AI interview consumer survey data analysis system based on artificial intelligence

Abstract

The application relates to the technical field of natural language processing, and discloses an AI interview consumption investigation data analysis system based on artificial intelligence. A system reads a response timestamp difference value and a word element ratio change quantity to generate a denoising coupling coefficient; a splicing condition limited string generates a questioning text and a reference polarity tensor; a base projection component matrix is generated by using the denoising coupling coefficient, the reference polarity tensor and a bias feature vector; a target net value vector is output by subtracting the base projection component matrix from an input feature vector; when a target net value vector norm meets a threshold value, a state operator is set to 0 to truncate a continuous multiplication operation and a state pruning instruction is issued. The application avoids intention misjudgment caused by interaction delay, offsets characteristic space induced components, and forcibly recovers server memory occupied by redundant dialogues through a 0 value level signal.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to an AI-based consumer survey data analysis system based on artificial intelligence. Background Technology

[0002] In current AI-driven consumer surveys, systems typically input the response text received from the front-end interactive terminal directly into a large language model for intent classification. However, in real-world multi-turn interaction scenarios, as the interaction cycle increases, users inevitably experience longer interaction times and reduced character input per transaction. Due to the lack of a decoupling mechanism for the underlying interactive behavior data, the large language model may misinterpret these longer interaction times and reduced character input per transaction as negative consumer intent, leading to an initial coordinate shift in the intent classification results.

[0003] To correct the initial coordinate offset, conventional methods typically involve forcibly appending external entity parameters to the query text. However, forcibly appending external entity parameters disrupts the basis orientation of the original query text in the multidimensional feature embedding space, introducing induced entanglement components into the input feature vector, causing subsequent natural language processing results to deviate further from the true target.

[0004] When scheduling dialogue tree branches in the session state machine, the deep neural network classifier relies on a flexible floating-point arithmetic mechanism. This mechanism can only output probability floating-point numbers approaching zero, not absolute zero. Because it cannot receive hardware trigger signals for absolute zero, the session state machine cannot accurately mark redundant dialogue branches as unreachable. The child nodes of redundant dialogue branches continuously occupy the volatile random access memory space of the backend natural language processing server, leading to unnecessary consumption of GPU memory and computing power. Summary of the Invention

[0005] This invention provides an AI-based consumer survey data analysis system to address the technical problems of existing systems, such as the inability to extract underlying interactive behavior data leading to misjudgment of intent in large language models, the destruction of the multidimensional feature embedding space due to rigid splicing of entity parameters, and the inability of flexible floating-point arithmetic mechanisms to output absolute 0 values, resulting in the inability of the session state machine to release backend server memory space.

[0006] In view of the above problems, the present invention provides an AI-based consumer survey data analysis system, comprising: The interaction state perception and parameter generation unit is configured to read the response timestamp difference and word quantity ratio change within adjacent interaction cycles, and use the response timestamp difference and word quantity ratio change to perform numerical calculations to generate a denoising coupling coefficient. An edge carrier injection unit is configured to receive a first entity parameter and a second entity parameter, selectively convert the second entity parameter into a conditional constraint string based on the entity word frequency determination result of the second entity parameter, and concatenate the conditional constraint string or the original form of the second entity parameter into the text containing the first entity parameter to generate a target question text, and output the reference polarity tensor corresponding to the target question text; The feature space stripping unit is configured to call an interface to convert the response text to the target question text into an input feature vector, obtain a preset bias feature vector, perform matrix multiplication on the result of scalar multiplication of the denoising coupling coefficient and the reference polarity tensor and the bias feature vector to generate a basis projection component matrix, and subtract the basis projection component matrix from the input feature vector in the feature space to output the target net value vector. The directed acyclic graph multiplicative circuit breaker unit is configured to assign the starting memory physical address of the target net value vector in the heap space to the feature load region pointer of at least one root node in a preset directed acyclic graph data structure, and configure a state operator for the at least one root node; extract the norm scalar value of the target net value vector; if the norm scalar value satisfies a preset threshold condition, change the value of the state operator to zero, generate a target score value through continuous multiplication operation; and in response to the value of the state operator being zero, output a state pruning instruction to the session state machine.

