Dynamic time sequence enhanced retrieval method and device, equipment and medium

Abstract

The invention relates to the technical field of natural language processing, can be applied to business system platforms of financial science and technology, medical treatment and health and the like, and discloses a dynamic time sequence enhanced retrieval method, device, equipment and medium, and the method comprises the following steps: carrying out analytic feature extraction on pre-acquired multi-modal data to obtain analytic feature data; performing timestamp extraction on the analyzed feature data to obtain a data timestamp; calculating an aging attention weight of the analysis feature data according to the data timestamp and preset query time; acquiring updated data in real time or regularly, and performing probabilistic dynamic feature extraction on the updated data to obtain probabilistic dynamic features; and performing duration dynamic fusion on the probabilistic dynamic features, the analysis feature data and the aging attention weight to obtain a fusion retrieval score, and screening an initial retrieval result according to the fusion retrieval score. And by introducing update data and multi-modal data for comprehensive analysis, the method adapts to dynamic change characteristics of the data, and the accuracy of a retrieval result is improved.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and in particular to a dynamic temporal enhanced retrieval method, device, equipment and medium. Background Art

[0002] Dynamic time series retrieval is an information retrieval technology that combines the time dimension of data with the dynamic change characteristics. Its core is to enable the system to more accurately filter, sort and present relevant information based on the time attributes and real-time dynamic characteristics of the data. It is widely used in industries such as financial technology and healthcare that are sensitive to data timeliness. For example, in the healthcare field, medical guidelines, drug clinical trial results, epidemic data, etc. are frequently updated, and the use of dynamic time series retrieval makes the retrieval results more accurate; in the stock, cryptocurrency and other markets in the financial technology field, dynamic time series retrieval is used to analyze price, trading volume and other data in real time to capture millisecond-level trading signals.

[0003] At present, traditional retrieval enhancement usually adopts static embedding methods, and some diachronic entity embedding models only support updates at regular time intervals. However, data updates are often sudden, such as policy adjustments during the epidemic period. If traditional retrieval methods are used, the retrieval results may be insufficiently accurate.

[0004] Therefore, in the face of growing retrieval needs, current retrieval methods urgently need to be improved to solve the problem that existing methods are unable to introduce updated data into analysis in real time, resulting in insufficient accuracy of retrieval results. Summary of the Invention

[0005] The present invention provides a dynamic time-series enhanced retrieval method, device, equipment and medium, which introduces updated data and multimodal data for comprehensive analysis, adapts to the dynamic change characteristics of data, and improves the accuracy of retrieval results.

[0006] In a first aspect, a dynamic temporal enhanced retrieval method is provided, comprising:

[0007] Performing analytical feature extraction on pre-acquired multimodal data to obtain analytical feature data;

[0008] Extracting a timestamp from the analyzed feature data to obtain a data timestamp;

[0009] Calculating the time-sensitive attention weight of the parsed feature data according to the data timestamp and the preset query time;

[0010] Acquire update data in real time or periodically, and perform probabilistic dynamic feature extraction on the update data to obtain probabilistic dynamic features;

[0011] Performing a diachronic dynamic fusion of the probabilistic dynamic features and the analytical feature data in combination with the temporal attention weight to obtain a fusion retrieval score, and screening initial retrieval results based on the fusion retrieval score;

[0012] The initial search results are optimized and adjusted according to preset user feedback to obtain target search results.

[0013] In a second aspect, a dynamic time sequence enhanced retrieval device is provided, comprising:

[0014] An extraction module is used to extract analytical features from pre-acquired multimodal data to obtain analytical feature data;

[0015] A timestamp extraction module, configured to extract timestamps from the parsed feature data to obtain data timestamps;

[0016] A calculation module, configured to calculate a time-sensitive attention weight of the parsed feature data based on the data timestamp and a preset query time;

[0017] An acquisition and extraction module is used to acquire update data in real time or periodically, and perform probabilistic dynamic feature extraction on the update data to obtain probabilistic dynamic features;

[0018] a fusion screening module, configured to perform a diachronic dynamic fusion of the probabilistic dynamic features and the analytical feature data in combination with the temporal attention weight to obtain a fusion retrieval score, and to screen the initial retrieval results according to the fusion retrieval score;

[0019] The adjustment module optimizes and adjusts the initial search results according to preset user feedback to obtain target search results.

[0020] In a third aspect, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned dynamic temporal enhancement retrieval method when executing the computer program.

[0021] In a fourth aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned dynamic temporal enhancement retrieval method are implemented.

[0022] In the solution implemented by the above-mentioned dynamic temporal enhanced retrieval method, device, computer equipment and storage medium, valuable features are extracted from multimodal data (such as images, text, audio, etc.), and time tags are given to the data to clarify the time point when the data is generated or updated, which helps to sort out the time context of the data and provide key information for analysis and processing based on the time dimension; by calculating the time-sensitive attention weight, the importance of recent data can be highlighted, the impact of outdated data can be weakened, the system can pay more attention to current valuable information, adapt to the dynamic changes of data, and improve the pertinence and effectiveness of information processing; real-time or periodic acquisition of updated data, timely capture of dynamic changes in data, so that the model can better cope with the dynamic and random nature of data, and improve its responsiveness to new information; comprehensive consideration of the historical characteristics, timeliness and real-time dynamic changes of data can screen out results that better meet user needs and improve the quality and relevance of retrieval results. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0024] Figure 1 This is a schematic diagram of an application environment of a dynamic time series enhanced retrieval method according to an embodiment of the present invention;

