Target knowledge data identification and extraction method based on depth algorithm optimization

Abstract

The invention discloses a target knowledge data recognition and extraction method optimized by a depth algorithm, relates to the technical field of knowledge extraction, and solves the problem of overlong subsequent content output time caused by relatively slow target data comparison process. Text semantics are converted into quantifiable geometric features, compared with traditional keyword matching, implicit association (such as synonyms and context dependence) between semantics can be captured, recognition features are mapped to a two-dimensional coordinate system, a target object is rapidly matched through the Euclidean distance, the efficiency is improved by 10-20 times compared with full-amount traversal query, clustering analysis based on historical extracted data is achieved, and the query efficiency is improved. Through calculation of a mean value time column and a feature value TT, the optimal computing power configuration is automatically identified, and for time feature sorting of different extraction objects, tasks which are short in time consumption are preferentially processed, so that the queue is prevented from being blocked by long tasks.

Description

Technical Field

[0001] The present invention relates to the technical field of knowledge extraction, and in particular to a method for identifying and extracting target knowledge data using deep algorithm optimization. Background Art

[0002] With the rapid development of big data and artificial intelligence technologies, efficiently identifying and extracting target knowledge from massive amounts of unstructured data (such as text, images, audio, etc.) has become a core requirement in fields such as knowledge management and intelligent decision-making. Patent application CN110689497B discloses a target recognition-based data extraction device and method. The device comprises: a candlelight identification device for performing candlelight target recognition on a merged image based on the geometric shape of candlelight, thereby issuing a candlelight acquisition command when a candlelight target is identified, and otherwise issuing a candlelight non-acquisition command; and a value setting device for, upon receiving the candlelight acquisition command, determining a corresponding white balance based on the number of candlelight targets in the merged image and selecting a compensation value corresponding to the white balance value as a reference compensation value to be sent to the notebook front-end imaging device. The target recognition-based data extraction device and method of this invention have clear logic and are simple to operate. By determining the corresponding white balance based on the number of candlelight targets in the image and selecting a compensation value corresponding to the white balance value as a reference compensation value, adaptive adjustment of the image signal is achieved.

