Textile quality risk identification and evaluation method based on natural language evaluation

Abstract

The application belongs to the field of textile quality risk analysis, and discloses a textile quality risk identification and evaluation method based on natural language evaluation, which comprises the following steps: first, establishing a textile quality risk identification domain knowledge base; then, using a large language model combined with the textile quality risk identification domain knowledge base to extract features from textile quality risk evaluation text data and obtain feature indexes; finally, determining the weights of the textile quality risk indexes, calculating the textile quality risk score according to the extracted feature indexes, dividing the risk level according to the risk score, and outputting the risk evaluation result; the textile quality risk identification domain knowledge base comprises a textile quality risk feature dictionary having a mapping relationship between textile quality risk indexes and textile quality risk feature words, as well as an emotional word dictionary and a degree adverb dictionary. The application can further realize the evaluation of the quality risk of textiles, and has the advantages of low cost, fast detection speed, and potential risk exploration.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of textile quality risk analysis, and particularly relates to a textile quality risk identification and evaluation method based on natural language evaluation. BACKGROUND

[0002] Textile products are an important part of daily life, and their quality is related to the consumer's use experience and health and safety. Traditional quality detection methods mainly rely on laboratory testing and expert evaluation, which can provide objective quality data, but it is difficult to reflect the feelings and problems of consumers in actual use.

[0003] With the development of the Internet and the Internet of Things, online text data has become an important basis for evaluating product quality. However, current textile quality risk assessment mainly relies on traditional quality inspection methods, which has obvious limitations:

[0004] First, it relies on human experience, and quality inspectors assess risks through sampling detection and subjective judgment, which is not only inefficient, but also easily affected by individual experience differences;

[0005] Second, structured data analysis is limited. Existing quality management systems are mostly based on numerical detection indicators such as color fastness and breaking strength, and it is difficult to effectively process unstructured text data such as consumer reviews and social media;

[0006] Third, the semantic association in the field of textile quality risk is weak, and existing technologies cannot deeply associate natural language descriptions with industry technical indicators, resulting in poor interpretability of risk assessment results.

[0007] To solve the above problems, a patent with the authorization announcement number CN115187065B discloses a clothing quality analysis system based on an e-commerce platform. The system processes and analyzes the size, material and color of the target clothing sold by the target clothing seller within a predetermined period by extracting the corresponding review information and after-sales information, and then analyzes and judges the quality of the clothing sold by the target clothing seller within a predetermined period. However, the system only considers size, material and color when evaluating the quality of the product based on reviews, and cannot detect other quality problems of clothing products. In addition, in terms of quality evaluation, the system only considers the number of quality keywords in the review as the basis for evaluation, and does not perform sentiment analysis on the review. The accuracy of the risk assessment results obtained by the system is poor.

[0008] Therefore, it is of great significance to develop a textile quality risk identification and evaluation method based on natural language evaluation to solve the above problems. SUMMARY

[0009] The purpose of the present application is to solve the problems existing in the prior art, provide a textile quality risk identification and evaluation method based on natural language evaluation, which aims to collect multi-source text data, use a semantic analysis model for feature extraction and sentiment polarity analysis, establish a textile quality knowledge graph, and calculate a quality risk coefficient, and finally generate a multi-dimensional risk evaluation report. This method can identify and evaluate the quality risk of textiles and provide detection opinions for regulatory departments.

[0010] To achieve the above purpose, the technical solution adopted by the present application is as follows:

[0011] A textile quality risk identification and evaluation method based on natural language evaluation, first establishes a textile quality risk identification domain knowledge base; then uses a large language model combined with the textile quality risk identification domain knowledge base to extract features from textile quality risk evaluation text data, obtains feature indicators, including quality risk features, sentiment words and degree adverbs; finally determines the weight of the textile quality risk index (i.e. weights the textile quality risk index), calculates the textile quality risk score according to the extracted feature indicators, and divides the risk level according to the risk score, and outputs the risk evaluation result;

[0012] The textile quality risk identification domain knowledge base includes a textile quality risk feature dictionary with a mapping relationship between textile quality risk indicators and textile quality risk feature words, as well as a sentiment word dictionary and a degree adverb dictionary. The sentiment word dictionary includes sentiment words and their sentiment word polarity, and the degree adverb dictionary includes degree adverbs and their degree adverb scores.

