Method for correcting dialogue intention information of old people based on deep neural network

Abstract

The invention discloses a method for correcting dialogue intention information of old people based on a deep neural network, and particularly relates to the technical field of dialogue optimization, and the method comprises the steps: collecting and cleaning multi-topic multi-round dialogue data, marking multi-round dialogue corpora, and unifying the format; performing feature extraction on the labeled corpus by using the multi-intention recognition sub-model to generate a multi-intention representation vector capable of representing each intention; screening context content related to the target intention in combination with a hierarchical attention mechanism of user context dynamic weight adjustment; processing context content related to the target intention through gating operation based on multiple rounds of interaction history, and reserving semantic fragments with reference value; performing parallel analysis on the reference semantic fragments based on a multi-task learning structure of a deep neural network, performing classification processing on different intentions, and calculating scores; and executing multi-intention fusion correction operation according to classification and scoring results, and outputting intention information capable of distinguishing each intention.

Description

Technical Field

[0001] The present invention relates to the technical field of dialogue optimization, and more specifically, to a method for correcting dialogue intention information of an elderly person based on a deep neural network. Background Art

[0002] Current dialogue systems mainly understand user intent through keyword matching or simple intent recognition algorithms. These systems perform well in dealing with daily conversations, but have obvious limitations when facing elderly users; elderly conversations usually involve multiple complex and context-specific intent expressions, such as health consultation, family arrangements, or life advice, which may appear at the same time in a conversation. In addition, the language expressions of elderly users often contain vague, implicit, or incomplete information, and the strong focus on health status and family background makes it difficult for traditional dialogue systems to meet the actual needs of elderly users.

[0003] In the existing technology, it is difficult to effectively distinguish and accurately correct the information of each intent in a conversation where multiple intents are expressed in parallel, which can easily lead to confusion and deviation of intent, and thus cannot meet the needs of elderly users. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method for correcting the conversation intention information of the elderly based on a deep neural network to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: The method for correcting the conversation intention information of the elderly based on a deep neural network includes the following steps: S1: Collect and clean multi-round dialogue data containing multiple topics, annotate the multi-round dialogue data and unify the format; S2: Use the multi-intent recognition sub-model to extract topic features from the annotated multi-round dialogue data and generate a multi-intent representation vector that can represent each intent; S3: A hierarchical attention mechanism combined with dynamic weight adjustment of user context is used to perform association calculation on multiple intent representation vectors to filter out contextual content related to the target intent. S4: Process the contextual content related to the target intent through gating operations based on multi-round interaction history, and retain reference semantic fragments that are of reference value for intent correction; S5: Based on the multi-task learning structure of deep neural network, the reference semantic fragments are analyzed in parallel, different intents are classified and processed respectively, and the score of each intent is calculated; S6: Perform multi-intent fusion correction operation based on the results of classification processing and intent scoring, and output intent information that can distinguish each intent.

[0006] In a preferred embodiment, collecting and cleaning multi-round dialogue data containing multiple topics, annotating the multi-round dialogue data and unifying the format, specifically includes: Obtain multi-round conversation data from multi-channel data sources; Clean the multi-round conversation data, including removing noise data, incomplete data, and meaningless conversation records; Use predefined annotation rules to annotate the cleaned multi-turn dialogue data. The annotation rules include annotation of topic categories, intent categories, and dialogue turn information. The labeled multi-round dialogue data is formatted in a unified manner, including the normalization of data fields, the unified representation of time series, and the normalization of the order of dialogue rounds.

[0007] In a preferred embodiment, the multi-intention recognition sub-model is used to extract topic features from the annotated multi-round dialogue corpus to generate a multi-intention representation vector that can represent each intention, specifically including: By inputting the pre-trained multi-intent recognition sub-model into the annotated multi-round dialogue data, the text content of each round of dialogue is analyzed and the semantic information related to the topic features is extracted; In the multi-intent recognition sub-model, the semantic analysis module is used to extract features from the text content of each round of dialogue. The extracted features include semantic features of topic categories and semantic features of intent categories. Use the embedding vector generation module to embed the extracted semantic features and generate initial vectors that can represent topic categories and intent categories; The feature fusion module of the multi-intention recognition sub-model is used to fuse the initial vectors generated in each round of dialogue to obtain a multi-intention representation vector that can comprehensively represent multiple intent categories and topic categories.

[0008] In a preferred embodiment, a hierarchical attention mechanism combined with dynamic weight adjustment of user context is used to perform association calculation on multiple intent representation vectors to filter context content related to the target intent, specifically including: The multi-intent representation vectors are processed in layers according to the predefined attention levels, and each layer assigns initial weights to different dimensional features in the user context information; Dynamically adjust the weight distribution of attention at each layer based on the historical conversation content and semantic features of the current user context information; In each layer of the attention mechanism, the multi-intent representation vectors are weighted according to the adjusted weight distribution to obtain the weight distribution of different intent categories in each layer; Through the weighted results of the multi-layer attention mechanism and the contextual information related to the target intent, the semantic fragments with the strongest correlation with the current target intent are screened out; The semantic fragments that are most relevant to the current target intent are processed uniformly to generate a contextual content representation that can characterize the target intent.

