An intelligent quality inspection system based on natural language models

Abstract

This invention provides an intelligent quality inspection system based on a natural language model, relating to the field of data processing technology. The system includes: calculating the bounding area of ​​a polygon in a vector space using a shoelace formula to obtain a semantic distribution area parameter; calculating a semantic aggregation coefficient based on the ratio of the semantic distribution area parameter to a preset benchmark area; using the semantic aggregation coefficient to perform weighted fusion processing on the word embedding vectors within a semantic analysis unit to generate an optimized semantic feature matrix; inputting the optimized semantic feature matrix into a Transformer-based deep learning classification model to extract deep semantic features and obtain a quality assessment result; and generating a final quality inspection report based on the quality assessment result and a preset quality threshold. This invention improves the accuracy of quality assessment and efficiently generates a final quality inspection report that meets the preset threshold.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an intelligent quality inspection system based on a natural language model. Background Technology

[0002] In common intelligent quality inspection scenarios such as customer service dialogue quality inspection and e-commerce product review quality assessment, automated quality assessment based on natural language processing (NLP) technology has been widely used. Taking customer service dialogue quality inspection on e-commerce platforms as an example, the platform needs to determine whether customer service has accurately responded to users' after-sales requests. Many common solutions rely on pre-trained language models to convert text into vector representations, and then use classification models to complete quality judgments, such as determining whether customer service dialogues belong to the category of effectively resolving requests, or whether product reviews are genuine experience feedback.

[0003] However, such methods may struggle to fully capture the inherent semantic integrity of text content that is semantically related but scattered in its expression. For example, in a customer service dialogue scenario for home appliances, a user might first mention that the washing machine they bought last week is particularly noisy during the spin cycle, then ask if they can get a replacement instruction manual, and later add that there is still an abnormal noise after the spin cycle ends. The core demands of these statements all revolve around the abnormal spin cycle function of the washing machine. However, because other irrelevant questions are interspersed in the middle, and the descriptions of the abnormality are scattered in different parts of the dialogue, the vectors generated by the pre-trained model may confuse key information such as the abnormal noise during the spin cycle and the loud noise with the semantics of requesting a replacement instruction manual, resulting in an insufficiently focused representation of the core demands. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an intelligent quality inspection system based on natural language models, which can improve the accuracy of quality assessment and efficiently generate final quality inspection reports that meet preset thresholds.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, an intelligent quality inspection system based on a natural language model includes:

[0007] The processing module is used to preprocess the collected raw text data to obtain a normalized text sequence;

[0008] The mapping module is used to input the normalized text sequence into the pre-trained word embedding model and generate the corresponding set of word embedding vectors through vector space mapping;

[0009] The analysis module is used to analyze the set of word embedding vectors and construct a semantic topology structure. Based on the semantic topology structure, a region growing algorithm is used to divide semantic regions and generate a set of semantic analysis units.

[0010] The optimization module is used to select key feature vectors to construct a vector space polygon based on the spatial distribution of word vectors within the semantic analysis unit set; calculate the enclosing area of ​​the vector space polygon using the shoelace formula to obtain the semantic distribution area parameter; calculate the semantic aggregation coefficient based on the ratio of the semantic distribution area parameter to the preset benchmark area; and use the semantic aggregation coefficient to perform weighted fusion processing on the word embedding vectors within the semantic analysis unit to generate the optimized semantic feature matrix.

[0011] The evaluation module is used to input the optimized semantic feature matrix into the Transformer-based deep learning classification model, extract deep semantic features, and obtain quality evaluation results.

[0012] The generation module is used to generate the final quality inspection report based on the quality assessment results and preset quality thresholds.

[0013] In a second aspect, a computing device includes:

[0014] One or more processors;

[0015] A storage device for storing one or more programs that, when executed by one or more processors, enable the one or more processors to implement the system.

[0016] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the system.

[0017] The above-described solution of the present invention has at least the following beneficial effects:

[0018] By constructing polygons in vector space and calculating using the shoelace formula, semantic distribution features are quantified and semantic aggregation coefficients are obtained. This enables weighted fusion of word embedding vectors, highlighting core semantic features and weakening redundant information, making the semantic feature matrix of the input classification model more focused on key quality inspection dimensions. The weighted processing based on semantic aggregation coefficients can adaptively adjust the importance of different word vectors, improving the targeting of deep semantic feature extraction and enhancing the model's adaptability to complex text scenarios. The entire process from raw text preprocessing, semantic analysis, quality assessment to report generation requires no manual intervention, replacing the quality inspection mode that relies on manual judgment and shortening the quality inspection cycle. Combined with preset quality thresholds, standardized quality inspection reports are automatically generated, making the quality inspection results more accurate. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of an intelligent quality inspection system based on a natural language model, provided by an embodiment of the present invention.

[0020] Figure 2This is a flowchart illustrating the process of analyzing a set of word embedding vectors to construct a semantic topology structure, and then using a region growing algorithm to divide semantic regions based on the semantic topology structure to generate a set of semantic analysis units, as provided in an embodiment of the present invention. Detailed Implementation

[0021] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0022] like Figure 1 As shown, an embodiment of the present invention proposes an intelligent quality inspection system based on a natural language model, comprising:

[0023] The processing module is used to preprocess the collected raw text data to obtain a normalized text sequence;

[0024] The mapping module is used to input the normalized text sequence into the pre-trained word embedding model and generate the corresponding set of word embedding vectors through vector space mapping;

[0025] The analysis module is used to analyze the set of word embedding vectors and construct a semantic topology structure. Based on the semantic topology structure, a region growing algorithm is used to divide semantic regions and generate a set of semantic analysis units.

[0026] The optimization module is used to select key feature vectors to construct a vector space polygon based on the spatial distribution of word vectors within the semantic analysis unit set; calculate the enclosing area of ​​the vector space polygon using the shoelace formula to obtain the semantic distribution area parameter; calculate the semantic aggregation coefficient based on the ratio of the semantic distribution area parameter to the preset benchmark area; and use the semantic aggregation coefficient to perform weighted fusion processing on the word embedding vectors within the semantic analysis unit to generate the optimized semantic feature matrix.

[0027] The evaluation module is used to input the optimized semantic feature matrix into the Transformer-based deep learning classification model, extract deep semantic features, and obtain quality evaluation results.

[0028] The generation module is used to generate the final quality inspection report based on the quality assessment results and preset quality thresholds.