[0003] For the target knowledge data extraction process, generally according to the corresponding instructions, the existence of the associated target object is confirmed, and then the associated target object is extracted. This part of the extraction process requires a comprehensive scan of the database, and adopts a method of data similarity comparison and verification before and after. The corresponding comparison process is relatively slow, which will cause the subsequent output time to be too long and cannot achieve a better target knowledge data identification and extraction process. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a target knowledge data identification and extraction method optimized by a deep algorithm, which solves the problem that the target data comparison process is relatively slow, resulting in excessively long subsequent content output time.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a target knowledge data identification and extraction method optimized by a deep algorithm, comprising the following steps: Step 1: Based on the input index target, the associated feature source data is confirmed. Based on the semantic vector associated with the feature source data, the identification feature is locked. Based on the identification feature, the object is locked and extracted from the cloud database; The specific method of locking the corresponding identification features of feature source data is as follows: Based on the input index target, confirm the feature source data associated with the index target, where the feature source data is preset data; According to the preset semantic vector map, the semantic vector associated with each single data in the corresponding feature source data is confirmed, and according to the data sorting method of the feature source data, several semantic vectors are sorted to confirm the semantic vector set belonging to this feature source data. The total number of vectors in the semantic vector set is recorded as G i , where i represents different index targets, and then confirm the angle between adjacent vectors in the semantic vector set: combine the end point of the previous vector in the adjacent vector with the initial point of the next vector to confirm the angle J between the two vectors i-k , where k represents different adjacent vectors, and then the several angles J associated with several adjacent vectors i-k Perform mean processing to confirm the feature mean angle JJ belonging to the corresponding semantic vector set i ; Use: SS i =G i ×C1+JJ i ×C2 confirms the parameter feature SS associated with the corresponding feature source data i , where C1 and C2 are preset fixed coefficient factors, and are based on the total number of vectors G i And parameter characteristics SS i Generate identification features (G i , SS i ); The specific method for locking and extracting objects from the cloud database based on identification features is as follows: Confirm different target objects from the cloud database, then confirm the source data associated with each target object. Based on the data sorting method of the single source data, confirm the semantic vector set associated with the single source data. Use the same confirmation method as the identification feature of the feature source data to confirm the identification features associated with the single source data. Generate a set of two-dimensional coordinate systems, whose horizontal coordinate axis is the number of vectors and vertical coordinate axis is the parameter feature. Confirm the identification point in the two-dimensional coordinate system based on the identification feature associated with the feature source data, and record the confirmed identification point as the standard point. Then, for the identification features associated with different single-segment source data in the target object, lock the data point in the two-dimensional coordinate system and record the straight-line distance Lq associated with different data points and the standard point, where q represents different data points. If there is a data point with Lq≤Y1, then this target object is recorded as the extraction object, where Y1 is the preset value. Otherwise, no mark is made. Processing several target objects in sequence, identifying and extracting objects from the several target objects and confirming them; Step 2: Based on the determined extraction object, identify the extraction features associated with the corresponding extraction object from the historical data associated with the corresponding extraction object. The specific method is as follows: Based on the determined extraction object, confirm the historical data that existed when the corresponding extraction object was extracted, identify the computing power characteristics associated with the single extraction process from the historical data, and the process time associated with the single extraction process; Based on the confirmed hash rate characteristics and process times, average the different process times associated with the same hash rate characteristics to determine the average time associated with the corresponding hash rate characteristics. Processing is performed based on the mean time associated with different computing power features: sort the associated mean times in ascending order according to the computing power features to generate a mean time column, and confirm the features of the mean times in the mean time column: mark the corresponding mean time as STz, record the mean time group before STz as QTz, and record the mean time group after STz as HTz. Use: (QTz-STz)+STz+(HTz-STz)=TT to lock the eigenvalue TT associated with the corresponding mean time, and then select the minimum value from the eigenvalues ​​TT associated with different mean times. Record the mean time associated with the minimum value as the extraction time, and record the computing power features associated with the selected time as the extracted computing power. The confirmed extraction computing power and extraction time are the extraction features associated with the corresponding extraction object; Step 3: Based on the different extraction features confirmed for different extraction objects and the total computing power held by the corresponding platform, the extraction logic of different extraction objects is confirmed, and multiple groups of extraction objects are extracted and displayed based on the confirmed extraction logic. The specific method is as follows: The total computing power held by the corresponding platform is calibrated as Zk. Then, based on the different extraction features associated with different extraction objects, the extraction time is determined from the extraction features. Several extraction objects are sorted from the smallest to the largest extraction time to generate an extraction object sequence. Then, based on the extraction computing power associated with the extraction features of the corresponding extraction objects, the extraction computing power associated with the first group of extraction objects is prioritized, and the corresponding computing power is allocated from the total computing power to the first group of extraction objects. Then, based on the computing power SL1 allocated to the first group of extraction objects, the computing power difference between the extraction computing power of the first group of extraction objects and the second group of extraction objects is determined. The computing power difference = the extraction computing power of the second group of extraction objects - the extraction computing power of the first group of extraction objects. Then, the computing power corresponding to the computing power difference is allocated to the second group of extraction objects. If the computing power difference is less than 0, no allocation is performed: After the previous group of extraction objects is extracted, the extraction computing power of the previous group will be automatically allocated to the next group of extraction objects. The computing power of the current extraction object will be adjusted based on the computing power used by the previous group of extraction computing power and the computing power difference associated with the current extraction object. The extraction logic belonging to the extraction object sequence is generated, and the extraction is performed according to the generated extraction logic.

