An information retrieval system and method based on semantic normalization

Abstract

The application discloses an information retrieval system and method based on semantic normalization, and relates to the field of artificial intelligence information technology, which comprises collecting semantic query records input by a user and performing preliminary semantic analysis to generate structured data; based on the structured data, entity disambiguation is performed by using a knowledge graph, an abstract class is generated by a neural network, calibration and dynamic weight adjustment are performed, and an entity abstract class with high confidence and a confidence score are generated; the entity abstract class and the confidence score are combined with a user context, an action-value function is calculated through a value network, and an optimal action is selected by using a greedy algorithm; semantic normalization mapping is performed according to the optimal action, an intermediate expression is obtained by using a meta-symbol dynamic generator; index retrieval and multi-dimensional sorting are performed based on the intermediate expression, and a sorted retrieval result list is generated. The application solves the problem of semantic fragmentation of multi-modal queries, and realizes deep semantic alignment and dynamic weight calibration of heterogeneous data.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence information technology, and in particular to an information retrieval system and method based on semantic normalization. Background Technology

[0002] Since the mid-20th century, information retrieval technology has evolved from keyword-based Boolean models to semantically aware intelligent systems. Early methods, such as Vector Space Models (VSM) and Latent Semantic Indexing (LSI), relied on statistical features and linear algebraic transformations but failed to capture deep semantic relationships. With breakthroughs in Natural Language Processing (NLP), machine learning techniques such as Support Vector Machines (SVM) and Random Forests were introduced for query classification and ranking optimization. In recent years, the deep learning revolution has brought about the Transformer architecture and pre-trained language models (such as BERT and GPT). These models achieve context-aware semantic representations through self-attention mechanisms, improving the accuracy of query understanding and document matching. Simultaneously, the integration of knowledge graphs provides structured knowledge support, entity disambiguation, and relational reasoning. Multimodal retrieval has also become a cutting-edge direction, processing text, speech, and image inputs through cross-modal neural networks (such as CLIP and ViLBERT) to achieve heterogeneous data alignment.

[0003] Despite advancements in existing technologies, shortcomings remain in two key areas. First, in multimodal query processing, insufficient semantic fusion between modalities leads to biased intent understanding. Traditional systems process text, speech, and image queries through independent pipelines and then simply concatenate the results, ignoring intermodal interaction effects and contextual dependencies. Second, limited adaptability prevents real-time responses to user feedback. Existing RL methods are mostly based on offline training and static strategies, resulting in long model update cycles and a lack of dynamic adjustment in the explore-exploitation balancing mechanism, leading to performance degradation in changing environments. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an information retrieval method based on semantic normalization to solve the problems of insufficient modal interaction and limited real-time response capability in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an information retrieval method based on semantic normalization, comprising: collecting semantic query records input by the user and performing preliminary semantic parsing to generate structured data; based on the structured data, using a knowledge graph for entity disambiguation, generating abstract classes through a neural network, and performing calibration and dynamic weight adjustment to generate high-confidence entity abstract classes and confidence scores; combining the entity abstract classes and confidence scores with the user context, calculating the action-value function through a value network, and utilizing... - A greedy algorithm selects the optimal action; semantic normalization mapping is performed based on the optimal action, and a meta-symbol dynamic generator is used to obtain intermediate representations; index retrieval and multi-dimensional sorting are performed based on the intermediate representations to generate a sorted list of search results; user interaction with the search results list is monitored, user feedback is collected, and the parameters of the value network are updated.

[0008] As a preferred embodiment of the information retrieval method based on semantic normalization described in this invention, the semantic query records include text query records, voice query records, and image query records.

[0009] As a preferred embodiment of the semantic normalization-based information retrieval method of the present invention, the specific steps for generating structured data are as follows:

[0010] Collect semantic queries input by users via text, voice, and images, perform multimodal preprocessing, and generate a uniformly encoded query string;

[0011] Perform multi-dimensional semantic parsing on the query string, obtain a candidate parsing set, and calculate the semantic consistency index;

[0012] Uncertainty assessment is performed on the candidate parsing set and semantic consistency index, and a decision dataset is generated.

[0013] Entities, attributes, and relationships are extracted from the decision dataset, and structured data is generated through triple assembly and standardized format conversion.

[0014] As a preferred embodiment of the semantic normalization-based information retrieval method of the present invention, the steps of using structured data, knowledge graphs for entity disambiguation, and neural networks to generate abstract classes are as follows.

[0015] Multi-source knowledge graph fusion is performed on structured data to generate entity feature vectors;

[0016] Entity disambiguation is performed on the entity feature vectors, and the semantic correlation between entities is obtained by using a multi-hop attention mechanism. Based on the semantic correlation, clustering and filtering are performed to generate a disambiguated entity set.

[0017] Based on the disambiguated entity set, an abstract class is generated using a reinforcement learning framework, and the optimal abstract path is selected through a policy network to calculate the initial confidence level.

[0018] As a preferred embodiment of the semantic normalization-based information retrieval method of the present invention, the specific steps for performing calibration and dynamic weight adjustment to generate high-confidence entity abstract classes and confidence scores are as follows.

[0019] Based on the abstract class and the initial confidence level, a cross-modal attention mechanism is used to calibrate and dynamically adjust the weights to generate a calibrated confidence score.

[0020] Based on the calibrated confidence scores, the boundaries of abstract classes are dynamically adjusted through action space exploration and reward mechanisms to generate high-confidence entity abstract classes and confidence scores.

