A language semantic analysis method based on bilateral fractional order neural network

Abstract

The application belongs to the technical field of natural language processing, and discloses a language semantic analysis method based on a bilateral fractional order neural network, which comprises the following steps: step 101, constructing a bilateral fractional order operator fusion system oriented by semantic features; the fusion weight of the bilateral fractional order operator is set according to the importance proportion of the time sequence features, the syntax features and the context features in the semantic analysis task; step 102, constructing a bilateral fractional order neural network model driven by a semantic analysis task; the first-order partial derivative of gradient calculation is replaced by bilateral fractional order partial derivative, the interlayer connection weight is updated through bilateral fractional order calculus operation, and the weight iteration is completed based on the corresponding bilateral fractional order gradient descent algorithm; and step 103, data preprocessing and model training, which are used for language semantic analysis. The application improves the semantic analysis accuracy and precision of complex semantic scenes, and optimizes the semantic feature extraction capability and context adaptation performance of the model.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, specifically relating to a language semantic analysis method based on a two-sided fractional neural network. Background Technology

[0002] Language semantic analysis is one of the core tasks of natural language processing. Its purpose is to accurately capture the semantic information, syntactic structure, and contextual relationships of text, providing support for downstream applications such as machine translation, sentiment analysis, and intelligent question answering. With the development of neural network technology, various neural network models are widely used in semantic analysis tasks, among which adaptive chaotic neural networks are a typical example.

[0003] In adaptive chaotic neural networks, fractional neural networks, as an extension of integer neural networks, can comprehensively extract temporal semantic features of language text by introducing fractional calculus theory. However, there are still unresolved technical problems in existing semantic analysis applications: when using fractional neural networks to process semantic analysis tasks in complex semantic scenarios such as long texts, transitional semantics, and ambiguous semantics, the update of interlayer connection weights in existing technologies mostly uses integer gradient descent or one-sided fractional operators, which can only utilize historical gradient information before the current time step. Therefore, it is very easy to get trapped in local optima, resulting in insufficient semantic analysis accuracy. Furthermore, the corresponding weights are not designed in conjunction with the core feature attributes of the semantic analysis task, which means that even when using two-sided fractional operators, the differences in the importance of temporal features, syntactic features, and contextual features in different semantic analysis tasks will be ignored. This results in poor adaptability of operator fusion and semantic features, making it difficult for the semantic analysis results to achieve sufficient semantic analysis accuracy and adaptability. Summary of the Invention

[0004] The purpose of this invention is to provide a language semantic analysis method based on a two-sided fractional neural network, in order to solve the technical problem that existing technologies can only utilize historical gradient information before the current time and lack weight design to adapt the two-sided fractional operators to the semantic analysis task, which makes it difficult to achieve sufficient semantic analysis accuracy and adaptability for semantic analysis tasks in complex semantic scenarios.

[0005] The language semantic analysis method based on a two-sided fractional neural network includes the following steps: Step 101: Construct a semantic feature-oriented two-sided fractional operator fusion system; wherein, the fusion weight of the two-sided fractional operators is set based on the importance ratio of temporal features, syntactic features, and contextual features in the semantic analysis task, and a dynamic and interpretable fusion weight rule is established. Step 102: Construct a two-sided fractional neural network model driven by semantic analysis task; wherein, based on the two-sided fractional operator fusion system constructed in Step 101, the first-order partial derivative of gradient calculation is replaced with the two-sided fractional partial derivative, the inter-layer connection weights are updated through two-sided fractional calculus operations, and the weight iteration is completed based on the corresponding two-sided fractional gradient descent algorithm. Step 103: Data preprocessing and model training. The trained two-sided fractional neural network model is used for language semantic analysis.

[0006] Preferably, step 101 includes the following steps: (1) Clarify the mathematical definitions and operational rules of left-side fractional operators, right-side fractional operators and double-sided fractional operators, and initially construct a double-sided fractional calculus system; The expression for a two-sided fractional calculus is:

[0007] in, It is of fractional order. , It is a left-hand fractional operator. It is a right-hand fractional operator. It is a two-sided fractional operator. This is a positional parameter within the text. For in position The semantic feature function at the location, It is a function The integration interval, For the integration interval The integral variables used for traversal within the inner circle, ω1 and ω2, are the fusion weights of the fractional-order operators on the left and right sides, respectively. function The first derivative, It is a gamma function; (2) Quantitatively calculate the proportion of importance of semantic core features based on the specific semantic analysis task; (3) Initial setting of fusion weights based on the proportion of feature importance; (4) Dynamic adjustment of fusion weights based on text type.

