An opinion propagation prediction method and system based on natural language processing

Abstract

This invention relates to the fields of natural language processing and online software update technology, and particularly to a method and system for predicting public opinion propagation based on natural language processing. The method includes: allocating multi-core shared contiguous physical memory pages on computing nodes and dividing them into a lock-free spin update main region and a first remote memory access mirror backup region; collecting text stream data from a distributed public opinion monitoring network and extracting unregistered words; constructing a spatiotemporal cross-node public opinion evolution tensor and determining suddenly popular public opinion terms; obtaining orthogonal public opinion semantic vectors through dimensionality reduction using Riemannian manifold tensor projection; overwriting the orthogonal public opinion semantic vectors to absolute physical memory addresses using transaction synchronization extension instructions and pushing them to the second remote memory access mirror backup region on remote nodes; updating the first-level static hash routing table within the natural language model and performing forward public opinion evolution inference operations. This invention solves the technical problem of memory fragmentation and system crashes caused by high-frequency online updates of the model triggered by suddenly popular terms.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and online software update technology, and in particular to a method and system for predicting public opinion dissemination based on natural language processing. Background Technology

[0002] Natural Language Processing (NLP) technology is widely used in distributed real-time public opinion monitoring and evolutionary prediction applications. Related public opinion propagation prediction methods utilize NLP models to analyze the characteristics of online text. The frequent generation of emerging trending terms in the online environment causes pre-deployed NLP models to fail to accurately identify new words. To maintain the accuracy of public opinion prediction and analysis, the prediction system needs to perform online updates to the NLP model. The high frequency of emerging trending terms triggers frequent online updates to the NLP model, requiring frequent allocation and release of physical memory resources on the underlying computing nodes. Frequent allocation and release of physical memory resources generates a large number of discontinuous free memory blocks, which accumulate over time, leading to memory fragmentation. Continued deterioration of memory fragmentation exhausts available contiguous physical memory, causing the public opinion prediction node to stop processing data and ultimately resulting in a system crash. Existing public opinion monitoring technologies have significant shortcomings and fail to address the technical pain point of memory fragmentation in public opinion prediction nodes and system crashes caused by frequent online updates of the NLP model triggered by emerging trending terms. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for predicting public opinion dissemination based on natural language processing. This invention solves the technical problem that the high-frequency online updates of natural language models triggered by sudden popular phrases cause memory fragmentation of public opinion prediction nodes and lead to system crashes.

[0004] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: Firstly, the present invention provides a method for predicting public opinion dissemination based on natural language processing, comprising: Multi-core shared contiguous physical memory pages are allocated on the computing node. The multi-core shared contiguous physical memory pages are divided into a lock-free spin update main area and a first remote memory access mirror backup area. The target feature is embedded in a slot in the lock-free spin update main area. Collect text stream data from multiple physical nodes in a distributed public opinion monitoring network, and extract unregistered words from the text stream data; Extract the spatiotemporal propagation characteristics of unregistered words across different physical nodes within a preset time window, and construct a spatiotemporal cross-node public opinion evolution tensor; When the temporal and spatial cross-node public opinion evolution tensor exceeds the preset public opinion outbreak warning threshold, it is determined that unregistered words are transformed into suddenly popular public opinion terms. Construct a context co-occurrence network of sudden popular public opinion terms within text stream data, perform dimensionality reduction operation on the context co-occurrence network to obtain evolution momentum scalar and orthogonal public opinion semantic vectors, so that the byte length of the orthogonal public opinion semantic vectors matches the single throughput bit width of the cache; Select inactive target feature embedding slots from the lock-free spin update master region, and use evolutionary momentum scalar to generate transaction synchronization extension instructions; The transaction synchronization extension instruction is used to overwrite the orthogonal public opinion semantic vector to the physical memory absolute address corresponding to the target feature embedding slot, and the remote memory access protocol is triggered synchronously to push the orthogonal public opinion semantic vector to the second remote memory access mirror backup area of ​​the remote distributed public opinion prediction node. The physical memory absolute address is updated to the first-level static hash routing table inside the pre-deployed natural language model, and the orthogonal public opinion semantic vector is read from the physical memory absolute address using the updated first-level static hash routing table; By integrating the base network parameters of a natural language model with orthogonal public opinion semantic vectors, forward public opinion evolution inference operations are performed on text stream data, and the prediction results of cross-node public opinion propagation trajectory are output.

[0005] Furthermore, the public opinion propagation prediction method based on natural language processing of the present invention allocates multi-core shared contiguous physical memory pages on computing nodes, and divides the multi-core shared contiguous physical memory pages into a lock-free spin-update primary region and a first remote memory access mirror backup region, including: Query the physical node mapping associations of a non-uniform memory access architecture; Based on the physical node mapping association, generate a contiguous physical memory page allocation request that includes the specified physical node; Receive the physical memory page base address pointer returned based on the contiguous physical memory page allocation request; Using the physical memory page base address pointer, the lock-free spin update main area and the first remote memory access mirror backup area are divided at a specified physical node, and the target feature is embedded in the lock-free spin update main area with preset target feature embedding slots; Read the starting physical address corresponding to the physical memory page base address pointer; Calculate the modulo result of the starting physical address relative to the width of the compute node's data cache line; Based on the modulo result, the starting physical address is offset to align with the boundary of the compute node's data cache line; By using the page locking system call instruction, the memory page table attribute bits corresponding to the shared contiguous physical memory pages of multiple cores are modified to a non-swapable state.

[0006] Furthermore, the public opinion propagation prediction method based on natural language processing of the present invention extracts the spatiotemporal propagation characteristics of unregistered words across different physical nodes within a preset time window, and constructs a spatiotemporal cross-node public opinion evolution tensor, including: The text stream data is stored in the receiving buffer area, the text stream data in the receiving buffer area is read, word segmentation is performed on the text stream data, and candidate words for segmentation are obtained. Extract known words from the basic dictionary and perform hash comparison between the candidate words and the known words; Filter the segmentation candidate words that did not match the known words into the unregistered words and add them to the unregistered words queue. Record the time difference between the transmission of unregistered words across different physical nodes in the unregistered word queue, and obtain the propagation delay time value within the preset time window; Count the frequency of unregistered words appearing in local areas within a preset time window; Convert the propagation delay time value into a propagation delay time vector between nodes; Convert the frequency values ​​of local regions into local outbreak frequency vectors; By splicing the propagation delay time vector between nodes and the local outbreak frequency vector, a spatiotemporal cross-node public opinion evolution tensor is generated.

[0007] Furthermore, the public opinion propagation prediction method based on natural language processing of the present invention constructs a contextual co-occurrence network of suddenly popular public opinion terms within text stream data, including: Using suddenly popular public opinion terms as the central words, the context text sequence is extracted from the text stream data according to the preset character span; Extract related words from the context text sequence, excluding the central word; The absolute positional interval between statistically related keywords and suddenly popular public opinion terms; Generate an adjacency matrix using absolute positional interval values; By using associated words as graph nodes and the values ​​in the adjacency matrix as graph edge weights, a context co-occurrence network is constructed.