[0007] The technical solution provided in this application has at least the following technical effects: The system generates a denoising coupling coefficient by reading the difference in response timestamps and the change in the ratio of word counts within adjacent interaction cycles. The system establishes a quantization model using the physical timestamps and character length changes at the underlying communication layer, avoiding misjudgments of intent caused by excessive semantic analysis of the phenomenon of longer interaction delays and reduced single input characters. This provides a numerical benchmark for subsequent algebraic mapping of multidimensional feature embedding space.

[0008] By obtaining a preset bias feature vector, the system performs matrix multiplication on the result of scalar multiplication of the denoising coupling coefficient and the reference polarity tensor with the bias feature vector to generate a basis projection component matrix. The system then subtracts the basis projection component matrix from the input feature vector in the feature space. The system uses algebraic subtraction to physically cancel the induced entanglement component caused by the conditional string, thus maintaining the basis orientation of the original question text in the feature space.

[0009] By extracting the norm scalar value of the target net value vector and performing a Boolean comparison with the threshold condition, the value of the state operator is changed to 0, and the target score value is generated through continuous multiplication. The system utilizes the mathematical property of discrete binary 0 value absorption to truncate the flexible floating-point output path of the deep neural network. In response to the value of the state operator being 0, the system outputs a state pruning instruction to the session state machine, triggering the session state machine to mark the child node containing the threshold condition identifier as unreachable and issue a memory release call instruction, thus reclaiming the volatile random access memory space of the backend natural language processing server. Attached Figure Description

[0010] Figure 1 An architecture diagram of an AI-based interview-based consumer survey data analysis system provided for embodiments of the present invention; Figure 2 A flowchart illustrating the AI-based consumer survey data analysis method provided in this embodiment of the invention. Detailed Implementation

[0011] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them.

[0012] Example: Please refer to Figure 1 and Figure 2 The hardware operating environment of the AI-based consumer survey data analysis system includes a front-end interactive terminal, a back-end natural language processing (NLP) server, and a session state machine deployed within the back-end NLP server. The front-end interactive terminal and the back-end NLP server establish a bidirectional data connection via a network communication interface. The back-end NLP server has an embedded application programming interface (API). When performing natural language generation tasks, the back-end NLP server faces physical memory constraints; the key-value cache usage increases linearly with the interaction cycle and faces the risk of memory overflow.

[0013] As the interactive data stream between the front-end interactive terminal and the back-end natural language processing server begins transmission, the AI-based interview consumer survey data analysis system enters the control loop. The interaction state perception and parameter generation unit reads the difference in response timestamps and the change in the ratio of word counts within adjacent interaction cycles, performs numerical calculations to generate a denoising coupling coefficient. After the denoising coupling coefficient is output to memory, the edge carrier injection unit receives the first entity parameter and the second entity parameter, generates the target question text, and outputs a reference polarity tensor. The response text to the target question text is converted into an input feature vector via an application programming interface. The feature space stripping unit performs multiplication and subtraction calculations on the input feature vector based on the denoising coupling coefficient and the reference polarity tensor, outputting a target net value vector. The directed acyclic graph multiplicative circuit breaker unit maps the target net value vector to a preset directed acyclic graph data structure, configures state operators for at least one root node in the directed acyclic graph data structure, generates a target score value through continuous multiplication operations, and sends a state pruning instruction to the session state machine to execute an unreachable state marking action, completing the memory computing power release loop.

[0014] After the interaction state awareness and parameter generation unit acquires thread control, the underlying data packet headers in the server communication log are parsed. The interaction state awareness and parameter generation unit extracts the current response timestamp and the previous response timestamp from the interaction data stream. The interaction state awareness and parameter generation unit performs a subtraction calculation on the current response timestamp and the previous response timestamp to obtain the initial response difference. The interaction state awareness and parameter generation unit performs a Boolean comparison between the initial response difference and a preset truncation constant. If the initial response difference is greater than the truncation constant, the interaction state awareness and parameter generation unit assigns the response timestamp difference to the truncation constant; if the initial response difference is less than or equal to the truncation constant, the interaction state awareness and parameter generation unit assigns the response timestamp difference to the initial response difference.