[0008] Preferably, step (2) includes the following steps: (2.1) Complete the task adaptability definition of three types of semantic features, specifically including: Temporal characteristics: refers to the sequential order of words and the progressive relationship of emotional expression in the commentary text; Contextual features: These refer to the relationship between current words and subsequent words in a commentary text, and the logic of emotional shifts. Grammatical features: These refer to the collocation of parts of speech, syntactic structure, and the binding relationship between modifiers and the objects being evaluated in the commentary text; (2.2) Define the manually labeled dataset based on the task adaptability of the three types of semantic features, and use the SHAP interpretability analysis algorithm combined with the random forest feature contribution evaluation to complete the quantitative calculation of the importance ratio of the three types of semantic features.

[0009] Preferably, step (3) establishes a quantization mapping rule from feature importance to fusion weight based on the importance ratios of the three types of semantic features obtained by the above quantization, and calculates the initial fusion weights of the two-sided fractional operators; the calculation formula for the initial fusion weights of the two-sided fractional operators is as follows:

[0010] in, The proportion of importance of time-series features. The importance percentage of context features. ω1 and ω2 represent the proportion of grammatical feature importance, while ω1 and ω2 are the initial fusion weights of the fractional operators on the left and right sides, respectively.

[0011] Preferably, in step (4), for different types of input text in e-commerce reviews, the fusion weight is dynamically adjusted within a preset range based on the real-time changes in the importance of their semantic features, so as to achieve real-time adaptation of the weight to the features of the input text.

[0012] Preferably, step 102 includes the following steps: (1) Set up a task-oriented network structure, select the BP neural network as the basic framework, and set the number of layers and neurons of the input layer, hidden layer and output layer; (2) Introduce two-sided fractional partial derivatives for weight update. Based on the two-sided fractional operator fusion system constructed in step 101, replace the first-order partial derivative of gradient calculation with two-sided fractional partial derivatives. (3) Set the activation functions for the hidden layer and the output layer; (4) Determine the weight initialization and update methods and the loss function.

[0013] Preferably, in step (4), the weights are initialized using a Xavier normal distribution and updated using a two-sided fractional gradient descent algorithm. The weight update formula is as follows:

[0014] in, The learning rate is used; cross-entropy is chosen as the loss function. , The weights before the update. For the updated weights, It is a two-sided fractional operator. It is a function The integration interval, It is of fractional order.

[0015] Preferably, step 103 includes the following steps: (1) Preprocess the language text data. First, perform word segmentation, stop word removal, and part-of-speech tagging on the text. Then, use word embedding technology to extract a comprehensive semantic feature vector by combining temporal, grammatical, and contextual features. (2) Divide the comprehensive semantic feature vector into training set and test set according to the proportion, and use the SMOTE algorithm to process the data of a small number of categories; (3) Set training parameters and training strategies, select an optimizer for training, suppress overfitting through early stopping and regularization, and dynamically adjust the learning rate to optimize model performance.

[0016] Preferably, the language semantic analysis method based on a two-sided fractional neural network further includes step 104: introducing a context feedback mechanism to achieve adaptive optimization across the entire context.

[0017] Preferably, in step 104, the step of introducing a semantic context feedback mechanism includes: Step 1: Construct a semantic context monitoring module to collect contextual semantic information, contextual association data, and analysis result feedback data of the text in real time; Step 2: Evaluate the semantic analysis accuracy and context fit of the current model based on monitoring data. When the accuracy is lower than the preset threshold or the fit does not meet the requirements, trigger the parameter adjustment mechanism. Step 3: Dynamically adjust the fractional order of the two-sided fractional neural network The learning rate and number of iterations are adjusted to optimize the model's semantic feature extraction capabilities and context adaptation performance.