[0029] In the embodiments of the present invention, through the construction of vector space polygons and the calculation of the shoelace formula, the semantic distribution features are quantified and the semantic aggregation coefficient is obtained, so as to achieve the weighted fusion of word embedding vectors, highlight the core semantic features, weaken the redundant information, and make the semantic feature matrix input into the classification model more focused on the key quality inspection dimensions; the weighted processing based on the semantic aggregation coefficient can adaptively adjust the importance of different word vectors, improve the pertinence of deep semantic feature extraction, and enhance the adaptability of the model to complex text scenarios; the whole process from the preprocessing of the original text, semantic analysis, quality assessment to report generation does not require manual intervention, replaces the quality inspection mode that relies on manual judgment, and shortens the quality inspection cycle; combines the preset quality threshold to automatically generate a standardized quality inspection report, and at the same time makes the quality inspection results more accurate.

[0030] In a preferred embodiment of the present invention, the collected original text data is preprocessed to obtain a standardized text sequence, which specifically includes: identifying redundant information in the original text, setting expressions that are repeated continuously 3 times or more as redundant, such as the repeated "Hello, do you have any other questions?" in the customer service conversation for 3 times, only retaining the first occurrence, and deleting the subsequent repeated content; processing colloquial abbreviations and Internet terms, converting "dear" to "you", "haven't got" to "haven't", and directly deleting Internet terms without clear semantics such as "yyds"; unifying the text format is to convert all letters to lowercase, and converting Chinese numerals such as "one, two, three", English numerals such as "two, three" to Arabic numerals 123; finally, checking the coherence of the processed text. If deleting Internet terms causes semantic breaks, such as the original sentence "This refrigerator yyds cools fast", after deleting "yyds", supplement "very" and adjust it to "This refrigerator cools very fast" to obtain a coherent standardized text sequence.

[0031] In this embodiment, by refining the preprocessing rules, clarifying the determination criteria for redundant information and the coherence adjustment method, it is possible to avoid excessive deletion of valid information, and make the text format regular and the semantics complete.

[0032] In a preferred embodiment of the present invention, the standardized text sequence is input into a pre-trained word embedding model, and a corresponding set of word embedding vectors is generated through vector space mapping, which may include:

[0033] Process the normalized text sequence to obtain purified text data; perform lexical segmentation on the purified text data based on a preset dictionary to obtain a basic vocabulary sequence; filter special characters and stop words in the basic vocabulary sequence to obtain a valid vocabulary sequence, specifically including: judging whether the character encoding conforms to the UTF-8 standard, identifying unresolvable garbled characters such as 大 caused by inconsistent encoding, scanning the text character by character and deleting all of them; then processing spaces and line breaks, only retaining 1 space for 2 or more consecutive spaces, deleting all consecutive meaningless line breaks (non-paragraph separators) within a paragraph, and retaining 1 line break for 2 or more consecutive line breaks between paragraphs to ensure the text format is regular, and obtaining purified text data after processing.

[0034] The preset dictionary contains common words and combinations in customer service conversations and e-commerce review scenarios. The basic vocabulary covers after-sales, refund, quality, good reviews, etc. The combined vocabulary includes after-sales service, quality problems, refund applications, etc. The occurrence frequency of newly added words in the scenario is counted monthly, and words and combinations with ≥10 occurrences are added to the dictionary for update; during segmentation, starting from the beginning of the text, the longest word combination in the dictionary is preferentially matched. For example, when a text appears with problems related to after-sales service, directly match after-sales service as a single segmentation unit; if a text segment cannot be matched with any word or combination in the dictionary (such as a user-created ice washer), it is split into single characters wash, ice, machine as temporary segmentation units, and at the same time marked as to be verified. Subsequently, manual sampling inspection is carried out to optimize the splitting rules until the entire text is segmented to form a basic vocabulary sequence; unifying the special character format is to convert full-width punctuation marks such as “!”“?” into half-width and then traverse and remove them one by one together with symbols such as “@”“#”“$”; then filter stop words. Stop words include function words such as “的”, “了”, “在”, “是” and high-frequency words such as “哦”, “呢”, “啊” that are meaningless in the quality inspection scenario. The preset stop word list is adjusted monthly according to the quality inspection data for newly added high-frequency meaningless words (such as “嗯嗯”, “哦哦”). Query this list to delete the stop words in the basic vocabulary sequence, and the remaining words form a valid vocabulary sequence.

[0035] Match and query each word in the valid vocabulary sequence with the vocabulary list of the pre-trained word embedding model to obtain the corresponding matching word vector; process the words that cannot be matched to obtain a segmented word vector combination, specifically including: collecting 1 million pieces of training data, including 400,000 pieces of customer service conversation texts covering 10 categories such as home appliances and clothing, 300,000 pieces of e-commerce review texts, 300,000 pieces of news and blog general language materials; after segmenting the training data according to the segmentation rules in step 102, first filter out single-character meaningless words (such as “啊呀”) with an occurrence frequency <2 in the segmentation results, and then remove duplicates from the remaining vocabulary sequence, retain the repeated words in different texts and record their occurrence scenarios (such as after-sales appears in customer service conversations, quality appears in e-commerce reviews) to obtain a clean training vocabulary sequence.

[0036] When training the GloVe model, the process of constructing the global word co-occurrence matrix is ​​as follows: Based on the cleaned training vocabulary sequence, a context window size of 5 is set, meaning the co-occurrence statistics for each center word cover the 5 words to its left and 5 words to its right. The number of times any two words within the window appear simultaneously is used as the initial value of the co-occurrence matrix. A weight decay rule is then applied to this initial value. The distance between the center word and other words within the window is k, where k is an integer from 1 to 5, corresponding to weights of 1.0, 0.5, 0.333, 0.25, and 0.2 (weight values ​​range from 0.2 to 1.0). The weights are then multiplied by the initial co-occurrence count to obtain the weighted global word co-occurrence matrix. The model training parameters are set as follows: the vector dimension is 300, the initial word vectors are randomly initialized using a normal distribution with a mean of 0 and a variance of 0.01, and the initial value of each dimension is between -0.04 and 0.04; the minimum word frequency is 5, and only words that appear ≥5 times in the co-occurrence matrix are retained for training; the batch size is 1024, and 1024 co-occurrence word pairs are processed in each iteration; the initial learning rate is 0.05, and the learning rate is multiplied by 0.5 and decayed after every 5 iterations. The learning rate is 0.025 in the 6th iteration and 0.0125 in the 11th iteration, with the overall learning rate ranging from 0.0125 to 0.05; the number of iterations is 15.