[0025] Figure 2 This is a flow chart of a dynamic time series enhanced retrieval method according to an embodiment of the present invention;

[0026] Figure 3 It is a structural diagram of a dynamic time series enhanced retrieval device according to one embodiment of the present invention;

[0027] Figure 4 is a structural diagram of a computer device according to an embodiment of the present invention;

[0028] Figure 5 It is another structural schematic diagram of a computer device in one embodiment of the present invention. DETAILED DESCRIPTION

[0029] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0030] The embodiment of the present invention provides a dynamic time series enhanced retrieval method, which can be applied in the following situations: Figure 1 In an application environment, the client communicates with the server through a network. The server can perform analytical feature extraction on pre-acquired multimodal data to obtain analytical feature data; perform timestamp extraction on the analytical feature data to obtain data timestamp; calculate the time-sensitive attention weight of the analytical feature data based on the data timestamp and the preset query time; obtain updated data in real time or periodically, perform probabilistic dynamic feature extraction on the updated data to obtain probabilistic dynamic features; dynamically fuse the probabilistic dynamic features and the analytical feature data with the time-sensitive attention weight to obtain a fused retrieval score, and filter the initial retrieval results based on the fused retrieval score; optimize and adjust the initial retrieval results based on preset user feedback to obtain target retrieval results, and feed the target retrieval results back to the client. The present invention provides a dynamic time series enhanced retrieval device, which, for the target retrieval result business, introduces updated data and multimodal data for comprehensive analysis to adapt to the dynamic change characteristics of the data and improve the accuracy of the retrieval results. The client can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, and portable wearable devices. The server can be implemented as an independent server or a server cluster consisting of multiple servers. The present invention is described in detail below through specific examples.

[0031] See also Figure 2 As shown, Figure 2 A flow chart of a dynamic time series enhanced retrieval method provided by an embodiment of the present invention includes the following steps:

[0032] S1. Perform analytical feature extraction on pre-acquired multimodal data to obtain analytical feature data.

[0033] In an embodiment of the present invention, the analytical feature extraction refers to the process of extracting abstract features that can characterize the essential semantics, structure or regularity of data from multimodal data (such as images, text, audio, etc.) through algorithm or model analysis.

[0034] Specifically, multimodal data includes text data, image data, audio data, and video data. For text data, abstract information such as semantics, grammar, and emotions are extracted; for image data, visual semantics such as objects, scenes, and textures are identified; for audio data, sound frequency, semantic content, and emotional features are analyzed; and for video data, the temporal dynamic features of visual frames and audio are integrated.

[0035] In specific medical and health scenarios, by analyzing the visual features of medical images (such as lung texture and nodules) and combining them with symptom keywords in medical record text, auxiliary diagnosis of pneumonia can be achieved with an accuracy rate exceeding the average level of radiologists.

[0036] In the financial technology scenario, by analyzing transaction timing characteristics (such as the frequency of remote logins) and biometric characteristics (dynamic features of facial recognition), the accuracy of fraud identification is increased to 99.9%, and the misjudgment rate is less than 0.01%.

[0037] In an embodiment of the present invention, extracting analytical features from pre-acquired multimodal data to obtain analytical feature data includes:

[0038] Perform standardized cleaning on the pre-acquired multimodal data to obtain standard cleaned data;

[0039] Performing submodal feature extraction on the standard cleaning data to obtain submodal features;

[0040] The sub-modal features are converted into the same dimensional space, and the conversion results are spliced ​​into analytical feature data.

[0041] In an embodiment of the present invention, the standardized cleaning refers to the process of eliminating noise, inconsistency and redundant information in multimodal data through regularized processing to make it conform to a unified format standard. The submodal feature extraction refers to the use of a dedicated algorithm to extract vector representations that can characterize the core features of different types of data (text, images, audio, etc.). The same-dimensional space conversion refers to mapping feature vectors of different modalities to feature spaces of the same dimension to eliminate dimensional differences to support fusion.

[0042] Specifically, for missing information in modalities such as text, images, and audio, mean filling, mode filling, or model prediction is used to complete it. Statistical methods or machine learning models are used to identify and eliminate outlier data, and the format of text, image, audio, and other modal data is unified and normalized. For text modalities, keywords are extracted through word segmentation, part-of-speech tagging, and named entity recognition. Traditional methods such as TF-IDF and bag-of-words models are used to extract semantic features of text. For image modal data, convolutional neural networks are used to extract basic features such as edges, textures, and colors. Target detection is then used to locate key areas, and feature vectors are generated through pooling operations to extract high-level semantic features. For audio data, Mel-frequency cepstral coefficients are used to extract acoustic features. LSTM / CNN is used to analyze temporal patterns. Optical flow is used to extract motion features from video data. 3D convolutional neural networks are used to analyze spatiotemporal relationships. Statistical methods or feature engineering are used to generate numerical feature vectors for structured data.

[0043] In detail, feature vectors of different dimensions (such as 768 dimensions for text and 1024 dimensions for images) are compressed or expanded to the same dimension through fully connected layers, linear transformations, or autoencoders. Contrastive learning or cross-modal attention is used to keep the features of different modalities consistent at the semantic level. They are directly concatenated in modal order (such as text features + image features + structured features) to generate parsed feature data.