[0018] The technical advantages of this invention are as follows: This invention constructs a semantic analysis task-driven bilateral fractional-order neural network model based on bilateral fractional-order calculus. The number of neurons in each layer is set according to the semantic analysis task, and the inter-layer connection weights are updated through bilateral fractional-order calculus operations. A customized bilateral fractional-order gradient descent algorithm is used to complete weight iteration, overcoming the limitation of existing technologies that can only utilize unidirectional semantic information for weight updates. By integrating the ability of left-side fractional-order calculus to trace historical semantic information and the ability of right-side fractional-order calculus to predict future semantic associations, this invention can accurately depict the temporal dynamic evolution of language text. Even in complex semantic scenarios, it can avoid the model prematurely converging to a local optimum, effectively improving the accuracy and precision of semantic analysis in complex semantic scenarios. Simultaneously, this invention constructs a bilateral fractional-order calculus theoretical system through a semantic feature-guided linear weighted fusion rule. The fusion weights are set based on the importance ratio of temporal features, syntactic features, and contextual features in the semantic analysis task, and can be dynamically adjusted according to the text type, optimizing the model's semantic feature extraction capability and context adaptation performance. Attached Figure Description

[0019] Figure 1 is a schematic diagram of the basic process of a language semantic analysis method based on a two-sided fractional neural network according to the present invention; Figure 2 shows a comparison of the semantic analysis accuracy of the present invention (two-sided fractional BP network model, TFBP) and the left and right fractional BP network models (LFBP, RFBP) on the training set when the learning rate is 0.1. Figure 3 shows a comparison of the semantic analysis accuracy of the present invention with the left and right fractional BP network models on the test set when the learning rate is 0.1. Figure 4 shows a comparison of the semantic analysis accuracy of the present invention with the left and right fractional-order BP network models under different learning rates under the Sigmoid activation function. Figure 5 shows a comparison of the semantic analysis accuracy of the present invention with the left and right fractional-order BP network models under the ReLU activation function, based on different learning rates. Figure 6 is a performance comparison chart of the present invention and the existing fractional gradient descent model of fractional-order BP network. Detailed Implementation

[0020] The following detailed description of the embodiments, with reference to the accompanying drawings, will further illustrate the specific implementation of the present invention, in order to help those skilled in the art to have a more complete, accurate, and in-depth understanding of the inventive concept and technical solution of the present invention.

[0021] While existing technologies employ fractional neural networks for semantic analysis tasks, their structural design and weight update mechanisms are not deeply adapted to the task. The number of neurons lacks clear task orientation, leading to insufficient feature extraction or model redundancy and overfitting due to network structure mismatch. Furthermore, the updating of inter-layer connection weights often relies on integer-order gradient descent or one-sided fractional operators, utilizing only historical gradient information prior to the current moment. This fails to simultaneously incorporate future semantic associations after the current moment, making it difficult to accurately depict the temporal dynamic evolution of language text from "foreword - current semantics - subsequent extension." Consequently, in long texts, transitional semantics, and ambiguous semantic scenarios, weight update convergence accuracy is low, easily getting trapped in local optima, hindering further improvement in semantic analysis accuracy.

[0022] As for the application of two-sided fractional neural networks, the existing fusion schemes of two-sided fractional operators only use simple linear superposition of fixed weights. They do not combine the core feature attributes of semantic analysis tasks to design weights, ignore the differences in the importance of temporal features, syntactic features, and contextual features in different semantic analysis tasks, and cannot dynamically adjust according to different text types such as formal text, spoken text, and professional domain text. This results in extremely poor adaptability of operator fusion to semantic features, and the semantic analysis accuracy fluctuates greatly under different tasks and different text scenarios. Therefore, the semantic analysis accuracy of the model is still insufficient and lacks generalization ability.

[0023] like Figures 1-6 As shown, to address the technical problems encountered when applying fractional-order neural networks in semantic analysis tasks, this invention provides a language semantic analysis method based on a two-sided fractional-order neural network. In this embodiment, it is used for sentiment semantic analysis of e-commerce product reviews. The language semantic analysis method includes the following steps: Step 101: Construct a semantic feature-guided two-sided fractional operator fusion system.

[0024] In this step, the mathematical definitions and operation rules of the left-side fractional operator, the right-side fractional operator, and the bilateral fractional operator are first clarified. Then, the bilateral fractional operator fusion system is constructed through the semantic feature-guided linear weighted fusion rule.

[0025] In this step, the fusion weights of the two-sided fractional-order operators are set based on the importance ratios of temporal features, syntactic features, and contextual features in the semantic analysis task, and dynamic and interpretable fusion weight rules are established. This dynamically adjusts the fusion weights according to the text type, thereby addressing the core deficiency of existing technologies where fixed-weight fusion of two-sided operators results in poor adaptability to semantic features.

