A deep learning-based multi-layer semantic feature extraction method

Abstract

This invention discloses a deep learning-based method for extracting multi-layer semantic features, specifically relating to the field of multi-layer semantic feature extraction technology. The method includes: acquiring semantic sample data, standardizing it, and mapping it to multiple semantic levels; performing semantic representation calculations on each semantic level to determine optional cross-layer connections; applying hierarchical order constraints to the optional cross-layer connections to construct a set of candidate semantic paths, which serve as candidate paths in an ant colony algorithm; initializing pheromone values ​​for each candidate path to form a candidate path set and a corresponding pheromone set; and using cross-layer semantic transmission paths as evaluable objects and continuously selecting, strengthening, and eliminating semantic paths based on prediction error feedback, thereby dynamically adjusting the generation process of multi-layer semantic features according to changes in data semantics, thus solving the problem of hidden semantic distortion caused by static solidification of semantic paths.

Description

Technical Field

[0001] This invention relates to the field of multi-layer semantic feature extraction technology, and more specifically, to a multi-layer semantic feature extraction method based on deep learning. Background Technology

[0002] Existing deep learning-based multi-layer semantic feature extraction methods typically feed input data into a pre-defined hierarchical computation structure, output hierarchical semantic representations at different network layers, and update parameters in reverse during the training phase by predicting the error between the output and the labeled values ​​to obtain semantic representation capabilities. Subsequently, during the runtime phase, the parameters are fixed and semantic features are output based on the result of a single forward computation. In this technical system, the semantic transmission relationship between layers is usually determined by the network structure and training results. Once training is completed, the inter-layer connections and their semantic transmission paths are basically stable, and semantic computation can only be completed along the predetermined path during the runtime phase. In practical applications, the vocabulary used, contextual combinations, business terms and collection sources of input data will change slowly over time. Such changes often do not immediately cause a significant drop in overall accuracy, but they will first change the triggering method of the underlying features and gradually affect the semantic transmission between layers, causing the semantic transmission path that was originally effective on the training data to lose support in the running environment. Since the computation is still completed along a stable path during the runtime phase, the high-level semantic output may remain formally stable, thus masking the fact that the underlying support has changed. As a result, the system continues to output multi-layered semantic features that appear stable on the surface but whose semantic basis has shifted. It is evident that existing multi-layer semantic feature extraction methods treat semantic paths as static structures formed during the training phase, lacking a mechanism for continuous selection, reinforcement, and elimination of semantic paths during training and operation. This results in semantic output following the old path even when semantic support conditions change, leading to hidden and cumulative semantic distortion. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a multi-layer semantic feature extraction method based on deep learning. By taking the cross-layer semantic transmission path as an evaluable object and continuously selecting, strengthening, and eliminating semantic paths based on prediction error feedback, the generation process of multi-layer semantic features is dynamically adjusted with changes in data semantics, thereby solving the problem of hidden semantic distortion caused by the static solidification of semantic paths.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a multi-layer semantic feature extraction method based on deep learning, comprising: S1. Obtain semantic sample data, perform normalization processing, and map it into multiple semantic levels; perform semantic representation calculation on each semantic level to determine optional cross-level connections; S2. Apply hierarchical order constraints to optional cross-layer connections, construct a set of candidate semantic paths, and use them as candidate paths in the ant colony algorithm; initialize pheromone values ​​for each candidate path to form a set of candidate paths and a corresponding set of pheromones. S3. Based on semantic sample data, perform semantic representation calculation along cross-layer connections to generate semantic representations corresponding to different semantic levels; perform path selection on the candidate path set according to the pheromone value corresponding to the candidate path to determine the target semantic path and output the prediction result. S4. Obtain the target value corresponding to the prediction result, calculate the error between the prediction result and the target value, and synchronously update the calculation parameters corresponding to each semantic level of the target semantic path; based on the contribution of the target semantic path to the reduction of error, update the pheromone values ​​corresponding to the target semantic path and other semantic paths. S5. Continuously input semantic sample data and repeatedly execute the update process so that the calculation parameters corresponding to the semantic level and the pheromone values ​​corresponding to the candidate path gradually converge under the joint action, forming a hierarchical semantic calculation structure. S6. Based on the hierarchical semantic computing structure, candidate paths with pheromone values ​​in the preset upper limit range are determined as stable semantic paths, and the multi-layer semantic representations covered by the stable semantic paths are output as multi-layer semantic feature extraction results, while completing the formation of the deep learning computing structure.

[0005] In a preferred embodiment, in S1, the process of determining optional cross-layer connections includes: S1-1. Take each semantic sample in the semantic sample data as the current semantic sample to be processed, read the original values ​​of each semantic field in the current semantic sample to be processed, and convert the original values ​​of each semantic field into a preset data representation format. When the original value of any semantic field does not meet the data representation format requirements, the currently processed semantic sample is marked as an invalid semantic sample and removed from subsequent calculations; S1-2. For the currently processed semantic samples that have not been eliminated, read the preset value range corresponding to each semantic field, and compare the value of the semantic field with the corresponding preset value range item by item. When the value of a semantic field is within the corresponding preset value range, the value of the semantic field remains unchanged; when the value of a semantic field exceeds the corresponding preset value range, the value of the semantic field is pruned to the boundary value of the corresponding preset value range to form standardized semantic sample data. S1-3. Using standardized semantic sample data as input data for the first semantic level, perform semantic representation calculation for the first semantic level to form the semantic output corresponding to the first semantic level. The execution process of semantic representation computation includes: For each semantic field contained in the normalized semantic sample data, read the corresponding semantic field value; For each semantic field, the calculation parameters corresponding to the current semantic level are read, and the semantic field value is multiplied by the corresponding calculation parameters to obtain the field contribution value of the corresponding semantic field in the current semantic level. Perform a step-by-step summation operation on all field contribution values ​​formed within the current semantic level to obtain a linear combination result for the current semantic level; Determine whether the linear combination result falls within the preset semantic output value range. If the linear combination result exceeds the preset semantic output value range, the linear combination result is clipped to the boundary value of the corresponding value range. If the linear combination result is within the preset semantic output value range, the linear combination result remains unchanged. The linear combination result after judgment and processing is determined as the semantic output corresponding to the current semantic level.

[0006] In a preferred embodiment, in S1, the process of determining optional cross-layer connections further includes: S1-4. For any two semantic levels with different level identifiers, read the semantic output corresponding to the semantic level with the smaller level identifier and the semantic output corresponding to the semantic level with the larger level identifier, respectively. While keeping the semantic representation calculation method unchanged, the semantic output corresponding to the semantic level with the smaller hierarchical identifier is used as the input data, and the semantic representation calculation is re-executed to obtain the corresponding calculation output result; When the difference between the calculated output and the semantic output corresponding to the semantic level with the larger hierarchical identifier is less than or equal to the preset difference threshold, it is recorded as a candidate cross-layer connection from the semantic level with the smaller hierarchical identifier to the semantic level with the larger hierarchical identifier. S1-5. Perform hierarchical order verification on all candidate cross-level connections in the records: For each candidate cross-layer connection, read the two semantic levels associated with the corresponding candidate cross-layer connection, and temporarily designate the semantic level with the smaller level identifier as the source level, and the semantic level with the larger level identifier as the target level. Determine whether there is a connection relationship from the target level to the source level among the candidate cross-layer connections, or whether there are multiple candidate cross-layer connections that are connected end to end to form a closed connection path that starts from any semantic level and eventually returns to the corresponding semantic level; When there is no connection relationship from the target level to the source level and no closed connection path is formed, the corresponding candidate cross-level connection is marked as a reachable cross-level connection. All candidate cross-layer connections marked as reachable cross-layer connections are determined as optional cross-layer connections with semantic association potential.