[0009] In a preferred embodiment, the specific formula for dynamically adjusting the weight distribution of attention in each layer is as follows: ;in, Indicates the adjusted Layer attention weights, Indicates The initial weights of the layers, Represents user context information and The similarity of the multi-intention representation vector input of the layer, Represents user context information, Indicates The multi-intent representation vector input of the layer.

[0010] In a preferred embodiment, the context content related to the target intent is processed by a gating operation based on multiple rounds of interaction history, and reference semantic fragments with reference value for intent correction are retained, specifically including: Extract user multi-round interaction history information from the context content related to the target intent, including conversation turns, context semantic fragments, and time series features associated with the target intent; Perform feature screening on the extracted historical information of multiple rounds of user interactions, assign an initial weight to each piece of historical information based on its relevance to the target intent, and remove historical information whose relevance to the target intent is lower than a preset threshold; On the basis of allocating the initial weights, the contextual content related to the target intent is screened through the gating operation, which specifically includes dynamically adjusting the gating threshold according to the weight of each piece of historical information to screen out the contextual content that meets the gating conditions; The context content filtered out by the gating operation is normalized, including time series sorting and content structuring of the filtered context semantic fragments, and the normalized context content is stored as a reference semantic fragment.

[0011] In a preferred embodiment, a multi-task learning structure based on a deep neural network analyzes the reference semantic fragments in parallel, classifies different intents and calculates the score of each intent, specifically including: Preprocess the reference semantic fragments according to their contextual relevance and time series characteristics; Construct a multi-task learning structure of a deep neural network, which includes intent classification and intent scoring tasks. The intent classification task is used to identify the intent category corresponding to the reference semantic fragment, and the intent scoring task is used to calculate the weight score of each intent. In the multi-task learning structure, the pre-processed reference semantic fragments are processed in parallel, and the intent classification task classifies each semantic fragment to determine the intent category to which it belongs; In the intent scoring task, each intent category is scored based on the semantic features and contextual relationships of the reference semantic fragments. The scoring criteria include the degree of association between the semantic fragments and the target intent and the time series characteristics. The results of intent classification and intent scoring are integrated to generate the classification label and scoring value for each reference semantic fragment.

[0012] In a preferred embodiment, a multi-intent fusion correction operation is performed according to the results of the classification process and the intent scoring, and the intent information capable of distinguishing each intent is output, specifically including: Receive classification labels and scoring values ​​generated by classification processing and intent scoring results, aggregate the classification labels according to intent categories, and initially form a classification set of intent categories; Based on the intent categories and their ratings in the classification set, sort the intent categories in descending order of ratings, and select the intent category with the highest rating as the primary intent; Perform semantic supplementation on the intent categories with lower scores but certain relevance, and merge them with the main intent to form a comprehensive correction result of multiple intents; The semantic information of the main intent and related intent is fused through the multi-intention fusion algorithm to generate intent information that can distinguish each intent; The fused and corrected intent information is stored in a structured manner, and a clear classification result that can distinguish between the main intent and related intent is output.

[0013] Technical effects and advantages of the method for correcting elderly conversation intention information based on deep neural network in the present invention: 1. The present invention can effectively solve the intention confusion and deviation problems caused by the parallel expression of multiple intentions in the conversations of the elderly through multi-task learning and hierarchical attention mechanism based on deep neural networks; by collecting and cleaning multi-round conversation data of the elderly on multiple topics, and using the multi-intention recognition sub-model to extract features from the annotated corpus, a multi-intention representation vector that can represent each intention is generated, which lays the foundation for subsequent intent classification and correction. The hierarchical attention mechanism combined with dynamic weight adjustment of user context can accurately screen out contextual content related to the target intent, avoid interference of irrelevant information on the intent correction process, and thus improve the system's adaptability to complex contexts.

[0014] 2. Through gating operations based on multi-round interaction history, the contextual content related to the target intent is processed, and semantic fragments with reference value for intent correction are retained. At the same time, the multi-task learning structure of the deep neural network is used to perform parallel analysis on the reference semantic fragments, classify and process different intents, and calculate the scoring results, thereby realizing the dynamic fusion and correction operations of multiple intents. By outputting clear and structured intent information, it can effectively distinguish each intent, improve the system's processing accuracy for parallel expressions of multiple intents, significantly optimize the interactive experience of elderly people's conversations, and provide reliable support for intelligent decision-making in scenarios such as health consultation and family arrangements. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a schematic diagram of a method for correcting elderly people's conversation intention information based on a deep neural network according to the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] Example: Figure 1 A method for correcting elderly people's conversation intention information based on a deep neural network is provided, which includes the following steps: S1: Collect and clean multi-round dialogue data containing multiple topics, annotate the multi-round dialogue data and unify the format.