[0026] This step specifically includes the following steps: (1) Clarify the mathematical definitions and operational rules of left-side fractional operators, right-side fractional operators and double-sided fractional operators, and initially construct a double-sided fractional calculus system.

[0027] (1.1) Define the fractional calculus on the left side:

[0028] in It is a fractional order; in this embodiment, a fractional order is selected. Capture demand by adapting the temporal features of sentiment semantics in e-commerce reviews. It is a left-hand fractional operator. It is a function The integration interval, This is a positional parameter within the text. For in position The semantic feature function at the location, For the integration interval The integral variable that is iterated over inside. function The nth derivative, This is a gamma function.

[0029] In particular, when The fractional calculus expression on the left side is: .

[0030] (1.2) Define the fractional calculus on the right-hand side:

[0031] in It is of fractional order. It is a right-hand fractional operator. It is a function The integration interval.

[0032] In particular, when The fractional calculus expression on the right is: .

[0033] (1.3) Based on the theory of left-hand and right-hand fractional calculus, we define two-handed fractional calculus:

[0034] in It is of fractional order. It is a two-sided fractional operator. It is a function The integration interval.

[0035] In particular, when The expression for a two-sided fractional calculus is:

[0036] Simplifying the above formula, we get:

[0037] Where ω1 and ω2 are the fusion weights of the fractional-order operators on the left and right sides, respectively. function The first derivative; It is a function The integration interval is thus determined. This completes the initial construction of the two-sided fractional calculus system.

[0038] (2) Quantitatively calculate the proportion of importance of semantic core features based on the specific semantic analysis task.

[0039] For the e-commerce product review sentiment analysis task in this embodiment, the definitions and task suitability of the three types of semantic features are first clarified. Then, the importance ratio is quantitatively calculated through an interpretable algorithm to provide a unique objective basis for setting the fusion weight. Specifically, the steps include the following: (2.1) Complete the task adaptability definition of three types of semantic features, specifically including: Temporal features: These refer to the sequential order of words and the progressive relationship of sentiment expression in the comment text. They determine the model's ability to capture the progressive logic of "previous product description → current sentiment evaluation" and are the core basic features of e-commerce comment sentiment analysis.

[0040] Contextual features refer to the relationship between the current words and the words in the following text in the comment text, as well as the logic of the emotional transition. This determines the model's ability to capture long-distance associations such as "current positive evaluation → subsequent negative transition" and "current product parameters → subsequent user experience". It is the core differentiating feature of e-commerce comment sentiment analysis.

[0041] Grammatical features: These refer to the word collocation, syntactic structure, and binding relationship between modifiers and the evaluation object in the review text. They determine the model's ability to accurately match "evaluation subject - sentiment vocabulary" and avoid matching errors such as "positive reviews for products and negative reviews for logistics". They are the core auxiliary features of e-commerce review sentiment analysis.

[0042] (2.2) Quantitative calculation of the proportion of feature importance for the three types of semantic features.

[0043] This step targets the large number (10,000) of e-commerce reviews collected in this embodiment. Based on the task adaptability definition of the three types of semantic features, a manually labeled dataset is defined. The importance percentage of the three types of semantic features is quantitatively calculated using the SHAP interpretability analysis algorithm combined with random forest feature contribution evaluation. The specific steps are as follows: (2.2.1) Using manually labeled “positive / negative” sentiment polarity labels as the target variable, and temporal features, grammatical features, and contextual features as explanatory variables, a random forest baseline classification model was trained to fit the influence of the three types of semantic features on the sentiment classification results.

[0044] (2.2.2) The SHAP algorithm is used to calculate the marginal contribution values ​​of the three types of semantic features to the sentiment classification results, eliminate collinearity interference between features, and normalize the contribution values ​​to obtain the importance ratio of the three types of semantic features in the sentiment analysis task of this embodiment. The results obtained in this embodiment are: temporal features account for 45% of the importance, contextual features account for 40% of the importance, and syntactic features account for 15% of the importance.

[0045] (3) Initial setting of fusion weights based on the proportion of feature importance.