[0037] During training, a squared loss function with weighted factors is used to calculate the error. Specifically, a weighted factor f(Cmn) is defined, where Cmn is the actual co-occurrence count of the center word m and the context word n ​​in the co-occurrence matrix, and the weight saturation value is set to 100 (fixed value 100). When Cmn = 0, f(Cmn) = 0; when 1 ≤ Cmn < 100, f(Cmn) = Cmn ÷ 100, and the result is rounded to three decimal places; when Cmn ≥ 100, f(Cmn) = 1. Through this calculation method, the weighted factor value is fixed between 0 and 1. The 300-dimensional vector wm of the center word m, the 300-dimensional vector wn of the context word n, and the initial center word bias bm are obtained. The initial values ​​are: wm (value 0.01), context word bias bn (initial value 0.01), and global bias b (initial value 0.001); calculate the dot product of wm and wn, i.e., multiply the first dimension of wm by the first dimension of wn, add the second dimension of wm by the second dimension of wn, and so on until all 300 dimensions are multiplied and summed; then add bm, bn, and b to the dot product result to obtain the predicted value; take the natural logarithm of the actual co-occurrence frequency Cmn (take 0 when Cmn = 0), and calculate the difference between the predicted value and the natural logarithm; square the difference and multiply it by the weight factor f(Cmn) to obtain the loss value of a single co-occurring word pair; sum the individual loss values ​​of 1024 samples in the batch to obtain the total batch loss, and keep the result to 4 decimal places.

[0038] Calculate the gradients of each parameter: the gradient of wm is 2×f(Cmn)×(predicted value - lnCmn)×wn, the gradient of wn is 2×f(Cmn)×(predicted value - lnCmn)×wm, the gradients of bm and bn are both 2×f(Cmn)×(predicted value - lnCmn), and the gradient of b is 2×f(Cmn)×(predicted value - lnCmn). Update each parameter by subtracting the current iteration learning rate × the corresponding gradient from its current value, retaining 4 decimal places after the update. Stop training when the total batch loss fluctuation is ≤0.001 for 3 consecutive iterations. After training, analyze the generated... Word vector normalization involves calculating the L2 norm of each word vector, which is the square root of the sum of the squares of the values ​​in each dimension of the vector. Then, each dimension value is divided by the L2 norm. After normalization, the values ​​in each dimension range from -1 to 1. The vocabulary list includes valid words that appear at least 5 times in the training data and have been manually reviewed. During the manual review, entries without actual semantic meaning, such as garbled words and combinations of special symbols, are excluded. Each valid word corresponds to a normalized 300-dimensional vector. Each word in the valid word sequence is compared with the vocabulary list in turn. If a matching entry exists, the corresponding 300-dimensional vector is taken as the matching word vector.

[0039] Unregistered words include emerging product names (such as "Smart Sweeping Robot Pro") and user-created expressions (such as "washes not clean and wastes water"). Words are segmented according to semantic segmentation rules; for example, "Smart Sweeping Robot Pro" is segmented into "smart," "sweeping," "robot," and "Pro," while "washes not clean and wastes water" is segmented into "wash," "get," "not," "clean," "still," and "wastes water." User-created words without clear roots or affixes (such as "refrigerator and washing machine") are segmented by single character, and the segmentation rules are recorded. These words are summarized monthly to optimize the pre-defined dictionary; for example, "refrigerator and washing machine" is added to the dictionary as an abbreviation for refrigerator and washing machine. Each segmented word is matched with the vocabulary of the pre-trained word embedding model, and the corresponding vectors are extracted to form a segmented word vector combination.

[0040] The segmented word vectors are combined and averaged to obtain alternative word vectors for out-of-vocabulary words. All matching and alternative word vectors are then integrated to obtain an initial word vector set. This involves: traversing all vectors in the segmented word vector combination, sequentially summing the values ​​of each dimension from dimension 1 to 300, treating any dimension as zero in extreme cases; dividing the sum of each dimension by the number of vectors in the combination, and rounding the result to four decimal places to obtain the average value for each dimension; using these average values ​​to form a 300-dimensional alternative word vector, with dimensions consistent with the vectors in the pre-trained word embedding model; establishing a word-vector mapping index table, labeling the source of each vector (matching word vector, alternative word vector); and arranging the vectors corresponding to each word in the original order of the valid word sequence. For example, in a valid word sequence like "washing machine, spin-drying, loud noise, replacement, instruction manual," the words "washing machine," "spin-drying," etc., correspond to matching word vectors, while "replace" serves as the alternative word vector for out-of-vocabulary words. These vectors are then combined in order to form the initial word vector set.

[0041] The initial word vectors are processed to unify their dimensions, resulting in word vectors with standard dimensions. Based on the weighted calculation of adjacent word vectors, out-of-vocabulary words in the standard-dimensional word vectors are augmented to obtain a complete set of word vectors. This complete set of word vectors is then arranged and combined according to the original text order to generate a set of word embedding vectors. Specifically, this involves checking the dimension of each vector; if a vector has 299 dimensions or less than 300 dimensions, dimensions with a value of 0 are added to the end of the vector until it reaches 300 dimensions; if a vector has 305 dimensions or more than 300 dimensions, the first 300 dimensions (preserving the core semantic information of the vector) are extracted using principal component analysis as a new vector; after adjustment, all vectors are 300 dimensions, resulting in word vectors with standard dimensions.

[0042] The substitution word vector and its position in the standard dimension word vector are identified by an index table. If the substitution word vector is at the beginning of the sequence, only the vector of the next word is taken as the reference vector; if it is at the end of the sequence, only the vector of the first word is taken as the reference vector; if it is in the middle of the sequence, the vectors of the first and last words are taken as the reference vector; if the sequence contains only one substitution word vector (without preceding or following words), then the vector itself is taken as the reference vector and the weighting coefficient is set to 1.0; the preceding reference vector is multiplied by 0.6 and the following reference vector is multiplied by 0.4, and the corresponding dimension values ​​of the two are added together to obtain the weighted vector. Add the weighted vector to the corresponding dimension value of the original substitute word vector and divide by 2 (keeping the result to 4 decimal places) to obtain the completed out-of-vocabulary word vector; replace the original substitute word vector with the completed vector, leaving the other vectors unchanged, to obtain the complete word vector set; store the complete word vector set in the format of word index and vector value, ensuring that the vector arrangement order is consistent with the word appearance order in the original text, such as the original text having the dehydration sound loud before the supplementary instruction manual, the corresponding vector also maintains this order; connect the vectors sequentially to form an ordered vector sequence, which is the word embedding vector set.

[0043] This embodiment clarifies the category coverage and quantity of training data, the specific conditions for stopping model training, and conducts manual review of the vocabulary to make the vector representation of pre-trained words embedded in the model more in line with the quality inspection scenario, reducing semantic bias caused by general corpora. The reference basis and subsequent optimization mechanism for out-of-vocabulary word splitting rules, combined with vector averaging precision control, make the alternative vectors of out-of-vocabulary words closer to the true semantics, avoiding semantic breaks.