[0021] As a preferred embodiment of the semantic normalization-based information retrieval method described in this invention, the method combines entity abstract classes and confidence scores with user context, calculates the action-value function through a value network, and utilizes... The greedy algorithm selects the optimal action, and the specific steps are as follows.

[0022] Based on entity abstract classes and confidence scores, combined with user context, the numerical range is normalized using the min-max scaling method, and one-hot encoding is used to process categorical variables to generate a structured feature matrix.

[0023] The structured feature matrix is ​​subjected to feature fusion and dimensionality compression to generate a low-dimensional state representation vector with context awareness.

[0024] Based on the state representation vector, the action-value function is calculated through a value network;

[0025] use Greedy algorithms select the optimal action from the action-value function.

[0026] As a preferred embodiment of the semantic normalization-based information retrieval method of the present invention, the specific steps for obtaining the intermediate representation are as follows:

[0027] Perform in-depth analysis and type classification of the optimal action to generate executable mapping configuration parameters;

[0028] Based on executable mapping configuration parameters, load the corresponding semantic mapping rules and generate dynamic mapping rule objects;

[0029] Perform semantic normalization mapping on the dynamic mapping rule object to generate a preliminary mapping semantic structure;

[0030] The initial mapping semantic structure is dynamically enhanced and its consistency is verified by a metasymbol dynamic generator to generate an intermediate representation.

[0031] As a preferred embodiment of the semantic normalization-based information retrieval method of the present invention, the specific steps for generating the sorted retrieval result list are as follows:

[0032] Based on standardized semantic representation, construct a multi-source index query task;

[0033] Perform parallel index retrieval on the index query task to obtain a candidate result set;

[0034] The candidate result set is sorted in multiple dimensions using a multi-dimensional scoring function and a dynamic weight adjustment mechanism to generate multi-dimensional sorted data.

[0035] Dynamic weight optimization and diversity re-sorting are performed on multi-dimensional sorted data to generate a sorted list of search results.

[0036] As a preferred embodiment of the semantic normalization-based information retrieval method of the present invention, the specific steps of monitoring user interaction with the search results list, collecting user feedback, and updating the parameters of the value network are as follows:

[0037] Real-time monitoring of user interaction with the sorted search results list, generating interaction event data;

[0038] Based on interaction event data, multimodal user feedback is collected and a feedback dataset is generated;

[0039] Based on the feedback dataset, the parameters of the value network are updated using a reinforcement learning algorithm;

[0040] The interaction event data includes click events, dwell time, and scroll depth.

[0041] Secondly, this invention provides an information retrieval system based on semantic normalization, comprising a semantic parsing module, a knowledge enhancement module, an action selection module, a semantic generation module, a retrieval ranking module, and a learning optimization module. The semantic parsing module collects user-input semantic query records and performs preliminary semantic parsing to generate structured data. The knowledge enhancement module, based on the structured data, uses a knowledge graph for entity disambiguation, generates abstract classes through a neural network, and performs calibration and dynamic weight adjustment to generate high-confidence entity abstract classes and confidence scores. The action selection module combines the entity abstract classes and confidence scores with the user context, calculates the action-value function through a value network, and utilizes… - A greedy algorithm selects the optimal action; a semantic generation module performs semantic normalization mapping based on the optimal action and uses a dynamic meta-symbol generator to obtain intermediate representations; a retrieval and ranking module performs index retrieval and multi-dimensional ranking based on the intermediate representations to generate a ranked list of retrieval results; and a learning and optimization module monitors user interaction with the retrieval results list, collects user feedback, and updates the parameters of the value network.

[0042] The beneficial effects of this invention are as follows: by using multi-source knowledge graph fusion and cross-modal attention mechanisms, it solves the semantic fragmentation problem of multimodal queries, and achieves deep semantic alignment and dynamic weight calibration of heterogeneous data; it generates multimodal feedback datasets in real time based on user interaction behavior, and dynamically updates the parameters of the value network through online reinforcement learning algorithms, thus constructing a real-time feedback loop from user interaction to model update and achieving continuous self-optimization of parameters. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of an information retrieval method based on semantic normalization.

[0045] Figure 2 A flowchart for generating structured data.

[0046] Figure 3 A flowchart for generating high-confidence entity abstract classes and confidence scores.

[0047] Figure 4 A flowchart for selecting the optimal action. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0050] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0051] Reference Figures 1-4 This is one embodiment of the present invention, which provides an information retrieval method based on semantic normalization, including the following steps:

[0052] S1. Collect the semantic query records input by the user and perform preliminary semantic parsing to generate structured data.

[0053] Collect semantic queries input by users via text, voice, and images, perform multimodal preprocessing, and generate a uniformly encoded query string.

[0054] The specific process includes: when collecting semantic queries input by the user via text, receiving the natural language text typed by the user in the input box; when collecting semantic queries input by the user via voice, receiving the voice signal emitted by the user and capturing the voice signal in the form of a digital audio stream; when collecting semantic queries input by the user via image, receiving the image file uploaded or captured by the user, obtaining the content in the form of a pixel matrix, and directly performing character encoding standardization processing on the semantic queries input by text, converting the semantic queries input by voice into corresponding text sequences using speech recognition, and extracting the contained object, scene, or text information of the semantic queries input by image using image semantic understanding methods and converting it into a natural language description; after completing the preprocessing of each modality, the semantic queries input by text, the text sequences converted from the semantic queries input by voice, and the natural language descriptions generated from the semantic queries input by image are semantically aligned and fused to generate a unified encoded query string.