[0007] In a preferred embodiment, in S2, the process of forming the candidate path set and the corresponding pheromone set includes: S2-1. Based on optional cross-layer connections, read the source and target levels associated with each optional cross-layer connection, and arrange the optional cross-layer connections in ascending order of semantic level identifiers to form a set of cross-layer connection sequences that satisfy the hierarchical order constraints. S2-2. For the set of cross-layer connection sequences, starting from the first semantic level, the optional cross-layer connections between adjacent semantic levels are concatenated layer by layer in ascending order of the level identifier. When the target level of a cross-layer connection has the same level identifier as the source level of another cross-layer connection, the two cross-layer connections are connected end to end to form a candidate semantic path that gradually points from the lower semantic level to the higher semantic level. All candidate semantic paths are aggregated to form a candidate semantic path set, and the number of semantic levels covered and the number of cross-level connections contained in each candidate semantic path are recorded. S2-3. For each candidate semantic path formed, check whether the corresponding candidate semantic path starts from the first semantic level and whether it continuously covers at least two different semantic levels. When a candidate semantic path satisfies the condition of starting from the first semantic level and continuously covering at least two different semantic levels, the corresponding candidate semantic path will be retained as a valid semantic path. Otherwise, the corresponding candidate semantic path will be removed from the candidate semantic path set; S2-4. Each valid semantic path that is retained is used as a candidate path in the ant colony algorithm, and a unique path identifier is assigned to each candidate path. For each candidate path, initialize the pheromone value corresponding to the candidate path; S2-5. For each candidate path, the corresponding pheromone value is used as the initial pheromone. For different candidate paths, the number of semantic level coverage or the number of cross-layer connections are read and normalized based on the number. The initial pheromone value is scaled proportionally according to the relative size of the number of candidate paths in all candidate paths, so that the pheromone values ​​corresponding to different candidate paths fall into a uniform numerical range, thereby forming a candidate path set and a corresponding pheromone set.

[0008] In a preferred embodiment, in S3, the process of outputting the prediction result includes: S3-1. For each candidate path in the candidate path set, according to the rule in the ant colony algorithm that the path selection tendency is represented by the pheromone value, read the pheromone value corresponding to the candidate path and compare the pheromone value with the pheromone values ​​corresponding to the other candidate paths one by one. Candidate paths are sorted in descending order of pheromone value, and the target path selection process is executed from the sorting results: The candidate path with the highest pheromone value in the ranking results is selected as the first target semantic path. If the preset target path number threshold is greater than one, the subsequent candidate paths are selected sequentially along the ranking results, and the number of selected target paths is updated after each selection until the number of selected target paths reaches the preset target path number threshold. The selected candidate paths are marked as target semantic paths, and the path identifiers corresponding to the target semantic paths are written into the semantic representation calculation record for this time, thereby determining the target semantic paths participating in this semantic representation calculation. S3-2. After determining the target semantic path, read the cross-layer connection relationships that are connected sequentially in the target semantic path, and read the corresponding source level and target level for each cross-layer connection. According to the connection relationship between the beginning and end of the cross-layer connection, the source level and the target level are traversed in order to form the semantic level sequence covered by the target semantic path; The ascending order of the semantic level identifiers in the semantic level sequence will be used as the execution order of the semantic representation calculation in this operation.

[0009] In a preferred embodiment, in S3, the process of outputting the prediction result further includes: S3-3. Using the normalized semantic sample data corresponding to the current semantic sample as input data, the semantic representation calculation of each semantic level is executed sequentially according to the determined hierarchical execution order. At each semantic level, the semantic output from the previous semantic level is read as the input data for the current semantic level. Semantic representation calculation is performed based on the calculation parameters corresponding to the current semantic level to form the semantic output corresponding to the current semantic level. The semantic output corresponding to the current semantic level is used as the input data for the next semantic level, until the semantic representation calculation of all semantic levels covered by the target semantic path is completed; S3-4. During the execution of semantic representation computation, semantic representation computation is only allowed to be performed along the cross-layer connections contained in the target semantic path; After completing the semantic representation calculation along the target semantic path, the semantic output corresponding to the final semantic level covered by the target semantic path is read and determined as the prediction result of the current semantic sample.

[0010] In a preferred embodiment, in S4, the process of updating the pheromone values ​​corresponding to the target semantic path and the other semantic paths includes: S4-1. For the prediction result of the current semantic sample, read the target value corresponding to the current semantic sample; perform subtraction operation on the prediction result and the corresponding target value item by item to obtain the difference between the prediction result and the target value; perform absolute value operation on each difference to obtain the corresponding error component. All error components are aggregated to form the prediction error of the current semantic sample in this semantic representation calculation; S4-2. Read the semantic level sequence corresponding to the target semantic path, and determine all semantic levels contained in the semantic level sequence as the target semantic level range for this calculation parameter update; S4-3. For each semantic level within the target semantic level range, read the semantic output and corresponding calculation parameters formed in this semantic representation calculation for the corresponding semantic level. Based on each error component contained in the prediction error, the error components are scaled according to a preset ratio to obtain the initial parameter adjustment amount; Determine whether the initial parameter adjustment amount falls within the preset threshold range; when the initial parameter adjustment amount exceeds the preset threshold range, limit the initial parameter adjustment amount to the boundary value of the corresponding threshold range; when the initial parameter adjustment amount is within the preset threshold range, directly use the initial parameter adjustment amount as the parameter adjustment amount corresponding to the semantic level. For each calculation parameter, read the original calculation parameter value, subtract the corresponding parameter adjustment amount from the original calculation parameter value, and obtain the updated calculation parameter value; The updated calculation parameter values ​​are written back to the calculation parameters of the corresponding semantic level to complete the update of the calculation parameters of the semantic level corresponding to the target semantic path.

[0011] In a preferred embodiment, in S4, the process of updating the pheromone values ​​corresponding to the target semantic path and the other semantic paths further includes: S4-4. Before updating the calculation parameters, save the prediction error corresponding to the current semantic sample; after completing the update of the calculation parameters, re-execute the semantic representation calculation based on the same semantic sample and the same target semantic path, and obtain the updated prediction error according to the error calculation process. The difference between the prediction error before and after the update is used to obtain the reduction in prediction error. The reduction in prediction error is then used as the contribution of the target semantic path to the reduction in prediction error. Based on the reduction in prediction error, a scaling operation is performed according to a preset ratio to obtain the initial pheromone increment; Determine whether the initial pheromone increment falls within a preset threshold range; when the initial pheromone increment exceeds the preset threshold range, limit the initial pheromone increment to the boundary value of the corresponding threshold range; when the initial pheromone increment is within the preset threshold range, directly use the initial pheromone increment as the pheromone increment corresponding to the target semantic path. The pheromone increment is added to the original pheromone value to obtain the updated target semantic path pheromone value. S4-5. For the remaining candidate paths in the candidate path set that were not selected as the target semantic path, read the corresponding pheromone values; perform a reduction operation on the original pheromone values ​​according to the preset attenuation ratio to obtain the attenuated pheromone values. Determine whether the decayed pheromone value falls within a preset threshold range; when the decayed pheromone value exceeds the preset threshold range, limit the decayed pheromone value to the boundary value of the corresponding threshold range; when the decayed pheromone value is within the preset threshold range, directly use the decayed pheromone value as the updated pheromone value of the corresponding candidate path.