[0015] After the write operation of the response timestamp difference is completed, the underlying byte-to-byte encoding segmenter is awakened to process the input text of adjacent interaction cycles. The byte-to-byte encoding segmenter converts the input text of adjacent interaction cycles into discrete character arrays. The interaction state awareness and parameter generation unit performs traversal statistics on the character arrays to obtain the number of unique tokens and the total number of tokens. The interaction state awareness and parameter generation unit calculates the ratio of the number of unique tokens to the total number of tokens as the current period token ratio. The interaction state awareness and parameter generation unit retrieves the baseline token ratio extracted from the first interaction cycle, which is stored in memory. The interaction state awareness and parameter generation unit calculates the difference between the current period token ratio and the baseline period token ratio, and outputs the change in the token count ratio.

[0016] After the change in the response timestamp difference and the ratio of word counts are both extracted to the arithmetic register, the floating-point feature fusion instruction is triggered. The interactive state perception and parameter generation unit performs a division calculation on the response timestamp difference using a preset constant. The interactive state perception and parameter generation unit obtains a preset negative sensitivity coefficient, multiplies the negative sensitivity coefficient by the change in the ratio of word counts, and inputs it into the natural exponential function to calculate the exponential feature value. The interactive state perception and parameter generation unit performs an addition calculation on the quotient of the response timestamp difference divided by the constant and the exponential feature value to obtain the comprehensive feature index. The interactive state perception and parameter generation unit compares the comprehensive feature index with a preset center threshold, and performs a nonlinear activation mapping operation on the result of the difference comparison to output a denoising coupling coefficient. The nonlinear activation mapping operation is controlled by the mathematical property of the bidirectional saturation limit, so that the output value of the denoising coupling coefficient is clamped within a closed interval formed by the preset lower limit and upper limit of the coefficient, and exhibits a smooth convergence characteristic when approaching the boundary of the closed interval. The comprehensive feature index, as a dimensionless floating-point value, constitutes the input variable of the nonlinear activation mapping operation.

[0017] With the denoising coupling coefficients residing in the system's shared memory address, the edge carrier injection unit gains thread scheduling privileges from the central processing unit. The edge carrier injection unit calls the entity recognition interface to perform a lexical scan on the externally input probe task queue. The entity recognition interface initializes a character sliding window of a preset length in memory. The entity recognition interface controls the character sliding window to traverse the character array in the probe task queue in single-step increments, extracting local character sequences within the sliding window. The entity recognition interface performs a key-value comparison between the local character sequence and the second entity parameter hash table pre-loaded in static random access memory. If a match is found, the entity recognition interface increments the memory accumulator by 1, and the value of the memory accumulator after traversal is output as the entity term frequency. The edge carrier injection unit loads the entity term frequency and a preset term frequency threshold into the arithmetic logic unit for a Boolean comparison. If the entity term frequency is greater than the preset term frequency threshold, the second entity parameter is converted into a conditional string; if the entity term frequency is not greater than the preset term frequency threshold, the second entity parameter is retained in its original form.

[0018] The edge carrier injection unit performs memory address addressing to read the pointer to the text containing the first entity parameter. The edge carrier injection unit concatenates the conditional string or primitive form into the text containing the first entity parameter to generate the target query text, and writes the target query text to the output buffer. The edge carrier injection unit remains suspended until a write completion hardware interrupt signal is detected from the system bus. In response to the write completion hardware interrupt signal, the edge carrier injection unit performs character segmentation on the target query text, using the segmented characters as addressing indexes to read the static word vector mapping dictionary pre-stored in read-only memory across the bus, obtaining the independent 1D floating-point vector corresponding to each character. The edge carrier injection unit sends the independent 1D floating-point vector to the vector register, performs matrix concatenation along the sequence length dimension to generate a 2D matrix structure, and writes the 2D matrix structure as the reference polarity tensor corresponding to the target query text into the tensor register.