[0055] Perform multi-dimensional semantic parsing on the query string, obtain a candidate parsing set, and calculate the semantic consistency index. The expression is:

[0056] ;

[0057] in, Let represent the semantic consistency index calculated in the candidate parsing set P, where P represents the candidate parsing set, n represents the total number of candidate parsings, and i represents the index number of the candidate parsing. This represents the weight coefficient of the i-th candidate parsing set. Let R represent the candidate parses of the i-th candidate parse set, and let R represent the reference parses. Candidate parsing Semantic similarity between and reference parsing R, Candidate parsing Information entropy This represents the smoothing constant.

[0058] The specific process includes analyzing the content of the query string from multiple dimensions such as syntactic structure, semantic roles, and contextual association when performing multi-dimensional semantic parsing on the query string, identifying multiple possible semantic interpretation paths, generating corresponding semantic expressions based on multiple semantic interpretation paths, forming a candidate parsing set, and calculating the semantic consistency index based on the semantic similarity between the reference parsing and each candidate parsing, the information entropy of each candidate parsing, and the weight coefficient assigned to each candidate parsing, combined with a smoothing constant.

[0059] It should be noted that the weight coefficients of the candidate parsing set are numerical values ​​assigned based on the credibility or importance of the candidate parsing in multi-dimensional semantic parsing, and are generated through an uncertainty assessment process.

[0060] Reference parsing is a standard semantic representation generated based on knowledge graphs and contextual information, used to measure the semantic accuracy of candidate parsings.

[0061] The smoothing constant is a fixed small amount added to the semantic consistency index calculation to avoid the denominator being zero or the value being unstable. Its value is obtained through optimization on the validation set.

[0062] Semantic similarity is a numerical metric that measures how close two semantic expressions are in meaning. It is obtained by encoding candidate parsings and reference parsings into semantic vectors and then calculating the cosine similarity between the vectors, or by using the matching score of a pre-trained language model.

[0063] When calculating the semantic consistency index, each candidate parser participates in the formula calculation through quantifiable indicators such as its corresponding semantic vector representation, weight coefficient, semantic similarity with the reference parser, and information entropy. These quantifiable indicators are derived from the semantic content of the candidate parser through existing technical methods, thereby achieving indirect numerical calculation.

[0064] Uncertainty assessment is performed on the candidate parsing set and semantic consistency index, and a decision dataset is generated.

[0065] The specific process includes comprehensively judging the candidate parsing set based on its consistency performance in multi-dimensional semantic parsing, its semantic similarity to the reference parsing, and the information entropy of the candidate parsing itself. Among these, candidate parsings with higher semantic similarity and lower information entropy are considered to have higher semantic stability and confidence levels. Combining the alignment degree with the reference parsing reflected by the semantic consistency index, the uncertainty of each candidate parsing is judged using information entropy and semantic fluctuation measure (for example, when a candidate parsing generates significantly different semantic expressions under different syntax parsing paths, or when the semantic vectors are discretely distributed in multiple inferences, the semantic fluctuation measure value is high, indicating greater uncertainty). Based on the uncertainty assessment results, candidate parsings with higher reliability and their corresponding semantic consistency indices are selected, and the evaluated candidate parsings and relevant indicators are integrated to form a decision dataset.

[0066] Semantic fluctuation metric refers to an indicator that measures the degree of semantic expression change of candidate parsing in multi-dimensional semantic parsing. It is derived from the difference in semantic distribution of the candidate parsing set under different parsing paths or contexts, and is obtained through the variance or dispersion of semantic vectors or parsing results.

[0067] Entities, attributes, and relationships are extracted from the decision dataset, and structured data is generated through triple assembly and standardized format conversion.

[0068] The specific process includes matching the semantic content in the decision dataset with the entity and relation patterns in the knowledge graph, identifying objects with independent semantic identities as entities, identifying content describing entity characteristics or states as attributes, identifying semantic associations between entities or between entities and attributes as relations, and combining the extracted entities, attributes, and relations into triples in the form of subject-verb-object. The content in the triples is then processed for terminology unification, format alignment, and syntax standardization to complete the standardized format conversion and generate structured data.

[0069] It should be noted that structured data includes entity words, template words, and contextual features.

[0070] S2. Based on structured data, entity disambiguation is performed using knowledge graphs. Abstract classes are generated through neural networks, and calibration and dynamic weight adjustment are performed to generate high-confidence entity abstract classes and confidence scores.

[0071] Multi-source knowledge graph fusion is performed on structured data to generate entity feature vectors.

[0072] The specific process includes aligning entities in structured data with knowledge graphs from multiple sources, finding knowledge nodes corresponding to entities and their associated attributes and relationships in each knowledge graph, integrating semantic information from different knowledge graphs, including entity type, description, contextual association, and topological features, mapping these heterogeneous but complementary knowledge representations to a unified vector space, and aggregating multi-source semantic features through vector concatenation, weighted averaging, or attention mechanisms to generate entity feature vectors that comprehensively reflect the semantic characteristics of entities in multi-source knowledge graphs.