[0018] S2: Use the multi-intent recognition sub-model to extract topic features from the annotated multi-round dialogue data and generate a multi-intent representation vector that can represent each intent.

[0019] S3: A hierarchical attention mechanism combined with dynamic weight adjustment of user context is used to perform association calculations on multi-intent representation vectors to filter out contextual content related to the target intent.

[0020] S4: The contextual content related to the target intent is processed through gating operations based on multi-round interaction history, and reference semantic fragments with reference value for intent correction are retained.

[0021] S5: The multi-task learning structure based on deep neural network analyzes the reference semantic fragments in parallel, classifies different intents separately and calculates the score of each intent.

[0022] S6: Perform multi-intent fusion correction operation based on the results of classification processing and intent scoring, and output intent information that can distinguish each intent.

[0023] Collect and clean multi-round dialogue data containing multiple topics, annotate multi-round dialogue data and unify the format, including: Obtain multi-turn conversation data from multi-channel data sources: Acquire multi-round conversation data from multiple data sources, including social media platforms, health consultation systems, home voice assistant devices, and elderly care systems. The selection of data sources should cover the main topics that the elderly often deal with, such as health, family, and daily life, to ensure the diversity and coverage of the data. The acquired data includes text conversations, speech-to-text records, and their corresponding conversation metadata, such as conversation time, participant roles, and conversation scenarios.

[0024] Clean the multi-round conversation data to remove noise data, incomplete data, and meaningless conversation records: Remove noisy data: Filter texts with incomplete semantics, a large number of spelling errors, or no practical meaning, such as garbled characters and meaningless records consisting of a single letter, through rules or machine learning models; Eliminate incomplete data: Delete conversation records that are missing key information (such as conversation turns, conversation content, timestamps, etc.) or are insufficient in length; Delete meaningless conversation records: Identify and remove common sentence patterns that have no actual semantics, such as irrelevant conversations consisting of phrases such as "um" and "okay".

[0025] The cleaned multi-turn dialogue data is annotated using predefined annotation rules. The annotation rules include annotation of topic categories, intent categories, and dialogue turn information: The annotation rules need to be defined in advance and formulated by domain experts. The annotation includes the following: Topic category labeling: label topics such as health, family, and daily life according to the content of the conversation; Intent category annotation: classify the explicit or implicit intent in each round of conversation, such as asking about health conditions or discussing family activities; Dialogue turn information annotation: Add a turn number to each dialogue round to ensure that subsequent steps can accurately parse the timing relationship of the dialogue.

[0026] The annotation work can be completed by combining manual annotation and automatic annotation. The automatic annotation part relies on semantic analysis algorithms to make preliminary classifications of common intentions, while manual annotation performs fine-grained processing on special or complex corpora.

[0027] The annotated multi-round dialogue data is formatted in a unified manner. The format unification includes the normalization of data fields, the unified representation of time series, and the normalization of the order of dialogue rounds: Normalization of data fields: Ensure that the fields of each record are complete and structured, such as the unified definition and naming of the conversation turn field, topic category field, and intent category field; Unified representation of time series: standardize the timestamp format in all conversation records, for example, convert the "year-month-day-hour-minute" format to ISO standard time representation; Normalize the sorting of conversation turns: Sort multiple conversation records according to the chronological order of the conversation to ensure that the data can accurately reflect the natural order and contextual relationship of the conversation.

[0028] The multi-intent recognition sub-model is used to extract topic features from the annotated multi-round dialogue data to generate a multi-intent representation vector that can represent each intent, including: By inputting the pre-trained multi-intention recognition sub-model into the annotated multi-round dialogue data, the text content of each round of dialogue is analyzed and the semantic information related to the topic features is extracted: Based on the obtained annotated multi-turn dialogue corpus, it is imported into the pre-trained multi-intent recognition sub-model as input data. The multi-turn dialogue corpus contains text content such as clearly annotated topic categories, intent categories, and turn information.

[0029] The multi-intent recognition sub-model uses semantic parsing algorithms to perform word segmentation, grammatical analysis, and semantic association processing on the text to ensure accurate extraction of implicit semantic information in each round of dialogue corpus. The text content parsing steps include: parsing each round of dialogue in the corpus separately to retain the semantic integrity of the round information; classifying the parsed results according to topic categories and intent categories to provide support for subsequent module processing; and recording the text semantic connections between rounds for multi-round context alignment to ensure the relevance of multi-intent semantics.

[0030] In the multi-intent recognition sub-model, the semantic analysis module is used to extract features from the text content of each round of dialogue. The extracted features include semantic features of the topic category and semantic features of the intent category: The topic classifier in the semantic analysis module is used to classify the text content of each round of conversation according to the predefined topic classification system (such as health, family, and daily life), and extract the feature information of each topic category. The extraction process is based on the training results of the word embedding model, and the semantic features of keywords or phrases related to the topic category are recorded.