[0019] As the target query text is sent to the front-end interactive terminal, the terminal generates a response text and sends it back to the back-end natural language processing server. After the feature space stripping unit detects that the response text has been written to the receiving buffer, it calls the application programming interface (API) to upload the response text to the embedded neural network layer. The embedded neural network layer converts the response text into a multi-dimensional floating-point array, generating an input feature vector in the memory space of the feature space stripping unit. When the input feature vector is loaded into the cache, the feature space stripping unit performs a feature vector stripping operation in the multi-dimensional feature embedding space. This operation is controlled by the mathematical properties of orthogonal projection decoupling, using the reference polarity tensor as the basis direction, and inversely eliminating the induced entanglement components generated in the input feature vector along the projection normal. The specific algebraic mapping formula executed is as follows: in, The target net value vector; The input feature vector; These are the noise reduction coupling coefficients; is the reference polarity tensor; e is the bias eigenvector; and t is the time marker of the current interaction cycle.

[0020] The feature space stripping unit retrieves the reference polarity tensor and denoising coupling coefficients, and obtains a preset bias feature vector. The feature space stripping unit performs scalar multiplication on the reference polarity tensor using the denoising coupling coefficients, and then performs matrix multiplication on the result of the scalar multiplication with the bias feature vector to generate a basis projection component matrix aligned with the dimensions of the input feature vector. The feature space stripping unit performs floating-point subtraction in the vector processor, subtracting the basis projection component matrix from the input feature vector to output the target net value vector. By involving the bias feature vector in the matrix multiplication operation, the system handles the heterogeneous dimensional mapping problem between vectors and tensors, filtering out background noise components at the multi-dimensional feature embedding space level.

[0021] As the target net value vector is placed into the memory block of the feature space stripping unit, the directed acyclic graph (DAG) multiplicative circuit breaker unit takes over system bus control. The DAG multiplicative circuit breaker unit reads the DAG data structure containing at least one root node, and pre-writes threshold conditions into the memory address space of at least one root node. The DAG multiplicative circuit breaker unit obtains the starting physical memory address of the target net value vector in the heap space and assigns this address to the feature payload area pointer of at least one root node through a memory pointer addressing and mounting operation. The DAG multiplicative circuit breaker unit allocates a dedicated register and configures the state operator for at least one root node.

[0022] The directed acyclic graph (DAG) multiplicative circuit breaker unit extracts the floating-point array from the target net value vector. Within the arithmetic logic unit (ALU), it performs a square root operation on the floating-point array to extract the norm scalar value of the target net value vector. The DAG multiplicative circuit breaker unit then sends the norm scalar value and a threshold condition to a hardware comparator for Boolean comparison. If the norm scalar value meets the preset threshold condition, the value of the state operator in the dedicated register is changed to 0; if the norm scalar value does not meet the preset threshold condition, the value of the state operator remains 1. The DAG multiplicative circuit breaker unit inputs the target net value vector to a preset linear classification network layer via the system bus. The linear classification network layer performs matrix multiplication on the target net value vector and outputs a feature score array. The DAG multiplicative circuit breaker unit extracts the feature score array and performs a weighted summation with preset feature weights.

[0023] The directed acyclic graph multiplicative circuit breaker unit performs a continuous multiplicative aggregation operation on the state operators corresponding to at least one root node and the weighted summation result. This continuous multiplicative aggregation operation is controlled by the mathematical property of discrete binary zero-value absorption; when the value of any state operator is 0, a truncation mechanism is triggered, causing the output of the target score to be zero. The continuous multiplicative aggregation operation includes a feature count parameter, which corresponds in dimension to the number of output channels of the last fully connected layer in a linear classification network.

[0024] When the value of the state operator changes to 0, the directed acyclic graph (DAG) multiplicative circuit breaker unit generates a state pruning instruction containing a threshold condition identifier. The DAG multiplicative circuit breaker unit sends the state pruning instruction to the session state machine via the underlying communication interface. The session state machine parses the state pruning instruction, reads the threshold condition identifier, and traverses the memory stack to retrieve the node structure pointers. Based on the state pruning instruction, the session state machine marks the structure pointers corresponding to the child nodes containing the threshold condition identifier as unreachable. The session state machine sends a memory release call instruction to the operating system kernel, releasing the memory heap space occupied by the dialogue branch attached to the unreachable child node.