[0044] like Figure 2 As shown, in another preferred embodiment of the present invention, the word embedding vector set is analyzed to construct a semantic topology structure; based on the semantic topology structure, a region growing algorithm is used to divide semantic regions and generate a set of semantic analysis units, which may include:

[0045] Density clustering analysis is performed on the word embedding vector set to identify the core semantic node set. Based on this core semantic node set, a semantic topology is constructed. The semantic similarity between nodes in the semantic topology is calculated to generate a semantic association weight matrix. Specifically, the DBSCAN algorithm is used, with the radius parameter set to 0.5. This value is determined based on the distance distribution statistics of semantically similar vectors in a quality inspection scenario; two vectors with a distance less than 0.5 are considered to have reachable density. The minimum sample size is set to 3 to filter isolated low-association vectors; a vector is considered a core point when the number of vectors with a distance less than 0.5 around it is no less than 3. All vectors are first L2 normalized by dividing each dimension value of each vector by its L2 norm. The square root of the sum of the squares of the values ​​of each dimension of the vector is obtained. After normalization, the values ​​of each dimension range from -1 to 1. For the normalized vector, the cosine similarity of each vector to all other vectors is calculated. In the calculation, the corresponding dimension values ​​of the two vectors are multiplied and then summed, and the result is rounded to four decimal places. The distance between the vectors is obtained by subtracting the similarity from 1. The distance ranges from 0 to 2. All vectors are initialized to an unvisited state. Each unvisited vector is traversed, and its distance to all other vectors is calculated. The number of vectors with a distance less than 0.5 is counted. If the number is not less than 3, the vector is marked as a core point and included in the core semantic node set, and is also marked as visited. If the number is less than 3, it is marked as a non-core point (boundary point or noise point). All core points are combined into a core semantic node set.

[0046] Using each core semantic node as a node in the graph, the cosine similarity between any two nodes is calculated after L2 normalization of the node vectors. The calculation method is to multiply the corresponding dimension values ​​of the two vectors and then sum them, and the result is rounded to four decimal places. If the similarity is greater than 0.6, an edge is established between the two nodes, and the weight of the edge is set to the similarity value, with the weight ranging from 0.6001 to 1.0000. If the similarity is less than or equal to 0.6, no edge is established. This method forms a semantic topology containing nodes and edges. Nodes correspond to core semantics, and the weights of edges reflect the strength of semantic associations between nodes. For each node, its semantic similarity with all nodes, including itself, is calculated. The similarity between a node and itself is fixed at 1, and the result is rounded to four decimal places. According to the order of nodes in the core semantic node set, these similarity values ​​are filled into a two-dimensional array. The row and column indices of the two-dimensional array correspond to the order of the core semantic nodes, respectively. The value in the i-th row and j-th column of the array is the semantic similarity between the i-th node and the j-th node. This two-dimensional array is the semantic association weight matrix, and the values ​​in the matrix range from 0.0000 to 1.0000.

[0047] Based on the semantic association weight matrix, a region growing algorithm is used to expand the semantic region and form an initial semantic partition. The initial semantic partition is then subjected to boundary optimization to obtain an optimized semantic partition. Specifically, this includes: selecting unassigned nodes from the core semantic node set, prioritizing the node with the smallest variance in vector dimension values ​​as the initial seed node. The variance is calculated as the sum of the squares of the differences between each dimension value and the mean of the vector divided by the number of dimensions (300). The smaller the variance, the more stable the semantics. In the semantic association weight matrix, unassigned nodes with a similarity greater than 0.7 to the seed node are found and included in the current growing region, while recording the order of inclusion. The newly included nodes are used as new seeds, and the process of finding unassigned nodes with a similarity greater than 0.7 is repeated until no new nodes can be included, and the current region is taken as an initial semantic partition. Seed nodes are reselected from the unassigned nodes, and the above expansion process is repeated until all core nodes are assigned to the corresponding initial semantic partitions.

[0048] Calculate the average value of each dimension of all node vectors in each initial partition. The calculation method is to sum the corresponding dimension values ​​of all nodes in the partition and divide by the number of nodes, keeping the result to four decimal places. These average values ​​form the center vector of the partition. Calculate the Euclidean distance between each node in the partition and the center vector. The calculation method is to take the square root of the sum of the squares of the differences between the corresponding dimension values ​​of the two vectors, keeping the result to four decimal places. The node with the largest distance is determined as the boundary node. Check the similarity of each boundary node to all nodes in other partitions. If the similarity of a boundary node to a node in another partition is greater than the similarity of the boundary node to the center vector of its own partition, the boundary node is adjusted to the corresponding partition. Repeat the above operations until the affiliation of all boundary nodes no longer changes, resulting in the optimized semantic partition.

[0049] Based on a preset semantic density threshold, optimized semantic partitions are filtered to obtain semantic partitions that meet the density requirements. Each semantic partition that meets the density requirements is divided into a semantic analysis unit, generating a set of semantic analysis units. Specifically, this includes: the preset semantic density threshold requires each partition to contain no less than 5 nodes; the number of nodes in each optimized semantic partition is counted; if the number of nodes is greater than or equal to 5, the partition meets the density requirements; if the number of nodes is less than 5, the Euclidean distance between the center vectors of the partition and all adjacent partitions is calculated, and the semantic association weights between partitions are obtained (within the range of 0.6001 to 1.0000), and the partition with the closest and most semantically related elements is selected. The adjacent partition with the highest semantic association weight is merged with the low-density partition. After merging, the center vector of the new partition is recalculated by summing the corresponding dimension values ​​of all nodes after merging and dividing by the total number of nodes to ensure semantic consistency. Finally, semantic partitions that meet the density requirements are obtained. Each partition that meets the density requirements contains a set of semantically closely related nodes, corresponding to a set of related semantics in the original text. It is treated as an independent semantic analysis unit, and the original text fragments corresponding to the unit are labeled, such as the related expressions of loud dehydration sound and abnormal dehydration noise. These semantic analysis units are arranged in the order of their appearance in the original text to form a set of semantic analysis units.

[0050] In this embodiment, boundary nodes are determined by Euclidean distance and their affiliation is strictly adjusted according to similarity, so that semantic partitioning can accurately aggregate scattered related semantics and solve the problem of insufficient semantic integrity. When merging low-density partitions, the center vector distance and semantic association weight are referenced to avoid semantic chaos after merging and ensure that the semantic density of each semantic analysis unit is sufficient.