[0006] The present invention provides a method for identifying and extracting target knowledge data using a deep algorithm optimization. Compared with the existing technology, it has the following advantages: This method converts text semantics into quantifiable geometric features by constructing a semantic vector set and calculating the feature mean angle of adjacent vectors. Compared with traditional keyword matching, it can capture implicit associations between semantics (such as synonyms and contextual dependencies), map recognition features to a two-dimensional coordinate system, and quickly match target objects using Euclidean distance, improving efficiency by 10-20 times compared to full-text traversal queries. Based on cluster analysis of historical extraction data, the optimal computing power configuration is automatically identified through mean time column and eigenvalue TT calculation. The time characteristics of different extraction objects are sorted, and short-time tasks are prioritized to avoid long tasks blocking the queue. Through the three-layer architecture of "semantic quantification-historical optimization-dynamic allocation", the efficiency of the entire link from data identification to resource scheduling is improved. It is particularly suitable for real-time knowledge extraction scenarios of large-scale unstructured data, and provides high-performance underlying technical support for enterprise-level knowledge management systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 Schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0008] 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 the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0009] Example 1: Please refer to Figure 1 , this application provides a target knowledge data identification and extraction method optimized by a deep algorithm, comprising the following steps: Step 1: Based on the input index target, the associated feature source data is confirmed. Based on the semantic vector associated with the feature source data, the identification feature is locked. Then, based on the identification feature, the extraction object is locked from the cloud database. Specifically, the corresponding index target is input by the operator, and each index target has preset feature source data, which is prepared in advance by relevant personnel. Different feature source data have different semantic vectors. Based on the data features of the corresponding feature source data confirmed by the corresponding semantic vector, the data features associated with other objects in the cloud database are confirmed. According to the specific identification confirmation process, the extraction object to be extracted is identified from the cloud database; The specific method of locking the corresponding identification features of the feature source data is as follows: Based on the input index target, confirm the feature source data associated with the index target, where the feature source data is preset data, and different index targets are associated with different preset feature source data; According to the preset semantic vector map, the semantic vector associated with each single data in the corresponding feature source data is confirmed, and according to the data sorting method of the feature source data, several semantic vectors are sorted to confirm the semantic vector set belonging to this feature source data. The total number of vectors in the semantic vector set is recorded as G i , where i represents different index targets, and then confirm the angle between adjacent vectors in the semantic vector set: combine the end point of the previous vector in the adjacent vector with the initial point of the next vector to confirm the angle J between the two vectors i-k , where k represents different adjacent vectors, and then the several angles J associated with several adjacent vectors i-k Perform mean processing to confirm the feature mean angle JJ belonging to the corresponding semantic vector set i , the confirmation direction of each adjacent vector is the same, and the angle is confirmed in the clockwise direction; Use: SS i =G i ×C1+JJ i ×C2 confirms the parameter feature SS associated with the corresponding feature source data i , where C1 and C2 are preset fixed coefficient factors, and are based on the total number of vectors G i And parameter characteristics SS i Generate identification features (G i , SS i ), the specific value is determined by the operator based on experience, C1 is generally 0.587, and C2 is generally 0.413; The specific method for locking and extracting objects from the cloud database based on identification features is as follows: Identify different target objects from the cloud database, then confirm the source data associated with each target object. Based on the data sorting method of the single-segment source data (with delimiters, each single-segment data is separated by the corresponding delimiter), confirm the semantic vector set associated with the single-segment source data. Use the same confirmation method as the identification feature of the feature source data to confirm the identification features associated with the single-segment source data. Generate a set of two-dimensional coordinate systems, the horizontal axis of which is the number of vectors, and the vertical axis is the parameter feature. Based on the identification features associated with the feature source data, the identification points are confirmed in the two-dimensional coordinate system, and the confirmed identification points are recorded as standard points. Then, for the identification features associated with different single-segment source data in the target object, the data points are locked in the two-dimensional coordinate system, and the straight-line distance Lq associated with different data points and the standard points is recorded, where q represents different data points. If there is a data point with Lq≤Y1, the target object is recorded as the extraction object, otherwise, no mark is made, where Y1 is a preset value, and its specific value is determined by the operator based on experience. The straight-line distance Lq associated with the two points is confirmed using the Pythagorean theorem. The identification features associated with the proposed feature source data are (G i , SS i ), the identification features associated with the corresponding data points are (T1q, Q1q), using: Lock the linear distance Lq associated with the two corresponding points; Processing several target objects in sequence, identifying and extracting objects from the several target objects and confirming them; Specifically, different index targets have different feature source data (which can be understood as feature codes, or feature data needed when the corresponding target is running). According to the semantic vector associated with each data, the corresponding semantic vector set is confirmed, and based on the angle and change characteristics generated between the corresponding semantic vectors, the corresponding angle and the associated feature mean angle can be confirmed. Then, according to the number of corresponding vectors and the determined feature mean angle, the corresponding parameter feature can be locked. Different target objects also have several different single-source data, and each single-source data has different identification features. According to the distance between the corresponding points, the specific confirmation of the extracted object can be made from the target object, so that the corresponding extracted object can be quickly locked, and then the corresponding extracted object can be extracted and displayed accordingly.