[0031] After the topic feature extraction is completed, the text intent category of each round of dialogue is classified through the intent classifier. The classification system of intent categories includes explicit intent (such as asking questions directly) and implicit intent (such as expressing emotions, asking questions in return, etc.). The extraction of semantic features is based on the deep semantic matching algorithm, which extracts important semantic features by calculating the correlation weights between text fragments and each intent category.

[0032] The extracted semantic features include the semantic representation of word vectors, context-related time series information, and the correspondence between topic categories and intent categories.

[0033] Use the embedding vector generation module to embed the extracted semantic features and generate initial vectors that can represent topic categories and intent categories: The embedding model is used to process the semantic features of the topic category and generate the embedding vector of the topic category. The embedding model is based on the pre-trained word embedding algorithm and generates a fixed-length high-dimensional vector based on the input semantic features. The semantic features of the intent category are embedded in the same way to generate the embedding vector of the intent category. The time series information of the conversation context is taken into account during the processing to maintain the coherence of the semantic representation.

[0034] The embedding vector formula is as follows: ;in, The comprehensive embedding vector is a fusion representation obtained by weighted calculation of the semantic feature vectors of the topic category and the intent category, which is used for subsequent multi-intent analysis; is the weight matrix of the semantic features of the topic category, which indicates the weight ratio of the topic category features when calculating the comprehensive embedding vector, and is dynamically adjusted by the multi-intent recognition sub-model during the training process; It is the vector representation of the semantic features of the topic category, which means the high-dimensional vector obtained after the semantic features corresponding to the topic category are processed by the embedding vector generation module, and is used to reflect the topic information involved in the conversation; is the weight matrix of the semantic features of the intent category, which indicates the importance weight of the intent category features in the calculation of the comprehensive embedding vector. The model will dynamically adjust the value of this matrix based on the training data. It is a vector representation of the semantic features of the intent category, which is converted into a high-dimensional representation through an embedding vector generation module and is used to characterize the semantic features of the intent category in the conversation.

[0035] The weight matrix in the embedding vector formula is dynamically updated by the semantic matching model through the gradient descent method during the training process to ensure that the weights can reflect the importance of different semantic features.

[0036] The feature fusion module of the multi-intention recognition sub-model is used to fuse the initial vectors generated in each round of dialogue to obtain a multi-intention representation vector that can comprehensively represent multiple intent categories and topic categories: Use the weighted formula to perform weighted summation on the embedding vectors of the topic category and intent category: ;in, Represents the final generated multi-intent representation vector; and are the fusion weight parameters of topic category and intent category, which are automatically adjusted by the feature fusion module; It is the global semantic feature vector of the multi-round dialogue context, calculated by the semantic alignment module.

[0037] The multi-intention representation vector after fusion calculation is normalized to ensure that the distribution range of the final vector in the high-dimensional space meets the input requirements of subsequent processing.

[0038] It is worth noting that the multi-intention recognition sub-model consists of a semantic analysis module, an embedding vector generation module, and a feature fusion module. Each module has a clear division of labor and works in coordination. The semantic analysis module is responsible for parsing the input annotated multi-round dialogue corpus, which includes a word segmentation unit, a grammatical parsing unit, and a semantic matching unit. It performs word segmentation, grammatical structure analysis, and semantic feature extraction on the text content in turn. The embedding vector generation module vectorizes the features extracted by the semantic analysis module, which includes a word embedding sub-unit, a semantic embedding sub-unit, and a feature combination sub-unit. It generates high-dimensional vector representations of topic categories and intent categories through a pre-trained model. The feature fusion module further performs fusion calculations on the embedding vectors, which includes a weighted fusion sub-unit, a normalization sub-unit, and a context alignment sub-unit. It performs weighted calculations on the embedding vectors of topic categories and intent categories through predefined weights, and performs semantic alignment in combination with the context information of multi-round dialogues. Finally, it generates a comprehensive representation vector that can represent multiple intent categories and multiple topic categories. Through orderly collaboration between modules, the multi-intention recognition sub-model gradually converts the annotated multi-round dialogue corpus into a multi-intention representation vector that can be used in subsequent steps.

[0039] A hierarchical attention mechanism combined with dynamic weight adjustment of user context is used to perform association calculation on multiple intent representation vectors to filter context content related to the target intent, including: The multi-intent representation vectors are layered according to the predefined attention levels, and each layer assigns initial weights to different dimensional features in the user context information: The initial weights are allocated according to the dimensional features of the contextual information. Each dimensional feature includes topic category information, intent category information, and time series features related to the conversation context. In the hierarchical processing process, the attention level of each layer is defined as an independent computing unit. Each layer receives the input of multiple intent representation vectors and performs feature screening for specific dimensional features.

[0040] Through this hierarchical processing, the multi-intent representation vector is decomposed into multiple sub-vectors with specific dimensional features, which serve as inputs for subsequent weight adjustment and weighted calculation.