[0025] To prevent singularity overflow in floating-point operations, the interactive state perception and parameter generation unit preloads clamping boundary parameters. During subtraction, the lower limit of the coefficient blocks invalid operation paths where the vector space offset is lower than the system's floating-point noise floor; the upper limit of the coefficient blocks the input feature vector coordinates in high-dimensional space from penetrating the origin and causing coordinate sign reversal. The interactive state perception and parameter generation unit uses a closed interval and a hardware limiter to restrict the output floating-point value to within the closed interval. Before the activation mapping operation is triggered, the interactive state perception and parameter generation unit retrieves the slope constant and the center threshold. The center threshold is used as the subtraction input for the floating-point operation unit to perform a difference comparison calculation with the comprehensive feature exponent. The slope constant is used as the multiplier and the difference calculated by the difference comparison is used to perform scalar multiplication calculation, controlling the smoothness of the activation mapping curve and preventing the level signal from jumping and oscillating within adjacent interaction cycles.

[0026] As the comprehensive feature index approaches the physical boundary of 0, the subtraction of the central threshold and the comprehensive feature index generates a negative difference component. The natural exponential function unit calculates the negative difference component and outputs an exponential denominator term approaching the limit of positive infinity. The floating-point divider receives the exponential denominator term and performs division, causing the fractional output value to converge to 0. The floating-point adder adds the lower limit of the coefficients to the fractional output value. The floating-point values ​​of the denoising coupling coefficients are fixed at the lower limit of the coefficients. The lower limit of the coefficients cuts off the path of indiscriminate scaling multiplication, maintaining the original geometric coordinate structure of the input feature vector. When the front-end data flow causes the comprehensive feature index to approach the limit of positive infinity, the subtraction of the central threshold and the comprehensive feature index outputs a positive difference component. The natural exponential function unit operates on the positive difference component and outputs an exponential multiplier approaching 0. The fractional output value approaches the numerical difference between the upper and lower limits of the coefficients. The floating-point adder adds the numerical difference to the lower limit of the coefficients. The output value of the denoising coupling coefficients is enveloped within the upper limit of the coefficients, preventing the input feature vector from crossing the origin of the multidimensional feature embedding space.

[0027] When the system environment cannot call the underlying word segmenter to obtain the change in the ratio of word count, the interrupt controller captures the interactive input event sent by the front-end interactive terminal. The interaction state perception and parameter generation unit extracts the hardware timestamp corresponding to the interactive input event, subtracts adjacent hardware timestamps to generate a time interval set, and writes it into a contiguous address block. The arithmetic logic unit performs discrete variance calculation on the time interval set to generate a discrete variance floating-point value. The arithmetic logic unit uses the discrete variance floating-point value to replace the change in the ratio of word count and performs mixed addition and division calculations with the difference of the response timestamp. To trigger the equivalent operator delimitation logic of the activation mapping operation, the interaction state perception and parameter generation unit calls the logarithmic calculation module with boundary constraints to replace the natural exponential function unit. The arithmetic logic unit extracts the floating-point difference between the comprehensive feature index and the center threshold and inputs it into the linear rectification operator. The linear rectification operator filters out the negative value part and outputs a positive activation dynamic parameter. The logarithmic calculation module performs logarithmic operation on the positive activation dynamic parameter and outputs a logarithmic floating-point value. The hardware limiter truncates logarithmic floating-point values ​​greater than the upper limit of the coefficient to the upper limit of the coefficient, and raises logarithmic floating-point values ​​less than the lower limit of the coefficient to the lower limit of the coefficient, thus achieving bidirectional saturation limit constraint.