[0046] This step, based on the importance proportions of the three semantic features obtained through quantization, establishes a quantization mapping rule from feature importance to fusion weights, calculates the initial fusion weights of the two-sided fractional operators, and ensures that the mapping rule perfectly matches the semantic function anchoring of the operators. The formula for calculating the initial fusion weights of the two-sided fractional operators is as follows:

[0047] in, The proportion of importance of time-series features. The proportion of importance of context features. ω1 and ω2 represent the proportion of grammatical feature importance, while ω1 and ω2 are the initial fusion weights of the fractional-order operators on the left and right sides, respectively.

[0048] The rules for calculating the initial fusion weights include: the higher the proportion of temporal features, the higher the weight of the left-hand operator, thus strengthening the ability to trace historical semantics; the higher the proportion of contextual features, the higher the weight of the right-hand operator, thus strengthening the ability to predict future semantics; and the higher the proportion of syntactic features, the closer the weight is to the 0.5 equilibrium value, thus strengthening the ability to capture global syntactic structures.

[0049] In this embodiment, the quantification results of the importance ratio of the three types of semantic features are as follows: , , Substitute this into the formula for calculating the initial fusion weight to obtain the initial fusion weight.

[0050] The initial weights perfectly match the characteristic requirements of the sentiment analysis task in this embodiment: temporal features account for the highest proportion, therefore the left-hand operator has a higher weight, strengthening the ability to trace historical sentiment progression information, while also taking into account the ability to predict future sentiment shifts corresponding to contextual features. Comparative experiments have verified that these initial weights are superior to the conventional fixed equal-weight weights in existing technologies. , When applied directly, compared with existing technologies, the sentiment analysis accuracy of this embodiment is improved by 6.2%, the number of convergence iterations is reduced by 31%, and the convergence speed, training efficiency and analysis accuracy are significantly improved.

[0051] (4) Dynamic adjustment of fusion weights based on text type.

[0052] For different types of input text in e-commerce reviews, the fusion weights are dynamically adjusted within a preset range based on the real-time changes in the importance of their semantic features, achieving real-time adaptation between the weights and the features of the input text. In this embodiment, the specific adjustment rules and effects are as follows: For short reviews on beauty and fashion that are highly colloquial: These texts use a lot of online slang, have strong emotional shifts in context, and prominent transitional features. Based on real-time feature importance calculations, the proportion of contextual features increased to 48%, temporal features decreased to 38%, and grammatical features accounted for 14%. The corresponding fusion weights were adjusted accordingly. , This enhances the ability of the right-side operator to predict future emotional shifts. Tests showed that, after dynamic adjustment, the accuracy of sentiment analysis for colloquial text improved by 8.7% compared to fixed weights. For detailed and rigorous parameter reviews of 3C digital products and home appliances: This type of text features a standardized grammatical structure, a prominent sequential progression in product parameter descriptions, and a clearly defined evaluator. Based on real-time feature importance calculations, the proportion of sequential features increased to 52%, contextual features decreased to 33%, and grammatical features accounted for 15%. Correspondingly, the fusion weights were adjusted accordingly. , This strengthens the ability of the left-side operator to trace historical parameter descriptions and evaluations. Tests showed that, after dynamic adjustment, the accuracy of matching the subject of professional parameter text evaluation improved by 9.3% compared to fixed weights. For long-text in-depth evaluation comments: These texts are lengthy, have prominent long-distance contextual relationships, and progressively developing emotional expressions. After real-time feature importance calculation, the proportion of contextual features is increased to 55%, temporal features to 30%, and grammatical features to 15%. The corresponding fusion weights are adjusted accordingly. , This enhances the ability of the right-side operator to predict semantic associations over long distances. After testing, the accuracy of capturing semantic associations in long texts improved by 10.2% compared to fixed weights after dynamic adjustment.

[0053] After the dynamic adjustment of the fusion weights, this step completes the construction of the two-sided fractional operator fusion system.

[0054] Step 102: Construct a two-sided fractional neural network model (TFBP) driven by semantic analysis task.

[0055] This step replaces the first-order partial derivative of the gradient calculation with the fractional partial derivative, constructs a two-sided fractional neural network model driven by the semantic analysis task, sets the number of neurons in each layer according to the semantic analysis task, updates the inter-layer connection weights through two-sided fractional calculus operations, and completes the weight iteration based on the corresponding two-sided fractional gradient descent algorithm, breaking through the limitation of existing technology that weight updates can only utilize unidirectional semantic information.

[0056] This step specifically includes the following steps: (1) Set up a task-oriented network structure.