[0012] In a preferred embodiment, in S5, the process of forming the hierarchical semantic computation structure includes: S5-1. After completing a semantic representation calculation and pheromone update for the current semantic sample, read the next semantic sample from the semantic sample data as the new current semantic sample. For a new current semantic sample, the normalization process and cross-layer connection recognition process in S1 are executed sequentially to form the corresponding normalized semantic sample data and optional cross-layer connections; S5-2. For a new current semantic sample, based on the currently saved candidate path set and the corresponding pheromone value, repeat the target semantic path selection process in S3; after determining the target semantic path, perform semantic representation calculation along the target semantic path to form a new prediction result; For the new prediction results, repeat the error calculation and calculation parameter update process in S4-1 to S4-3 to continuously adjust the calculation parameters corresponding to the semantic level of the target semantic path under the condition of multiple semantic samples. S5-3. After completing the calculation parameter update, repeat the pheromone update process in S4-4 and S4-5. Based on the contribution of the target semantic path to the reduction of prediction error under different semantic sample conditions, continuously enhance or decay the pheromone value corresponding to each candidate path in the candidate path set. S5-4. During the process of continuously inputting multiple semantic samples and repeatedly executing S5-1 to S5-3, when it is detected that the update magnitude of the calculation parameters is continuously lower than the preset parameter change threshold within the preset number of rounds, and the ranking result of the pheromone values ​​in the candidate path set remains within the preset pheromone value change magnitude threshold range during the continuous update process, it is determined that the hierarchical semantic calculation structure has reached the convergence state and a stable hierarchical semantic calculation structure is formed. Once the hierarchical semantic computation structure reaches convergence, the update operations on computation parameters and pheromone values ​​are stopped.

[0013] In a preferred embodiment, in S6, the process of forming the deep learning computational structure includes: S6-1. After the hierarchical semantic computation structure reaches a convergence state, fix the computation parameters in the hierarchical semantic computation structure and fix the pheromone value corresponding to each candidate path in the candidate path set. Based on a fixed set of candidate paths and their corresponding pheromone values, the candidate paths in the set are sorted according to their pheromone values. Candidate paths with pheromone values ​​within a preset upper limit range are identified as stable semantic paths; S6-2. Obtain the semantic samples to be processed for multi-layer semantic feature extraction. Repeat the normalization process in S1 for the semantic samples to be processed to form the corresponding normalized semantic sample data. For standardized semantic sample data, semantic representation computation is performed sequentially along the semantic hierarchy covered by the determined stable semantic path, based on the established stable semantic path. At each semantic level, the semantic output from the previous semantic level is read as the input data for the current semantic level. Semantic representation calculation is performed based on fixed calculation parameters to form the semantic output of the corresponding semantic level. This process continues until the semantic representation calculation of all semantic levels covered by the stable semantic path is completed. S6-3. Read the semantic output corresponding to each semantic level covered by the stable semantic path, and collect the semantic output of each semantic level according to the order of the semantic level identifier to form a multi-level semantic feature containing multi-level semantic representation. Output the multi-level semantic feature as the final multi-level semantic feature extraction result. Based on stable semantic paths, fixed computational parameters, and the cross-layer semantic representation computational order determined by the stable semantic paths, a deep learning computational structure for multi-layer semantic feature extraction is formed.

[0014] The technical effects and advantages of this invention are as follows: This scheme continuously selects, strengthens, and eliminates cross-layer semantic paths during training and operation, so that multi-layer semantic features no longer rely on static semantic paths formed during the training phase, but rather on dynamic, stable, and effective semantic transmission paths based on prediction error feedback. This fundamentally avoids the problem of hidden semantic distortion caused by outputting semantics along the old path when the input semantic support conditions change slowly. By incorporating cross-layer semantic transmission relationships into learnable and evaluable objects, and by organizing cross-layer connections into candidate semantic paths and introducing a pheromone mechanism, the validity of connections between semantic levels is verified by data-driven methods rather than by pre-set network structures or one-time training results, thereby reducing the long-term interference of artificial structural assumptions on multi-layer semantic modeling results. A clear error attribution channel is established between semantic representation computation and path evaluation. By comparing the changes in prediction error of the same path before and after parameter update, the true contribution of semantic path to error reduction is quantified. This puts computation parameter update and semantic path enhancement in the same feedback process, avoiding the problem of parameter optimization and semantic delivery effectiveness being separated in existing methods. By using the magnitude of parameter update and the stability of pheromone ranking as joint convergence criteria, a stable hierarchical semantic computation structure is formed under the condition of multiple semantic samples. This makes the final multi-layer semantic features not only converge numerically, but also have structural stability at the semantic transmission path level, thereby enhancing the interpretability and reliability of the model in the long-term operating environment. In the feature extraction stage, semantic representation calculations are performed on new semantic samples based on stable semantic paths and fixed computational parameters, so that the source of the output multi-layer semantic features is clear and the path is determined. Moreover, the semantic support capability of the path has been repeatedly verified in historical samples, thereby improving the adaptability and engineering practical value of multi-layer semantic features in scenarios of business migration and slow evolution of data distribution. Attached Figure Description

[0015] Figure 1 This is a flowchart outlining the method steps of the present invention. Detailed Implementation

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

[0017] Refer to the instruction manual appendix Figure 1 An embodiment of the present invention provides a multi-layer semantic feature extraction method based on deep learning, comprising: S1. Obtain semantic sample data, perform normalization processing, and map it into multiple semantic levels; perform initial semantic representation calculation on each semantic level, and determine the optional cross-layer connections with semantic association possibilities based on the semantic representation output results between different semantic levels. S2. Apply hierarchical order constraints to the optional cross-layer connections, construct a candidate semantic path set based on the optional cross-layer connections, and use the candidate semantic path set as the candidate path in the ant colony algorithm; initialize pheromone values ​​for each candidate path to form a candidate path set and a corresponding pheromone set, which are used to characterize the selection tendency of different cross-layer semantic paths. S3. Based on semantic sample data, perform semantic representation calculation along the cross-layer connections constrained by the candidate path set to generate semantic representations corresponding to different semantic levels. During the semantic representation calculation process, perform path selection on the candidate path set according to the pheromone value corresponding to the candidate path to determine the target semantic path participating in this semantic representation calculation, and output the prediction result based on the semantic representation of the semantic level covered by the target semantic path. S4. Obtain the target value corresponding to the prediction result, calculate the error between the prediction result and the target value, and synchronously update the calculation parameters corresponding to the semantic level covered by the target semantic path; based on the contribution of the target semantic path to the reduction of the prediction error, update the pheromone values ​​corresponding to the target semantic path and other semantic paths. S5. Continuously input semantic sample data, repeatedly execute the semantic representation calculation, calculation parameter update and pheromone value update process, so that the calculation parameters corresponding to each semantic level and the pheromone values ​​corresponding to the candidate paths gradually converge under the joint action, forming a stable hierarchical semantic calculation structure. S6. Based on a stable hierarchical semantic computation structure, candidate paths with pheromone values ​​in a preset upper limit range are determined as stable semantic paths, and the multi-layer semantic representations covered by the stable semantic paths are output as multi-layer semantic feature extraction results. At the same time, the formation of a deep learning computation structure for multi-layer semantic feature extraction is completed.