[0044] In an embodiment of the present invention, sub-modal feature extraction can ensure that the key information of each modality is captured to the maximum extent, providing a rich original feature basis for cross-modal fusion, while reducing cross-modal noise coupling and improving feature purity. The same-dimensional space conversion can construct a unified semantic space, providing a mathematical basis for cross-modal association analysis.

[0045] In an embodiment of the present invention, heterogeneous data is converted into a comprehensive representation with semantic consistency through sub-modal intensive cultivation + cross-modal fusion. This process not only retains the unique value of each modality, but also creates a new information dimension through correlation analysis.

[0046] S2. Extracting a timestamp from the analyzed feature data to obtain a data timestamp.

[0047] In the embodiment of the present invention, the timestamp extraction refers to the process of identifying and extracting time information related to data generation or event occurrence from multimodal analysis feature data, and converting it into a standardized time format.

[0048] Specifically, time coordinates are assigned to data to clarify the correspondence between data and time, and a "bridge" is built to connect the temporal associations of multimodal data, so that cross-modal data can be aligned and analyzed in the time dimension.

[0049] In medical and health scenarios, associating CT image timestamps of different dates with blood test timestamps can analyze the temporal relationship between lesion development and indicator changes. At the same time, outdated diagnosis and treatment data can be filtered out by timestamps to ensure that the model is based on the latest medical knowledge.

[0050] In financial risk control, abnormal time series patterns can be discovered through transaction timestamps, and combined with geographic location characteristics, the risk of credit card fraud can be identified and the time correlation of transactions can be determined.

[0051] In the embodiment of the present invention, extracting the timestamp from the parsed feature data to obtain the data timestamp includes:

[0052] Extracting a time field from the parsed feature data;

[0053] Normalize the time field to obtain a time standard format and determine the time accuracy;

[0054] The timestamp of the parsed feature data is extracted according to the time specification format and the time precision to obtain a data timestamp.

[0055] In an embodiment of the present invention, extracting the time field refers to identifying and separating all time-related information fragments from the parsed feature data, format normalization refers to converting the extracted time field into a unified and standard time expression form, and determination refers to clarifying its time accuracy (such as year, month, day, hour, minute, second, millisecond level, etc.).

[0056] Specifically, regular expressions and time parsing libraries are used to match time keywords in natural language, and explicit time fields are identified and extracted. For example, from the medical record "The patient was admitted to the hospital at 9:30 am on May 15, 2025", "May 15, 2025" and "9:30 am" are extracted. For the inference and extraction of implicit time fields, for example, the StudyDate and StudyTime tags of the CT image DICOM file are read, and "20250520" and "103000" are extracted. At the same time, context-related inference can be used to generate logical inference time based on the data.

[0057] Furthermore, time texts in different formats are converted into a standard format. For example, the text 2025 / 5 / 23 is standardized as 2025-05-23, and times in different time zones are converted to UTC (Coordinated Universal Time) or a specified time zone to avoid time zone ambiguity. The granularity of the time field is analyzed (such as only containing the date "2025-05-23", or containing hours, minutes, and seconds "2025-05-2310:30:00"), and time fields with insufficient precision are completed according to the rules (such as automatically associating time with only hours, minutes, and seconds with the current date). For example, if the time field "14:30:00" is known to belong to "2025-05-23", it is completed to "2025-05-2314:30:00". Finally, the timestamps of multimodal data are extracted according to the requirements of standardization and precision.

[0058] In the embodiment of the present invention, through format normalization processing, cross-source time ambiguity is eliminated, and spatiotemporal alignment of multimodal data is achieved. At the same time, the time accuracy boundary is clarified, the granularity requirements of different scenarios are adapted, a unified time language is constructed, and computing efficiency and compatibility are improved.

[0059] In the embodiment of the present invention, timestamp extraction not only solves the time alignment problem of cross-modal data, but also enables the data to have the ability of time series analysis, laying the foundation for subsequent applications such as time series pattern mining, dynamic prediction, and event tracing.

[0060] S3. Calculate the time-sensitive attention weight of the parsed feature data according to the data timestamp and the preset query time.

[0061] In the embodiment of the present invention, the calculation refers to measuring the importance of data in the time dimension.

[0062] Specifically, based on the data timestamp and the preset query time, the newness of the data is measured by the time difference and converted into a weight value between 0 and 1 (or other range of values). The specific formula is as follows:

[0063]

[0064] Among them, sim(q,d i ) represents the semantic similarity (cosine value) between the query content and the document, which is calculated by the fine-tuned DeBERTa model; φΔt) represents the time decay function, defined as φ(Δt)=e -γ|Δt| , where γ = 0.05 / day (derived from the half-life statistics of insurance policies), λ represents a trainable parameter with an initial value of 0.7 (a priori in the insurance field), which is dynamically adjusted based on user click feedback, q represents the query content, and t q Indicates a timestamp.

[0065] In the specific scenario of medical health, non-monotonic adjustments are made through the timeliness weights of patients' historical medical records. For example, the newer the chronic disease data (blood pressure records), the higher the weight, while vaccination records may be valid for a long time (the weight decays very slowly).

[0066] In financial scenarios, "highly timely" financial news data is automatically filtered based on weights. For example, in the recommendation system, when a user queries for "latest technology news", only data with a timeliness weight > 0.5 (that is, the time difference is within the preset range) is retained, filtering out old news.