[0013] The calculation formula of the textile quality risk score is as follows:

[0014] P=S×O×E;

[0015] Where S represents the risk degree score of the textile quality risk indicator to be calculated, S=10×s, s is the weight value of the textile quality risk indicator; O represents the frequency of occurrence of the textile quality risk indicator, O=10×(ia / in), ia is the number of negative text data containing the textile quality risk indicator, and in is the number of text data of the total quality risk text; E value is the negative text sentiment quantization value containing the textile quality risk indicator, E =-5×Ea×Eb, Ea is the sentiment score, Ea=-1 (i.e. Ea= the sum of negative text sentiment quantization values containing the textile quality risk indicator / total negative text), and Eb is the degree adverb score, Eb= the sum of degree adverb scores in negative text containing the textile quality risk indicator / total negative text.

[0016] As a preferred technical solution:

[0017] The textile quality risk index refers to the detection index of the textile detection standards GB 18401-2010, GB 31701-2015, GB 21500-2008, GB 18383-2007, GB 20400-2006, GB 30585-2014, GB 25038-2010 and GB 18587-2001 for detecting textiles, including dyeing fastness, formaldehyde content, odor, breaking strength, etc.

[0018] The extraction process of the textile quality risk feature word is as follows: first, collect the textile quality risk evaluation text data (that is, collect user comment text data from e-commerce platforms, social media and after-sales systems through the Python Scrapy framework and API interface), then use the SimHash algorithm to remove duplicate texts, use the regular expression of the re module in Python to filter special symbols, and delete short texts with a string length less than 5, then use a large language model to assist in identifying invalid comments (including advertisements), and only keep the sentences containing textile quality descriptions to form a corpus; then input the text data in the corpus into the large language model, generate candidate feature words through instructions, find out potential feature words by using the mask prediction function of BERT (specifically: the BERT model is directly called from the HuggingFace Transformers library, and its mechanism is to first mask the candidate feature words in the text and replace them with a special [MASK] mark, then input the masked text into the BERT model, and finally predict the masked word based on the unmasked context), and use the found potential feature words as candidate features, combine the Word2Vec embedding vector, and calculate the relevance of all words in the corpus to the candidate feature words; finally, select the top n candidate feature words with the highest relevance as extended feature words, remove duplicate words, and combine the professional terms in the textile detection standards to manually filter the extended feature words, remove unreasonable extended words, and obtain the textile quality risk feature words.

[0019] The relevance of all words in the corpus to the candidate feature words is calculated by combining the Word2Vec embedding vector, and the specific process is as follows: first, use the corpus and the gensim library in python to train the Word2Vec model, then use the trained Word2Vec model to obtain the vector representation of the candidate feature words, and finally traverse all the words in the corpus, calculate the cosine similarity with the candidate feature words, and sort them in descending order of relevance.

[0020] The textile quality risk feature dictionary with the mapping relationship between the textile quality risk indicators and the textile quality risk characteristic words is obtained by matching the textile quality risk characteristic words and the textile quality risk indicators through manual screening.

[0021] The textile quality risk identification and evaluation method based on natural language evaluation as described above uses a large language model combined with a textile quality risk identification domain knowledge base to extract features from textile quality risk evaluation text data, and the specific process is as follows:

[0022] (1) Langchain is used to read the text in the textile quality risk identification domain knowledge base, and the text is divided into chunks (text blocks). The meaning of chunk is to split long text into small sections to improve the accuracy of search recall;

[0023] (2) The text divided into chunks is processed by embedding vectorization using a text2vec-large-chinese Chinese embedding model, and vector data is obtained. Then, the vector data is stored in the vector database using the FAISS vector database tool;