[0008] Furthermore, the public opinion propagation prediction method based on natural language processing of the present invention performs dimensionality reduction operation on the context co-occurrence network to obtain an evolutionary momentum scalar and orthogonal public opinion semantic vectors, such that the byte length of the orthogonal public opinion semantic vectors matches the single-pass throughput bit width of the cache, including: Project the context co-occurrence network onto a high-dimensional Riemannian manifold space; Calculate the geodesic distance parameters between graph nodes in a context co-occurrence network within a high-dimensional Riemannian manifold space; Using the geodesic distance parameter, Riemannian manifold tensor projection dimensionality reduction operation is performed on the context co-occurrence network along the tangent space direction to output the initial semantic vector; During the Riemannian manifold tensor projection dimensionality reduction operation, the gradient of the dimensionality reduction loss rate change is extracted to generate the evolutionary momentum scalar. The initial semantic vector is truncated to obtain an orthogonal public opinion semantic vector, such that the total byte length of the orthogonal public opinion semantic vector is equal to the single throughput bit width of the cache.

[0009] Furthermore, the public opinion propagation prediction method based on natural language processing of the present invention utilizes evolutionary momentum scalars to generate transaction synchronization extension instructions, and uses these instructions to overwrite orthogonal public opinion semantic vectors to the physical memory absolute address corresponding to the target feature embedding slot, including: Extract the target feature from the lock-free spin update master region and embed it into the preset base address of the slot; The evolution momentum scalar is added to the preset base address to calculate the corresponding physical offset address, and the physical offset address is determined as the absolute physical memory address corresponding to the target feature embedding slot. Encapsulate the absolute physical memory address and orthogonal public opinion semantic vector into a transaction synchronization extension instruction; Execute the transaction synchronization extension instruction to write the orthogonal public opinion semantic vector to the absolute address of physical memory; A network interrupt signal is generated by utilizing the data write level flip change that occurs at the absolute address of physical memory.

[0010] Furthermore, the public opinion propagation prediction method based on natural language processing of the present invention synchronously triggers a remote memory access protocol to push orthogonal public opinion semantic vectors to the second remote memory access mirror backup area of ​​the remote distributed public opinion prediction node, including: In response to a network interruption signal, read the orthogonal public opinion semantic vector stored in the absolute address of physical memory; Add a destination address header to the orthogonal public opinion semantic vector and encapsulate it into a remote memory access protocol network data frame; The remote memory access protocol network data frames are sent to remote distributed public opinion prediction nodes using network communication links. Receive network data frames from Remote Memory Access Protocol (RMAP) and extract orthogonal public opinion semantic vectors. The extracted orthogonal public opinion semantic vectors are written to the second remote memory access mirror backup area of ​​the remote distributed public opinion prediction node using a zero-copy mechanism.

[0011] Furthermore, the public opinion propagation prediction method based on natural language processing of the present invention updates the absolute physical memory address to the first-level static hash routing table within the pre-deployed natural language model, including: Receive cross-node parameter synchronization confirmation events; In response to cross-node parameter synchronization confirmation events, perform hash operations on sudden trending public opinion terms to generate string identifiers; Read the absolute physical memory address where the orthogonal public opinion semantic vector is stored; Combine string identifiers with absolute physical memory addresses to construct key-value pairs and update data; By using a single-step atomic compare and swap instruction, the key-value pair update data is overwritten into the first-level static hash routing table inside the natural language model; Generate an address path ready status flag using the updated first-level static hash routing table.

[0012] Furthermore, the public opinion propagation prediction method based on natural language processing of the present invention integrates the base network parameters of the natural language model with orthogonal public opinion semantic vectors, performs forward public opinion evolution inference operations on text stream data, and outputs cross-node public opinion propagation trajectory prediction results, including: In response to the address path ready status flag, the received text stream data is input into the natural language model; Using the updated first-level static hash routing table, query the string identifiers corresponding to the sudden and popular public opinion phrases included in the text stream data; Using the absolute physical memory address corresponding to the string identifier in the first-level static hash routing table, the orthogonal public opinion semantic vector is read; An evolutionary inference network is constructed by integrating orthogonal public opinion semantic vectors with the base network parameters of a natural language model. By using an evolutionary inference network to perform forward public opinion evolutionary inference matrix multiplication on text stream data, the prediction results of cross-node public opinion propagation trajectory are calculated.

[0013] Secondly, the present invention provides a public opinion propagation prediction system based on natural language processing, applied to the public opinion propagation prediction system based on natural language processing as described above, comprising: The memory allocation module is used to allocate multi-core shared contiguous physical memory pages on the computing node, divide the multi-core shared contiguous physical memory pages into a lock-free spin update main area and a first remote memory access mirror backup area, and embed slots with preset target features in the lock-free spin update main area; The vocabulary extraction module is used to collect text stream data from the distributed public opinion monitoring network and extract unregistered words from the text stream data. The tensor construction module is used to extract the spatiotemporal propagation characteristics of unregistered words within a preset time window and construct a spatiotemporal cross-node public opinion evolution tensor. The status determination module is used to determine whether unregistered words are transformed into suddenly popular public opinion terms when the tensor of public opinion evolution across nodes in time and space exceeds the preset public opinion outbreak warning threshold. The dimensionality reduction module is used to construct a context co-occurrence network of sudden popular public opinion terms within the text stream data. It performs dimensionality reduction operations on the context co-occurrence network to obtain the evolution momentum scalar and orthogonal public opinion semantic vectors, so that the byte length of the orthogonal public opinion semantic vectors matches the single throughput bit width of the cache. The instruction generation module is used to select target feature embedding slots that are inactive from the lock-free spin update master region and generate transaction synchronization extension instructions using evolution momentum scalars. The physical overwrite module is used to overwrite the orthogonal public opinion semantic vector to the physical memory absolute address corresponding to the target feature embedding slot using transaction synchronization extension instructions, and synchronously trigger the remote memory access protocol to push the orthogonal public opinion semantic vector to the second remote memory access mirror backup area of ​​the remote distributed public opinion prediction node. The routing update module is used to update the physical memory absolute address to the first-level static hash routing table inside the pre-deployed natural language model, and use the updated first-level static hash routing table to read the orthogonal public opinion semantic vector from the physical memory absolute address. The evolutionary inference module is used to fuse the base network parameters of the natural language model with orthogonal public opinion semantic vectors, perform forward public opinion evolutionary inference operations on text stream data, and output cross-node public opinion propagation trajectory prediction results.