[0067] In an embodiment of the present invention, the step of calculating the time-sensitive attention weight of the parsed feature data according to the data timestamp and the preset query time includes:

[0068] Obtaining user query content, and converting the user query content and the parsing feature data into a query semantic vector and a parsing semantic vector respectively;

[0069] Calculating the semantic similarity between the query semantic vector and the parsed semantic vector;

[0070] Performing time difference processing on the data timestamp and the preset query time to obtain a time difference result, and determining the time attenuation factor of the analyzed feature data according to the time difference result;

[0071] The temporal attention weight of the parsed feature data is configured according to the time decay factor and the semantic similarity result.

[0072] In an embodiment of the present invention, the conversion refers to converting text / multimodal features into dense vectors (semantic vectors) in a high-dimensional space through a deep learning model, so that semantically similar content is closer in the vector space. The calculation refers to measuring the similarity of two semantic vectors through a mathematical measurement method. The determination of the attenuation factor refers to constructing a function that increases and decreases with time based on the time difference between the data timestamp and the query time.

[0073] Specifically, the user query content is segmented, and contextual semantic features are extracted through a multi-layer neural network to output a vector of fixed dimension. If it is text data, the process is similar to query conversion, and the domain characteristics of the feature data (such as medical and financial terms) need to be considered. A domain-adapted pre-trained model may be used. If it is multimodal features such as images / audio, the features need to be extracted through models such as convolutional neural networks, and then converted into a semantic vector of the same dimension as the text vector through a cross-modal mapping network (such as CLIP). The semantic similarity between the query semantic vector and the parsed semantic vector can be calculated using methods such as cosine similarity, Euclidean distance, and dot product similarity.

[0074] Furthermore, the exponential decay model is used to calculate the time decay factor of the analytical feature data. The specific formula is as follows:

[0075] decay=e -λ·Δt

[0076] Where Δt is the time difference, λ is the decay coefficient, and decay is the time decay factor.

[0077] Furthermore, the temporal attention weight = semantic similarity × time decay factor. For example, if the semantic similarity is 0.8 and the decay factor is 0.223, the weight is 0.8×0.223≈0.178.

[0078] In an embodiment of the present invention, the "data recency" is converted into a calculable value through the time difference, making timeliness a measurable feature; an attenuation factor is automatically generated based on the time difference to ensure that recent data has a greater weight in the calculation and adapt to real-time requirements; data priority is dynamically adjusted through the timeliness weight, which is suitable for time-sensitive fields such as news, social networking, and risk control.

[0079] In an embodiment of the present invention, the semantic value and temporal value of multimodal data are quantified into a unified weight system to provide data support for the decision-making of the intelligent system.

[0080] S4. Acquire update data in real time or periodically, and perform probabilistic dynamic feature extraction on the update data to obtain probabilistic dynamic features.

[0081] In an embodiment of the present invention, the acquisition refers to pulling updated data from a data source at a preset frequency (real time or timed) through a data acquisition system, and the probabilistic dynamic feature extraction refers to extracting features with probabilistic representation from dynamically updated data through statistical learning or deep learning methods.

[0082] Specifically, a streaming processing framework is adopted to continuously monitor changes in data sources through long polling and other methods. The data is transmitted to the processing end within seconds after it is generated, or data is pulled in batches at a fixed period based on the scheduling system. For the updated data obtained, sliding windows and time series models are used to extract features that change over time, and then statistical modeling, Bayesian inference, and deep learning probabilistically are performed to obtain probabilistic dynamic features.

[0083] In the embodiment of the present invention, the extracting probabilistic dynamic features from the update data to obtain probabilistic dynamic features includes:

[0084] Extracting semantic features from the update data to obtain update data features;

[0085] Obtaining the mean vector and covariance matrix of the entity embedding, performing a linear transformation on the updated data features and the mean vector to obtain an updated mean vector;

[0086] Converting the covariance matrix into a matrix vector, performing linear transformation on the updated data features and the matrix vector to obtain an updated covariance matrix;

[0087] Performing trace judgment on the updated covariance matrix to obtain a judgment result, and combining the updated data feature with the updated mean vector and the updated covariance matrix according to the judgment result to generate a probabilistic dynamic feature.

[0088] In an embodiment of the present invention, the semantic feature extraction refers to extracting a feature vector that can represent its semantic meaning from the updated data through natural language processing or a deep learning model. The acquisition refers to obtaining the statistical features embedded in the entity (such as users, products, text paragraphs) from historical data or a pre-trained model. The splicing linear transformation refers to splicing the updated data features with the mean vector / matrix vector and generating a new feature vector or matrix through a linear transformation. The trace judgment refers to calculating the trace of the covariance matrix (the sum of the diagonal elements) to judge the validity or information content of the matrix, which is used to decide whether to integrate the new data features to update the probabilistic features.

[0089] Specifically, the updated data includes text, images, videos, and structured data. The text is segmented and stop words are removed. The BERT model is used to map the text into a low-dimensional dense vector. The visual features of the image / video data are extracted through CNN, and then converted into semantic vectors through a fully connected layer. The structured data is normalized, the categorical data is encoded or embedded, and then fused into a semantic vector through linear transformation.