[0073] It should be noted that multi-source semantic features refer to specific information about the same entity extracted from knowledge graphs from multiple sources. This includes the entity's type, descriptive text, associated attributes, adjacency relationships, and topological structure in each knowledge graph, as well as its original or encoded representation. These features originate from different knowledge graphs and are heterogeneous and complementary.

[0074] Semantic features are comprehensive semantic representations formed by fusing multi-source semantic features. They reflect the overall semantic essence or abstract semantic identity of an entity in a cross-knowledge graph context. They are the result description after fusion, rather than the original input features.

[0075] Entity disambiguation is performed on the entity feature vectors. A multi-hop attention mechanism is used to obtain the semantic correlation between entities. Based on the semantic correlation, clustering and filtering are performed to generate a disambiguated entity set.

[0076] The specific process includes matching entity names or identifiers in structured data with nodes in a multi-source knowledge graph that have the same surface form but different unique identifiers to obtain all possible corresponding real objects as candidate entities. Then, a multi-hop attention mechanism is used to propagate along the relational path in the knowledge graph to obtain the semantic relevance between candidate entities through context and neighborhood structure. Based on the degree of semantic relevance, candidate entities are clustered, and semantically similar candidate entities are grouped into the same cluster. Candidate entities that do not meet the semantic consistency criteria within or between clusters are filtered out, and the most representative candidate entities in each semantic cluster are retained to generate a disambiguated entity set.

[0077] It should be noted that the multi-hop attention mechanism is a method for multi-step reasoning along relational paths in a knowledge graph. It consists of multiple sequentially connected attention steps. Each step obtains attention weights based on the current entity and its adjacency relationships, selects the neighboring entity most relevant to the query semantics as the input for the next step, and achieves multi-hop propagation through continuous repetition. The multi-hop attention mechanism obtains semantic relevance by weighted aggregation of entity and relation embedding vectors in the knowledge graph. Its components include entity embedding, relation embedding, attention scoring function, and multi-step state update unit.

[0078] Based on the disambiguated entity set, an abstract class is generated using a reinforcement learning framework. The optimal abstraction path is then selected via a policy network, and the initial confidence score is calculated. The expression is:

[0079] ;

[0080] in, Let exp represent the initial confidence level for choosing action a given state vector s, where s represents the state vector, a represents the action, and exp represents the natural exponential function. This represents the immediate reward function for choosing action a given state vector s. Indicates temperature parameter, Let A represent any action in action space A, where A represents action space. This indicates that an action is performed under the state vector. The instant reward function, Represents the entropy sensitivity coefficient. Describes the policy function in the state vector s Information entropy Represents the policy function. This represents the maximum entropy.

[0081] The specific process includes mapping each entity and semantic feature in the disambiguated entity set to a state vector in the reinforcement learning framework, defining a series of possible abstract operations as actions in the action space, evaluating the expected reward of each action under the current state vector through the policy network, and selecting the abstract path that maximizes the long-term cumulative reward as the optimal abstract path based on the evaluation results. The optimal abstract path corresponds to a subset of entities with semantic commonalities, thus forming an abstract class. At the same time, the preliminary confidence is calculated based on the action probability distribution output by the policy network and the immediate reward of the selected action.

[0082] It should be noted that the temperature parameter is a hyperparameter used to adjust the smoothness of the probability distribution of the output actions of the policy network, and it is obtained by tuning using the validation set during the training process in the reinforcement learning framework.

[0083] The immediate reward function is derived in the reinforcement learning framework based on the semantic matching degree between the state vector and the selected action, the rationality of the abstract path, and the structural support in the knowledge graph. It is obtained through reward rules or supervision signals based on user feedback.

[0084] The entropy sensitivity coefficient is a hyperparameter used to adjust the influence of the policy function's information entropy on the initial confidence level. It is obtained by tuning based on validation performance during reinforcement learning training.

[0085] The training process of the reinforcement learning framework begins with initializing the parameters of the policy network and transforming the disambiguated entity set into a sequence of state vectors. Under each state vector, the policy network outputs the selection probability of each action in the action space. Based on the selection probability, an action is sampled and executed, thereby determining an abstract path and generating a corresponding abstract class. Subsequently, the immediate reward brought by the action is calculated according to the immediate reward function, and a cumulative reward signal is formed by combining the expected reward of subsequent states. The initial confidence is constructed using this cumulative reward signal and the information entropy of the policy function. The parameters of the policy network are updated using the gradient ascent method to maximize the weighted sum of the expected reward and the entropy regularization term. The above interaction, evaluation and update process is repeated until the policy network converges, and finally a reinforcement learning framework that can stably select the optimal abstract path is obtained.

[0086] Based on the abstract class and the initial confidence level, a cross-modal attention mechanism is used for calibration and dynamic weight adjustment to generate a calibrated confidence score.

[0087] The specific process includes, based on an abstract class and preliminary confidence, calibration and dynamic weight adjustment are performed through a cross-modal attention mechanism. The abstract class is converted into a semantic vector representation. Simultaneously, the preliminary confidence from text, speech, and image modalities are mapped to their corresponding confidence vectors. The semantic vector of the abstract class is used as the query vector, and the confidence vectors of each modality are used as key and value vectors. The dot product similarity between the query vector and each key vector is obtained, and the similarity is scaled and normalized to obtain the attention weights corresponding to each modality. The confidence vectors of each modality are weighted and aggregated with their corresponding attention weights to obtain a fused weighted confidence representation. Finally, a linear transformation or activation function is applied to the weighted confidence representation to generate a calibrated confidence score.