[0014] Beneficial effects of this invention: Natural language models (NLPs) frequently trigger dynamic physical memory allocation and release operations when faced with high-frequency online updates of emerging trending public opinion terms. This frequent allocation and release of physical memory resources generates a large number of discontinuous free memory blocks, which accumulate over time and cause memory fragmentation. This invention pre-allocates multi-core shared contiguous physical memory pages on the computing node and divides them into a lock-free spin update main area and a first remote memory access mirror backup area. Target feature embedding slots are preset in the lock-free spin update main area, and orthogonal public opinion semantic vectors are overwritten to the absolute physical memory address corresponding to the target feature embedding slot using transaction synchronization extension instructions. By pre-setting a fixed capacity of feature embedding slots on the underlying computing node and using an absolute address direct overwrite mechanism, the system-level dynamic memory allocation and release instructions during the NLP update process are completely blocked. The prediction system can complete the parameter loading of emerging trending public opinion terms without changing the physical memory allocation topology of the underlying computing node. The aforementioned static physical memory partitioning mechanism, in conjunction with hardware-level transaction write features, eliminates the objective conditions for memory fragmentation at the underlying physical mechanism level, completely solving the technical defect of memory fragmentation in public opinion prediction nodes caused by high-frequency online updates, which leads to system crashes.

[0015] When massive concurrent network text stream data floods into natural language processing nodes, hardware bus transmission congestion and disk read / write latency can easily occur. To eliminate physical transmission bottlenecks, this invention constructs a context co-occurrence network of sudden popular public opinion phrases within the text stream data and performs Riemannian manifold tensor projection dimensionality reduction operations to obtain orthogonal public opinion semantic vectors with byte lengths matching the single-pass throughput bit width of the cache. The byte length of the orthogonal public opinion semantic vectors achieves physical width alignment with the single-pass throughput bit width of the cache, ensuring that the parameter writing process perfectly matches the processor's cache line prefetching mechanism and effectively preventing additional clock cycle consumption caused by addressing across cache boundaries. Simultaneously, the system uses page locking system call instructions to modify the memory page table attribute bits corresponding to multi-core shared contiguous physical memory pages to a non-swappable state. Modifying the memory page table attribute bits to a non-swappable state forces the operating system to abandon virtual memory page replacement operations. The aforementioned data structure dimensionality reduction alignment features and the underlying physical memory locking features of the operating system work synergistically to significantly reduce disk physical input / output latency consumption and greatly improve the data throughput efficiency of the underlying hardware processing pipeline for the spatiotemporal cross-node public opinion evolution tensor.

[0016] When distributed computing clusters perform cross-node parameter synchronization and forward inference, conventional software-level read-write mutex mechanisms and transmission control protocol network packet communication introduce significant operating system context switching overhead. This invention utilizes the data write level flipping change of the physical memory absolute address to generate a network interrupt signal, synchronously triggering a remote memory access protocol to push orthogonal public opinion semantic vectors to the second remote memory access mirror backup area of ​​the remote distributed public opinion prediction node. This push mechanism bypasses the remote operating system kernel protocol stack using a low-level zero-copy mechanism, achieving cross-node physical memory direct write communication. Subsequently, the system executes a single-step atomic instruction to update the physical memory absolute address to the pre-deployed first-level static hash routing table within the natural language model. The combination of the single-step atomic instruction and the first-level static hash routing table establishes a memory-resident addressing path, enabling the natural language model to achieve lock-free fast reading across the virtual memory paging mapping layer when fusing base network parameters and orthogonal public opinion semantic vectors. This deeply integrated hardware and software mechanism avoids thread blocking caused by software lock contention, achieving low-latency cross-node public opinion propagation trajectory prediction while maintaining a constant zero increase in memory fragmentation across all physical nodes. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0018] Figure 1This is a flowchart illustrating a public opinion propagation prediction method based on natural language processing according to the present invention. Detailed Implementation

[0019] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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. The various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0020] Firstly, please refer to Figure 1 The present invention provides a method for predicting public opinion dissemination based on natural language processing, comprising: Step 1: Allocate multi-core shared contiguous physical memory pages on the computing node, divide the multi-core shared contiguous physical memory pages into a lock-free spin update master area and a first remote memory access mirror backup area, and embed slots with preset target features in the lock-free spin update master area; Step 2: Collect text stream data from multiple physical nodes in the distributed public opinion monitoring network and extract the unregistered words included in the text stream data; Step 3: Extract the spatiotemporal propagation characteristics of unregistered words across different physical nodes within a preset time window, and construct a spatiotemporal cross-node public opinion evolution tensor; Step 4: When the temporal cross-node public opinion evolution tensor is greater than the preset public opinion outbreak warning threshold, it is determined that the unregistered words are transformed into sudden popular public opinion terms. Step 5: Construct a context co-occurrence network of sudden popular public opinion terms within the text stream data, perform dimensionality reduction operation on the context co-occurrence network to obtain the evolution momentum scalar and orthogonal public opinion semantic vectors, so that the byte length of the orthogonal public opinion semantic vectors matches the single throughput bit width of the cache. Step 6: Select the target feature embedding slot that is inactive from the lock-free spin update master region, and generate transaction synchronization extension instructions using evolutionary momentum scalar; Step 7: Use the transaction synchronization extension instruction to overwrite the orthogonal public opinion semantic vector to the physical memory absolute address corresponding to the target feature embedding slot, and synchronously trigger the remote memory access protocol to push the orthogonal public opinion semantic vector to the second remote memory access mirror backup area of ​​the remote distributed public opinion prediction node. Step 8: Update the physical memory absolute address to the first-level static hash routing table inside the pre-deployed natural language model, and use the updated first-level static hash routing table to read the orthogonal public opinion semantic vector from the physical memory absolute address; Step 9: Integrate the base network parameters of the natural language model with the orthogonal public opinion semantic vector, perform forward public opinion evolution inference operation on the text stream data, and output the prediction results of cross-node public opinion propagation trajectory.

[0021] To more clearly demonstrate the core technical logic of this invention, the main process of the method is first described in detail: In the scenario of distributed real-time public opinion monitoring and evolution simulation, the distributed computing cluster faces the need to process massive concurrent network text stream data. The system kernel pre-queries the physical node mapping association of the underlying non-uniform memory access architecture, and triggers the operating system memory management subsystem to generate a continuous physical memory page allocation request containing the specified physical node by reading the node distance topology parameters of the hardware layer. After receiving the physical memory page base address pointer returned by the underlying hardware based on the continuous physical memory page allocation request, the system uses the physical memory page base address pointer to divide the lock-free spin update main area and the first remote memory access mirror backup area on the specified physical node, and presets the target feature embedding slot in the lock-free spin update main area.

[0022] To improve the hardware cache hit efficiency for subsequent feature reading, the logic control unit reads the starting physical address corresponding to the physical memory page base address pointer and calculates the modulo result of the starting physical address relative to the width of the compute node's data cache line. Based on the modulo result, a physical absolute address offset is applied to the starting physical address, aligning it with the boundary of the compute node's data cache line, where the data cache line width is specifically set to 64 bytes. Subsequently, the system uses a page locking system call instruction to modify the page table attribute bits corresponding to the multi-core shared contiguous physical memory pages to a non-swappable state, thereby blocking disk read / write latency caused by the operating system performing virtual memory paging operations at the physical hardware level.