[0051] In a preferred embodiment of the present invention, based on the spatial distribution of word vectors within the semantic analysis unit set, key feature vectors are selected to construct a vector space polygon; the enclosing area of ​​the vector space polygon is calculated using the shoelace formula to obtain the semantic distribution area parameter; the semantic aggregation coefficient is calculated based on the ratio between the semantic distribution area parameter and the preset benchmark area; the word embedding vectors within the semantic analysis unit set are weighted and fused using the semantic aggregation coefficient to generate an optimized semantic feature matrix, which may include:

[0052] Based on the spatial coordinates of word vectors within the semantic analysis unit set, the distribution boundary of each word vector is calculated to obtain the boundary feature vector set. According to the spatial positional relationship of the boundary feature vector set, a convex polygon structure is constructed, specifically including: traversing all word vectors within the semantic analysis unit set, each word vector having 300-dimensional spatial coordinates; for each dimension, comparing the values ​​of all word vectors in that dimension, recording the word vector with the largest value and the word vector with the smallest value; if multiple word vectors in the same dimension have the same value and are either the largest or smallest, recording them all; summarizing the maximum and minimum value word vectors of all 300 dimensions, removing duplicate vectors through complete vector value matching, and the remaining vectors form the boundary feature vector set. For example, in the semantic analysis unit for abnormal washing machine spin-drying, the maximum value of word vectors such as spin-drying, loud noise, and abnormal noise in the 50th dimension is borne by the abnormal noise vector, and the minimum value is borne by the spin-drying vector. These vectors will be included in the boundary feature vector set.

[0053] The Graham scan algorithm is used to process the boundary feature vector set. First, PCA is performed to reduce the dimensionality of the 300-dimensional boundary feature vectors. The covariance matrix of the 300-dimensional vectors is calculated by subtracting the mean of each dimension from all vector values ​​to obtain the deviation vectors. All deviation vectors form the deviation matrix, and the covariance matrix is ​​the transpose of the deviation matrix multiplied by the deviation matrix and divided by the number of vectors minus one. Eigenvalue decomposition is performed on the covariance matrix to obtain 300 eigenvalues ​​and their corresponding eigenvectors. The two eigenvectors with the largest eigenvalues ​​are selected as principal components. Each 300-dimensional boundary vector is multiplied by these two principal component vectors and summed to obtain the 2-dimensional coordinates (x-value and y-value) of each vector, which are used as planar coordinates. The starting point is determined in the dimensionality-reduced planar coordinates. The y-values ​​of all vectors are traversed, and the vector with the smallest y-value is selected. If multiple vectors have the same y-value, their x-values ​​are compared, and the vector with the smallest x-value is selected as the starting point.

[0054] Calculate the spatial angles between other vectors and the starting point. For each non-starting point vector, calculate the coordinate difference between the vector and the starting point, i.e., dx equals the x-value of the vector minus the x-value of the starting point, and dy equals the y-value of the vector minus the y-value of the starting point. Use the arctangent function atan2(dy,dx) to calculate the angles (the result ranges from -180 degrees to 180 degrees), and convert the negative angles to a range of 0 to 360 degrees by adding 360 degrees. Sort the non-starting point vectors in ascending order of their angles to form an ordered vector list, and place the starting point at the beginning of the list.

[0055] Construct a sequence of convex polygon vertices, initialize a stack structure, and push the first two vectors of the ordered vector list onto the stack. Starting from the third vector, traverse the list, taking the second-to-last vector from the top of the stack as p1, the top vector as p2, and the currently traversed vector as p3. Calculate the difference between vector p2 and p1 (x-value of p2 minus x-value of p1, y-value of p2 minus y-value of p1) to obtain vector a. Calculate the difference between vector p3 and p2 (x-value of p3 minus x-value of p2, y-value of p3 minus y-value of p2) to obtain vector b. Calculate the cross product of vector a and vector b, i.e., the cross product of vector a and vector b. The x-value is multiplied by the y-value of vector b, minus the y-value of vector a multiplied by the x-value of vector b. If the cross product is less than 0 (indicating the broken line turns right), the top vector p2 is popped from the stack, and the calculation is repeated until the cross product is greater than or equal to 0. If the cross product is greater than or equal to 0 (indicating the broken line turns left or is collinear), the current vector p3 is pushed onto the stack. After traversal, the remaining vectors in the stack form the vertices of the convex polygon in the order they were pushed onto the stack. If the number of vectors is less than 3, 2 vectors form a line segment (considered a degenerate convex polygon), and 1 vector forms a point. The planar coordinates of these vertices are arranged in the order in the stack, forming an ordered vertex coordinate sequence.

[0056] Based on the ordered vertex coordinate sequence of the convex polygon structure, a coordinate data set is obtained. The sum of the cross products of adjacent coordinates in the coordinate data set is calculated to obtain the original area value. The original area value is then subjected to absolute value processing and normalization to obtain the semantic distribution area parameter. Specifically, this includes: extracting the two-dimensional coordinates (x and y values) of each vertex from the ordered vertex coordinate sequence, storing them sequentially in the coordinate data set, and setting the next vertex after the last vertex as the first vertex to close the polygon; for each pair of adjacent vertices (v1 and v2), the x-value of v1 is multiplied by the y-value of v2 to obtain product 1, and then the x-value of v2 is multiplied by the y-value of v1 to obtain product 2. Subtracting product 2 from product 1 yields the product of the pair of adjacent vertices. The cross product of vertices is calculated; the cross products of all adjacent vertex pairs are added together to obtain the original area value, which may be positive or negative depending on the vertex arrangement order; the absolute value of the original area value is calculated to ensure that the area is positive; the preset normalization benchmark is the maximum distribution area of ​​semantic analysis units of the same type, which is determined by statistically analyzing the maximum distribution area of ​​all semantic analysis units in the same scenario (such as home appliance after-sales service) in the past 3 months; the absolute value of the original area is divided by the benchmark to obtain the normalized value, and the value is rounded to four decimal places as the semantic distribution area parameter. The larger the parameter value, the more dispersed the word vector is in space. For example, the parameter of the washing machine spin-drying abnormal unit is 0.3521, indicating that its distribution dispersion is moderate.

[0057] The initial aggregation coefficient is obtained by calculating the ratio of the semantic distribution area parameter to the preset benchmark area. The initial aggregation coefficient is then logarithmically transformed to obtain the intermediate aggregation coefficient. The intermediate aggregation coefficient is then validated to obtain the semantic aggregation coefficient. Specifically, the preset benchmark area is the average distribution area of ​​similar semantic analysis units, determined by statistically analyzing the average distribution area of ​​1000 samples from the same scene over the past three months. If the benchmark area is 0, the initial aggregation coefficient is set to 0. The natural logarithm (base e) of the initial aggregation coefficient is taken; if the initial aggregation coefficient is 0, the intermediate aggregation coefficient is set to 0. The intermediate aggregation coefficient is then validated to ensure it falls within the range of -5 to 5. If it is less than -5, it is set to -5; if it is greater than 5, it is set to 5, thus obtaining the semantic aggregation coefficient.