[0088] It should be noted that the cross-modal attention mechanism is used in this process to measure the semantic relevance between the abstract class and the initial confidence of different modalities, and to dynamically allocate the fusion weights of each modality. The cross-modal attention mechanism is derived from the attention calculation method for handling multimodal alignment and fusion in the existing technology.

[0089] Based on the calibrated confidence scores, the boundaries of abstract classes are dynamically adjusted through action space exploration and reward mechanisms to generate high-confidence entity abstract classes and confidence scores.

[0090] The specific process includes defining a series of actions in the action space to expand, shrink, or migrate the boundaries of abstract classes; evaluating the reliability of the current abstract class based on the calibrated confidence score; and constructing a reward signal by combining user feedback or semantic consistency indicators. Under the reinforcement learning framework, the action space is explored and utilized to select actions that can improve the overall confidence score and perform boundary adjustments. After each adjustment, the aggregate confidence score of entities within the abstract class is reacquired, and the boundary changes that improve the confidence score are retained. The process is iteratively optimized until convergence, generating entity abstract classes with high confidence scores and confidence scores.

[0091] S3. Combine entity abstract classes and confidence scores with user context, calculate the action-value function through a value network, and utilize... - Greedy algorithms select the optimal action.

[0092] Based on entity abstract classes and confidence scores, combined with user context, the numerical range is normalized using the min-max scaling method, and one-hot encoding is used to process categorical variables to generate a structured feature matrix.

[0093] The specific process includes reading the content of each field from the entity abstract class, confidence score, and user context item by item, and classifying them according to the data type and semantic attributes of the field content: identifying the confidence score and attributes in the user context that can be represented as continuous numerical values, such as query time interval, historical click count, and dwell time, as numerical features; identifying the category name of the entity abstract class, the user's region, device type, and query intent category, which have discrete symbolic values, as categorical variables; and matching and judging each field according to the preset feature type rules to complete the extraction of numerical features and categorical variables.

[0094] It should be noted that the feature type rules are obtained through domain knowledge or historical data statistical analysis based on the pre-defined classification criteria of entity abstract class, confidence score, and the semantic definition and data format of each field in the user context.

[0095] The structured feature matrix is ​​subjected to feature fusion and dimensionality compression to generate a low-dimensional state representation vector with context awareness.

[0096] The specific process includes semantically fusing numerical features and categorical variables from different sources in the structured feature matrix through vector concatenation or weighted combination, aligning user context information with entity abstract classes and confidence scores in a unified representation space, and using principal component analysis to reduce the dimensionality of the fused high-dimensional features, retaining the feature components that contribute most to semantic discrimination, removing redundancy and noise, and generating a low-dimensional state representation vector that contains joint information of user context and query semantics.

[0097] Based on the state representation vector, the action-value function is calculated through a value network, and its expression is:

[0098] ;

[0099] in, This represents the action-value function for choosing action a given state vector s. R represents the output layer weight vector, and R represents the modified linear unit activation function. This represents the hidden layer weight matrix. This represents the hidden layer bias vector. This represents the output layer bias vector. Indicates the scaling factor. This represents the transpose of the state vector s.

[0100] The specific process includes concatenating the state representation vector with each action in the action space to form a joint state-action input vector. This joint state-action input vector is then fed into a value network containing hidden and output layers. The hidden layer performs a linear transformation on the input and processes it through a modified linear unit activation function. The output layer further performs a linear transformation and adds a bias term. Simultaneously, a normalized inner product term between the state representation vector and the action vector is added to the output, along with a scaling factor, to obtain the action-value function value for each action under the state representation vector.

[0101] It should be noted that the scaling factor is a hyperparameter used to adjust the influence of the inner product term of the state representation vector and the action vector on the action-value function, and is obtained by tuning it during training based on the performance of the validation set.

[0102] use Greedy algorithms select the optimal action from the action-value function.

[0103] The specific process involves dynamically adjusting an exploration probability between zero and one based on policy performance during training. A random number is generated at each decision point. If the random number is less than... Then, a random action is selected from the action space for exploration. If the random number is greater than or equal to... Then, the action with the largest action-value function value is selected as the optimal action for utilization, thereby achieving a balance between exploring unknown actions and utilizing known high-value actions, and the selected action is output as the optimal action.

[0104] S4. Perform semantic normalization mapping based on the optimal action, and use the metasymbol dynamic generator to obtain the intermediate representation.

[0105] Perform in-depth analysis and type classification of the optimal action to generate executable mapping configuration parameters.

[0106] The specific process includes representing the optimal action as a natural language instruction or structured operation description, and performing syntactic parsing and keyword matching on the verbs, objects, and modifiers in the action pattern library to identify the core operation type. For example, when the optimal action contains verbs such as "replace with" or "map to" followed by a specific entity name, it is determined to be an entity replacement type; when it contains verbs such as "expand" or "associate" and involves a relation path, it is determined to be a relation expansion type; when it contains expressions such as "limit," "filter," or "only contain" and points to an attribute value, it is determined to be an attribute constraint type. This analysis process relies on existing natural language processing methods such as part-of-speech tagging, dependency parsing, and rule matching, combined with the action semantic labels learned in the action-value function training phase, to finally determine the action category to which the optimal action belongs.

[0107] It should be noted that the action pattern library is a set of rules containing predefined action types and corresponding keywords, syntactic structures, and semantic templates, which are obtained through manual annotation and summarization from historical interaction logs.