[0023] The underlying network interface continuously collects text stream data from multiple physical nodes in the distributed public opinion monitoring network and stores the text stream data in a receive buffer area mapped in memory. The word segmentation process reads the text stream data in the receive buffer area and performs word segmentation based on an affix dictionary to obtain segmentation candidate words. After extracting known words from the basic dictionary, the system performs a hash comparison between the segmentation candidate words and the known words, filtering out segmentation candidate words that do not match the known words as unregistered words and adding them to the unregistered word queue.

[0024] For newly generated words in the queue that are awaiting testing, the system records the time difference between the transmission of these words across different physical nodes and calculates the propagation delay time within a preset time window, specifically configured as 60 seconds. The statistics engine summarizes the frequency values ​​of the unregistered words in local areas within the preset time window. Using vectorization, the system converts the propagation delay time values ​​into inter-node propagation delay time vectors and the local area frequency values ​​into local burst frequency vectors. By concatenating the inter-node propagation delay time vectors and the local burst frequency vectors according to fixed dimensional coordinates, a spatiotemporal cross-node public opinion evolution tensor is generated, which is used to comprehensively quantify the spatial diffusion lag and temporal burst density of newly generated words in the public opinion monitoring network.

[0025] The state determination module compares the comprehensive scalar norm of the spatiotemporal cross-node public opinion evolution tensor with a preset public opinion outbreak warning threshold. When the spatiotemporal cross-node public opinion evolution tensor is greater than the preset public opinion outbreak warning threshold, it determines that the unregistered words have broken through the normal natural language distribution pattern and transformed into suddenly popular public opinion terms. Subsequently, the feature extraction module uses the suddenly popular public opinion terms as the central words, extracts the context text sequence related to the central words from the text stream data according to a preset character span, and extracts the related words within the context text sequence other than the central words.

[0026] By statistically analyzing the absolute positional intervals between related words and emerging trending phrases in public opinion, the system generates an adjacency matrix using the inverse mapping relationship of these absolute positional intervals. This ensures that smaller absolute positional intervals correspond to larger weights within the adjacency matrix. Finally, by using related words as graph nodes and combining them with the values ​​in the adjacency matrix as graph edge weights, a contextual co-occurrence network of emerging trending phrases within the text stream data is constructed.

[0027] Given that conventional linear dimensionality reduction struggles to capture the complex nonlinear semantic shifts of sudden network contexts, the computational unit projects the context co-occurrence network onto a high-dimensional Riemannian manifold space and calculates the geodesic distance parameters between the graph nodes contained in the context co-occurrence network within this space. The geodesic distance parameters objectively reflect the shortest path for the nonlinear evolution of lexical semantics in the curved manifold geometry. Using these geodesic distance parameters, the system performs Riemannian manifold tensor projection dimensionality reduction on the context co-occurrence network along the tangent space direction of the high-dimensional Riemannian manifold space, outputting mutually independent and orthogonal initial semantic vectors.

[0028] During the dimensionality reduction operation of the Riemannian manifold tensor projection, the computation unit extracts the gradient of the dimensionality reduction loss rate change in real time to generate an evolutionary momentum scalar, which quantifies the active rate at which sudden popular public opinion semantics permeates into different contexts over time. Subsequently, the system truncates the floating-point bits at the end of the initial semantic vector to obtain an orthogonal public opinion semantic vector, making the total byte length of the orthogonal public opinion semantic vector equal to the single-pass throughput width of the cache (64 bytes), thereby achieving physical width alignment between subsequent parameter writing and the CPU hardware-level cache prefetch mechanism.

[0029] During the feature update and cross-node synchronization phase, the update process extracts the preset base address of the target feature embedding slot from the lock-free spin update master region. It then adds the evolution momentum scalar as an address offset factor to the preset base address to calculate and determine the physical memory absolute address corresponding to the target feature embedding slot. After encapsulating the physical memory absolute address and the orthogonal public opinion semantic vector into a transaction synchronization extension instruction, the central processing unit executes the transaction synchronization extension instruction. Utilizing the hardware transaction memory mechanism, it writes the orthogonal public opinion semantic vector to the physical memory absolute address in a lock-free atomic state, thereby avoiding lock contention blocking at the operating system level at the physical hardware layer.

[0030] As the data write level changes with the absolute address of physical memory, a network interrupt signal is synchronously generated on the underlying bus. The underlying network interface card (NIC) hardware controller responds to the network interrupt signal, reads the orthogonal public opinion semantic vector stored in the absolute address of physical memory, adds a network layer destination address header pointing to the remote node, and encapsulates it into a Remote Memory Access Protocol (RMAP) network data frame. Using a high-speed fiber optic network communication link, the local NIC sends the RMAP network data frame to the remote distributed public opinion prediction node. The NIC in the remote distributed public opinion prediction node receives the network data frame and extracts the orthogonal public opinion semantic vector. Utilizing the zero-copy mechanism of NIC pass-through processing to bypass the remote operating system kernel protocol stack, it writes the orthogonal public opinion semantic vector into the second remote memory access mirror backup area of ​​the remote distributed public opinion prediction node.

[0031] After the remote distributed public opinion prediction node completes the write to the backup memory, it returns a confirmation message via the network communication link. The local computing node receives the cross-node parameter synchronization confirmation event and performs a hash operation on the strings contained in the sudden popular public opinion terms to generate a 32-bit string identifier. The routing management process reads the physical memory absolute address where the orthogonal public opinion semantic vector is stored, and combines the string identifier with the physical memory absolute address to construct key-value pairs to update the data. The central processing core uses single-step atomic comparison and exchange instructions to write the key-value pair update data into the first-level static hash routing table inside the pre-deployed natural language model without locking, and then uses the updated first-level static hash routing table to generate an addressing path ready status indicator.

[0032] The natural language model responds to the addressing path readiness status flag, inputting the text stream data received from the network into the computation graph architecture. Using the updated first-level static hash routing table, the inference engine queries the corresponding string identifiers based on the emerging trending public opinion terms contained in the text stream data, and reads the orthogonal public opinion semantic vectors across the virtual memory paging mapping layer according to the physical memory absolute addresses mapped in the first-level static hash routing table. During the forward propagation of the computation graph, the inference engine integrates the orthogonal public opinion semantic vectors with the base network parameters of the natural language model to construct an evolutionary inference network with the ability to recognize new vocabulary. Finally, the evolutionary inference network performs forward public opinion evolutionary inference matrix multiplication on the text stream data, outputting a cross-node public opinion propagation trajectory prediction result containing the future distribution of public opinion heat and evolution direction.