[0058] The semantic aggregation coefficient is used to weight the word embedding vectors within the semantic analysis unit set to obtain an enhanced word vector set. The enhanced word vector set is then reorganized according to its original dimensions to generate an optimized semantic feature matrix. Specifically, the semantic aggregation coefficient ranges from -5 to 5, and is converted into a weight factor by adding the absolute value of the coefficient to 1. Therefore, the weight factor ranges from 1 to 6 (e.g., the absolute value of the coefficient -2.1 is 2.1, and 1 plus 2.1 equals 3.1). The 300 dimension values ​​of each word embedding vector are multiplied by the weight factor, and the product is rounded to four decimal places to obtain the weighted word vector. All the weighted word vectors are arranged in their original order to form an enhanced word vector set. The number of rows is determined by the number of vectors in the enhanced word vector set (usually 5 to 20), and the number of columns is determined by 300 dimensions. The dimension values ​​of each vector are then filled into the corresponding positions in the matrix (the first vector corresponds to the first row, and the first dimension corresponds to the first column) to generate the optimized semantic feature matrix.

[0059] In this embodiment, the calculation of semantic aggregation coefficients incorporates historical benchmark data and range verification, enabling the weighted processing to highlight the core semantics while avoiding excessive amplification of noise, thus enhancing the stability of the features.

[0060] In a preferred embodiment of the present invention, the optimized semantic feature matrix is ​​input into a Transformer-based deep learning classification model to extract deep semantic features and obtain a quality assessment result, which may include:

[0061] The optimized semantic feature matrix is ​​input into the Transformer encoder to calculate the attention-enhanced features. The attention-enhanced features are then subjected to layer normalization to obtain normalized features. These normalized features are then input into a feedforward neural network for nonlinear transformation to obtain deep semantic encoding. Specifically, the Transformer encoder contains six encoding layers, each consisting of a multi-head self-attention mechanism and a feedforward neural network. The weights of each layer are initialized through pre-training (the pre-training data consists of 1 million customer service dialogues and e-commerce review texts, with 50 iterations during training, using the cross-entropy loss function, stopping when the loss value fluctuates ≤0.001 for 5 consecutive iterations). The multi-head self-attention mechanism has eight heads, transforming the optimized semantic feature matrix into a query matrix, key matrix, and value matrix through three linear transformation matrices. Each head contains subqueries, subkeys, and subvalues. The matrix dimensions are 38 (for the first 4 heads) and 37 (for the last 4 heads, 300 = 38 × 4 + 37 × 4). To calculate the attention score for each head, the subquery matrix and the transpose of the subkey matrix are multiplied (row × column corresponding elements are multiplied and summed) to obtain the original score matrix. Each element is divided by the square root of 8 (approximately 2.828), and then normalized using the softmax function (each element is taken as its natural exponent, divided by the sum of the natural exponents of all elements in that row, and the result is rounded to 4 decimal places). The normalized attention score is multiplied by the subvalue matrix (row × column corresponding elements are multiplied and summed) to obtain the output for each head. The outputs of the 8 heads are concatenated and subjected to a linear transformation (weights initialized as before) to obtain the multi-head attention output. The multi-head attention output is then added element-wise to the input feature matrix of that layer (corresponding position values ​​are added) to obtain the attention enhancement feature.

[0062] The mean and variance of the attention enhancement features are calculated along the feature dimension, which has 300 columns. Each row represents one word vector, and the number of elements is the total number of vectors (consistent with the number of enhanced word vectors). The mean is the sum of all elements in each feature column divided by the number of elements, rounded to four decimal places. The variance is the sum of the squared differences between each element in each feature column and the mean of that column, divided by (number of elements + 0.00001), rounded to four decimal places. Each feature value is subtracted from the mean of that feature column and then divided by the square root of the variance of that feature column to obtain the normalized feature. A linear transformation is performed on the normalized features. The weight matrix is ​​300×300 (range [-0.01, 0.01]), and the bias vector is 300-dimensional (range [0, 0.02]). That is, the feature value is multiplied by its corresponding weight and then the bias is added to obtain the normalized feature, ensuring that different... The samples have consistent feature distributions. The feedforward neural network contains two linear transformations. The first layer has a weight matrix of 300×1024 (range [-0.01, 0.01]) and a bias vector of 1024 dimensions (range 0.1±0.01), mapping the 300-dimensional features to 1024 dimensions (each element is the sum of the 300-dimensional feature multiplied by the corresponding weight and the bias). The ReLU activation function is used to process the 1024-dimensional features, setting values ​​less than 0 to 0 and keeping values ​​greater than or equal to 0 unchanged. The second layer has a weight matrix of 1024×300 (range [-0.01, 0.01]) and a bias vector of 300 dimensions (range 0.1±0.01), mapping the 1024-dimensional features back to 300 dimensions (calculated in the same way as the first layer). The 300-dimensional features are added element-wise to the normalized features to obtain the deep semantic encoding.

[0063] The deep semantic encoding is processed to obtain stable semantic features; these stable semantic features are then input into a global pooling layer for feature compression to obtain a global semantic vector; finally, the global semantic vector is input into a fully connected classification layer for linear transformation to obtain the original classification score. Specifically, this involves: applying dropout to the deep semantic encoding with a dropout rate of 10% (fixed range); generating a mask matrix of the same dimension as the deep semantic encoding using a Bernoulli distribution (probability parameter p = 0.1), where the mask matrix elements are either 0 or 1, with 10% of the elements being 0; and multiplying the deep semantic encoding element-wise by the mask matrix (with a specific position for each 0 element). (The eigenvalues ​​are set to 0, and the positional eigenvalues ​​of elements with 1 remain unchanged). During each forward propagation of the model, the mask matrix is ​​randomly generated to avoid the model from over-relying on specific features. The features after dropout processing are the stable semantic features. The global pooling layer processes each of the 300 dimensions of the stable semantic features, and the number of vectors is the total number of word vectors in the semantic analysis unit (denoted as n). For the first dimension, the values ​​of the n vectors in this dimension are added together and then divided by n. The result is rounded to four decimal places to obtain the average value of the first dimension. The average values ​​of the remaining 299 dimensions are calculated in the same way to form a 300-dimensional vector. This vector is the global semantic vector.