[0028] The underlying data analysis system server physically integrates a multi-core CPU, a graphics processing unit (GPU) accelerator card, volatile random access memory (RAM), and non-volatile flash memory chips. The network communication interface receives external message protocols via a bus, extracts the payload data from the message protocol, and writes the payload data into the heap space of the volatile RAM. Machine code instruction sets are programmed into the non-volatile flash memory chips. With the issuance of process mounting instructions, the interactive state awareness and parameter generation unit, edge carrier injection unit, feature space stripping unit, and directed acyclic graph (DAG) multiplicative fuse unit are loaded into the cache. The code logic is converted into very long instruction words (VLIW). The GPU accelerator card opens a single-instruction-multi-data-stream concurrent channel within its vector register, executing tensor scaling multiplication calculations and feature space stripping matrix subtraction calculations in parallel. In response to the Boolean comparison result, the DAG multiplicative fuse unit issues a memory overwrite interrupt request to the multi-core CPU. The multi-core CPU suspends the thread and writes the 0-level value of the state operator into the address segment allocated for the DAG data structure in the volatile RAM. The machine instruction set is stored in a non-volatile computer-readable storage medium, and the internal arrangement of the non-volatile computer-readable storage medium represents the binary level encoded stream of the data analysis method.

[0029] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based consumer survey data analysis system, characterized in that, include: The interaction state perception and parameter generation unit is configured to read the response timestamp difference and word quantity ratio change within adjacent interaction cycles, and use the response timestamp difference and word quantity ratio change to perform numerical calculations to generate a denoising coupling coefficient. An edge carrier injection unit is configured to receive a first entity parameter and a second entity parameter, selectively convert the second entity parameter into a conditional constraint string based on the entity word frequency determination result of the second entity parameter, and concatenate the conditional constraint string or the original form of the second entity parameter into the text containing the first entity parameter to generate a target question text, and output the reference polarity tensor corresponding to the target question text; The feature space stripping unit is configured to call an interface to convert the response text to the target question text into an input feature vector, obtain a preset bias feature vector, perform matrix multiplication on the result of scalar multiplication of the denoising coupling coefficient and the reference polarity tensor and the bias feature vector to generate a basis projection component matrix, and subtract the basis projection component matrix from the input feature vector in the feature space to output the target net value vector. The directed acyclic graph multiplicative circuit breaker unit is configured to assign the starting memory physical address of the target net value vector in the heap space to the feature load region pointer of at least one root node in a preset directed acyclic graph data structure, and configure a state operator for the at least one root node; extract the norm scalar value of the target net value vector, and if the norm scalar value satisfies a preset threshold condition, change the value of the state operator to zero, and generate a target score value through continuous multiplication operation; In response to the value of the state operator being zero, a state pruning instruction is output to the session state machine.

2. The AI-based consumer survey data analysis system based on artificial intelligence according to claim 1, characterized in that, The interaction state perception and parameter generation unit is specifically configured as follows: Extract the current response timestamp from the interactive data stream and subtract the previous response timestamp to obtain the response timestamp difference; The input text of adjacent interaction cycles is segmented into words. The ratio of the number of non-repeating words to the total number of words is calculated as the word ratio of the current period. A preset benchmark word ratio is obtained. The difference between the current word ratio and the benchmark word ratio is calculated to obtain the change in the word ratio. The response timestamp difference is divided using a preset constant to obtain a preset negative sensitivity coefficient. The negative sensitivity coefficient is multiplied by the change in the ratio of word counts and then input into a natural exponential function to obtain an exponential feature value. The quotient of the division is added to the exponential feature value to obtain a comprehensive feature index. The comprehensive feature index is used as the input variable for calculating the denoising coupling coefficient.

3. The AI-based consumer survey data analysis system based on artificial intelligence according to claim 2, characterized in that, The specific execution logic for the interaction state perception and parameter generation unit to generate the denoising coupling coefficient is as follows: The comprehensive feature index is compared with the preset center threshold, and a nonlinear activation mapping operation is performed on the obtained difference to output the denoising coupling coefficient. The nonlinear activation mapping operation is controlled by the mathematical property of the bidirectional saturation limit, which clamps the output value of the denoising coupling coefficient within a closed interval formed by the preset lower limit and upper limit of the coefficient, and exhibits smooth convergence characteristics when approaching the boundary of the closed interval.

4. The AI-based consumer survey data analysis system based on artificial intelligence according to claim 2, characterized in that, The interaction state perception and parameter generation unit is further configured to: Extract the time interval set of interactive input events, calculate the discrete variance of the time interval set, and use the discrete variance to replace the change in the ratio of word counts to calculate the comprehensive feature index.