[0018] In S1, the process of determining optional cross-layer connections includes: S1-1. Using each semantic sample in the semantic sample data as the current semantic sample to process, read the original values ​​of each semantic field in the current semantic sample and convert the original values ​​of each semantic field into a preset data representation format. The original values ​​of the semantic fields refer to the field values ​​directly read from the semantic sample data without undergoing data type conversion, value correction, or interval constraint processing. Converting to the preset data representation format means that for each semantic field, according to the preset data type requirements corresponding to the semantic field, the original values ​​are converted into numerical values, encoded values, or vectorized values, so that the converted field values ​​can be used as input data for subsequent semantic representation calculations. When the original value of any semantic field does not meet the data representation format requirements, the currently processed semantic sample is marked as an invalid semantic sample and removed from subsequent calculations; where not meeting the data representation format requirements means that when the original value of any semantic field results in a null value, an illegal character result, or a result that exceeds the representation range allowed by the preset data type after performing numerical conversion, encoding conversion, or vectorization conversion, the currently processed semantic sample is marked as an invalid semantic sample and removed from subsequent calculations; S1-2. For the currently processed semantic samples that have not been eliminated, read the preset value range corresponding to each semantic field, and compare the value of the semantic field with the corresponding preset value range item by item; where semantic field refers to each independent data dimension or attribute item that constitutes the semantic sample data, and each semantic field is used to represent a semantic related information in the semantic sample data. When the value of a semantic field is within the corresponding preset value range, the semantic field value remains unchanged; when the value of a semantic field exceeds the corresponding preset value range, the semantic field value is pruned to the boundary value of the corresponding preset value range to form normalized semantic sample data; where pruning to the boundary value of the corresponding preset value range means that when the semantic field value is less than the lower limit of the preset value range, the semantic field value is replaced with the lower limit value of the preset value range; when the semantic field value is greater than the upper limit of the preset value range, the semantic field value is replaced with the upper limit value of the preset value range. S1-3. Using standardized semantic sample data as input data for the first semantic level, perform semantic representation calculation for the first semantic level to form the semantic output corresponding to the first semantic level. The execution process of semantic representation computation includes: For each semantic field contained in the normalized semantic sample data, read the corresponding semantic field value; For each semantic field value, the calculation parameter corresponding to the current semantic level is read, and the semantic field value is multiplied with the corresponding calculation parameter to obtain the field contribution value of the corresponding semantic field under the current semantic level. The calculation parameter corresponding to the current semantic level refers to the numerical adjustment amount that is pre-set for the current semantic level or gradually updated during multiple semantic sample training processes. It is used to adjust the numerical influence strength of different semantic field values ​​in the semantic representation calculation of the current semantic level. Perform a step-by-step summation operation on all field contribution values ​​formed within the current semantic level to obtain a linear combination result for the current semantic level; Determine whether the linear combination result falls within the preset semantic output value range. If the linear combination result exceeds the preset semantic output value range, the linear combination result is clipped to the boundary value of the corresponding value range. If the linear combination result is within the preset semantic output value range, the linear combination result remains unchanged. The linear combination result after judgment and processing is determined as the semantic output corresponding to the current semantic level.

[0019] In S1, the process of determining optional cross-layer connections also includes: S1-4. For any two semantic levels with different level identifiers, read the semantic output corresponding to the semantic level with the smaller level identifier and the semantic output corresponding to the semantic level with the larger level identifier, respectively. While keeping the semantic representation calculation method unchanged, the semantic output corresponding to the semantic level with the smaller hierarchical identifier is used as the input data, and the semantic representation calculation is re-executed to obtain the corresponding calculation output result; When the difference between the calculated output and the semantic output corresponding to the semantic level with the larger hierarchical identifier is less than or equal to the preset difference threshold, it is recorded as a candidate cross-layer connection from the semantic level with the smaller hierarchical identifier to the semantic level with the larger hierarchical identifier. S1-5. Perform hierarchical order verification on all candidate cross-level connections in the records: For each candidate cross-layer connection, read the two semantic levels associated with the corresponding candidate cross-layer connection, and temporarily designate the semantic level with the smaller level identifier as the source level, and the semantic level with the larger level identifier as the target level. Determine whether there is a connection relationship from the target level to the source level among the candidate cross-layer connections, or whether there are multiple candidate cross-layer connections that are connected end to end to form a closed connection path that starts from any semantic level and eventually returns to the corresponding semantic level; When there is no connection relationship from the target level to the source level and no closed connection path is formed, the corresponding candidate cross-level connection is marked as a reachable cross-level connection. All candidate cross-layer connections marked as reachable cross-layer connections are identified as optional cross-layer connections with semantic association potential, which are used for the construction of subsequent candidate semantic paths and the constraint of cross-layer semantic representation computation paths.

[0020] In S2, the process of forming the candidate path set and the corresponding pheromone set includes: S2-1. Based on optional cross-layer connections, read the source and target levels associated with each optional cross-layer connection, and arrange the optional cross-layer connections in ascending order of semantic level identifiers to form a set of cross-layer connection sequences that satisfy the hierarchical order constraints. S2-2. For the set of cross-layer connection sequences, starting from the first semantic level, the optional cross-layer connections between adjacent semantic levels are concatenated layer by layer in ascending order of the level identifier. When the target level of a cross-layer connection has the same level identifier as the source level of another cross-layer connection, the two cross-layer connections are connected end to end to form a candidate semantic path that gradually points from the lower semantic level to the higher semantic level. All candidate semantic paths are aggregated to form a candidate semantic path set, and the number of semantic levels covered and the number of cross-level connections contained in each candidate semantic path are recorded. S2-3. For each candidate semantic path formed, check whether the corresponding candidate semantic path starts from the first semantic level and whether it continuously covers at least two different semantic levels. When a candidate semantic path satisfies the condition of starting from the first semantic level and continuously covering at least two different semantic levels, the corresponding candidate semantic path will be retained as a valid semantic path. Otherwise, the corresponding candidate semantic path will be removed from the candidate semantic path set; S2-4. Each valid semantic path that is retained is used as a candidate path in the ant colony algorithm, and a unique path identifier is assigned to each candidate path. For each candidate path, initialize the pheromone value corresponding to the candidate path to characterize the candidate path's tendency to be selected in subsequent semantic representation calculations; where initialization means assigning a numerical pheromone scalar to each candidate path and writing the pheromone scalar into the path record corresponding to the candidate path as the initial reference value for subsequent path selection and pheromone updates. S2-5. For each candidate path, the corresponding pheromone value is used as the initial pheromone. For different candidate paths, the number of semantic level coverages or the number of cross-layer connections are read and normalized based on the number. The initial pheromone value is scaled proportionally according to the relative size of the number of candidate paths in all candidate paths, so that the pheromone values ​​corresponding to different candidate paths fall within a uniform numerical range, thereby forming a candidate path set and a corresponding pheromone set. The candidate path set is used to constrain the selection of cross-layer connections in subsequent semantic representation calculations, and the pheromone set is used to characterize the selection tendency of each candidate path. The proportional scaling refers to reading the upper and lower limits of the number corresponding to all candidate paths, mapping the number of each path to the upper and lower limits, so that the initial pheromone value falls within a preset pheromone value range.

[0021] In S3, the process of outputting the prediction results includes: S3-1. For each candidate path in the candidate path set, according to the rule in the ant colony algorithm that uses pheromone values ​​to represent path selection tendency, read the pheromone value corresponding to the candidate path and compare the pheromone value with the pheromone values ​​corresponding to the other candidate paths one by one. The rule in the ant colony algorithm that uses pheromone values ​​to represent path selection tendency means that the pheromone value corresponding to the candidate path is used as a path evaluation metric that reflects the historical effectiveness of the candidate path. The larger the pheromone value, the higher the effective contribution of the candidate path to the prediction result in the previous semantic representation calculation, and thus the higher the selection priority in the subsequent path selection process. Candidate paths are sorted in descending order of pheromone value, and the target path selection process is executed from the sorting results: The candidate path with the highest pheromone value in the ranking results is selected as the first target semantic path. If the preset target path number threshold is greater than one, the subsequent candidate paths are selected sequentially along the ranking results, and the number of selected target paths is updated after each selection until the number of selected target paths reaches the preset target path number threshold. The selected candidate paths are marked as target semantic paths, and the path identifiers corresponding to the target semantic paths are written into the semantic representation calculation record for this time, thereby determining the target semantic paths participating in this semantic representation calculation. S3-2. After determining the target semantic path, read the cross-layer connection relationships that are connected sequentially in the target semantic path, and read the corresponding source level and target level for each cross-layer connection. According to the connection relationship between the beginning and end of the cross-layer connection, the source level and the target level are traversed in order to form the semantic level sequence covered by the target semantic path; The ascending order of the semantic level identifiers in the semantic level sequence will be used as the execution order of the semantic representation calculation in this operation.