[0090] Furthermore, we collect the historical embedding vectors of a certain type of entity, average the dimensions of the vectors, and obtain the mean vector μ. For each entity embedding vector, we calculate its deviation from the mean vector. The specific formula is as follows:

[0091]

[0092] Among them, x i represents each entity embedding vector, x i -μ represents the deviation of the entity embedding vector from the mean vector, Cov ij Represents the correlation between the i-th dimension and the j-th dimension features, concatenates the updated data feature vector and the mean vector by dimension, and then calculates the updated mean vector new_μ through the weight matrix w1 and bias b1. The specific formula is as follows:

[0093] new_μ=W1·[f;μ]+b1

[0094] Furthermore, the covariance matrix Cov is expanded into a one-dimensional vector v by row (or column), the updated data feature vector f and the vector v are concatenated into [f; v], and the updated covariance matrix is ​​calculated by the weight matrix w2 and the bias b2. The specific formula is as follows:

[0095] new_v=W2·[f;v]+b2

[0096] Furthermore, the trace of the updated covariance matrix new_v is calculated. The specific formula is as follows:

[0097]

[0098] Among them, Cov ii It represents the correlation between the features of the i-th dimension and the i-th dimension, and the trace value reflects the sum of the characteristic variances of the matrix (the larger the trace, the more dispersed the feature distribution).

[0099] Furthermore, the threshold τ can be set based on historical data statistics:

[0100] If tr(new_Cov)≥τ, it is considered that the matrix contains sufficient semantic feature distribution information and the updated mean and covariance are retained;

[0101] If tr(new_Cov)≤τ, it means that the matrix does not contain enough information (possibly due to data noise or feature sparsity), and some updates are adjusted or discarded.

[0102] In the embodiment of the present invention, performing trace judgment on the updated covariance matrix to obtain a judgment result includes:

[0103] Calculating a trace of the updated covariance matrix, and determining a comparison result between the trace and a preset threshold;

[0104] If the trace is greater than a preset threshold, the updated data is audited, and the updated data is reacquired based on the audit result;

[0105] If the trace is smaller than a preset threshold, the updated data feature is combined with the updated mean vector and the updated covariance matrix to generate a probabilistic dynamic feature.

[0106] In an embodiment of the present invention, the judgment refers to the numerical value between the judgment trace and a preset threshold value, and the review refers to a systematic inspection of the legality, accuracy, completeness, compliance and other dimensions of the updated data to determine whether the data meets the preset standards or business requirements.

[0107] Specifically, the trace judgment is performed on the updated covariance matrix, and the specific explanation of the judgment result is consistent with the above content, which will not be repeated here.

[0108] In an embodiment of the present invention, by splicing linear transformations, the mean and covariance can integrate the latest data features and adapt to the dynamic changes in data semantics. The updated mean and covariance constitute probability distribution parameters, characterizing the semantic center and discreteness of the features, and providing more robust input for subsequent tasks.

[0109] In the embodiment of the present invention, through probabilistic dynamic feature extraction, the temporal dimension dynamics and uncertainty of the data are quantified into computable probabilistic features, providing more comprehensive input for subsequent predictions and decisions.

[0110] S5. Perform temporal dynamic fusion on the probabilistic dynamic features and the analytical feature data in combination with the temporal attention weight to obtain a fusion retrieval score, and filter the initial retrieval results according to the fusion retrieval score.

[0111] In an embodiment of the present invention, the diachronic dynamic fusion refers to dynamically associating and weightedly fusing the features of different time dimensions with timeliness weights in a time series to form an indicator for comprehensively measuring the value of data. The screening refers to sorting and filtering the initial search results according to the fusion search score, retaining data with a score higher than the threshold or a high ranking, and eliminating low-value results.

[0112] Specifically, through diachronic dynamic fusion, real-time features and historical features are weighted and integrated according to the time dimension to form a comprehensive score, ensuring that the results meet both "semantic relevance" and "strong timeliness". The initial search results are screened based on the fusion score, outdated or irrelevant data are eliminated, and finally high-value search results are returned.

[0113] In the embodiment of the present invention, the step of combining the probabilistic dynamic features and the analytical feature data with the temporal attention weight to perform diachronic dynamic fusion to obtain a fusion retrieval score includes:

[0114] Weighting the parsed feature data according to the temporal attention weight to obtain a weighted feature;

[0115] Calculating the confidence weight of the probabilistic dynamic feature according to the covariance matrix, and performing gated fusion on the weighted feature and the probabilistic dynamic feature according to the confidence weight to obtain a fused feature vector;

[0116] Obtaining a balance coefficient of a preset field, and calculating a similarity between the user query content and the fused feature vector;

[0117] The calculated similarity is numerically operated with the temporal attention weight, the confidence weight and the balance coefficient to obtain a fusion retrieval score.

[0118] In an embodiment of the present invention, the weighting refers to linearly weighting the parsed feature data using the temporal attention weight, strengthening the features with strong temporality and semantic relevance, and weakening the influence of outdated features. The calculation refers to calculating the confidence of the probabilistic dynamic features through the covariance matrix and quantifying the reliability of the features. The gated fusion refers to constructing a "gated" switch through the confidence weight to dynamically control the fusion ratio of the weighted features and the probabilistic dynamic features. The numerical operation refers to integrating the similarity score, temporal weight, confidence weight, and balance coefficient into a single fusion retrieval score through mathematical operations as the basis for result sorting.

[0119] Specifically, assuming that the analytical feature data is a vector F = [f1, f2, ..., f n ], the temporal attention weight is the scalar ω (or weight vector w), and the vector F is multiplied by the scalar ω to obtain the weighted feature; the inverse of the trace of the covariance matrix (the sum of the diagonal elements) represents the confidence weight, and the smaller the trace, the higher the confidence.