[0041] According to the historical conversation content and semantic features of the current user context information, the weight distribution of each layer of attention is dynamically adjusted: First, the semantic features in the user's historical conversation content are extracted. These features include topic continuity, context semantic relevance, and potential contextual relationships of the target intent. The initial weight of each layer is dynamically adjusted based on the matching degree between the semantic features in the current user context information and the multi-intent representation vector. The specific formula for dynamic adjustment is as follows: ;in, Indicates the adjusted Layer attention weights, Indicates The initial weights of the layers, Represents user context information and The similarity of the multi-intention representation vector input of the layer, Represents user context information, Indicates The multi-intent representation vector input of the layer.

[0042] in, The calculation of is based on the cosine distance or dot product operation between the user semantic feature and a specific dimension of the representation vector, and the result is used as a dynamically adjusted scaling factor.

[0043] In each layer of the attention mechanism, the multi-intent representation vectors are weighted according to the adjusted weight distribution to obtain the weight distribution of different intent categories in each layer: Use the adjusted attention weights to weight the multi-intent representation vector of each layer input. The calculation formula is as follows: ;in, Indicates Feature output after layer weighted calculation; The weighted result retains the contextual content related to the target intent in each layer, and achieves hierarchical accumulation of semantic information through layer-by-layer weighting.

[0044] Through the weighted results of the multi-layer attention mechanism and the contextual information related to the target intent, the semantic fragments with the strongest correlation with the current target intent are screened out: The weight distribution of weighted results at each level is integrated, and the context information with the highest weight value is used as the candidate semantic segment with the strongest correlation with the target intent; the final matching value between the target intent and the candidate semantic segment is calculated, and the screening formula is as follows: ;in, represents the semantic fragments finally filtered out, Represent the output feature values ​​of the multi-layer attention weighted results, Indicates the number of layers of the attention mechanism.

[0045] Through the above filtering operation, it is ensured that the semantic relevance between the filtered semantic fragments and the current target intent is the strongest.

[0046] The semantic fragments that are most relevant to the current target intent are processed uniformly to generate a contextual content representation that can represent the target intent: The filtered semantic fragments are re-encoded to retain the semantic context; the encoding results are stored as a structured representation for input in subsequent steps.

[0047] Through the gating operation based on the multi-round interaction history, the context content related to the target intent is processed, and the reference semantic fragments with reference value for intent correction are retained, including: Extract the user's multi-round interaction history information from the context content related to the target intent, including the conversation turn, context semantic fragments, and time series features associated with the target intent: Conversation turn information: records the sequence number of each conversation round of the user to ensure that subsequent analysis can be processed according to the natural context; Contextual semantic fragments: Extract semantic fragments that are directly or indirectly related to the target intent, such as extracting "medication time" and "physical symptoms" in health consultation scenarios; Time series features: record the time information of each round of dialogue, which is used for subsequent analysis of the temporal relationship of semantic fragments.

[0048] The extracted multi-round interaction history information is organized in a structured format to ensure that each piece of information contains a complete round number, semantic fragment, and timestamp, providing basic data for subsequent feature screening and gating operations.

[0049] The extracted historical information of multiple rounds of user interactions is subjected to feature screening, and an initial weight is assigned to each piece of historical information according to its relevance to the target intent. The historical information whose relevance to the target intent is lower than the preset threshold is eliminated: According to the characteristics of the target intent, the semantic relevance between the historical information of multiple rounds of interactions and the target intent is analyzed. For example, in the "family arrangement" scenario, if the target intent is "determine the activity time", the historical information containing keywords such as time and activity name is prioritized.

[0050] An initial weight is assigned to each piece of historical information, and the weight is assigned based on the degree of semantic matching. For example, historical information containing target intent keywords or semantic features is assigned a higher weight; content whose semantic fragments have no obvious correlation with the target intent is assigned a lower weight.

[0051] Eliminate historical information with weights below the preset threshold, and retain contextual content that is more closely related to the target intent. For example, for the target intent "health consultation", retain "recent medication records" and "symptom descriptions", and exclude irrelevant content such as "weather inquiries".

[0052] On the basis of assigning the initial weights, the contextual content related to the target intent is screened through the gating operation, which includes dynamically adjusting the gating threshold according to the weight of each piece of historical information to screen out the contextual content that meets the gating conditions: Set the initial gating conditions and dynamically adjust the gating threshold according to the weight value of historical information. For example, if the target intent is clear and highly relevant (such as "determine the specific time"), the gating threshold is set higher to retain only the most relevant contextual content; if the target intent is vague (such as "health status consultation"), the gating threshold is set lower to retain more potentially relevant information.

[0053] Apply gating conditions to filter contextual content and only retain information that meets the gating conditions. For example, when filtering semantic fragments of the user's health status, only contextual content that meets the target intent, such as "blood pressure records" and "medication status", is retained, and other information that does not meet the conditions is eliminated.

[0054] The filtered context content is organized into candidate semantic fragments, each of which contains complete turn information, time series features, and semantic features.

[0055] The context content filtered out by the gating operation is normalized, including time series sorting and content structuring of the filtered context semantic fragments, and the normalized context content is stored as a reference semantic fragment: Sort the selected semantic fragments in chronological order to ensure the temporal logical coherence of the content.