[0022] In S3, the process of outputting the prediction results also includes: S3-3. Using the normalized semantic sample data corresponding to the current semantic sample as input data, the semantic representation calculation of each semantic level is executed sequentially according to the determined hierarchical execution order. At each semantic level, the semantic output from the previous semantic level is read as the input data for the current semantic level. Semantic representation calculation is performed based on the calculation parameters corresponding to the current semantic level to form the semantic output corresponding to the current semantic level. The semantic output corresponding to the current semantic level is used as the input data for the next semantic level, until the semantic representation calculation of all semantic levels covered by the target semantic path is completed; S3-4. During the semantic representation computation process, semantic representation computation is only allowed to be performed along the cross-layer connections contained in the target semantic path. When a cross-layer connection is detected to be not part of the target semantic path, the corresponding cross-layer connection is prohibited from participating in the current semantic representation computation process, thereby ensuring that the semantic representation computation is strictly constrained by the target semantic path. After completing the semantic representation calculation along the target semantic path, the semantic output corresponding to the final semantic level covered by the target semantic path is read and determined as the prediction result of the current semantic sample, which is used for subsequent error calculation and pheromone update.

[0023] In S4, the process of updating the pheromone values ​​corresponding to the target semantic path and the other semantic paths includes: S4-1. For the prediction result of the current semantic sample, read the target value corresponding to the current semantic sample; perform subtraction operation on the prediction result and the corresponding target value item by item to obtain the difference between the prediction result and the target value; perform absolute value operation on each difference to obtain the corresponding error component; where the target value refers to the known semantic result or expected semantic output corresponding to the current semantic sample, which is used as a reference benchmark for the semantic representation calculation result of the current semantic sample. All error components are aggregated to form the prediction error of the current semantic sample in this semantic representation calculation, which is used to characterize the degree of prediction bias. S4-2. Read the semantic level sequence corresponding to the target semantic path, and determine all semantic levels contained in the semantic level sequence as the target semantic level range for this calculation parameter update; S4-3. For each semantic level within the target semantic level range, read the semantic output and corresponding calculation parameters formed in this semantic representation calculation for the corresponding semantic level. Based on each error component contained in the prediction error, the error components are scaled according to a preset ratio to obtain the initial parameter adjustment amount; Determine whether the initial parameter adjustment amount falls within the preset threshold range; when the initial parameter adjustment amount exceeds the preset threshold range, limit the initial parameter adjustment amount to the boundary value of the corresponding threshold range; when the initial parameter adjustment amount is within the preset threshold range, directly use the initial parameter adjustment amount as the parameter adjustment amount corresponding to the semantic level. For each calculation parameter, read the original calculation parameter value, subtract the corresponding parameter adjustment amount from the original calculation parameter value, and obtain the updated calculation parameter value; The updated calculation parameter values ​​are written back to the calculation parameters of the corresponding semantic level to complete the update of the calculation parameters of the semantic level corresponding to the target semantic path.

[0024] In S4, the process of updating the pheromone values ​​corresponding to the target semantic path and the other semantic paths also includes: S4-4. Before updating the calculation parameters, save the prediction error corresponding to the current semantic sample; after completing the update of the calculation parameters, re-execute the semantic representation calculation based on the same semantic sample and the same target semantic path, and obtain the updated prediction error according to the error calculation process. The difference between the prediction error before and after the update is used to obtain the reduction in prediction error. The reduction in prediction error is then used as the contribution of the target semantic path to the reduction in prediction error. Based on the reduction in prediction error, a scaling operation is performed according to a preset ratio to obtain the initial pheromone increment; where performing the scaling operation according to the preset ratio means multiplying the reduction in prediction error by the preset ratio to obtain the change in pheromone. Determine whether the initial pheromone increment falls within a preset threshold range; when the initial pheromone increment exceeds the preset threshold range, limit the initial pheromone increment to the boundary value of the corresponding threshold range; when the initial pheromone increment is within the preset threshold range, directly use the initial pheromone increment as the pheromone increment corresponding to the target semantic path. The pheromone increment is added to the original pheromone value to obtain the updated target semantic path pheromone value, thereby increasing the priority of the target semantic path in subsequent path selection processes. S4-5. For the remaining candidate paths in the candidate path set that were not selected as the target semantic path, read the corresponding pheromone values; perform a reduction operation on the original pheromone values ​​according to the preset attenuation ratio to obtain the attenuated pheromone values; where performing a reduction operation on the original pheromone values ​​according to the preset attenuation ratio means performing a multiplication operation between the original pheromone values ​​and the preset attenuation ratio to obtain the attenuated pheromone values. Determine whether the decayed pheromone value falls within a preset threshold range; when the decayed pheromone value exceeds the preset threshold range, limit the decayed pheromone value to the boundary value of the corresponding threshold range; when the decayed pheromone value is within the preset threshold range, directly use the decayed pheromone value as the updated pheromone value of the corresponding candidate path, and write the updated pheromone value back to the corresponding candidate path to reduce the priority of unselected paths in subsequent path selection processes.

[0025] In S5, the process of forming a hierarchical semantic computing structure includes: S5-1. After completing a semantic representation calculation and pheromone update for the current semantic sample, read the next semantic sample from the semantic sample data as the new current semantic sample. For a new current semantic sample, the normalization process and cross-layer connection recognition process in S1 are executed sequentially to form the corresponding normalized semantic sample data and optional cross-layer connections; S5-2. For a new current semantic sample, based on the currently saved candidate path set and the corresponding pheromone value, repeat the target semantic path selection process in S3; after determining the target semantic path, perform semantic representation calculation along the target semantic path to form a new prediction result; For the new prediction results, the error calculation and calculation parameter update process in S4-1 to S4-3 is repeated to continuously adjust the calculation parameters corresponding to the semantic level of the target semantic path under the condition of multiple semantic samples. The condition of multiple semantic samples refers to the continuous selection of multiple different semantic samples in the semantic sample data and the use of the semantic samples as the current semantic sample input in the order of their arrangement in the semantic sample data, so that the calculation parameter update process is repeatedly triggered under the driving force of the prediction error corresponding to multiple semantic samples. S5-3. After completing the calculation parameter update, repeat the pheromone update process in S4-4 and S4-5. Based on the contribution of the target semantic path to the reduction of prediction error under different semantic sample conditions, continuously enhance or decay the pheromone value corresponding to each candidate path in the candidate path set. S5-4. During the process of continuously inputting multiple semantic samples and repeatedly executing S5-1 to S5-3, when it is detected that the update magnitude of the calculation parameters is continuously lower than the preset parameter change threshold within the preset number of rounds, and the ranking result of the pheromone values ​​in the candidate path set remains within the preset pheromone value change magnitude threshold range during the continuous update process, it is determined that the hierarchical semantic calculation structure has reached the convergence state and a stable hierarchical semantic calculation structure is formed. After the hierarchical semantic computation structure reaches convergence, the update operations on computation parameters and pheromone values ​​are stopped; the stable hierarchical semantic computation structure and the pheromone values ​​corresponding to each candidate path in the candidate path set are used as the final computation basis for subsequent multi-layer semantic feature extraction.