[0024] (3) A large language model is selected, the textile quality risk identification domain knowledge base is injected, and the instruction fine-tuning technology is adopted to construct a multi-task prompt template containing textile quality feature word extraction, sentiment analysis and degree discrimination. The template content includes extracting textile quality risk characteristic words, textile quality risk indicators, emotional polarity (i.e. emotional score, divided into positive and negative, with scores assigned as 1 and -1), degree adverbs and degree adverb scores from textile quality risk evaluation text data, and requires the large language model to read, match and process text data from the vector database; among them, the degree adverb extraction extracts the degree adverb according to the degree adverb table and outputs the degree score of the degree adverb;

[0025] (4) The multi-task prompt template containing textile quality feature word extraction, sentiment analysis and degree discrimination constructed in step (3) is used to extract features from textile quality risk evaluation text data.

[0026] The textile quality risk identification and evaluation method based on natural language evaluation as described above divides the risk level according to the risk score, specifically: P≥x2 belongs to high risk, x1≤P<x2 belongs to higher risk, P<x1 belongs to general risk, where x2>x1, when the risk score P of the quality risk indicator is within the risk level range, it is considered that the quality risk indicator exists corresponding risk.

[0027] The textile quality risk identification and evaluation method based on natural language evaluation as described above, x1=64, x2=180. Each index is classified into three levels, and is divided according to the risk acceptance degree of the user. In the present application, the risk degree score of the quality risk index is divided into: 1~3 for general risk, 4~5 for higher risk, and 6~10 for high risk. The occurrence frequency of the quality risk index is divided into: 1~3 for general risk, 4~6 for higher risk, and 7~10 for high risk. The text sentiment quantification value of the quality risk index is divided into: 1~3 for general risk, 4~6 for higher risk, and 7~10 for high risk. Here, the interval boundary value of the higher risk is taken as the value of x1 and x2, x1=4×4×4=64, x2=5×6×6=180.

[0028] Advantages:

[0029] (1) The present application extracts the feature indexes of textile quality risk through text data and text processing technology, performs sentiment analysis and degree adverb extraction, and identifies the quality risk of textiles, further realizes the evaluation of the quality risk of textiles, and the textile quality risk identification and evaluation method based on natural language evaluation has the advantages of low cost, fast detection speed and potential risk exploration.

[0030] (2) The present application establishes a textile quality risk identification field knowledge base, uses a large language model to extract features from text data, calculates the textile quality risk score, realizes the evaluation of the textile quality risk level, processes and analyzes the online textile related text data, and further analyzes and judges the product quality risk of online textiles, provides a key sampling direction for the existing supervision and sampling method of the current supervision department, and realizes the analysis of textile quality risk from the natural language field. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 It is a flowchart of the textile quality risk identification and evaluation method based on natural language evaluation of the present application;

[0032] Figure 2 It is a schematic diagram of the large language model prompt template of the present application. DETAILED DESCRIPTION

[0033] The present application will be further described in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present application and not to limit the scope of the present application. In addition, it should be understood that those skilled in the art can make various modifications or modifications to the present application after reading the content taught by the present application, and these equivalent forms also fall within the scope defined by the claims attached hereto.

[0034] A textile quality risk identification and evaluation method based on natural language evaluation, as shown in Figure 1 The steps are as follows:

[0035] (1) Establish a textile quality risk identification field knowledge base;

[0036] The textile quality risk identification field knowledge base includes a textile quality risk feature dictionary with a mapping relationship between textile quality risk indicators and textile quality risk feature words, as well as an emotional word dictionary and a degree adverb dictionary; wherein:

[0037] The textile quality risk indicators refer to the detection indicators for textile detection in the textile detection standards GB 18401—2010, GB 31701—2015, GB 21500—2008, GB 18383—2007, GB 20400—2006, GB 30585-2014, GB 25038-2010 and GB 18587-2001;

[0038] The emotional word dictionary is based on the Dalian University of Technology Chinese Emotional Lexicon Ontology, which selects part of the words in the emotional lexicon ontology and adds new emotional words. The emotional polarity of each emotional word is reclassified to obtain the emotional word dictionary. The positive polarity emotional words are assigned a value of 1, and the negative polarity emotional words are assigned a value of -1;