[0056] The semantic fragments are associated with their corresponding timestamps, turn numbers, and contextual semantic features to generate structured data representations. For example, the "symptom description" is stored together with its corresponding conversation time and turn information to form standardized data that is easy to process later.

[0057] The normalized context content is stored as a reference semantic fragment for subsequent intent correction steps. For example, in a health consultation scenario, the reference semantic fragment may include "symptom changes described by the user" and "medication recommendations given by the doctor".

[0058] The multi-task learning structure based on deep neural network analyzes the reference semantic fragments in parallel, classifies different intents and calculates the score of each intent, including: Preprocess the reference semantic fragments according to their contextual relevance and time series features: Receive the reference semantic segments filtered and normalized by the previous stage, extract the contextual relevance and time series features of each semantic segment, and ensure that the preprocessed data contains the following: Contextual relevance: records the semantic relevance between the reference semantic segment and the target intent. For example, the segments related to health intent contain keywords such as “taking medicine” and “symptoms”. Time series features: Timestamps are marked according to the chronological order of occurrence of semantic fragments to preserve time continuity in subsequent analysis.

[0059] The extracted contextual relevance information and time series features are bound to the reference semantic fragments to generate structured input data. Each record of the input data should contain the text content of the semantic fragment, semantic relevance information, timestamp, and target intent context label.

[0060] The generated input data is normalized to ensure that the time series features and contextual relevance of different semantic fragments are in a unified numerical range so that they can be received and processed by the deep neural network.

[0061] Construct a multi-task learning structure of a deep neural network, which includes intent classification and intent scoring tasks. The intent classification task is used to identify the intent category corresponding to the reference semantic fragment, and the intent scoring task is used to calculate the weight score of each intent: Intent classification task module: used to identify the intent category corresponding to the reference semantic fragment, for example, the health consultation fragment is classified as "symptom description" or "treatment advice"; Intent scoring task module: used to calculate the score of each intent category. The scoring result reflects the matching degree and importance between the semantic fragment and the target intent.

[0062] The input layer of the network structure receives the preprocessed reference semantic fragment data, including the semantic fragment content, contextual relevance information and time series features. Each layer of the network processes different features of the input data.

[0063] In the multi-task learning structure, the preprocessed reference semantic fragments are processed in parallel, and the intent classification task classifies each semantic fragment to determine the intent category to which it belongs: In the intent classification task module, the semantic fragment content is classified through a deep neural network, and the intent category of the fragment is output based on the contextual relevance and time series features. The classification operation determines the semantic category based on the keywords and contextual information in the fragment, for example, classifying "feeling dizzy recently" as "symptom description".

[0064] In the intent scoring task module, the input reference semantic fragments are scored. The scoring criteria include the semantic relevance between the semantic fragment and the target intent, the temporal correlation, and the importance of the fragment in multiple rounds of dialogue.

[0065] To ensure the synchronization of classification results and scoring calculations, network weights are shared between task modules. For example, the same feature extraction layer can be reused for the semantic feature extraction part, reducing computational overhead and improving processing efficiency.

[0066] In the intent scoring task, each intent category is scored based on the semantic features and contextual relationships of the reference semantic fragment. The scoring is based on the degree of association between the semantic fragment and the target intent and the time series characteristics: The intent score is calculated based on multi-dimensional features, including: Semantic relevance score: the degree of match between the content of the reference semantic segment and the target intention semantic template; Temporal relevance score: the temporal proximity between the time of the segment occurrence and the time of execution of the target intention; Context match score: How well the snippet content corresponds to the target intent context label.

[0067] In the intent scoring task, each intent category is scored by analyzing the semantic features and contextual relationships of the reference semantic fragments. The scoring specifically includes the following calculation process: First, the association score between the semantic fragment and the target intent is calculated based on the semantic matching algorithm. This score measures the degree of match between the semantic features of the fragment and the target intent template; second, the time correlation score is calculated based on the time series features to measure the proximity between the occurrence time of the fragment and the execution time of the target intent; finally, the context matching score is calculated in combination with the context information to evaluate the semantic consistency between the fragment and the multi-round dialogue context. Finally, the scores of each dimension are weighted and summed according to the predefined weight coefficient to generate the scoring result for each intent category.

[0068] The multi-dimensional scoring results are weighted and integrated according to the predefined weight coefficients, and the importance of each dimension can be dynamically adjusted according to the actual needs of the target intent. For example, in a health consultation scenario, time relevance may be more important than context matching.

[0069] Integrate the results of intent classification and intent scoring to generate the classification label and scoring value for each reference semantic fragment: Each reference semantic fragment is assigned a classification label, which clearly states the intent category to which the fragment belongs; the score corresponding to each fragment reflects its importance and relevance to the target intent.

[0070] The integrated results are stored as structured output data for subsequent intent modification and decision processing steps. For example, in a health consultation application, the integrated results may include snippets and their scores in the "symptom description" category, which are used to prioritize high-scoring content for recommendation to doctors.