[0120] Furthermore, a gating signal, i.e., a gating vector g, is generated by the confidence weight c, g = σ(c)∈[0,1], where σ is a Sigmoid function that maps the weight to a gating value between 0 and 1. The formula for calculating the fused feature vector is as follows:

[0121] Fusion feature vector = g×D+(1-g)×W

[0122] Among them, the weighted feature is W and the probabilistic dynamic feature is D;

[0123] Furthermore, the user query content is converted into a semantic vector, and the cosine similarity is calculated with the fused feature vector to obtain a similarity score in the range of [-1,1]. The fused retrieval score = similarity score × [0.6 × time-sensitive attention weight + 0.4 × dynamic feature confidence weight], where 0.6 and 0.4 are the preset balance coefficients in the insurance field (which can be dynamically adjusted through user feedback).

[0124] S6. Optimize and adjust the initial search results according to preset user feedback to obtain target search results.

[0125] In an embodiment of the present invention, the optimization adjustment refers to modifying the order, content or relevance of the initial search results based on a preset user feedback mechanism (such as click, rating, favorite and other behavioral data) to make it more in line with the user's real needs.

[0126] Specifically, user feedback data is collected. For example, feedback includes: ratings actively provided by users (such as 1-5 stars), labels of "relevant / irrelevant", text comments, etc. If a result is clicked by a large number of users or has a high rating, its ranking in the result list will be improved; if a result is frequently marked as "irrelevant" or returns immediately after clicking (high bounce rate), its ranking will be lowered or even eliminated; the relevance score of the result is recalculated through feedback data, results with high negative feedback are eliminated, and similar but not retrieved content is supplemented.

[0127] In the embodiment of the present invention, real needs are accurately matched through user feedback, so as to avoid the disconnection between search results and intentions. The ranking is adjusted in real time according to click, rating and other behaviors to adapt to changes in demand, aggregate a large amount of user feedback, and strengthen common needs.

[0128] It can be seen that in the above scheme, for the target retrieval result business, the pre-acquired multimodal data is subjected to analytical feature extraction to obtain analytical feature data; the analytical feature data is subjected to timestamp extraction to obtain data timestamp; the temporal attention weight of the analytical feature data is calculated according to the data timestamp and the preset query time; the updated data is acquired in real time or at regular intervals, and the updated data is subjected to probabilistic dynamic feature extraction to obtain probabilistic dynamic features; the probabilistic dynamic features and the analytical feature data are combined with the temporal attention weight to perform diachronic dynamic fusion to obtain a fused retrieval score, and the initial retrieval results are screened according to the fused retrieval score; the initial retrieval results are optimized and adjusted according to preset user feedback to obtain the target retrieval results, and the updated data and multimodal data are introduced for comprehensive analysis to adapt to the dynamic change characteristics of the data and improve the accuracy of the retrieval results.

[0129] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0130] In one embodiment, a dynamic time sequence enhancement retrieval device is provided, and the dynamic time sequence enhancement retrieval device corresponds one-to-one to a dynamic time sequence enhancement retrieval method in the above embodiment. Figure 3 As shown, the dynamic time series enhanced retrieval device includes an extraction module 101, a timestamp extraction module 102, a conversion calculation module 103, an acquisition extraction module 104, a fusion screening module 105, and an adjustment module 106. The functional modules are described in detail as follows:

[0131] Extraction module 101, used to extract analytical features from pre-acquired multimodal data to obtain analytical feature data;

[0132] The timestamp extraction module 102 is used to extract the timestamp of the analyzed feature data to obtain a data timestamp;

[0133] A calculation module 103 is used to calculate the time-sensitive attention weight of the analyzed feature data according to the data timestamp and the preset query time;

[0134] An acquisition and extraction module 104 is configured to acquire update data in real time or periodically, and perform probabilistic dynamic feature extraction on the update data to obtain probabilistic dynamic features;

[0135] A fusion and screening module 105 is configured to perform a diachronic dynamic fusion of the probabilistic dynamic features and the analytical feature data in combination with the temporal attention weight to obtain a fusion retrieval score, and to screen initial retrieval results based on the fusion retrieval score;

[0136] The adjustment module 106 optimizes and adjusts the initial search results according to preset user feedback to obtain target search results.

[0137] In one embodiment, the extraction module 101, when performing analytical feature extraction on pre-acquired multimodal data to obtain analytical feature data, is configured to:

[0138] Perform standardized cleaning on the pre-acquired multimodal data to obtain standard cleaned data;

[0139] Performing submodal feature extraction on the standard cleaning data to obtain submodal features;

[0140] The sub-modal features are converted into the same dimensional space, and the conversion results are spliced ​​into analytical feature data.

[0141] In one embodiment, the timestamp extraction module 102, when extracting the timestamp from the parsed feature data to obtain the data timestamp, is configured to:

[0142] Extracting a time field from the parsed feature data;

[0143] Normalize the time field to obtain a time standard format and determine the time accuracy;

[0144] The timestamp of the parsed feature data is extracted according to the time specification format and the time precision to obtain a data timestamp.

[0145] In one embodiment, the calculation module 103, when calculating the temporal attention weight of the parsed feature data according to the data timestamp and the preset query time, is configured to:

[0146] Obtaining user query content, and converting the user query content and the parsing feature data into a query semantic vector and a parsing semantic vector respectively;

[0147] Calculating the semantic similarity between the query semantic vector and the parsed semantic vector;

[0148] Performing time difference processing on the data timestamp and the preset query time to obtain a time difference result, and determining the time attenuation factor of the analyzed feature data according to the time difference result;

[0149] The temporal attention weight of the parsed feature data is configured according to the time decay factor and the semantic similarity result.