[0026] In S6, the process of forming the deep learning computational architecture includes: S6-1. After the hierarchical semantic computation structure reaches convergence, fix the computation parameters in the hierarchical semantic computation structure and fix the pheromone values ​​corresponding to each candidate path in the candidate path set as the computation basis for the final multi-layer semantic feature extraction stage; where fixing means that no update operation is performed on the computation parameters and pheromone values ​​in the subsequent multi-layer semantic feature extraction process. Based on a fixed set of candidate paths and their corresponding pheromone values, the candidate paths in the set are sorted according to their pheromone values. Candidate paths with pheromone values ​​within a preset upper limit range are identified as stable semantic paths and used for subsequent multi-layer semantic feature extraction. S6-2. Obtain the semantic samples to be processed for multi-layer semantic feature extraction. Repeat the normalization process in S1 for the semantic samples to be processed to form the corresponding normalized semantic sample data, which serves as the input data for the multi-layer semantic feature extraction stage. For standardized semantic sample data, semantic representation computation is performed sequentially along the semantic hierarchy covered by the determined stable semantic path, based on the established stable semantic path. At each semantic level, the semantic output from the previous semantic level is read as the input data for the current semantic level. Semantic representation calculation is performed based on fixed calculation parameters to form the semantic output of the corresponding semantic level. This process continues until the semantic representation calculation of all semantic levels covered by the stable semantic path is completed. S6-3. Read the semantic output corresponding to each semantic level covered by the stable semantic path, and collect the semantic output of each semantic level according to the order of the semantic level identifier to form a multi-level semantic feature containing multi-level semantic representation. Output the multi-level semantic feature as the final multi-level semantic feature extraction result. Based on stable semantic paths, fixed computational parameters, and the cross-layer semantic representation computational order determined by the stable semantic paths, a deep learning computational structure for multi-layer semantic feature extraction is formed.

[0027] It should be noted that, including but not limited to, this approach does not first set a fixed multi-layer semantic structure and then train the parameters. Instead, it treats the transmission of cross-layer semantics as an object that needs to be selected, verified, and stabilized. Because of this, the approach does not regard the connections between semantic levels as predetermined relationships from the beginning, but rather gradually filters and solidifies them through a data-driven approach. In the first stage of the scheme formation, semantic sample data does not directly enter semantic representation computation, but first undergoes a clear normalization process. This processing is not for optimizing numerical values, but to ensure that all subsequent semantic representation computations are performed within a comparable, tailorable, and decidable numerical space. By performing format conversion, interval constraints, and anomaly removal on semantic fields, the scheme ensures that any subsequent semantic outputs have a consistent scale basis; otherwise, cross-layer comparison computation and path selection would lose their meaning. This is implemented in S1 through invalid semantic sample removal and the formation of normalized semantic sample data. Building on this, the scheme introduces the concept of semantic hierarchy, but does not assume that there must be a fixed semantic transitive relationship between the hierarchy. Instead, it allows each semantic hierarchy to perform semantic representation calculation independently. Each semantic representation calculation is explicitly broken down into: the numerical combination of the semantic field value and the calculation parameter of the current semantic hierarchy, the accumulation of the combination result, and the range pruning of the output value. The purpose of this is not to emphasize a certain specific calculation form, but to ensure that how the calculation parameter affects the semantic output is explicit, updatable, and consistent with the execution logic across all semantic hierarchies. Subsequently, the scheme does not stop at obtaining the semantic outputs of each semantic level, but uses these semantic outputs to back-verify whether there are stable semantic connections between the levels. By taking the semantic outputs of lower semantic levels as input again, performing a calculation while keeping the semantic representation calculation method unchanged, and comparing the differences with the existing semantic outputs of higher semantic levels, the scheme uses the similarity of the results to determine whether there are feasible cross-level semantic connections. Only those cross-level relationships that can be reproduced or approximated in numerical performance are eligible to enter the subsequent path construction stage, thereby avoiding the bias caused by artificially setting the semantic structure. After forming optional cross-layer connections, the scheme enters the second core stage, which is to organize the scattered cross-layer connections into semantic paths that can be selected and evaluated as a whole. This stage is not a simple enumeration of paths, but rather a selection of candidate semantic paths that truly transmit continuously from lower-level semantics to higher-level semantics through hierarchical order constraints, conditions for splicing the beginning and end, and verification of the starting level. At this point, each candidate semantic path is assigned an independent pheromone value. The essential function of this step is to transform the historical validity of semantic paths into a comparable and updatable quantity, providing a basis for subsequent path selection. Subsequently, during the actual execution of semantic representation computation, the scheme no longer propagates along all possible paths simultaneously. Instead, candidate paths are sorted and selected based on pheromone values, allowing only the selected target semantic path to participate in this semantic representation computation. This makes semantic representation computation no longer a passive propagation process on a fixed structure, but an active execution process constrained by path selection. The prediction result is thus naturally tied to the condition of which semantic path is selected, providing a clear boundary for subsequent error attribution and path evaluation. During the error and update phase, the scheme does not only use the error to adjust the computational parameters, but also uses it to evaluate the effectiveness of the semantic path itself. On the one hand, the prediction error drives the update of the computational parameters of each semantic level involved in the target semantic path, so that the semantic representation ability is gradually enhanced under the same path conditions. On the other hand, by comparing the changes in prediction error before and after the parameter update, the scheme directly quantifies whether the semantic path really helps to reduce the error, and accordingly enhances or diminishes the pheromone value of the corresponding path. In this way, the computational parameter update and the semantic path selection are bound in the same feedback process, rather than two independent processes. As semantic samples are continuously input and the above process is repeated, the adjustment range of the calculation parameters will gradually decrease, and the pheromone ranking of the candidate paths will tend to stabilize. The solution utilizes this phenomenon and uses the update range of calculation parameters and the stability of pheromone ranking as joint criteria in S5 to determine that the hierarchical semantic calculation structure has converged. What is obtained at this time is not an abstract model state, but a set of stable semantic paths, a set of fixed calculation parameters, and the cross-layer semantic representation calculation order determined by them. Finally, in the multi-layer semantic feature extraction stage, the scheme no longer performs any structure search or parameter update. Instead, it directly performs semantic representation calculation on new semantic samples to be processed based on the stabilized semantic path and computational parameters. By sequentially outputting the semantic representations of each semantic level along the stable semantic path and collecting them in hierarchical order, a multi-layer semantic feature that can simultaneously reflect different semantic abstraction levels is finally obtained. The semantic path and computational parameters corresponding to this output result have been repeatedly verified and screened under multi-semantic sample conditions, rather than being set all at once.