[0033] To further fully disclose the underlying operational rules of this invention, the key data processing path in the above steps is described below, combining specific mathematical derivation equations and numerical verification processes: After the system kernel reads the starting physical address corresponding to the physical memory page base address pointer, the logic control unit performs hardware-level address alignment calculation. The logic control unit calculates the modulo result of the starting physical address relative to the width of the compute node's data cache line, and performs a physical absolute address offset on the starting physical address based on the modulo result to align with the boundary of the compute node's data cache line. The underlying calculation formula for address offset alignment is expressed as:

[0034] In the underlying calculation formula for address offset alignment Represents the starting physical address after the offset. This represents the starting physical address corresponding to the physical memory page base address pointer. This represents the width of the data cache line in the compute node. Using the offset starting physical address, subsequent system execution of physical memory read / write instructions and the underlying processor's cache prefetch step size achieve physical width binding at the hardware level, eliminating the extra clock cycle overhead caused by the CPU addressing across cache boundaries from a low-level physics perspective. The statistics engine converts the propagation delay time value into an inter-node propagation delay time vector and the local region occurrence frequency value into a local outbreak frequency vector, concatenating them to generate a spatiotemporal cross-node public opinion evolution tensor. The state determination module calculates the comprehensive scalar norm based on the spatiotemporal cross-node public opinion evolution tensor. The underlying calculation formula for the comprehensive scalar norm is expressed as:

[0035] In the underlying formula for calculating the synthetic scalar norm, The comprehensive scalar norm representing the tensor of spatiotemporal cross-node public opinion evolution. This represents the total number of physical nodes included in the distributed public opinion monitoring network. This represents the physical node ID of the distributed public opinion monitoring network. The vector representing the propagation delay between nodes is numbered as follows: The propagation delay time value of the physical node. This represents the total number of local areas included in the distributed public opinion monitoring network. Logical ID representing a local area contained within a distributed public opinion monitoring network. The vector representing the frequency of local outbreaks is numbered as follows: The frequency of occurrence in local areas. The inverse square of the propagation delay time objectively reflects the high-speed burst characteristics of network communication between physical nodes, while the square of the frequency of occurrence in local areas objectively reflects the degree of clustering of words in independent data regions. The state determination module compares the comprehensive scalar norm of the spatiotemporal cross-node public opinion evolution tensor with the preset public opinion outbreak warning threshold. When the comprehensive scalar norm of the spatiotemporal cross-node public opinion evolution tensor is greater than the preset public opinion outbreak warning threshold, it is determined that the unregistered words have been transformed into suddenly popular public opinion terms.

[0036] After constructing the context co-occurrence network, the computation unit performs Riemannian manifold tensor projection dimensionality reduction on the network along the tangent space direction of the high-dimensional Riemannian manifold space, outputting independent initial semantic vectors with orthogonal properties. During the Riemannian manifold tensor projection dimensionality reduction operation, the computation unit extracts the gradient of the dimensionality reduction loss rate to generate an evolutionary momentum scalar. The underlying hardware mapping formula for calculating the evolutionary momentum scalar is expressed as:

[0037] In the underlying hardware mapping formula for calculating the evolutionary momentum scalar Represents the evolutionary momentum scalar. The scaling factor representing the manifold projection iteration step size allocated to the computation unit. This represents the dimensionality reduction loss rate during the Riemannian manifold tensor projection dimensionality reduction operation. Represents the initial semantic vector. This represents the gradient of the dimensionality reduction loss rate relative to the initial semantic vector. This represents the single-pass throughput bit width of the cache. The single-pass throughput bit width of the physical hardware cache is directly introduced into the Riemann tensor projection gradient calculation process, ensuring that the output evolutionary momentum scalar has underlying hardware bus bandwidth constraints in the initial calculation stage. This directly blocks physical memory access blockage caused by sudden surges in public opinion and ultra-high-frequency feature vector writes from the mathematical algorithm level. The inference engine integrates orthogonal public opinion semantic vectors with the base network parameters of the natural language model to construct an evolutionary inference network. This network is then used to perform forward public opinion evolutionary inference matrix multiplication on the text stream data. The hardware-level calculation formula for the forward public opinion evolutionary inference matrix multiplication is expressed as:

[0038] In the hardware-level calculation formula of the forward public opinion evolution reasoning matrix multiplication operation... This represents the prediction results of cross-node public opinion propagation trajectory. The nonlinear tensor activation function representing the deployment of the evolutionary inference network. The base network parameter matrix representing the natural language model. Represents orthogonal semantic vectors of public opinion. The network bias matrix represents the numerical value of the evolutionary inference network.

[0039] To verify the absolute feasibility of the underlying logic derivation and physical hardware data processing flow, specific underlying device operating parameters were input into the prediction system for computational derivation. The initial decimal value of the starting physical address corresponding to the physical memory page base address pointer was set to 1020, and the width of the computing node data cache line was set to 64 bytes. The logic control unit performed a modulo operation between 1020 and 64 to obtain a remainder of 60. Subtracting 60 from 64 yielded an address offset of 4. Adding 4 to the initial value of 1020 resulted in an offset starting physical address of 1024. Since 1024 is divisible by 64, it perfectly aligns with the computing node data cache line boundary. The system set the total number of physical nodes in the distributed public opinion monitoring network to 2, the total number of local regions to 2, the propagation delay time vectors between nodes to contain propagation delay times of 0.5 s and 0.25 s, and the local outbreak frequency vectors to contain the frequency of occurrence of local regions to contain 3 and 4 occurrences, respectively.

[0040] The statistics engine performs negative square operations on the values ​​0.5 and 0.25 respectively to obtain the values ​​4 and 16, and sums them to obtain the value 20. The statistics engine performs square operations on the values ​​3 and 4 respectively to obtain the values ​​9 and 16, and sums them to obtain the value 25. The values ​​20 and 25 are added together to obtain the value 45. The square root operation of the value 45 is used to calculate that the comprehensive scalar norm of the spatiotemporal cross-node public opinion evolution tensor is approximately 6.7.

[0041] The state determination module performs a precise numerical comparison between the value 6.7 and the pre-set threshold value 5.0 for public opinion outbreak warning. If the value 6.7 is greater than 5.0, the logic condition for converting unregistered words into suddenly popular public opinion terms is triggered. The scaling factor for the manifold projection iteration step size allocated to the computation unit is set to 0.01, the gradient of the extracted dimensionality reduction loss rate relative to the initial semantic vector is 1280, and the single-pass throughput bit width of the physical hardware cache is constant at 64 bits. The computation unit multiplies 0.01 by 1280 to obtain 12.8, and divides 12.8 by 64 to calculate the precise value of the evolutionary momentum scalar as 0.2.