[0150] In one embodiment, the acquisition and extraction module 104, when performing probabilistic dynamic feature extraction on the update data to obtain the probabilistic dynamic feature, is configured to:

[0151] Extracting semantic features from the update data to obtain update data features;

[0152] Obtaining the mean vector and covariance matrix of the entity embedding, performing a linear transformation on the updated data features and the mean vector to obtain an updated mean vector;

[0153] Converting the covariance matrix into a matrix vector, performing linear transformation on the updated data features and the matrix vector to obtain an updated covariance matrix;

[0154] Performing trace judgment on the updated covariance matrix to obtain a judgment result, and combining the updated data feature with the updated mean vector and the updated covariance matrix according to the judgment result to generate a probabilistic dynamic feature.

[0155] In one embodiment, when converting the projection result into a gated weight value through a nonlinear activation function and a preset bias parameter, it is used to:

[0156] Add the preset bias parameters to the projection result to obtain the added projection result;

[0157] Performing nonlinear mapping on the added projection result through a nonlinear activation function to obtain an initial gating weight value;

[0158] The initial gating weight value is normalized to obtain a gating weight value.

[0159] In one embodiment, when performing trace judgment on the updated covariance matrix and obtaining a judgment result, it is used to:

[0160] Calculating a trace of the updated covariance matrix and determining whether the trace is greater than a preset threshold;

[0161] If the trace is greater than a preset threshold, the updated data is audited, and the updated data is reacquired based on the audit result;

[0162] If the trace is less than or equal to a preset threshold, the updated data feature is combined with the updated mean vector and the updated covariance matrix to generate a probabilistic dynamic feature.

[0163] In one embodiment, the fusion screening module 105, when performing temporal dynamic fusion by combining the probabilistic dynamic features and the analytical feature data with the temporal attention weight to obtain a fusion retrieval score, is configured to:

[0164] Weighting the parsed feature data according to the temporal attention weight to obtain a weighted feature;

[0165] Calculating the confidence weight of the probabilistic dynamic feature according to the covariance matrix, and performing gated fusion on the weighted feature and the probabilistic dynamic feature according to the confidence weight to obtain a fused feature vector;

[0166] Obtaining a balance coefficient of a preset field, and calculating a similarity between the user query content and the fused feature vector;

[0167] The calculated similarity is numerically operated with the temporal attention weight, the confidence weight and the balance coefficient to obtain a fusion retrieval score.

[0168] The present invention provides a dynamic time series enhanced retrieval device, which, for a target retrieval result business, performs analytical feature extraction on pre-acquired multimodal data to obtain analytical feature data; performs timestamp extraction on the analytical feature data to obtain a data timestamp; calculates the time-sensitive attention weight of the analytical feature data based on the data timestamp and a preset query time; obtains updated data in real time or at a fixed time, performs probabilistic dynamic feature extraction on the updated data to obtain a probabilistic dynamic feature; dynamically fuses the probabilistic dynamic feature and the analytical feature data in combination with the time-sensitive attention weight to obtain a fused retrieval score, and screens initial retrieval results based on the fused retrieval score; optimizes and adjusts the initial retrieval results based on preset user feedback to obtain a target retrieval result, and adapts to the dynamic change characteristics of the data by introducing updated data and multimodal data for comprehensive analysis, thereby improving the accuracy of the retrieval results.

[0169] For the specific definition of a dynamic time series enhancement retrieval device, please refer to the definition of a dynamic time series enhancement retrieval method above, which will not be repeated here. The various modules in the above-mentioned dynamic time series enhancement retrieval device can be implemented in whole or in part by software, hardware, and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0170] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 4As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external client via a network connection. When the computer program is executed by the processor, it realizes the functions or steps on the server side of a dynamic timing enhancement retrieval method.

[0171] In one embodiment, a computer device is provided. The computer device may be a client, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, memory, network interface, display screen, and input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps on the client side of a dynamic time-series enhanced retrieval method.

[0172] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps are performed:

[0173] Performing analytical feature extraction on pre-acquired multimodal data to obtain analytical feature data;

[0174] Extracting a timestamp from the analyzed feature data to obtain a data timestamp;

[0175] Calculating the time-sensitive attention weight of the parsed feature data according to the data timestamp and the preset query time;

[0176] Acquire update data in real time or periodically, and perform probabilistic dynamic feature extraction on the update data to obtain probabilistic dynamic features;

[0177] Performing a diachronic dynamic fusion of the probabilistic dynamic features and the analytical feature data in combination with the temporal attention weight to obtain a fusion retrieval score, and screening initial retrieval results based on the fusion retrieval score;

[0178] The initial search results are optimized and adjusted according to preset user feedback to obtain target search results.

[0179] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0180] Performing analytical feature extraction on pre-acquired multimodal data to obtain analytical feature data;

[0181] Extracting a timestamp from the analyzed feature data to obtain a data timestamp;

[0182] Calculating the time-sensitive attention weight of the parsed feature data according to the data timestamp and the preset query time;

[0183] Acquire update data in real time or periodically, and perform probabilistic dynamic feature extraction on the update data to obtain probabilistic dynamic features;

[0184] Performing a diachronic dynamic fusion of the probabilistic dynamic features and the analytical feature data in combination with the temporal attention weight to obtain a fusion retrieval score, and screening initial retrieval results based on the fusion retrieval score;

[0185] The initial search results are optimized and adjusted according to preset user feedback to obtain target search results.