[0028] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-layer semantic feature extraction method based on deep learning, characterized in that, include: S1. Obtain semantic sample data, perform normalization processing, and map it into multiple semantic levels; Perform semantic representation computation on each semantic level to determine optional cross-level connections; S2. Apply hierarchical order constraints to optional cross-layer connections, construct a set of candidate semantic paths, and use them as candidate paths in the ant colony algorithm; initialize pheromone values ​​for each candidate path to form a set of candidate paths and a corresponding set of pheromones. S3. Based on semantic sample data, perform semantic representation calculation along cross-layer connections to generate semantic representations corresponding to different semantic levels; perform path selection on the candidate path set according to the pheromone value corresponding to the candidate path to determine the target semantic path and output the prediction result. S4. Obtain the target value corresponding to the prediction result, calculate the error between the prediction result and the target value, and synchronously update the calculation parameters corresponding to each semantic level of the target semantic path. Based on the contribution of the target semantic path to error reduction, update the pheromone values ​​corresponding to the target semantic path and other semantic paths. S5. Continuously input semantic sample data and repeatedly execute the update process so that the calculation parameters corresponding to the semantic level and the pheromone values ​​corresponding to the candidate path gradually converge under the joint action, forming a hierarchical semantic calculation structure. S6. Based on the hierarchical semantic computing structure, candidate paths with pheromone values ​​in the preset upper limit range are determined as stable semantic paths, and the multi-layer semantic representations covered by the stable semantic paths are output as multi-layer semantic feature extraction results, while completing the formation of the deep learning computing structure.

2. The deep learning-based multi-layer semantic feature extraction method according to claim 1, characterized in that: In S1, the process of determining optional cross-layer connections includes: S1-1. Take each semantic sample in the semantic sample data as the current semantic sample to be processed, read the original values ​​of each semantic field in the current semantic sample to be processed, and convert the original values ​​of each semantic field into a preset data representation format. When the original value of any semantic field does not meet the data representation format requirements, the currently processed semantic sample is marked as an invalid semantic sample and removed from subsequent calculations; S1-2. For the currently processed semantic samples that have not been eliminated, read the preset value range corresponding to each semantic field, and compare the value of the semantic field with the corresponding preset value range item by item. When the value of a semantic field is within the corresponding preset value range, the value of the semantic field remains unchanged; when the value of a semantic field exceeds the corresponding preset value range, the value of the semantic field is pruned to the boundary value of the corresponding preset value range to form standardized semantic sample data. S1-3. Using standardized semantic sample data as input data for the first semantic level, perform semantic representation calculation for the first semantic level to form the semantic output corresponding to the first semantic level. The execution process of semantic representation computation includes: For each semantic field contained in the normalized semantic sample data, read the corresponding semantic field value; For each semantic field, the calculation parameters corresponding to the current semantic level are read, and the semantic field value is multiplied by the corresponding calculation parameters to obtain the field contribution value of the corresponding semantic field in the current semantic level. Perform a step-by-step summation operation on all field contribution values ​​formed within the current semantic level to obtain a linear combination result for the current semantic level; Determine whether the linear combination result falls within the preset semantic output value range. If the linear combination result exceeds the preset semantic output value range, the linear combination result is clipped to the boundary value of the corresponding value range. If the linear combination result is within the preset semantic output value range, the linear combination result remains unchanged. The linear combination result after judgment and processing is determined as the semantic output corresponding to the current semantic level.

3. The deep learning-based multi-layer semantic feature extraction method according to claim 2, characterized in that: In S1, the process of determining optional cross-layer connections also includes: S1-4. For any two semantic levels with different level identifiers, read the semantic output corresponding to the semantic level with the smaller level identifier and the semantic output corresponding to the semantic level with the larger level identifier, respectively. While keeping the semantic representation calculation method unchanged, the semantic output corresponding to the semantic level with the smaller hierarchical identifier is used as the input data, and the semantic representation calculation is re-executed to obtain the corresponding calculation output result; When the difference between the calculated output and the semantic output corresponding to the semantic level with the larger hierarchical identifier is less than or equal to the preset difference threshold, it is recorded as a candidate cross-layer connection from the semantic level with the smaller hierarchical identifier to the semantic level with the larger hierarchical identifier. S1-5. Perform hierarchical order verification on all candidate cross-level connections in the records: For each candidate cross-layer connection, read the two semantic levels associated with the corresponding candidate cross-layer connection, and temporarily designate the semantic level with the smaller level identifier as the source level, and the semantic level with the larger level identifier as the target level. Determine whether there is a connection relationship from the target level to the source level among the candidate cross-layer connections, or whether there are multiple candidate cross-layer connections that are connected end to end to form a closed connection path that starts from any semantic level and eventually returns to the corresponding semantic level; When there is no connection relationship from the target level to the source level and no closed connection path is formed, the corresponding candidate cross-level connection is marked as a reachable cross-level connection. All candidate cross-layer connections marked as reachable cross-layer connections are determined as optional cross-layer connections with semantic association potential.

4. The deep learning-based multi-layer semantic feature extraction method according to claim 3, characterized in that: In S2, the process of forming the candidate path set and the corresponding pheromone set includes: S2-1. Based on optional cross-layer connections, read the source and target levels associated with each optional cross-layer connection, and arrange the optional cross-layer connections in ascending order of semantic level identifiers to form a set of cross-layer connection sequences that satisfy the hierarchical order constraints. S2-2. For the set of cross-layer connection sequences, starting from the first semantic level, the optional cross-layer connections between adjacent semantic levels are concatenated layer by layer in ascending order of the level identifier. When the target level of a cross-layer connection has the same level identifier as the source level of another cross-layer connection, the two cross-layer connections are connected end to end to form a candidate semantic path that gradually points from the lower semantic level to the higher semantic level. All candidate semantic paths are aggregated to form a candidate semantic path set, and the number of semantic levels covered and the number of cross-level connections contained in each candidate semantic path are recorded. S2-3. For each candidate semantic path formed, check whether the corresponding candidate semantic path starts from the first semantic level and whether it continuously covers at least two different semantic levels. When a candidate semantic path satisfies the condition of starting from the first semantic level and continuously covering at least two different semantic levels, the corresponding candidate semantic path will be retained as a valid semantic path. Otherwise, the corresponding candidate semantic path will be removed from the candidate semantic path set; S2-4. Each valid semantic path that is retained is used as a candidate path in the ant colony algorithm, and a unique path identifier is assigned to each candidate path. For each candidate path, initialize the pheromone value corresponding to the candidate path; S2-5. For each candidate path, the corresponding pheromone value is used as the initial pheromone. For different candidate paths, the number of semantic level coverage or the number of cross-layer connections are read and normalized based on the number. The initial pheromone value is scaled proportionally according to the relative size of the number of candidate paths in all candidate paths, so that the pheromone values ​​corresponding to different candidate paths fall into a uniform numerical range, thereby forming a candidate path set and a corresponding pheromone set.

5. The deep learning-based multi-layer semantic feature extraction method according to claim 4, characterized in that: In S3, the process of outputting the prediction results includes: S3-1. For each candidate path in the candidate path set, according to the rule in the ant colony algorithm that the path selection tendency is represented by the pheromone value, read the pheromone value corresponding to the candidate path and compare the pheromone value with the pheromone values ​​corresponding to the other candidate paths one by one. Candidate paths are sorted in descending order of pheromone value, and the target path selection process is executed from the sorting results: The candidate path with the highest pheromone value in the ranking results is selected as the first target semantic path. If the preset target path number threshold is greater than one, the subsequent candidate paths are selected sequentially along the ranking results, and the number of selected target paths is updated after each selection until the number of selected target paths reaches the preset target path number threshold. The selected candidate paths are marked as target semantic paths, and the path identifiers corresponding to the target semantic paths are written into the semantic representation calculation record for this time, thereby determining the target semantic paths participating in this semantic representation calculation. S3-2. After determining the target semantic path, read the cross-layer connection relationships that are connected sequentially in the target semantic path, and read the corresponding source level and target level for each cross-layer connection. According to the connection relationship between the beginning and end of the cross-layer connection, the source level and the target level are traversed in order to form the semantic level sequence covered by the target semantic path; The ascending order of the semantic level identifiers in the semantic level sequence will be used as the execution order of the semantic representation calculation in this operation.