[0057] A backpropagation (BP) neural network was selected as the basic framework (the two-sided fractional-order BP network model, abbreviated as TFBP), with the number of neurons in the input layer set to 128 (the dimension of the matching word embedding vector), one hidden layer with 64 neurons, and two neurons in the output layer. This structural configuration is based on the feature dimension requirements of sentiment analysis in this embodiment, avoiding the problems of insufficient feature extraction or model redundancy caused by arbitrarily setting the number of neurons in existing technologies.

[0058] (2) Introduce two-sided fractional partial derivatives for weight update.

[0059] Based on the bilateral fractional-order operator fusion system constructed in step 101, the first-order partial derivative of the gradient calculation is replaced with a bilateral fractional-order partial derivative. In this embodiment, the fractional order is set to α=0.5. This step enables the weight update mechanism to simultaneously combine historical and future semantic association information, breaking the limitation of existing technologies that can only utilize unidirectional (historical) gradient information.

[0060] (3) Set the activation functions for the hidden layer and the output layer.

[0061] The hidden layer uses the ReLU activation function, and the output layer uses the Sigmoid activation function to normalize the sentiment probability.

[0062] (4) Determine the weight initialization and update methods and the loss function.

[0063] The weights are initialized using a Xavier normal distribution and updated using a two-sided fractional gradient descent algorithm. The weight update formula is as follows:

[0064] in, The learning rate is used. Cross-entropy is chosen as the loss function. , The weights before the update. For the updated weights, It is a two-sided fractional operator. It is a function The integration interval, It is a fractional order used for training and optimizing the model of this method. The innovation of the bilateral fractional order neural network model lies in its ability to more accurately depict the temporal dynamic evolution from the preceding context to the current semantics, and then to the subsequent extension, effectively suppressing the risk of the model getting trapped in local optima.

[0065] Step 103: Data preprocessing and model training.

[0066] This step preprocesses the language text data by first segmenting the text, removing stop words, and tagging parts of speech. Then, it uses word embedding technology to extract comprehensive semantic features by combining temporal, grammatical, and contextual features. Next, it divides the training and test sets proportionally, addresses the class imbalance problem, sets initial parameters, uses cross-entropy as the loss function, and trains the model using an optimizer. It also suppresses overfitting through early stopping and regularization, and dynamically adjusts the learning rate to optimize model performance.

[0067] This step specifically includes the following steps: (1) Preprocess the language text data.

[0068] This step first preprocesses the text by segmenting, removing stop words, and tagging parts of speech. Then, it uses word embedding technology to extract a comprehensive semantic feature vector by combining temporal, grammatical, and contextual features.

[0069] Specifically, this step uses jieba segmentation to segment all comments, removes stop words based on the stop word list, and uses Stanford PO tool for part-of-speech tagging. At the same time, Word2Vec is used to convert the vocabulary into 128-dimensional word embedding vectors, extract the temporal semantics, grammatical structure, and contextual features of the comments, and concatenates them into a 256-dimensional comprehensive semantic feature vector.

[0070] (2) Divide the comprehensive semantic feature vector into training set and test set according to the proportion.

[0071] This step divides the 10,000 feature vectors into an 8:2 ratio, resulting in 8,000 training vectors and 2,000 test vectors. The SMOTE algorithm is then used to process the limited number of categories of comment data, completing the standardization process. The SMOTE algorithm effectively addresses the class imbalance problem.

[0072] (3) Set training parameters and training strategy, and select an optimizer for training.

[0073] Set training parameters: Set initial parameters Learning rate The number of iterations is 1000, and the batch size is 32.

[0074] The Adam optimizer was selected; an early stopping strategy was employed during training, with the learning rate varying according to the gradient every 300 epochs and the results output. Overfitting was suppressed through early stopping and regularization, and the learning rate was dynamically adjusted to optimize model performance.

[0075] Step 104: Introduce a contextual feedback mechanism to achieve adaptive optimization across the entire context.

[0076] This step introduces a semantic context feedback mechanism, builds a monitoring module, collects text context information, context features and analysis results feedback data in real time, establishes an evaluation system, sets core indicator thresholds, and when the indicators do not meet the standards, adaptively adjusts parameters such as score level, learning rate, and fusion weight within a preset range to adapt to the semantic analysis needs of different contexts.