[0108] Attribute constraint types refer to a type of action in semantic operations that imposes limiting conditions on the attribute values ​​of entities. Specifically, they include restrictions on the range of attribute values ​​(such as time intervals, upper and lower limits of values), the specification of specific attribute values ​​(such as the color being red, the brand being a certain manufacturer), requirements on the existence of attributes (such as the requirement to include price information or user ratings), and constraints on the logical relationships between attributes.

[0109] Based on the executable mapping configuration parameters, the corresponding semantic mapping rules are loaded, and a dynamic mapping rule object is generated.

[0110] The specific process includes parsing the action category, target entity, mapping direction, semantic granularity, and contextual constraints contained in the executable mapping configuration parameters, searching for semantic mapping rules that completely match the executable mapping configuration parameters in the pre-stored set of semantic mapping rules, the semantic mapping rules defining the transformation logic and constraints from the source semantic structure to the target semantic structure, binding the found semantic mapping rules with the executable mapping configuration parameters and instantiating them to generate dynamic mapping rule objects.

[0111] It should be noted that semantic mapping rules are preset rules that define the transformation logic and constraints between the source semantic structure and the target semantic structure, and are obtained by inductive extraction from the labeled semantic alignment data.

[0112] Perform semantic normalization mapping on the dynamic mapping rule object to generate a preliminary mapping semantic structure.

[0113] The specific process includes reading the source semantic structure, target semantic structure, transformation logic, and constraints contained in the dynamic mapping rule object, replacing, generalizing, or normalizing the semantic elements in the current query according to the mapping direction and semantic granularity specified by the dynamic mapping rule object. For example, synonyms are unified into standard terms, specific instances are merged into their respective abstract classes, or different expressions of relationships are aligned to standard relationships in the knowledge graph, generating a preliminary mapping semantic structure that conforms to the unified semantic expression specification.

[0114] The initial mapping semantic structure is dynamically enhanced and its consistency is verified by a metasymbol dynamic generator to generate an intermediate representation.

[0115] The specific process includes converting entities, attributes, and relationships in the initial mapped semantic structure into abstract expressions composed of metasymbols, where metasymbols represent semantic roles or logical types rather than specific words. Based on ontology constraints and semantic mapping rules in the knowledge graph, the combination of metasymbols is checked for logical consistency, such as the type compatibility of subject-verb-object structures and the legality of relational directions. After the consistency check passes, the metasymbol sequence is semantically expanded or refined according to contextual requirements and query intent, such as supplementing implicit attributes, deriving indirect relationships, or introducing default values. The enhanced and verified metasymbol sequence is then re-instantiated into a standardized semantic expression readable by natural language, generating an intermediate representation.

[0116] It should be noted that the metasymbol dynamic generator refers to the processing mechanism used to convert the initial mapping semantic structure into an abstract expression composed of metasymbols and to perform semantic enhancement and consistency verification. It consists of a metasymbol set, semantic role templates, ontology constraint rules, and symbol instantiation logic.

[0117] S5. Based on the intermediate representation, perform index retrieval and multi-dimensional sorting to generate a sorted list of search results.

[0118] Based on standardized semantic representation, construct a multi-source index query task.

[0119] The specific process includes parsing the entity, attribute, and relation triples contained in the standardized semantic representation, and generating query subtasks adapted to different data sources based on the content of the triples. Each query subtask performs syntax transformation for a specific data source format, such as converting triples into keyword queries suitable for inverted indexes, SPARQL queries suitable for knowledge graphs, or embedding similarity queries suitable for vector databases. All adapted query subtasks are combined to form a multi-source index query task covering multiple heterogeneous data sources.

[0120] Perform parallel index retrieval on the index query task to obtain a candidate result set.

[0121] The specific process involves breaking down the multi-source index query task according to the data source type and the division of query subtasks to obtain multiple independent subquery requests. In a database or search engine environment that supports parallel processing, the subquery requests are simultaneously distributed to different types of index structures such as inverted indexes, knowledge graph indexes, or vector indexes for concurrent execution. Each index structure returns matching local results according to its own retrieval mechanism. All local results are aggregated and integrated to form a complete candidate result set.

[0122] By using a multi-dimensional scoring function and a dynamic weight adjustment mechanism, the candidate result set is sorted in multiple dimensions to generate multi-dimensional sorted data.

[0123] The specific process includes obtaining semantic relevance score, confidence score, context matching score, and timeliness score for each candidate result based on standardized semantic representation and user context. Then, based on the semantic features of the current query and the user context content, the weights corresponding to each evaluation dimension are dynamically determined using preset weight mapping rules. The scores of each candidate result on each dimension are weighted and aggregated with the corresponding weights to obtain a comprehensive score for the candidate results. The candidate result set is then arranged from high to low according to the comprehensive score to generate multi-dimensional sorted data.

[0124] It should be noted that the weight mapping rule is a preset functional relationship that maps query semantic features and user context to the weights of each evaluation dimension, and is obtained through training on historical interaction data.

[0125] Dynamic weight optimization and diversity re-sorting are performed on multi-dimensional sorted data to generate a sorted list of search results.