[0064] The fully connected classification layer contains three neurons, corresponding to the three categories of qualified, need improvement, and unqualified. The weight matrix within the layer is 300×3 (initialized using a random normal distribution, mean 0, variance 0.01, and range [-0.03, 0.03]), and the bias vector is 3-dimensional (value 0.01±0.005). The product of the global semantic vector and the weight matrix is ​​calculated. That is, the score of the first neuron is the sum of the corresponding multiplications of the 300 elements of the global vector and the 300 elements of the first column of the weight matrix, plus the first element of the bias vector. The scores of the second and third neurons are calculated in the same way to obtain three original classification scores (retaining two decimal places), such as 2.1 for qualified, 1.5 for need improvement, and 0.8 for unqualified.

[0065] The original classification scores are processed using a normalized exponential function to obtain the quality assessment probability distribution. Based on this probability distribution, the category corresponding to the highest probability value is selected to obtain the quality assessment result, which includes:

[0066] In this embodiment, the activation function processing and parameter initialization refinement of the feedforward neural network enhance the nonlinear expressive power of semantic features.

[0067] In a preferred embodiment of the present invention, generating a final quality inspection report based on the quality assessment results and a preset quality threshold may include:

[0068] The quality assessment results are compared and analyzed with preset quality level thresholds to obtain quality level classification results. Based on the quality level classification results, combined with the key features of the semantic analysis unit set, a preliminary quality judgment conclusion is generated, specifically including: finding the maximum value among the three original classification scores (e.g., 2.1); subtracting the maximum value from each score to obtain an adjusted score (e.g., 0, -0.6, -1.3); taking the natural exponent (e raised to the power of the adjusted score) on the adjusted score to obtain three index values ​​(e.g., 1.0, 0.5488, 0.2776); and dividing each index value by the three index values. The sum (1.0 + 0.5488 + 0.2776 = 1.8264) is used to obtain the probability of each category (rounded to 4 decimal places, such as 0.5475, 0.3005, 0.1520). The category with the highest probability is selected as the quality assessment result. The preset threshold for pass is probability ≥ 0.7, the threshold for improvement is 0.3 ≤ probability < 0.7, and the threshold for failure is probability < 0.3. The highest probability value in the quality assessment result is compared with the threshold. If it is ≥ 0.7, the grade is pass; if it is between 0.3 and 0.7, it is for improvement; if it is < 0.3, it is failure. The quality grade classification result is obtained.

[0069] The top three words with the highest attention weights in the semantic analysis unit set are extracted as key features. Attention weights are calculated by associating the attention scores of the eight heads of the Transformer encoder with their corresponding words, adding the eight scores, dividing by 8, and rounding the result to three decimal places (e.g., the eight attention scores for "dehydration" are 0.81, 0.83, 0.82, 0.84, 0.80, 0.82, 0.83, and 0.81, totaling 6.56, divided by 8 gives 0.820). Words are sorted by weight from highest to lowest, and the top three are selected (e.g., "dehydration" 0.820, "abnormal sound" 0.750, "unresolved" 0.610). If the level is satisfactory, the conclusion is described as follows: the response handling for the top two key features in customer service dialogues meets quality standards, and the communication related to the third key feature has no obvious defects. The response to dehydration and abnormal noise meets quality standards, and there are no obvious defects in the communication regarding unresolved issues. If the level is "Needs Improvement," the conclusion is that there is room for optimization in the response to the top two key features in customer service conversations, and the efficiency of resolving issues related to the third key feature needs to be improved. For example, there is room for optimization in the response to dehydration and abnormal noise in customer service conversations, and the efficiency of resolving unresolved issues needs to be improved. If the level is "Unsatisfactory," the conclusion is that the response to the top two key features in customer service conversations does not meet quality requirements, and there are obvious defects in the handling of issues related to the third key feature. For example, the response to dehydration and abnormal noise in customer service conversations does not meet quality requirements, and there are obvious defects in the handling of unresolved issues. Through the above methods, a preliminary quality judgment conclusion that clearly points to specific problems is formed.

[0070] The preliminary quality assessment conclusions are correlated with historical quality data to obtain information on quality change trends. Based on this trend information, the preliminary quality assessment conclusions are optimized and adjusted to obtain the final quality assessment conclusions. Specifically, this includes: retrieving quality inspection results for similar semantic analysis units (such as appliance dehydration issues) from the past three months, compiling statistics weekly (12 weeks total), with a weekly sample size ≥ 100; calculating the percentage of items requiring improvement each week (number of samples requiring improvement / total number of samples × 100%, rounded to one decimal place); and calculating the difference between the average percentage of items requiring improvement in the current month and the average of the previous two months. If the difference is > 5% (based on historical fluctuations), further calculations are performed. If the deviation is twice the standard deviation, the trend is deteriorating; if the deviation is less than -5%, the trend is improving; if the deviation is within ±5%, the trend is stable, thus obtaining information on the trend of quality change. If the trend is deteriorating, supplement the preliminary conclusion with recommendations (selected from the database of effective historical rectification measures), such as recommending that the after-sales department strengthen training on handling abnormal washing machine spin-drying problems and add one special drill per week. If the trend is improving, supplement that the current handling process is effective and recommend maintaining and compiling it into standard scripts. If the trend is stable, supplement that the handling effect needs to be continuously monitored and 20 dialogues should be randomly checked for review each month. The adjusted conclusion is the final quality judgment conclusion.

[0071] The final quality assessment conclusions are organized and integrated according to the preset structural specifications to obtain structured quality inspection results. The completeness and consistency of the structured quality inspection results are verified to generate a final quality inspection report, which includes: the preset structural specifications contain five parts; the quality level must be marked with the specific category and corresponding probability (e.g., qualified, probability 0.72); three key features are listed and their corresponding attention weights (retaining two decimal places); the judgment basis cites the semantic aggregation coefficient (e.g., -2.1) and classification probability; trend analysis explains the average and direction of change of the percentage of areas to be improved each week over the past three months; improvement suggestions clearly identify the responsible department (e.g., after-sales training group) and specific measures; verification is performed to check whether any part is missing (if there are no improvement suggestions, add "no special suggestions" and maintain the current process); the logic of trend analysis and judgment basis is checked through automated scripts (e.g., the trend corresponding to the qualified conclusion should not be continuously deteriorating), and 10% of the reports are manually reviewed; after verification, the final quality inspection report is generated.

[0072] This embodiment, by incorporating optimization suggestions from a historical database of effective measures, makes the final conclusions more practical; the standardized structured report and verification mechanism ensure the integrity and logical consistency of the quality inspection results.