[0042] The system generates underlying transaction synchronization extension instructions with absolute physical location information based on the evolution momentum scalar value of 0.2 and the preset base address. It completes the lock-free overwriting of the absolute physical memory address by the orthogonal public opinion semantic vector, integrates and constructs the evolutionary inference network and outputs accurate cross-node public opinion propagation trajectory prediction results. It also completes the incremental clearing of physical memory fragments of natural language processing nodes in ultra-high concurrency scenarios.

[0043] Secondly, the present invention provides a public opinion propagation prediction system based on natural language processing, applied to the public opinion propagation prediction system based on natural language processing as described above, comprising: The memory allocation module is used to allocate multi-core shared contiguous physical memory pages on the computing node, divide the multi-core shared contiguous physical memory pages into a lock-free spin update main area and a first remote memory access mirror backup area, and embed slots with preset target features in the lock-free spin update main area; The vocabulary extraction module is used to collect text stream data from the distributed public opinion monitoring network and extract unregistered words from the text stream data. The tensor construction module is used to extract the spatiotemporal propagation characteristics of unregistered words within a preset time window and construct a spatiotemporal cross-node public opinion evolution tensor. The status determination module is used to determine whether unregistered words are transformed into suddenly popular public opinion terms when the tensor of public opinion evolution across nodes in time and space exceeds the preset public opinion outbreak warning threshold. The dimensionality reduction module is used to construct a context co-occurrence network of sudden popular public opinion terms within the text stream data. It performs dimensionality reduction operations on the context co-occurrence network to obtain the evolution momentum scalar and orthogonal public opinion semantic vectors, so that the byte length of the orthogonal public opinion semantic vectors matches the single throughput bit width of the cache. The instruction generation module is used to select target feature embedding slots that are inactive from the lock-free spin update master region and generate transaction synchronization extension instructions using evolution momentum scalars. The physical overwrite module is used to overwrite the orthogonal public opinion semantic vector to the physical memory absolute address corresponding to the target feature embedding slot using transaction synchronization extension instructions, and synchronously trigger the remote memory access protocol to push the orthogonal public opinion semantic vector to the second remote memory access mirror backup area of ​​the remote distributed public opinion prediction node. The routing update module is used to update the physical memory absolute address to the first-level static hash routing table inside the pre-deployed natural language model, and use the updated first-level static hash routing table to read the orthogonal public opinion semantic vector from the physical memory absolute address. The evolutionary inference module is used to fuse the base network parameters of the natural language model with orthogonal public opinion semantic vectors, perform forward public opinion evolutionary inference operations on text stream data, and output cross-node public opinion propagation trajectory prediction results.

[0044] As a further extension of the technical solution of the present invention, the following is a specific embodiment focusing on underlying hardware collaborative scheduling and bus transaction instruction verification: In distributed real-time public opinion monitoring and evolution simulation scenarios, the underlying hardware system needs to support physical layer access to massive concurrent text stream data and parameter synchronization across nodes. The compute node starts and enters the hardware initialization phase of the basic input / output system. The central processing unit (CPU) reads the physical node mapping association of the non-uniform memory access architecture from the hardware configuration register. Based on the physical node mapping association, the memory management unit sends a continuous physical memory page allocation request containing the specified physical node to the motherboard memory control unit in the corresponding physical region. The memory control unit allocates space in a specific storage granular group and returns a physical memory page base address pointer based on the continuous physical memory page allocation request via the system front-side bus. Using the physical memory page base address pointer, the CPU divides the lock-free spin update main region and the first remote memory access mirror backup region on the specified physical node, and presets target feature embedding slots in the lock-free spin update main region. The logic control unit reads the starting physical address corresponding to the physical memory page base address pointer, calculates the modulo result of the starting physical address relative to the compute node data cache line width, and performs a physical absolute address offset on the starting physical address based on the modulo result to align with the compute node data cache line boundary. The central processing unit sends a page lock system call instruction to the hardware page table register, modifying the memory page table attribute bits corresponding to the multi-core shared contiguous physical memory pages to a non-swapable state, thereby blocking the generation of disk paging interrupts from the underlying physical mechanism.

[0045] The network interface controller receives text stream data from multiple physical nodes in the distributed public opinion monitoring network and stores the text stream data in the receive buffer area through the direct memory access channel of the peripheral component interconnect bus. The instruction fetch unit of the central processing unit reads the text stream data from the receive buffer area and loads it into the general-purpose data register. The arithmetic logic unit performs word segmentation on the text stream data to obtain segmentation candidate words, extracts known words from the basic dictionary, performs hardware-level hash comparison between the segmentation candidate words and the known words, filters the segmentation candidate words that do not match the known words as unregistered words, and adds the unregistered words to the unregistered word queue. The timer module of the central processing unit records the difference in hardware clock cycles of unregistered words across different physical nodes in the unregistered word queue, and obtains the propagation delay time value within a preset time window. The floating-point unit in the central processing unit counts the frequency of unregistered words in local areas within the preset time window, uses the single instruction multiple data stream extended instruction set to convert the propagation delay time value into an inter-node propagation delay time vector, and converts the local area frequency value into a local burst frequency vector. The floating-point arithmetic unit concatenates the inter-node propagation delay time vector and the local outbreak frequency vector to generate a spatiotemporal cross-node public opinion evolution tensor. The hardware comparator compares the spatiotemporal cross-node public opinion evolution tensor with a preset public opinion outbreak warning threshold written to a preset register. When the spatiotemporal cross-node public opinion evolution tensor is detected to be greater than the preset public opinion outbreak warning threshold, the processor branch prediction unit determines that the unregistered words are transformed into sudden popular public opinion terms.

[0046] The central processing unit extracts context text sequences from the text stream data, using the trending public opinion phrase as the central word, and extracts related words within the context text sequence excluding the central word. The arithmetic logic unit calculates the absolute positional interval values ​​between related words and the trending public opinion phrase to generate an adjacency matrix. The related words are used as graph nodes, and the values ​​in the adjacency matrix are used as graph edge weights to construct a context co-occurrence network. The tensor processing core projects the context co-occurrence network onto a high-dimensional Riemannian manifold space, calculates the geodesic distance parameters between the graph nodes contained in the context co-occurrence network, and performs dimensionality reduction operations along the tangent space direction using the geodesic distance parameters. The tensor processing core outputs an initial semantic vector and extracts the gradient of the dimensionality reduction loss rate to generate an evolutionary momentum scalar. The initial semantic vector is truncated in the hardware output buffer register to obtain orthogonal public opinion semantic vectors. The byte length of the orthogonal public opinion semantic vector matches the single-throughput bit width of the cache, achieving equal-width data binding in the hardware processing pipeline.