[0039] The degree adverb dictionary is based on the Hownet degree adverb dictionary. The degree adverbs extracted from the comment data by manual method are normalized according to the scoring method of Hownet degree adverbs, with a value space of (0, 2]. All degree adverbs are divided into very weak (0.5), weak (0.8), strong (1.3), very strong (1.6), extremely strong (1.8), and strongest (2) in total six categories;

[0040] The extraction process of textile quality risk feature words is as follows:

[0041] (a) Collect textile quality risk evaluation text data (i.e. collect user comment text data from e-commerce platforms, social media and after-sales systems through Python Scrapy framework and API interface), then use SimHash algorithm to remove duplicate text, use regular expressions of re module in Python to filter special symbols, and delete short texts with length less than 5, then use large language model to assist in identifying invalid comments, and then determine whether it is a default template content through generative classification, only keep the sentences containing textile quality description to form a corpus, and finally use LTP segmentation tool for text segmentation and part-of-speech tagging;

[0042] (b) The BERT model is directly called from the Hugging Face Transformers library. Then, the tokenizer of the BERT model is used to process complex expressions such as compound words in “severe shrinkage after washing”. The text data processed by the tokenizer is then input into the large language model. Candidate feature words are generated through instructions. The BERT model is then called. The masking prediction function of the BERT model is used to mask the candidate feature words in the text and replace them with special [MASK] tags. The masked text is then input into the BERT model. Finally, the masked words (i.e., potential feature words) are predicted based on the unmasked context, and the found potential feature words are used as candidate features.

[0043] (c) Use the corpus and the gensim library in Python to train the Word2Vec model. Then, use the trained Word2Vec model to obtain the vector representation of the candidate feature words. Finally, traverse all the words in the corpus, calculate the cosine similarity with the candidate feature words, and sort them from high to low relevance.

[0044] (d) Select the n candidate feature words with the highest relevance as extended feature words, remove duplicate words, and then combine them with the professional terminology in the textile testing standards to manually filter the screened extended feature words, remove unreasonable extended words, and obtain textile quality risk feature words.

[0045] Finally, the above-mentioned textile quality risk feature words and textile quality risk indicators were matched by manual screening to establish a feature word table with the affiliation relationship between textile quality risk feature words and textile quality risk indicators. This is the textile quality risk feature dictionary with the mapping relationship between textile quality risk indicators and textile quality risk feature words. The first 5 quality risk indicators and quality risk feature words in the textile quality risk feature dictionary are shown in the table below:

[0046] ;

[0047] (2) Feature extraction was performed on the text data of textile quality risk assessment using a large language model combined with a knowledge base in the field of textile quality risk identification to obtain feature indicators, including quality risk features, sentiment words and degree adverbs. The specific process is as follows:

[0048] (I) Use Langchain to read the text in the knowledge base of textile quality risk identification and divide the text into chunks.

[0049] (II) Using the text2vec-large-chinese Chinese embedding model, the text segmented into chunks is embedded into a vector, and then the vector data is stored in the vector database using the FAISS vector database tool;

[0050] (III) Selecting a large language model GPT architecture, using a RAG enhanced retrieval generation method to inject a textile quality risk identification domain knowledge base, and using a command fine-tuning technology, a multi-task prompt template containing textile quality feature word extraction, sentiment analysis and degree discrimination is constructed as shown in Figure 2 The template content includes extracting textile quality risk feature words, textile quality risk indicators, sentiment polarity (i.e. sentiment score, divided into positive and negative, with scores assigned as 1 and -1 respectively) and degree adverbs and degree adverb scores from textile quality risk evaluation text data, and requiring the large language model to read, match and process the text data from the vector database; wherein:

[0051] Quality attribute feature word identification extracts a candidate feature word set from the natural language text evaluation data of the textile to be analyzed through feature extraction, filters the semantics based on the textile quality risk domain knowledge, extracts the textile quality risk feature words, and outputs the textile quality risk indicators, such as the comment "This T-shirt is [very] easy to <pilling>, (poor quality)", the extracted textile quality risk feature word is "pilling", and the corresponding textile quality risk indicator output by the textile quality risk feature word dictionary is "pilling performance";