5. The AI-based consumer survey data analysis system based on artificial intelligence according to claim 1, characterized in that, The edge carrier injection unit is specifically configured as follows: The entity recognition interface is called, and the entity word frequency of the second entity parameter is extracted by using a character sliding window and hash table key-value comparison. If the entity word frequency is greater than the preset word frequency threshold, the second entity parameter is converted into the conditional string; If the frequency of the entity word is not greater than the preset word frequency threshold, the second entity parameter is directly used in the original form to participate in the generation of the target question text; The conditional string or the second entity parameter of the original form is concatenated into the text containing the first entity parameter to generate the target question text and write it to the output buffer; in response to the write completion hardware interrupt signal returned by the system bus, the word embedding features corresponding to the target question text are extracted and the reference polarity tensor is generated by concatenating the dimension matrix.

6. The AI-based consumer survey data analysis system based on artificial intelligence according to claim 1, characterized in that, The specific calculation logic for the feature space stripping unit to subtract the basis projection component matrix from the input feature vector is as follows: Controlled by the mathematical properties of orthogonal projection decoupling, using the reference polarity tensor as the basis direction, the induced entanglement components generated in the input feature vector are eliminated inversely along the projection normal; the specific algebraic mapping formula executed is as follows: in, The target net value vector; The input feature vector; The denoising coupling coefficient is... Let be the reference polarity tensor; e be the bias feature vector; and t be the time identifier of the current interaction period.

7. The AI-based consumer survey data analysis system based on artificial intelligence according to claim 1, characterized in that, The directed acyclic graph multiplicative circuit breaker unit executes the following logic when calculating the target score: The target net value vector is input into a preset linear classification network layer to perform matrix multiplication calculation, and the feature score array is output. The state operator corresponding to the at least one root node is subjected to a continuous multiplication aggregation operation with the weighted sum of the feature score array based on preset feature weights to generate the target score value; The total number of features in the continuous multiplication aggregation operation is strictly corresponding in dimension to the number of output channels of the last fully connected layer of the linear classification network; the continuous multiplication aggregation operation is controlled by the mathematical property of discrete binary zero-value absorption. When the value of any state operator is zero, a forced truncation mechanism is triggered to make the output result of the target score value return to zero.

8. The AI-based consumer survey data analysis system based on artificial intelligence according to claim 1, characterized in that, The specific configuration for the directed acyclic graph multiplicative circuit breaker unit to output the state pruning instruction to the session state machine is as follows: When the value of the state operator changes to zero, the state pruning instruction containing the threshold condition identifier is generated; Based on the state pruning instruction, the child nodes in the session state machine that contain the threshold condition identifier are marked as unreachable.

9. A data analysis method for AI-based consumer interviews, characterized in that: The method includes the following steps: Read the response timestamp difference and word quantity ratio change within adjacent interaction cycles, and use the response timestamp difference and word quantity ratio change to perform numerical calculations to generate a denoising coupling coefficient; Receive a first entity parameter and a second entity parameter. Based on the entity word frequency determination result of the second entity parameter, selectively convert the second entity parameter into a conditional constraint string. Then, concatenate the conditional constraint string or the original form of the second entity parameter into the text containing the first entity parameter to generate the target question text. Output the reference polarity tensor corresponding to the target question text. The API call converts the response text to the target question text into an input feature vector, obtains a preset bias feature vector, performs matrix multiplication on the result of scalar multiplication of the denoising coupling coefficient and the reference polarity tensor and the bias feature vector to generate a basis projection component matrix, subtracts the basis projection component matrix from the input feature vector in the feature space, and outputs the target net value vector. Assign the starting memory physical address of the target net value vector in the heap space to the feature load region pointer of at least one root node in the preset directed acyclic graph data structure, and configure a state operator for the at least one root node. Extract the norm scalar value of the target net value vector. If the norm scalar value meets a preset threshold condition, change the value of the state operator to zero and generate the target score value through continuous multiplication. In response to the value of the state operator being zero, a state pruning instruction is output to the session state machine.

10. A non-volatile computer-readable storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the steps of the AI-based interview consumer survey data analysis method according to claim 9.

Images