[0047] The central processing unit (CPU) extracts the preset base address of the target feature embedding slot from the lock-free spin update master region. The arithmetic logic unit adds the evolution momentum scalar to the preset base address to calculate the corresponding physical offset address, and determines the physical offset address as the physical memory absolute address corresponding to the target feature embedding slot. The instruction decoding unit encapsulates the physical memory absolute address and the orthogonal public opinion semantic vector into a transaction synchronization extension instruction. The CPU executes the transaction synchronization extension instruction, and the processor's transaction memory control logic exclusively locks the corresponding level 1 hardware cache line, atomically writing the orthogonal public opinion semantic vector to the physical memory absolute address. A data write level flip occurs at the physical memory absolute address, generating a network interrupt signal on the high-speed serial computer expansion bus. The underlying network interface card (NIC) control chip captures the network interrupt signal, bypasses the operating system kernel, and reads the orthogonal public opinion semantic vector stored in the physical memory absolute address using physical addressing. The underlying NIC control chip adds a destination address header to the orthogonal public opinion semantic vector and encapsulates it into a remote memory access protocol (RMAP) network data frame. The network interface controller uses a fiber optic network communication link to send the RMAC network data frame to the remote distributed public opinion prediction node. The network card hardware of the remote distributed public opinion prediction node receives network data frames of the remote memory access protocol, extracts orthogonal public opinion semantic vectors, and uses a zero-copy mechanism to push the orthogonal public opinion semantic vectors into the second remote memory access mirror backup area.

[0048] The network interface controller receives a hardware response message from a remote distributed public opinion prediction node, triggering the motherboard control chip to generate a hardware interrupt for a cross-node parameter synchronization confirmation event. The central processing unit (CPU) responds to the cross-node parameter synchronization confirmation event, performs a hash operation on the emerging trending public opinion terms to generate string identifiers, and combines these string identifiers with physical memory absolute addresses to construct key-value pairs to update the data. The CPU dispatches a single-step atomic comparison and exchange instruction to the system memory bus, writing the key-value pair update data into the pre-deployed first-level static hash routing table within the natural language model without locking. The system memory bus returns a write success signal and triggers the state machine to generate an addressing path ready status flag. The natural language model responds to the addressing path ready status flag, and the CPU inputs the received text stream data into the computational graph architecture of the natural language model. The inference computational graph scheduling logic uses the updated first-level static hash routing table to query the string identifiers, and uses the physical memory absolute address corresponding to the string identifier in the first-level static hash routing table to bypass the virtual memory page table translation module, issuing a physical memory direct read instruction to read the orthogonal public opinion semantic vector. The tensor processing core integrates orthogonal public opinion semantic vectors with the base network parameters of a natural language model, and configures a hardware-accelerated evolutionary inference network structure. The tensor processing core utilizes the evolutionary inference network to perform forward public opinion evolutionary inference matrix multiplication on the text stream data, refreshes the completed data register state to main memory, and outputs the predicted results of cross-node public opinion propagation trajectories.

Claims

1. A method for predicting public opinion propagation based on natural language processing, characterized in that, include: Multi-core shared contiguous physical memory pages are allocated on the computing node. The multi-core shared contiguous physical memory pages are divided into a lock-free spin update main area and a first remote memory access mirror backup area. The target feature is embedded in a slot in the lock-free spin update main area. Collect text stream data from multiple physical nodes in a distributed public opinion monitoring network, and extract unregistered words from the text stream data; Extract the spatiotemporal propagation characteristics of unregistered words across different physical nodes within a preset time window, and construct a spatiotemporal cross-node public opinion evolution tensor; When the temporal and spatial cross-node public opinion evolution tensor exceeds the preset public opinion outbreak warning threshold, it is determined that unregistered words are transformed into suddenly popular public opinion terms. Construct a context co-occurrence network of sudden popular public opinion terms within text stream data, perform dimensionality reduction operation on the context co-occurrence network to obtain evolution momentum scalar and orthogonal public opinion semantic vectors, so that the byte length of the orthogonal public opinion semantic vectors matches the single throughput bit width of the cache; Select inactive target feature embedding slots from the lock-free spin update master region, and use evolutionary momentum scalar to generate transaction synchronization extension instructions; The transaction synchronization extension instruction is used to overwrite the orthogonal public opinion semantic vector to the physical memory absolute address corresponding to the target feature embedding slot, and the remote memory access protocol is triggered synchronously to push the orthogonal public opinion semantic vector to the second remote memory access mirror backup area of ​​the remote distributed public opinion prediction node. The physical memory absolute address is updated to the first-level static hash routing table inside the pre-deployed natural language model, and the orthogonal public opinion semantic vector is read from the physical memory absolute address using the updated first-level static hash routing table; By integrating the base network parameters of a natural language model with orthogonal public opinion semantic vectors, forward public opinion evolution inference operations are performed on text stream data, and the prediction results of cross-node public opinion propagation trajectory are output.

2. The public opinion propagation prediction method based on natural language processing according to claim 1, characterized in that, Allocate multi-core shared contiguous physical memory pages on the compute node, and divide the multi-core shared contiguous physical memory pages into a lock-free spin-up primary region and a first remote memory access mirror backup region, including: Query the physical node mapping associations of a non-uniform memory access architecture; Based on the physical node mapping association, generate a contiguous physical memory page allocation request that includes the specified physical node; Receive the physical memory page base address pointer returned based on the contiguous physical memory page allocation request; Using the physical memory page base address pointer, the lock-free spin update main area and the first remote memory access mirror backup area are divided at a specified physical node, and the target feature is embedded in the lock-free spin update main area with preset target feature embedding slots; Read the starting physical address corresponding to the physical memory page base address pointer; Calculate the modulo result of the starting physical address relative to the width of the compute node's data cache line; Based on the modulo result, the starting physical address is offset to align with the boundary of the compute node's data cache line; By using the page locking system call instruction, the memory page table attribute bits corresponding to the shared contiguous physical memory pages of multiple cores are modified to a non-swapable state.