[0126] The specific process includes: readjusting the weights of each evaluation dimension based on real-time user feedback signals or contextual changes; dynamically optimizing the weights to update the current comprehensive score according to the weight mapping rules; and combining a diversity control mechanism in the re-ranking stage to identify redundant items by recognizing the semantic similarity between candidate results and de-weighting or arranging highly similar results at intervals to ensure that the results cover different semantic subtopics. The results are then output in the order determined by the optimized comprehensive score and diversity constraints to generate a ranked list of search results.

[0127] It should be noted that the diversity control mechanism is a method used to reduce semantic redundancy in search results and improve the breadth of topic coverage. It is obtained by acquiring the semantic similarity between candidate results and combining it with the maximum marginal relevance criterion or submodular function optimization.

[0128] S6. Monitor user interaction with the search results list, collect user feedback, and update the parameters of the value network.

[0129] Real-time monitoring of user interactions with the sorted search results list generates interaction event data.

[0130] The specific process includes capturing behavioral signals generated by users during browsing, clicking, staying, skipping, or providing feedback, and recording these behavioral signals in a structured manner according to timestamps, operation types, target result identifiers, and contextual states to form interactive event data containing user intent and preference information.

[0131] Based on interaction event data, multimodal user feedback is collected and a feedback dataset is generated.

[0132] The specific process includes extracting implicit feedback signals from user interaction behavior in different interaction behaviors from interaction event data, including explicit preferences corresponding to click behavior, interest intensity reflected by dwell time, negative tendencies reflected by skipping operations (referring to the user's disinterest or dissatisfaction shown by skipping, quickly closing or ignoring a search result), and direct evaluations expressed by explicit ratings (such as five-star ratings, likes or dislikes) or annotations; simultaneously collecting multimodal feedback content such as text comments, voice commands and image tags associated with these interaction behaviors, and aligning and fusing the feedback signals of various modalities with the corresponding search result identifiers, timestamps and context states, and finally integrating them to form a feedback dataset with a unified structure and complementary modalities.

[0133] The parameters of the value network are updated using a reinforcement learning algorithm based on the feedback dataset.

[0134] The specific process includes extracting the ranking results, selected actions, and implicit reward signals corresponding to user interaction behaviors from the feedback dataset, and transforming them into triple samples of state vectors, actions, and actual rewards. The time-series difference error between the predicted value of the action-value function and the actual reward is obtained using these triple samples. Based on this time-series difference error, backpropagation updates are performed on the hidden layer weight matrix, output layer weight vector, hidden layer bias vector, and output layer bias scalar in the value network using gradient descent. During the update process, the scaling factor is simultaneously adjusted to optimize the contribution of the inner product term between the state vector and the action vector. Finally, the parameters of the value network gradually converge to a stable state that can accurately evaluate the value of actions.

[0135] This embodiment also provides an information retrieval system based on semantic normalization, including: a semantic parsing module, a knowledge enhancement module, an action selection module, a semantic generation module, a retrieval ranking module, and a learning optimization module. The semantic parsing module collects user-input semantic query records and performs preliminary semantic parsing to generate structured data. The knowledge enhancement module uses knowledge graphs to disambiguate entities based on the structured data, generates abstract classes through neural networks, and performs calibration and dynamic weight adjustment to generate high-confidence entity abstract classes and confidence scores. The action selection module combines the entity abstract classes and confidence scores with the user context, calculates the action-value function through a value network, and utilizes... - A greedy algorithm selects the optimal action; a semantic generation module performs semantic normalization mapping based on the optimal action and uses a dynamic meta-symbol generator to obtain intermediate representations; a retrieval and ranking module performs index retrieval and multi-dimensional ranking based on the intermediate representations to generate a ranked list of retrieval results; and a learning and optimization module monitors user interaction with the retrieval results list, collects user feedback, and updates the parameters of the value network.

[0136] In summary, this invention addresses the semantic fragmentation problem of multimodal queries by using multi-source knowledge graph fusion and cross-modal attention mechanisms, achieving deep semantic alignment and dynamic weight calibration of heterogeneous data. It also generates multimodal feedback datasets in real-time based on user interaction behavior and dynamically updates the parameters of the value network through online reinforcement learning algorithms, constructing a real-time feedback loop from user interaction to model update, thus achieving continuous self-optimization of parameters.