[0071] Based on the results of classification and intent scoring, a multi-intent fusion correction operation is performed to output intent information that can distinguish each intent, including: Receive the classification labels and scoring values ​​generated by the classification processing and intent scoring results, aggregate the classification labels according to the intent categories, and initially form a classification set of intent categories: Obtain the classification processing results and intent scoring results of the reference semantic fragments, including the classification label and scoring value corresponding to each semantic fragment; aggregate the classification labels according to the intent category, and classify the semantic fragments of the same category into a set, for example, store the fragments of the "health consultation" category in the health category set; retain the scoring value of each semantic fragment during the aggregation process to ensure that subsequent sorting can be completed based on the scoring information.

[0072] Based on the intent categories and their ratings in the classification set, sort the intent categories in descending order of ratings, and select the intent category with the highest rating as the primary intent: The intent categories in the classification set are sorted from high to low according to the score value, which indicates the degree of association between the semantic fragment and the target intent; the intent category with the highest score is selected as the main intent, which reflects the core expression needs of the user in the current context. For example, in the "medication consultation" scenario, the highest-scoring intent category may be "drug taking time"; other intent categories with lower scores but still have a certain relevance are marked as secondary intents for subsequent supplementary processing.

[0073] The semantic supplementation processing is performed on the intent categories with lower scores but certain relevance, and they are integrated with the main intent to form a comprehensive correction result of multiple intents: For the secondary intent category, analyze its semantic relevance to the primary intent. For example, in the health consultation scenario, use "symptom description" as a semantic supplement to the primary intent of "treatment advice"; merge the supplemented semantic fragments with the primary intent category to form a comprehensive correction result of multiple intents, ensuring that the corrected content contains both the information of the primary intent and the contextual supplementary information related to it; retain the key information of the secondary intent during the supplementation process to facilitate the fusion and storage of multiple intent content in subsequent steps.

[0074] The semantic information of the main intent and related intents is fused through the multi-intention fusion algorithm to generate intent information that can distinguish each intent: Based on the multi-intention fusion algorithm, the semantic information of the primary intention and the secondary intention is deeply integrated; the following factors are considered in the fusion process: Semantic consistency: Ensure that the primary intent and secondary intent are consistent in content expression. For example, in health consultation, ensure that the information of "medication advice" and "medication time" is logically coherent; Contextual relevance: Merge contextually relevant content from primary and secondary intents, such as aligning temporal information with semantic content.

[0075] After the fusion is completed, comprehensive semantic information is generated that can distinguish each intent, ensuring that the multi-intent information is independent and complementary to each other.

[0076] The fused and corrected intent information is stored in a structured manner, and a clear classification result that can distinguish between the main intent and related intent is output: The results of multi-intent fusion correction are stored in a structured manner, including: classification labels, semantic content and scoring information of each intent category; the output intent information can clearly distinguish between primary intent and secondary intent. For example, in the user's dialogue request, "medication time" is marked as the primary intent, and "drug name" and "medication frequency" are marked as secondary intents; the structured stored intent information will serve as input data for subsequent decision analysis or interactive response.

[0077] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0078] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or may be transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium may be a solid-state hard disk.

[0079] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0080] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0081] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0082] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0083] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0084] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0085] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0086] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for correcting elderly people's conversation intention information based on deep neural network, characterized in that: The steps include: S1: Collect and clean multi-round dialogue data containing multiple topics, annotate the multi-round dialogue data and unify the format; S2: Use the multi-intent recognition sub-model to extract topic features from the annotated multi-round dialogue data and generate a multi-intent representation vector that can represent each intent; S3: A hierarchical attention mechanism combined with dynamic weight adjustment of user context is used to perform association calculation on multiple intent representation vectors to filter out contextual content related to the target intent. S4: Process the contextual content related to the target intent through gating operations based on multi-round interaction history, and retain reference semantic fragments that are of reference value for intent correction; S5: Based on the multi-task learning structure of deep neural network, the reference semantic fragments are analyzed in parallel, different intents are classified and processed respectively, and the score of each intent is calculated; S6: Perform multi-intent fusion correction operation based on the results of classification processing and intent scoring, and output intent information that can distinguish each intent.

2. The method for correcting elderly people's conversation intention information based on deep neural network according to claim 1 is characterized in that: Collect and clean multi-round dialogue data containing multiple topics, annotate multi-round dialogue data and unify the format, including: Obtain multi-round conversation data from multi-channel data sources; Clean the multi-round conversation data, including removing noise data, incomplete data, and meaningless conversation records; Use predefined annotation rules to annotate the cleaned multi-turn dialogue data. The annotation rules include annotation of topic categories, intent categories, and dialogue turn information. The labeled multi-round dialogue data is formatted in a unified manner, including the normalization of data fields, the unified representation of time series, and the normalization of the order of dialogue rounds.