[0077] The steps for introducing the semantic context feedback mechanism include: Step 1: Construct a semantic context monitoring module to collect contextual semantic information, contextual association data, and analysis result feedback data of the text in real time; Step 2: Evaluate the semantic analysis accuracy and context fit of the current model based on monitoring data. When the accuracy is lower than the preset threshold or the fit does not meet the requirements, trigger the parameter adjustment mechanism. Step 3: Dynamically adjust the fractional order of the two-sided fractional neural network Parameters such as learning rate and number of iterations are used to optimize the model's semantic feature extraction capabilities and context adaptation performance.

[0078] In this embodiment, the specific execution process of this step is as follows: (1) Real-time collection of contextual information, linguistic features, and analysis results feedback data for newly added beauty reviews. This includes the core beauty product type (e.g., foundation, lipstick), skin type description (e.g., dry skin, oily skin), the contextual information of emotional words (e.g., "drying" in relation to foundation use), and the deviation data between model analysis results and manual annotation results. The collection frequency is real-time, with an indicator evaluation conducted every 100 newly added reviews.

[0079] (2) Set semantic analysis accuracy ≥90% and context fit ≥85% as evaluation thresholds. When the accuracy is lower than the threshold, adjust the parameters.

[0080] (3) Adjust the learning rate to 0.25 and 0.5 respectively, retrain the model, and output the results.

[0081] The semantic analysis task was performed on the model of this invention (two-sided fractional-order BP network model) and the existing left-side fractional-order BP network model and right-side fractional-order BP network model based on the test set and training set obtained in this embodiment. The semantic analysis accuracy of each model was statistically analyzed. The semantic analysis accuracy comparison chart of the model of this invention and similar models in the prior art is shown in the figure below. Figure 2-5 As shown.

[0082] The performance data of the model of this invention (improved model) and the fractional gradient descent model of the fractional-order BP network before improvement in the prior art are then analyzed, and the resulting performance data comparison chart is shown below. Figure 6 As shown, the performance data includes prediction accuracy, training convergence speed (number of iterations), mean squared error (MSE), and training time.

[0083] The above comparison demonstrates that this invention, based on the memory and hereditary characteristics of two-sided fractional calculus, constructs a two-sided fractional neural network model. By integrating the ability of the left-side fractional calculus to trace historical semantic information and the right-side fractional calculus to predict future semantic relationships, it comprehensively extracts the temporal semantic features of language text, providing a new, efficient, accurate, and universal method for the field of natural language processing. When applied to semantic analysis tasks in complex language environments, this invention achieves significantly higher accuracy than existing single-sided fractional BP network models, and also exhibits clear advantages in various performance metrics.

[0084] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.

Claims

1. A language semantic analysis method based on a two-sided fractional neural network, characterized in that, Includes the following steps: Step 101: Construct a semantic feature-oriented two-sided fractional operator fusion system; wherein, the fusion weight of the two-sided fractional operators is set based on the importance ratio of temporal features, syntactic features, and contextual features in the semantic analysis task, and a dynamic and interpretable fusion weight rule is established. Step 102: Construct a two-sided fractional neural network model driven by semantic analysis task; wherein, based on the two-sided fractional operator fusion system constructed in Step 101, the first-order partial derivative of gradient calculation is replaced with the two-sided fractional partial derivative, the inter-layer connection weights are updated through two-sided fractional calculus operations, and the weight iteration is completed based on the corresponding two-sided fractional gradient descent algorithm. Step 103: Data preprocessing and model training. The trained two-sided fractional neural network model is used for language semantic analysis.

2. The language semantic analysis method based on a two-sided fractional neural network according to claim 1, characterized in that, Step 101 includes the following steps: (1) Clarify the mathematical definitions and operational rules of left-side fractional operators, right-side fractional operators and double-sided fractional operators, and initially construct a double-sided fractional calculus system; The expression for a two-sided fractional calculus is: in, It is of fractional order. , It is a left-hand fractional operator. It is a right-hand fractional operator. It is a two-sided fractional operator. This is a positional parameter within the text. For in position The semantic feature function at the location, It is a function The integration interval, For the integration interval The integral variables used for traversal within the inner circle, ω1 and ω2, are the fusion weights of the fractional-order operators on the left and right sides, respectively. function The first derivative, It is a gamma function; (2) Quantitatively calculate the proportion of importance of semantic core features based on the specific semantic analysis task; (3) Initial setting of fusion weights based on the proportion of feature importance; (4) Dynamic adjustment of fusion weights based on text type.