6. The deep learning-based multi-layer semantic feature extraction method according to claim 5, characterized in that: In S3, the process of outputting the prediction results also includes: S3-3. Using the normalized semantic sample data corresponding to the current semantic sample as input data, the semantic representation calculation of each semantic level is executed sequentially according to the determined hierarchical execution order. At each semantic level, the semantic output from the previous semantic level is read as the input data for the current semantic level. Semantic representation calculation is performed based on the calculation parameters corresponding to the current semantic level to form the semantic output corresponding to the current semantic level. The semantic output corresponding to the current semantic level is used as the input data for the next semantic level, until the semantic representation calculation of all semantic levels covered by the target semantic path is completed; S3-4. During the execution of semantic representation computation, semantic representation computation is only allowed to be performed along the cross-layer connections contained in the target semantic path; After completing the semantic representation calculation along the target semantic path, the semantic output corresponding to the final semantic level covered by the target semantic path is read and determined as the prediction result of the current semantic sample.

7. The deep learning-based multi-layer semantic feature extraction method according to claim 6, characterized in that: In S4, the process of updating the pheromone values ​​corresponding to the target semantic path and the other semantic paths includes: S4-1. For the prediction result of the current semantic sample, read the target value corresponding to the current semantic sample; perform subtraction operation on the prediction result and the corresponding target value item by item to obtain the difference between the prediction result and the target value; perform absolute value operation on each difference to obtain the corresponding error component. All error components are aggregated to form the prediction error of the current semantic sample in this semantic representation calculation; S4-2. Read the semantic level sequence corresponding to the target semantic path, and determine all semantic levels contained in the semantic level sequence as the target semantic level range for this calculation parameter update; S4-3. For each semantic level within the target semantic level range, read the semantic output and corresponding calculation parameters formed in this semantic representation calculation for the corresponding semantic level. Based on each error component contained in the prediction error, the error components are scaled according to a preset ratio to obtain the initial parameter adjustment amount; Determine whether the initial parameter adjustment amount falls within the preset threshold range; when the initial parameter adjustment amount exceeds the preset threshold range, limit the initial parameter adjustment amount to the boundary value of the corresponding threshold range; when the initial parameter adjustment amount is within the preset threshold range, directly use the initial parameter adjustment amount as the parameter adjustment amount corresponding to the semantic level. For each calculation parameter, read the original calculation parameter value, subtract the corresponding parameter adjustment amount from the original calculation parameter value, and obtain the updated calculation parameter value; The updated calculation parameter values ​​are written back to the calculation parameters of the corresponding semantic level to complete the update of the calculation parameters of the semantic level corresponding to the target semantic path.

8. The deep learning-based multi-layer semantic feature extraction method according to claim 7, characterized in that: In S4, the process of updating the pheromone values ​​corresponding to the target semantic path and the other semantic paths also includes: S4-4. Before updating the calculation parameters, save the prediction error corresponding to the current semantic sample; after completing the update of the calculation parameters, re-execute the semantic representation calculation based on the same semantic sample and the same target semantic path, and obtain the updated prediction error according to the error calculation process. The difference between the prediction error before and after the update is used to obtain the reduction in prediction error. The reduction in prediction error is then used as the contribution of the target semantic path to the reduction in prediction error. Based on the reduction in prediction error, a scaling operation is performed according to a preset ratio to obtain the initial pheromone increment; Determine whether the initial pheromone increment falls within a preset threshold range; when the initial pheromone increment exceeds the preset threshold range, limit the initial pheromone increment to the boundary value of the corresponding threshold range; when the initial pheromone increment is within the preset threshold range, directly use the initial pheromone increment as the pheromone increment corresponding to the target semantic path. The pheromone increment is added to the original pheromone value to obtain the updated target semantic path pheromone value. S4-5. For the remaining candidate paths in the candidate path set that were not selected as the target semantic path, read the corresponding pheromone values; perform a reduction operation on the original pheromone values ​​according to the preset attenuation ratio to obtain the attenuated pheromone values. Determine whether the decayed pheromone value falls within a preset threshold range; when the decayed pheromone value exceeds the preset threshold range, limit the decayed pheromone value to the boundary value of the corresponding threshold range; when the decayed pheromone value is within the preset threshold range, directly use the decayed pheromone value as the updated pheromone value of the corresponding candidate path.

9. The deep learning-based multi-layer semantic feature extraction method according to claim 8, characterized in that: In S5, the process of forming a hierarchical semantic computing structure includes: S5-1. After completing a semantic representation calculation and pheromone update for the current semantic sample, read the next semantic sample from the semantic sample data as the new current semantic sample. For a new current semantic sample, the normalization process and cross-layer connection recognition process in S1 are executed sequentially to form the corresponding normalized semantic sample data and optional cross-layer connections; S5-2. For a new current semantic sample, based on the currently saved candidate path set and the corresponding pheromone value, repeat the target semantic path selection process in S3; after determining the target semantic path, perform semantic representation calculation along the target semantic path to form a new prediction result; For the new prediction results, repeat the error calculation and calculation parameter update process in S4-1 to S4-3 to continuously adjust the calculation parameters corresponding to the semantic level of the target semantic path under the condition of multiple semantic samples. S5-3. After completing the calculation parameter update, repeat the pheromone update process in S4-4 and S4-5. Based on the contribution of the target semantic path to the reduction of prediction error under different semantic sample conditions, continuously enhance or decay the pheromone value corresponding to each candidate path in the candidate path set. S5-4. During the process of continuously inputting multiple semantic samples and repeatedly executing S5-1 to S5-3, when it is detected that the update magnitude of the calculation parameters is continuously lower than the preset parameter change threshold within the preset number of rounds, and the ranking result of the pheromone values ​​in the candidate path set remains within the preset pheromone value change magnitude threshold range during the continuous update process, it is determined that the hierarchical semantic calculation structure has reached the convergence state and a stable hierarchical semantic calculation structure is formed. Once the hierarchical semantic computation structure reaches convergence, the update operations on computation parameters and pheromone values ​​are stopped.

10. The deep learning-based multi-layer semantic feature extraction method according to claim 9, characterized in that: In S6, the process of forming the deep learning computational architecture includes: S6-1. After the hierarchical semantic computation structure reaches a convergence state, fix the computation parameters in the hierarchical semantic computation structure and fix the pheromone value corresponding to each candidate path in the candidate path set. Based on a fixed set of candidate paths and their corresponding pheromone values, the candidate paths in the set are sorted according to their pheromone values. Candidate paths with pheromone values ​​within a preset upper limit range are identified as stable semantic paths; S6-2. Obtain the semantic samples to be processed for multi-layer semantic feature extraction. Repeat the normalization process in S1 for the semantic samples to be processed to form the corresponding normalized semantic sample data. For standardized semantic sample data, semantic representation computation is performed sequentially along the semantic hierarchy covered by the determined stable semantic path, based on the established stable semantic path. At each semantic level, the semantic output from the previous semantic level is read as the input data for the current semantic level. Semantic representation calculation is performed based on fixed calculation parameters to form the semantic output of the corresponding semantic level. This process continues until the semantic representation calculation of all semantic levels covered by the stable semantic path is completed. S6-3. Read the semantic output corresponding to each semantic level covered by the stable semantic path, and collect the semantic output of each semantic level according to the order of the semantic level identifier to form a multi-level semantic feature containing multi-level semantic representation. Output the multi-level semantic feature as the final multi-level semantic feature extraction result. Based on stable semantic paths, fixed computational parameters, and the cross-layer semantic representation computational order determined by the stable semantic paths, a deep learning computational structure for multi-layer semantic feature extraction is formed.

Images