[0010] Step 2: Based on the determined extraction object, identify the extraction features associated with the corresponding extraction object from the historical data associated with the corresponding extraction object, wherein the extraction features include extraction time and extraction computing power; Among them, the specific method of confirming the extraction features corresponding to different extraction objects is: Based on the determined extraction object, confirm the historical data that existed when the corresponding extraction object was extracted, identify the computing power characteristics associated with the single extraction process from the historical data, and the process time associated with the single extraction process; Based on the confirmed hash rate characteristics and process times, average the different process times associated with the same hash rate characteristics to determine the average time associated with the corresponding hash rate characteristics. If there is only a single process time, the corresponding process time is the average time. Processing is performed based on the mean time associated with different computing power features: sort the associated mean times in ascending order according to the computing power features to generate a mean time column, and confirm the features of the mean times in the mean time column: mark the corresponding mean time as STz, record the mean time group before STz as QTz, and record the mean time group after STz as HTz. Use: (QTz-STz)+STz+(HTz-STz)=TT to lock the eigenvalue TT associated with the corresponding mean time, and then select the minimum value from the eigenvalues ​​TT associated with different mean times. Record the mean time associated with the minimum value as the extraction time, and record the computing power features associated with the selected time as the extracted computing power. The confirmed extraction computing power and extraction time are the extraction features associated with the corresponding extraction object; Specifically, different extraction objects have different extraction situations during the extraction process. Different extraction situations are associated with different computing power values ​​and time values. Then, based on the specific change status of the corresponding computing power values ​​and time values, the relatively clustered extraction computing power and extraction time are confirmed, which facilitates the subsequent output of different extraction objects and ensures the output rate of the extraction objects, so that multiple groups of extraction objects can achieve the fastest output efficiency when outputting, and their output time can be fully reduced.

[0011] Step 3: Based on the different extraction features confirmed for different extraction objects and the total computing power held by the corresponding platform, the extraction logic of different extraction objects is confirmed, and the extraction and display of multiple groups of extraction objects are performed based on the confirmed extraction logic; The specific method for confirming the extraction logic is as follows: The total computing power held by the corresponding platform is calibrated as Zk. Then, based on the different extraction features associated with different extraction objects, the extraction time is determined from the extraction features. Several extraction objects are sorted from the smallest to the largest extraction time to generate an extraction object sequence. Then, based on the extraction computing power associated with the extraction features of the corresponding extraction objects, the extraction computing power associated with the first group of extraction objects is prioritized, and the corresponding computing power is allocated from the total computing power to the first group of extraction objects. Then, based on the computing power SL1 allocated to the first group of extraction objects, the computing power difference between the extraction computing power of the first group of extraction objects and the second group of extraction objects is determined. The computing power difference = the extraction computing power of the second group of extraction objects - the extraction computing power of the first group of extraction objects. Then, the computing power corresponding to the computing power difference is allocated to the second group of extraction objects. If the computing power difference is less than 0, no allocation is performed: After the previous set of extraction objects is extracted, the extraction computing power of the previous set will be automatically allocated to the next set of extraction objects. The computing power of the current extraction object will be adjusted based on the computing power used by the previous set of extraction objects and the computing power difference associated with the current extraction object. The extraction logic belonging to the extraction object sequence will be generated, and extraction will be carried out according to the generated extraction logic to ensure the extraction rate. Specifically, different extraction objects are sorted according to the corresponding extraction time, so that the corresponding object sequence can be confirmed; According to the extraction computing power associated with different extraction objects in the object sequence, it is planned to be {30, 25, 35, 40, 42, 38}. Then the first extraction object is allocated 30, the second extraction object is not allocated and directly uses 30, and the third extraction object is allocated a computing power difference of 5. Add 5 to the original computing power of 30 of the previous group of extraction objects, then the computing power difference allocated to the fourth extraction object is also 5, the fifth extraction object is 2, the sixth extraction object is not allocated, and so on, until all extraction objects are allocated, and the extraction rate of the extraction objects is fully guaranteed, and the extraction time is also reduced to a minimum, achieving the optimal extraction processing effect and ensuring the specific process of the extraction processing.

[0012] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0013] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A target knowledge data identification and extraction method optimized by deep algorithm, characterized in that: The following steps are involved: Step 1: Based on the input index target, the associated feature source data is confirmed. Based on the semantic vector associated with the feature source data, the identification feature is locked. Based on the identification feature, the object is locked and extracted from the cloud database; Step 2: Based on the determined extraction object, identify the extraction features associated with the corresponding extraction object from the historical data associated with the corresponding extraction object; Step 3: Based on the different extraction features confirmed by different extraction objects and the total computing power held by the corresponding platform, the extraction logic of different extraction objects is confirmed, and multiple groups of extraction objects are extracted and displayed based on the confirmed extraction logic.