[0186] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can be found in the relevant descriptions of the server side and the client side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0187] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0188] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0189] The above-described embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. If software tools or components other than those of the company appear in the application embodiments, they are merely used for illustration and do not represent actual use. Although the present invention has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above-mentioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A dynamic time series enhanced retrieval method, characterized in that: include: Performing analytical feature extraction on pre-acquired multimodal data to obtain analytical feature data; Extracting a timestamp from the analyzed feature data to obtain a data timestamp; Calculating the time-sensitive attention weight of the parsed feature data according to the data timestamp and the preset query time; Acquire update data in real time or periodically, and perform probabilistic dynamic feature extraction on the update data to obtain probabilistic dynamic features; Performing a diachronic dynamic fusion of the probabilistic dynamic features and the analytical feature data in combination with the temporal attention weight to obtain a fusion retrieval score, and screening initial retrieval results based on the fusion retrieval score; The initial search results are optimized and adjusted according to preset user feedback to obtain target search results.

2. The dynamic time series enhanced retrieval method according to claim 1, characterized in that: The extracting analytical features from the pre-acquired multimodal data to obtain analytical feature data includes: Perform standardized cleaning on the pre-acquired multimodal data to obtain standard cleaned data; Performing submodal feature extraction on the standard cleaning data to obtain submodal features; The sub-modal features are converted into the same dimensional space, and the conversion results are spliced ​​into analytical feature data.

3. The dynamic time series enhanced retrieval method according to claim 1, wherein: The extracting the timestamp from the parsed feature data to obtain the data timestamp includes: Extracting a time field from the parsed feature data; Normalize the time field to obtain a time standard format and determine the time accuracy; The timestamp of the parsed feature data is extracted according to the time specification format and the time precision to obtain a data timestamp.

4. The dynamic time series enhanced retrieval method according to claim 1, wherein: The calculating the time-sensitive attention weight of the parsed feature data according to the data timestamp and the preset query time includes: Obtaining user query content, and converting the user query content and the parsing feature data into a query semantic vector and a parsing semantic vector respectively; Calculating the semantic similarity between the query semantic vector and the parsed semantic vector; Performing time difference processing on the data timestamp and the preset query time to obtain a time difference result, and determining the time attenuation factor of the analyzed feature data according to the time difference result; The temporal attention weight of the parsed feature data is configured according to the time decay factor and the semantic similarity result.

5. The dynamic time series enhanced retrieval method according to claim 4, characterized in that: The extracting probabilistic dynamic features from the update data to obtain probabilistic dynamic features includes: Extracting semantic features from the update data to obtain update data features; Obtaining the mean vector and covariance matrix of the entity embedding, performing a linear transformation on the updated data features and the mean vector to obtain an updated mean vector; Converting the covariance matrix into a matrix vector, performing linear transformation on the updated data features and the matrix vector to obtain an updated covariance matrix; Performing trace judgment on the updated covariance matrix to obtain a judgment result, and combining the updated data feature with the updated mean vector and the updated covariance matrix according to the judgment result to generate a probabilistic dynamic feature.

6. The dynamic time series enhanced retrieval method according to claim 5, characterized in that: The performing trace judgment on the updated covariance matrix to obtain a judgment result includes: Calculating a trace of the updated covariance matrix and determining whether the trace is greater than a preset threshold; If the trace is greater than a preset threshold, the updated data is audited, and the updated data is reacquired based on the audit result; If the trace is less than or equal to a preset threshold, the updated data feature is combined with the updated mean vector and the updated covariance matrix to generate a probabilistic dynamic feature.

7. The dynamic time series enhanced retrieval method according to claim 5, characterized in that: The step of combining the probabilistic dynamic features and the analytical feature data with the temporal attention weight to perform diachronic dynamic fusion to obtain a fusion retrieval score includes: Weighting the parsed feature data according to the temporal attention weight to obtain a weighted feature; Calculating the confidence weight of the probabilistic dynamic feature according to the covariance matrix, and performing gated fusion on the weighted feature and the probabilistic dynamic feature according to the confidence weight to obtain a fused feature vector; Obtaining a balance coefficient of a preset field, and calculating a similarity between the user query content and the fused feature vector; The calculated similarity is numerically operated with the temporal attention weight, the confidence weight and the balance coefficient to obtain a fusion retrieval score.

8. A dynamic time sequence enhanced retrieval device, characterized in that: include: An extraction module is used to extract analytical features from pre-acquired multimodal data to obtain analytical feature data; A timestamp extraction module, configured to extract timestamps from the parsed feature data to obtain data timestamps; A calculation module, configured to calculate a time-sensitive attention weight of the parsed feature data based on the data timestamp and a preset query time; An acquisition and extraction module is used to acquire update data in real time or periodically, and perform probabilistic dynamic feature extraction on the update data to obtain probabilistic dynamic features; a fusion screening module, configured to perform a diachronic dynamic fusion of the probabilistic dynamic features and the analytical feature data in combination with the temporal attention weight to obtain a fusion retrieval score, and to screen the initial retrieval results according to the fusion retrieval score; The adjustment module optimizes and adjusts the initial search results according to preset user feedback to obtain target search results.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the dynamic temporal enhancement retrieval method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the dynamic temporal enhancement retrieval method according to any one of claims 1 to 7 is implemented.

Images