and all candidate intent trigger words in each community are identified as an intent trigger word cluster.

7. The method for mining elderly care intentions based on semantic association analysis according to claim 6, characterized in that, For each intent-triggered word cluster, based on the candidate intent-triggered words contained in the intent-triggered word cluster and the associated semantic subgraph corresponding to the intent-triggered word cluster, a feature vector of elderly care intent corresponding to the intent-triggered word cluster is constructed, including: For each intent trigger word cluster, the initial semantic weights of all candidate intent trigger words in the intent trigger word cluster are summed to obtain the cluster semantic weight value of the intent trigger word cluster; Extract the dependency relationship types of all edges from the associated semantic subgraph corresponding to the intent triggering word cluster, and count the frequency of each dependency relationship type in the associated semantic subgraph corresponding to the intent triggering word cluster. The dependency relationship type with the highest frequency is determined as the main dependency relationship type of the intent triggering word cluster. Extract the part-of-speech tags of all nodes from the associated semantic subgraph corresponding to the intent-triggered word cluster, and count the frequency of each part-of-speech tag in the associated semantic subgraph corresponding to the intent-triggered word cluster. The part-of-speech tag with the highest frequency is determined as the main part-of-speech tag of the intent-triggered word cluster. Extract the word vector of each node from the associated semantic subgraph corresponding to the intent-triggered word cluster, and perform average pooling on the word vectors of all nodes to obtain the cluster semantic center vector of the intent-triggered word cluster; The cluster semantic weight value, the type encoding corresponding to the main dependency relationship type, the tag encoding corresponding to the main part-of-speech tag, and the cluster semantic center vector are concatenated to obtain the elderly care intention feature vector corresponding to the intention triggering word cluster.

8. The method for mining elderly care intentions based on semantic association analysis according to claim 7, characterized in that, The cluster semantic weight value, the type encoding corresponding to the main dependency relationship type, the tag encoding corresponding to the main part-of-speech tag, and the cluster semantic center vector are concatenated to form a fixed-dimensional multidimensional feature vector using a vector concatenation method.

9. A method for mining elderly care intentions based on semantic association analysis according to claim 7, characterized in that, The process of inputting each elderly care intention feature vector into an intention classifier to obtain the elderly care intention category corresponding to the feature vector and outputting the elderly care intention category includes: Obtain multiple standard intent feature vectors corresponding to various preset elderly care intent categories, with each elderly care intent category corresponding to one standard intent feature vector; For each feature vector of elderly care intention, calculate the cosine similarity between the feature vector of elderly care intention and each standard intention feature vector to obtain the similarity sequence corresponding to the feature vector of elderly care intention. The category of elderly care intention corresponding to the standard intention feature vector with the maximum cosine similarity in the similarity sequence is determined as the category of elderly care intention corresponding to the elderly care intention feature vector. Sort all elderly care intention categories corresponding to the feature vectors of elderly care intentions according to the original order of occurrence of the intention trigger word clusters corresponding to each feature vector in the interaction data to be processed, and output the sorted elderly care intention categories.

10. A system for mining elderly care intentions based on semantic association analysis, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for mining elderly care intentions based on semantic association analysis as described in any one of claims 1 to 9.

Images