[0137] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An information retrieval method based on semantic normalization, characterized in that: include, Collect user-input semantic query records and perform preliminary semantic parsing to generate structured data; Based on structured data, entity disambiguation is performed using knowledge graphs. Abstract classes are generated through neural networks, and calibration and dynamic weight adjustment are applied to generate high-confidence entity abstract classes and confidence scores. The specific steps are as follows: Multi-source knowledge graph fusion is performed on structured data to generate entity feature vectors; Entity disambiguation is performed on the entity feature vectors, and the semantic correlation between entities is obtained by using a multi-hop attention mechanism. Based on the semantic correlation, clustering and filtering are performed to generate a disambiguated entity set. Based on the disambiguated entity set, an abstract class is generated using a reinforcement learning framework, and the optimal abstract path is selected through a policy network to calculate the initial confidence level. Based on the abstract class and initial confidence level, a cross-modal attention mechanism is used for calibration and dynamic weight adjustment to generate a calibrated confidence score. The specific process includes: Based on abstract classes and preliminary confidence scores, when calibrating and dynamically adjusting weights through a cross-modal attention mechanism, the abstract classes are converted into semantic vector representations. Simultaneously, the preliminary confidence scores from text, speech, and image modalities are mapped to their corresponding confidence vectors. The semantic vector of the abstract classes is used as the query vector, and the confidence vectors of each modality are used as key and value vectors. The dot product similarity between the query vector and each key vector is obtained, and the similarity is scaled and normalized to obtain the attention weights corresponding to each modality. The confidence vectors of each modality are weighted and aggregated with their corresponding attention weights to obtain a fused weighted confidence representation. A linear transformation or activation function is then applied to the weighted confidence representation to generate the calibrated confidence score. Based on the calibrated confidence scores, the boundaries of abstract classes are dynamically adjusted through action space exploration and reward mechanisms to generate high-confidence entity abstract classes and confidence scores; By combining entity abstract classes and confidence scores with user context, an action-value function is calculated through a value network, and then... The greedy algorithm selects the optimal action, and the specific steps are as follows. Based on entity abstract classes and confidence scores, combined with user context, the numerical range is normalized using the min-max scaling method, and one-hot encoding is used to process categorical variables to generate a structured feature matrix. The specific process includes: The system reads the content of each field from the entity abstract class, confidence score, and user context item by item, and classifies them according to the data type and semantic attributes of the field content: the confidence score and attributes in the user context that can be represented as continuous numerical values, such as query time interval, historical click count, and dwell time, are identified as numerical features; the category name of the entity abstract class, the user's region, device type, and query intent category, which have discrete symbolic values, are identified as categorical variables; the identification process matches and judges each field according to the preset feature type rules to complete the extraction of numerical features and categorical variables. The structured feature matrix is ​​subjected to feature fusion and dimensionality compression to generate a low-dimensional state representation vector with context awareness. Based on the state representation vector, the action-value function is calculated through a value network; use Greedy algorithms select the optimal action from the action-value function; Perform semantic normalization mapping based on the optimal action, and use a metasymbol dynamic generator to obtain intermediate representations; Based on the intermediate representation, index retrieval and multi-dimensional sorting are performed to generate a sorted list of search results. Monitor user interactions with the search results list, collect user feedback, and update the parameters of the value network.

2. The information retrieval method based on semantic normalization as described in claim 1, characterized in that: The semantic query records include text query records, voice query records, and image query records.

3. The information retrieval method based on semantic normalization as described in claim 2, characterized in that: The specific steps for generating structured data are as follows: Collect semantic queries input by users via text, voice, and images, perform multimodal preprocessing, and generate a uniformly encoded query string; Perform multi-dimensional semantic parsing on the query string, obtain a candidate parsing set, and calculate the semantic consistency index; Uncertainty assessment is performed on the candidate parsing set and semantic consistency index, and a decision dataset is generated. Entities, attributes, and relationships are extracted from the decision dataset, and structured data is generated through triple assembly and standardized format conversion.

4. The information retrieval method based on semantic normalization as described in claim 3, characterized in that: The specific steps for obtaining the intermediate representation are as follows: Perform in-depth analysis and type classification of the optimal action to generate executable mapping configuration parameters; Based on executable mapping configuration parameters, load the corresponding semantic mapping rules and generate dynamic mapping rule objects; Perform semantic normalization mapping on the dynamic mapping rule object to generate a preliminary mapping semantic structure; The initial mapping semantic structure is dynamically enhanced and its consistency is verified by a metasymbol dynamic generator to generate an intermediate representation.

5. The information retrieval method based on semantic normalization as described in claim 4, characterized in that: The specific steps for generating the sorted search results list are as follows. Based on standardized semantic representation, construct a multi-source index query task; Perform parallel index retrieval on the index query task to obtain a candidate result set; The candidate result set is sorted in multiple dimensions using a multi-dimensional scoring function and a dynamic weight adjustment mechanism to generate multi-dimensional sorted data. Dynamic weight optimization and diversity re-sorting are performed on multi-dimensional sorted data to generate a sorted list of search results.

6. The information retrieval method based on semantic normalization as described in claim 5, characterized in that: The specific steps for monitoring user interaction with the search results list, collecting user feedback, and updating the parameters of the value network are as follows: Real-time monitoring of user interaction with the sorted search results list, generating interaction event data; Based on interaction event data, multimodal user feedback is collected and a feedback dataset is generated; Based on the feedback dataset, the parameters of the value network are updated using a reinforcement learning algorithm; The interaction event data includes click events, dwell time, and scroll depth.

7. An information retrieval system based on semantic normalization, based on the information retrieval method based on semantic normalization as described in any one of claims 1 to 6, characterized in that: It includes a semantic parsing module, a knowledge enhancement module, an action selection module, a semantic generation module, a retrieval ranking module, and a learning optimization module. The semantic parsing module is used to collect user-input semantic query records and perform preliminary semantic parsing to generate structured data; The knowledge enhancement module is used to perform entity disambiguation based on structured data and knowledge graphs. It generates abstract classes through neural networks, performs calibration and dynamic weight adjustment, and generates high-confidence entity abstract classes and confidence scores. The action selection module combines entity abstract classes and confidence scores with user context, calculates the action-value function through a value network, and utilizes... - Greedy algorithm selects the optimal action; The semantic generation module is used to perform semantic normalization mapping based on the optimal action and to obtain intermediate representations using a dynamic metasymbol generator. The retrieval and sorting module is used to perform index retrieval and multi-dimensional sorting based on intermediate descriptions, and generate a sorted list of retrieval results. The learning optimization module is used to monitor user interaction with the search results list, collect user feedback, and update the parameters of the value network.

Images