[0073] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0074] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0075] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An intelligent quality inspection system based on a natural language model, characterized in that, include: The processing module is used to preprocess the collected raw text data to obtain a normalized text sequence; The mapping module is used to input the normalized text sequence into the pre-trained word embedding model and generate the corresponding set of word embedding vectors through vector space mapping; The analysis module is used to analyze the set of word embedding vectors and construct a semantic topology. Based on semantic topology, a region growing algorithm is used to divide semantic regions, generating a set of semantic analysis units, including: Density clustering analysis is performed on the word embedding vector set to identify the core semantic node set; based on the core semantic node set, a semantic topology structure is constructed; the semantic similarity between nodes in the semantic topology structure is calculated to generate a semantic association weight matrix; Based on the semantic association weight matrix, a region growing algorithm is used to expand the semantic region and form an initial semantic partition; the initial semantic partition is then subjected to boundary optimization to obtain an optimized semantic partition. The optimized semantic partitions are filtered based on a preset semantic density threshold to obtain semantic partitions that meet the density requirements; each semantic partition that meets the density requirements is divided into a semantic analysis unit to generate a set of semantic analysis units. The optimization module is used to construct a vector space polygon by selecting key feature vectors based on the spatial distribution of word vectors within the semantic analysis unit set; calculate the enclosing area of ​​the vector space polygon using the shoelace formula to obtain the semantic distribution area parameter; calculate the semantic aggregation coefficient based on the ratio of the semantic distribution area parameter to the preset benchmark area; and use the semantic aggregation coefficient to perform weighted fusion processing on the word embedding vectors within the semantic analysis unit to generate an optimized semantic feature matrix, including: Based on the spatial coordinates of word vectors within the semantic analysis unit set, the distribution boundary of each word vector is calculated to obtain the boundary feature vector set; according to the spatial positional relationship of the boundary feature vector set, a convex polygon structure is constructed. Based on the vertex coordinates of the convex polygon structure, the semantic distribution area parameter is calculated using the shoelace formula; the semantic distribution area parameter is then compared with the preset benchmark area to obtain the semantic aggregation coefficient. The word embedding vectors within the semantic analysis unit set are weighted using semantic aggregation coefficients to obtain an enhanced word vector set; the enhanced word vector set is then reorganized according to its original dimensions to generate an optimized semantic feature matrix. The evaluation module is used to input the optimized semantic feature matrix into the Transformer-based deep learning classification model, extract deep semantic features, and obtain quality evaluation results. The generation module is used to generate the final quality inspection report based on the quality assessment results and preset quality thresholds.

2. The intelligent quality inspection system based on a natural language model according to claim 1, characterized in that, The normalized text sequence is input into a pre-trained word embedding model, which generates a corresponding set of word embedding vectors through vector space mapping, including: The normalized text sequence is segmented into words to obtain a segmented sequence; the segmented sequence is then input into a pre-trained word embedding model to obtain an initial set of word vectors. The initial word vectors are processed to unify their dimensions to obtain standard-dimensional word vectors; the out-of-vocabulary words in the standard-dimensional word vectors are augmented based on the weighted calculation of adjacent word vectors to obtain a complete set of word vectors; the complete set of word vectors is arranged and combined according to the original text order to generate a set of word embedding vectors.

3. The intelligent quality inspection system based on a natural language model according to claim 2, characterized in that, The normalized text sequence is segmented into words to obtain a segmented sequence; The segmented word sequence is input into a pre-trained word embedding model to obtain an initial set of word vectors, including: The normalized text sequence is processed to obtain purified text data; the purified text data is segmented into words based on a preset dictionary to obtain a basic word sequence; special characters and stop words in the basic word sequence are filtered to obtain an effective word sequence. Each word in the effective word sequence is matched against the vocabulary of the pre-trained word embedding model to obtain the corresponding matching word vector; words that cannot be matched are processed to obtain a combination of segmented word vectors. The segmented word vectors are combined and averaged to obtain the alternative word vectors for out-of-vocabulary words; all matching word vectors and alternative word vectors are integrated to obtain the initial word vector set.

4. The intelligent quality inspection system based on a natural language model according to claim 3, characterized in that, Based on the vertex coordinates of the convex polygon structure, the semantic distribution area parameter is obtained by using the shoelace formula. The semantic distribution area parameter is compared with a preset benchmark area to obtain the semantic aggregation coefficient, including: Based on the ordered vertex coordinate sequence of the convex polygon structure, a coordinate data set is obtained; the sum of the cross products of adjacent coordinates in the coordinate data set is calculated to obtain the original area value; the original area value is processed by absolute value and normalized to obtain the semantic distribution area parameter. The initial aggregation coefficient is obtained by calculating the ratio of the semantic distribution area parameter to the preset benchmark area; the initial aggregation coefficient is then logarithmically transformed to obtain the intermediate aggregation coefficient; the intermediate aggregation coefficient is then verified to obtain the semantic aggregation coefficient.

5. The intelligent quality inspection system based on a natural language model according to claim 4, characterized in that, The optimized semantic feature matrix is ​​input into a Transformer-based deep learning classification model to extract deep semantic features, yielding quality evaluation results, including: The optimized semantic feature matrix is ​​input into the Transformer encoder to calculate the attention-enhanced features; the attention-enhanced features are then subjected to layer normalization to obtain normalized features; the normalized features are then input into the feedforward neural network for nonlinear transformation to obtain the deep semantic code. Deep semantic encoding is processed to obtain stable semantic features; the stable semantic features are input into a global pooling layer for feature compression to obtain a global semantic vector; the global semantic vector is input into a fully connected classification layer for linear transformation to obtain the original classification score. The original classification scores are processed by a normalized exponential function to obtain the quality assessment probability distribution; the category corresponding to the highest probability value is selected based on the quality assessment probability distribution to obtain the quality assessment result.

6. The intelligent quality inspection system based on a natural language model according to claim 5, characterized in that, Based on the quality assessment results and preset quality thresholds, a final quality inspection report is generated, including: The quality assessment results are compared and analyzed with the preset quality level thresholds to obtain the quality level classification results; based on the quality level classification results, combined with the key features of the semantic analysis unit set, a preliminary quality judgment conclusion is generated. The preliminary quality assessment conclusions are correlated with historical quality data to obtain information on quality change trends; the preliminary quality assessment conclusions are then optimized and adjusted based on the information on quality change trends to obtain the final quality assessment conclusions. The final quality judgment conclusions are organized and integrated according to the preset structural specifications to obtain structured quality inspection results; the completeness and consistency of the structured quality inspection results are verified to generate the final quality inspection report.

7. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the system as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the system as described in any one of claims 1 to 6.

Images