2. The target knowledge data identification and extraction method optimized by deep algorithm according to claim 1 is characterized in that: In step 1, the specific method of locking the identification feature corresponding to the feature source data is: Based on the input index target, confirm the feature source data associated with the index target, where the feature source data is preset data; According to the preset semantic vector map, the semantic vector associated with each single data in the corresponding feature source data is confirmed, and according to the data sorting method of the feature source data, several semantic vectors are sorted to confirm the semantic vector set belonging to this feature source data. The total number of vectors in the semantic vector set is recorded as G i , where i represents different index targets, and then confirm the angle between adjacent vectors in the semantic vector set: combine the end point of the previous vector in the adjacent vector with the initial point of the next vector to confirm the angle J between the two vectors i-k , where k represents different adjacent vectors, and then the several angles J associated with several adjacent vectors i-k Perform mean processing to confirm the feature mean angle JJ belonging to the corresponding semantic vector set i ; Use: SS i =G i ×C1+JJ i ×C2 confirms the parameter feature SS associated with the corresponding feature source data i , where C1 and C2 are preset fixed coefficient factors, and are based on the total number of vectors G i And parameter characteristics SS i Generate identification features (G i , SS i ).

3. The target knowledge data identification and extraction method optimized by deep algorithm according to claim 2 is characterized in that: In step 1, the specific method of locking and extracting the object from the cloud database based on the identification features is as follows: Confirm different target objects from the cloud database, then confirm the source data associated with each target object. Based on the data sorting method of the single source data, confirm the semantic vector set associated with the single source data. Use the same confirmation method as the identification feature of the feature source data to confirm the identification features associated with the single source data. Generate a set of two-dimensional coordinate systems, whose horizontal coordinate axis is the number of vectors and the vertical coordinate axis is the parameter feature. Confirm the identification point in the two-dimensional coordinate system based on the identification feature associated with the feature source data, and record the confirmed identification point as the standard point. Then, for the identification features associated with different single-segment source data in the target object, lock the data point in the two-dimensional coordinate system and record the linear distance Lq associated with different data points and the standard point, where q represents different data points. If there is a data point with Lq≤Y1, then this target object is recorded as the extraction object, where Y1 is a preset value. Process several target objects in sequence, identify and extract objects from the target objects, and confirm them.

4. The target knowledge data identification and extraction method optimized by deep algorithm according to claim 3 is characterized in that: If there is no data point where Lq≤Y1, no marking is performed.

5. The target knowledge data identification and extraction method optimized by deep algorithm according to claim 1 is characterized in that: In step 2, the specific method for confirming the extraction features corresponding to different extraction objects is: Based on the determined extraction object, confirm the historical data that existed when the corresponding extraction object was extracted, identify the computing power characteristics associated with the single extraction process from the historical data, and the process time associated with the single extraction process; Based on the confirmed hash rate characteristics and process times, average the different process times associated with the same hash rate characteristics to determine the average time associated with the corresponding hash rate characteristics. Processing is performed based on the mean time associated with different computing power characteristics: sort the associated mean times according to the computing power characteristics from small to large, generate a mean time column, and perform feature confirmation on the mean time in the mean time column: mark the corresponding mean time as STz, record the mean time group before STz as QTz, and record the mean time group after STz as HTz. Use: (QTz-STz)+STz+(HTz-STz)=TT to lock the characteristic value TT associated with the corresponding mean time, and then select the minimum value from the characteristic values ​​TT associated with different mean times, record the mean time associated with the minimum value as the extraction time, and record the computing power characteristics associated with the selected time as the extraction computing power. The confirmed extraction computing power and extraction time are the extraction features associated with the corresponding extraction object.

6. The target knowledge data identification and extraction method optimized by deep algorithm according to claim 1 is characterized in that: In step 3, the specific method for confirming the extraction logic is: The total computing power held by the corresponding platform is calibrated as Zk. Then, based on the different extraction features associated with different extraction objects, the extraction time is determined from the extraction features. Several extraction objects are sorted from the smallest to the largest extraction time to generate an extraction object sequence. Then, based on the extraction computing power associated with the extraction features of the corresponding extraction objects, the extraction computing power associated with the first group of extraction objects is prioritized, and the corresponding computing power is allocated from the total computing power to the first group of extraction objects. Then, based on the computing power SL1 allocated to the first group of extraction objects, the computing power difference between the extraction computing power of the first group of extraction objects and the second group of extraction objects is determined. The computing power difference = the extraction computing power of the second group of extraction objects - the extraction computing power of the first group of extraction objects. Then, the computing power corresponding to the computing power difference is allocated to the second group of extraction objects: After the previous group of extraction objects is extracted, the extraction computing power of the previous group will be automatically allocated to the next group of extraction objects. The computing power of the current extraction object will be adjusted based on the computing power used by the previous group of extraction computing power and the computing power difference associated with the current extraction object. The extraction logic belonging to the extraction object sequence is generated, and the extraction is performed according to the generated extraction logic.

7. The target knowledge data identification and extraction method optimized by deep algorithm according to claim 6 is characterized in that: If the computing power difference is less than 0, no allocation will be made.

Images