3. The public opinion propagation prediction method based on natural language processing according to claim 2, characterized in that, Extract the spatiotemporal propagation characteristics of unregistered words across different physical nodes within a preset time window, and construct a spatiotemporal cross-node public opinion evolution tensor, including: The text stream data is stored in the receiving buffer area, the text stream data in the receiving buffer area is read, word segmentation is performed on the text stream data, and candidate words for segmentation are obtained. Extract known words from the basic dictionary and perform hash comparison between the candidate words and the known words; Filter the segmentation candidate words that did not match the known words into the unregistered words and add them to the unregistered words queue. Record the time difference between the transmission of unregistered words across different physical nodes in the unregistered word queue, and obtain the propagation delay time value within the preset time window; Count the frequency of unregistered words appearing in local areas within a preset time window; Convert the propagation delay time value into a propagation delay time vector between nodes; Convert the frequency values ​​of local regions into local outbreak frequency vectors; By splicing the propagation delay time vector between nodes and the local outbreak frequency vector, a spatiotemporal cross-node public opinion evolution tensor is generated. 4.The public opinion propagation prediction method based on natural language processing of claim 3, wherein, Constructing a contextual co-occurrence network of suddenly popular public opinion terms within text stream data, including: Using suddenly popular public opinion terms as the central words, the context text sequence is extracted from the text stream data according to the preset character span; Extract related words from the context text sequence, excluding the central word; The absolute positional interval between statistically related keywords and suddenly popular public opinion terms; Generate an adjacency matrix using absolute positional interval values; By using associated words as graph nodes and the values ​​in the adjacency matrix as graph edge weights, a context co-occurrence network is constructed. 5.The public opinion propagation prediction method based on natural language processing according to claim 4, characterized in that, Dimensionality reduction is performed on the context co-occurrence network to obtain the evolution momentum scalar and orthogonal sentiment semantic vectors, such that the byte length of the orthogonal sentiment semantic vectors matches the single-pass throughput bit width of the cache, including: Project the context co-occurrence network onto a high-dimensional Riemannian manifold space; Calculate the geodesic distance parameters between graph nodes in a context co-occurrence network within a high-dimensional Riemannian manifold space; Using the geodesic distance parameter, Riemannian manifold tensor projection dimensionality reduction operation is performed on the context co-occurrence network along the tangent space direction to output the initial semantic vector; During the Riemannian manifold tensor projection dimensionality reduction operation, the gradient of the dimensionality reduction loss rate change is extracted to generate the evolutionary momentum scalar. The initial semantic vector is truncated to obtain an orthogonal public opinion semantic vector, such that the total byte length of the orthogonal public opinion semantic vector is equal to the single throughput bit width of the cache. 6.The public opinion propagation prediction method based on natural language processing according to claim 5, characterized in that, The transaction synchronization extension instructions are generated using evolutionary momentum scalars. These instructions then overwrite the orthogonal public opinion semantic vectors to the absolute physical memory addresses corresponding to the target feature embedding slots. This includes: Extract the target feature from the lock-free spin update master region and embed it into the preset base address of the slot; The evolution momentum scalar is added to the preset base address to calculate the corresponding physical offset address, and the physical offset address is determined as the absolute physical memory address corresponding to the target feature embedding slot. Encapsulate the absolute physical memory address and orthogonal public opinion semantic vector into a transaction synchronization extension instruction; Execute the transaction synchronization extension instruction to write the orthogonal public opinion semantic vector to the absolute address of physical memory; A network interrupt signal is generated by utilizing the data write level flip change that occurs at the absolute address of physical memory.

7. The public opinion propagation prediction method based on natural language processing according to claim 6, characterized in that, Synchronously triggering the remote memory access protocol pushes orthogonal public opinion semantic vectors to the second remote memory access mirror backup area of ​​the remote distributed public opinion prediction node, including: In response to a network interruption signal, read the orthogonal public opinion semantic vector stored in the absolute address of physical memory; Add a destination address header to the orthogonal public opinion semantic vector and encapsulate it into a remote memory access protocol network data frame; The remote memory access protocol network data frames are sent to remote distributed public opinion prediction nodes using network communication links. Receive network data frames from Remote Memory Access Protocol (RMAP) and extract orthogonal public opinion semantic vectors. The extracted orthogonal public opinion semantic vectors are written to the second remote memory access mirror backup area of ​​the remote distributed public opinion prediction node using a zero-copy mechanism. 8.The public opinion propagation prediction method based on natural language processing of claim 7, wherein, Update the physical memory absolute address to the first-level static hash routing table inside the pre-deployed natural language model, including: Receive cross-node parameter synchronization confirmation events; In response to cross-node parameter synchronization confirmation events, perform hash operations on sudden trending public opinion terms to generate string identifiers; Read the absolute physical memory address where the orthogonal public opinion semantic vector is stored; Combine string identifiers with absolute physical memory addresses to construct key-value pairs and update data; By using a single-step atomic compare and swap instruction, the key-value pair update data is overwritten into the first-level static hash routing table inside the natural language model; Generate an address path ready status flag using the updated first-level static hash routing table. 9.The public opinion propagation prediction method based on natural language processing of claim 8, wherein, By integrating the base network parameters of a natural language model with orthogonal public opinion semantic vectors, forward public opinion evolution inference operations are performed on text stream data, outputting cross-node public opinion propagation trajectory prediction results, including: In response to the address path ready status flag, the received text stream data is input into the natural language model; Using the updated first-level static hash routing table, query the string identifiers corresponding to the sudden and popular public opinion phrases included in the text stream data; Using the absolute physical memory address corresponding to the string identifier in the first-level static hash routing table, the orthogonal public opinion semantic vector is read; An evolutionary inference network is constructed by integrating orthogonal public opinion semantic vectors with the base network parameters of a natural language model. By using an evolutionary inference network to perform forward public opinion evolutionary inference matrix multiplication on text stream data, the prediction results of cross-node public opinion propagation trajectory are calculated.

10. A public opinion propagation prediction system based on natural language processing, applied to the public opinion propagation prediction method based on natural language processing according to any one of claims 1-9, characterized in that, include: The memory allocation module is used to allocate multi-core shared contiguous physical memory pages on the computing node, divide the multi-core shared contiguous physical memory pages into a lock-free spin update main area and a first remote memory access mirror backup area, and embed slots with preset target features in the lock-free spin update main area; The vocabulary extraction module is used to collect text stream data from the distributed public opinion monitoring network and extract unregistered words from the text stream data. The tensor construction module is used to extract the spatiotemporal propagation characteristics of unregistered words within a preset time window and construct a spatiotemporal cross-node public opinion evolution tensor. The status determination module is used to determine whether unregistered words are transformed into suddenly popular public opinion terms when the tensor of public opinion evolution across nodes in time and space exceeds the preset public opinion outbreak warning threshold. The dimensionality reduction module is used to construct a context co-occurrence network of sudden popular public opinion terms within the text stream data. It performs dimensionality reduction operations on the context co-occurrence network to obtain the evolution momentum scalar and orthogonal public opinion semantic vectors, so that the byte length of the orthogonal public opinion semantic vectors matches the single throughput bit width of the cache. The instruction generation module is used to select target feature embedding slots that are inactive from the lock-free spin update master region and generate transaction synchronization extension instructions using evolution momentum scalars. The physical overwrite module is used to overwrite the orthogonal public opinion semantic vector to the physical memory absolute address corresponding to the target feature embedding slot using transaction synchronization extension instructions, and synchronously trigger the remote memory access protocol to push the orthogonal public opinion semantic vector to the second remote memory access mirror backup area of ​​the remote distributed public opinion prediction node. The routing update module is used to update the physical memory absolute address to the first-level static hash routing table inside the pre-deployed natural language model, and use the updated first-level static hash routing table to read the orthogonal public opinion semantic vector from the physical memory absolute address. The evolutionary inference module is used to fuse the base network parameters of the natural language model with orthogonal public opinion semantic vectors, perform forward public opinion evolutionary inference operations on text stream data, and output cross-node public opinion propagation trajectory prediction results.

Images