3. The language semantic analysis method based on a two-sided fractional neural network according to claim 2, characterized in that, Step (2) includes the following steps: (2.1) Complete the task adaptability definition of three types of semantic features, specifically including: Temporal characteristics: refers to the sequential order of words and the progressive relationship of emotional expression in the commentary text; Contextual features: These refer to the relationship between current words and subsequent words in a commentary text, and the logic of emotional shifts. Grammatical features: These refer to the collocation of parts of speech, syntactic structure, and the binding relationship between modifiers and the objects being evaluated in the commentary text; (2.2) Define the manually labeled dataset based on the task adaptability of the three types of semantic features, and use the SHAP interpretability analysis algorithm combined with the random forest feature contribution evaluation to complete the quantitative calculation of the importance ratio of the three types of semantic features.

4. The language semantic analysis method based on a two-sided fractional neural network according to claim 2, characterized in that, Step (3) establishes a quantization mapping rule from feature importance to fusion weight based on the three types of semantic feature importance ratios obtained by the above quantization, and calculates the initial fusion weights of the two-sided fractional operators. The formula for calculating the initial fusion weights of the two-sided fractional operators is: in, The proportion of importance of time-series features. The importance percentage of context features. ω1 and ω2 represent the proportion of grammatical feature importance, while ω1 and ω2 are the initial fusion weights of the fractional operators on the left and right sides, respectively.

5. The language semantic analysis method based on a two-sided fractional neural network according to claim 2, characterized in that, In step (4), for different types of input text in e-commerce reviews, the fusion weight is dynamically adjusted within a preset range based on the real-time changes in the importance of their semantic features, so as to achieve real-time adaptation of the weight to the features of the input text.

6. The language semantic analysis method based on a two-sided fractional neural network according to claim 1, characterized in that, Step 102 includes the following steps: (1) Set up a task-oriented network structure, select the BP neural network as the basic framework, and set the number of layers and neurons of the input layer, hidden layer and output layer; (2) Introduce two-sided fractional partial derivatives for weight update. Based on the two-sided fractional operator fusion system constructed in step 101, replace the first-order partial derivative of gradient calculation with two-sided fractional partial derivatives. (3) Set the activation functions for the hidden layer and the output layer; (4) Determine the weight initialization and update methods and the loss function.

7. The language semantic analysis method based on a two-sided fractional neural network according to claim 6, characterized in that, In step (4), the weights are initialized using a Xavier normal distribution and updated using a two-sided fractional gradient descent algorithm. The weight update formula is as follows: in, The learning rate is used; cross-entropy is chosen as the loss function. , The weights before the update. For the updated weights, It is a two-sided fractional operator. It is a function The integration interval, It is of fractional order.

8. The language semantic analysis method based on a two-sided fractional neural network according to claim 1, characterized in that, Step 103 includes the following steps: (1) Preprocess the language text data. First, perform word segmentation, stop word removal, and part-of-speech tagging on the text. Then, use word embedding technology to extract a comprehensive semantic feature vector by combining temporal, grammatical, and contextual features. (2) Divide the comprehensive semantic feature vector into training set and test set according to the proportion, and use the SMOTE algorithm to process the data of a small number of categories; (3) Set training parameters and training strategies, select an optimizer for training, suppress overfitting through early stopping and regularization, and dynamically adjust the learning rate to optimize model performance.

9. The language semantic analysis method based on a two-sided fractional neural network according to claim 1, characterized in that, It also includes step 104: introducing a contextual feedback mechanism to achieve adaptive optimization across the entire context.

10. The language semantic analysis method based on a two-sided fractional neural network according to claim 9, characterized in that, In step 104, the step of introducing a semantic context feedback mechanism includes: Step 1: Construct a semantic context monitoring module to collect contextual semantic information, contextual association data, and analysis result feedback data of the text in real time; Step 2: Evaluate the semantic analysis accuracy and context fit of the current model based on monitoring data. When the accuracy is lower than the preset threshold or the fit does not meet the requirements, trigger the parameter adjustment mechanism. Step 3: Dynamically adjust the fractional order of the two-sided fractional neural network The learning rate and number of iterations are adjusted to optimize the model's semantic feature extraction capabilities and context adaptation performance.

Images