3. The method for correcting elderly people's conversation intention information based on deep neural network according to claim 1 is characterized in that: The multi-intent recognition sub-model is used to extract topic features from the annotated multi-round dialogue data to generate a multi-intent representation vector that can represent each intent, including: By inputting the pre-trained multi-intent recognition sub-model into the annotated multi-round dialogue data, the text content of each round of dialogue is analyzed and the semantic information related to the topic features is extracted; In the multi-intent recognition sub-model, the semantic analysis module is used to extract features from the text content of each round of dialogue. The extracted features include semantic features of topic categories and semantic features of intent categories. Use the embedding vector generation module to embed the extracted semantic features and generate initial vectors that can represent topic categories and intent categories; The feature fusion module of the multi-intention recognition sub-model is used to fuse the initial vectors generated in each round of dialogue to obtain a multi-intention representation vector that can comprehensively represent multiple intent categories and topic categories.

4. The method for correcting elderly people's conversation intention information based on deep neural network according to claim 1 is characterized in that: A hierarchical attention mechanism combined with dynamic weight adjustment of user context is used to perform association calculation on multiple intent representation vectors to filter context content related to the target intent, including: The multi-intent representation vectors are processed in layers according to the predefined attention levels, and each layer assigns initial weights to different dimensional features in the user context information; Dynamically adjust the weight distribution of attention at each layer based on the historical conversation content and semantic features of the current user context information; In each layer of the attention mechanism, the multi-intent representation vectors are weighted according to the adjusted weight distribution to obtain the weight distribution of different intent categories in each layer; Through the weighted results of the multi-layer attention mechanism and the contextual information related to the target intent, the semantic fragments with the strongest correlation with the current target intent are screened out; The semantic fragments that are most relevant to the current target intent are processed uniformly to generate a contextual content representation that can characterize the target intent.

5. The method for correcting elderly people's conversation intention information based on deep neural network according to claim 4 is characterized in that: The specific formula for dynamically adjusting the weight distribution of attention at each layer is as follows: ;in, Indicates the adjusted Layer attention weights, Indicates The initial weights of the layers, Represents user context information and The similarity of the multi-intention representation vector input of the layer, Represents user context information, Indicates The multi-intent representation vector input of the layer.

6. The method for correcting elderly people's conversation intention information based on deep neural network according to claim 1 is characterized in that: Through the gating operation based on the multi-round interaction history, the context content related to the target intent is processed, and the reference semantic fragments with reference value for intent correction are retained, including: Extract user multi-round interaction history information from the context content related to the target intent, including conversation turns, context semantic fragments, and time series features associated with the target intent; Perform feature screening on the extracted historical information of multiple rounds of user interactions, assign an initial weight to each piece of historical information based on its relevance to the target intent, and remove historical information whose relevance to the target intent is lower than a preset threshold; On the basis of assigning the initial weights, the contextual content related to the target intent is screened through the gating operation, which specifically includes dynamically adjusting the gating threshold according to the weight of each piece of historical information to screen out the contextual content that meets the gating conditions; The context content filtered out by the gating operation is normalized, including time series sorting and content structuring of the filtered context semantic fragments, and the normalized context content is stored as a reference semantic fragment.

7. The method for correcting elderly people's conversation intention information based on deep neural network according to claim 1 is characterized in that: The multi-task learning structure based on deep neural network analyzes the reference semantic fragments in parallel, classifies different intents and calculates the score of each intent, including: Preprocess the reference semantic fragments according to their contextual relevance and time series characteristics; Construct a multi-task learning structure of a deep neural network, which includes intent classification and intent scoring tasks. The intent classification task is used to identify the intent category corresponding to the reference semantic fragment, and the intent scoring task is used to calculate the weight score of each intent. In the multi-task learning structure, the pre-processed reference semantic fragments are processed in parallel, and the intent classification task classifies each semantic fragment to determine the intent category to which it belongs; In the intent scoring task, each intent category is scored based on the semantic features and contextual relationships of the reference semantic fragments. The scoring criteria include the degree of association between the semantic fragments and the target intent and the time series characteristics. The results of intent classification and intent scoring are integrated to generate the classification label and scoring value for each reference semantic fragment.

8. The method for correcting elderly people's conversation intention information based on deep neural network according to claim 1 is characterized in that: Based on the results of classification and intent scoring, a multi-intent fusion correction operation is performed to output intent information that can distinguish each intent, including: Receive classification labels and scoring values ​​generated by classification processing and intent scoring results, aggregate the classification labels according to intent categories, and initially form a classification set of intent categories; Based on the intent categories and their ratings in the classification set, sort the intent categories in descending order of ratings, and select the intent category with the highest rating as the primary intent; Perform semantic supplementation on the intent categories with lower scores but certain relevance, and merge them with the main intent to form a comprehensive correction result of multiple intents; The semantic information of the main intent and related intent is fused through the multi-intention fusion algorithm to generate intent information that can distinguish each intent; The fused and corrected intent information is stored in a structured manner, and a clear classification result that can distinguish between the main intent and related intent is output.

Images