[0052] Sentiment polarity analysis classifies each piece of text data according to the sentiment lexicon using a binary classification method, outputs the result of positive sentiment polarity as 1 or the result of negative sentiment polarity as -1, such as the comment "This T-shirt is [very] easy to <pilling>, (poor quality)", the extracted sentiment word is "poor quality", the sentiment polarity is judged to be negative according to the sentiment lexicon, and the output value of the sentiment polarity is "-1";

[0053] Degree adverb extraction extracts degree adverbs according to the degree adverb table and outputs the degree score of the degree adverb, such as the comment "This T-shirt is [very] easy to <pilling>, (poor quality)", the extracted degree adverb is "very", and the output degree score of the degree adverb is "2" according to the degree adverb dictionary;

[0054] (IV) Using the multi-task prompt template containing textile quality feature word extraction, sentiment analysis and degree discrimination constructed in step (III) to extract features from textile quality risk evaluation text data;

[0055] (3) Output the risk assessment result;

[0056] (3.1) Construct a judgment matrix M by AHP method, and weight vectors are calculated by eigenvalue method and consistency test through expert scoring method. After consistency test, the textile quality risk indicators are weighted; wherein:

[0057] The index weight is calculated according to the judgment matrix M, and the maximum eigenvalue λ_max and the corresponding eigenvector A are calculated. The eigenvalue and eigenvector of the judgment matrix M are solved in the form of solving the eigenvalue of the matrix in linear algebra. The eigenvector corresponding to the maximum eigenvalue is the eigenvector A, and the solving equation is M = λ_max A. The eigenvector A is the numerical value of the importance of each evaluation index. The weight value of each index can be obtained by normalization calculation;

[0058] The formula for consistency test is as follows:

[0059] CR = CI / RI;

[0060] CI = (λ_max-n) / (n-1);

[0061] In the formula, n is the dimension of the judgment matrix, CR is the random consistency ratio, CI is the consistency value, and RI is the average random consistency index value. When CR < 0.1, it indicates that the consistency of the judgment matrix is within the reasonable range of the internal structure of the judgment matrix;

[0062] The above method can be used to evaluate the quality risk of different textiles, and the index weight suitable for the product category can be obtained. For example, the risk of "714street" brand woven shorts product reviews on Tmall platform is evaluated. After the above steps, the textile quality risk indicators are "fiber content", "color fastness", "odor" and "bursting strength". According to the index, a judgment matrix is constructed, and the relative importance ratio of each index is given by textile practitioners and industry experts. The specific ratio is 1-9 scale, and the judgment matrix is shown in the following table:

[0063] The index weight is calculated and consistency test is performed, and the judgment matrix is shown in the following table:

[0064] ;

[0065] According to the above judgment matrix M and solving formula, the weight value of the fiber content of the shorts is 0.11, the weight value of the color fastness of the shorts is 0.255, the weight value of the odor of the shorts is 0.569, and the weight value of the bursting strength of the shorts is 0.066;

[0066] At the same time for testing the judgment matrix in the construction of internal logic of rationality, according to the consistency test formula calculation can be known, the above judgment matrix M is 4 order matrix, RI value takes 0.90, the CI of judgment matrix = 0.0195, CR = CI / RI = 0.0217 < 0.1, namely through the consistency test, the weight value of the fiber content of the short pants garment is 0.11, the weight value of the color fastness is 0.255, the weight value of the odor is 0.569, and the weight value of the bursting strength is 0.066;

[0067] (3.2) According to the extracted characteristic index, the textile quality risk score P is calculated; wherein the calculation formula of the textile quality risk score is as follows:

[0068] P=S×O×E;

[0069] Wherein, S represents the risk degree score of the textile quality risk index to be calculated, S = 10 × s, s is the weight value of the textile quality risk index; O represents the occurrence frequency of the textile quality risk index, O = 10 × (ia / in), ia is the number of text data containing the textile quality risk index, and in is the number of text data of the total quality risk text; E value is the text sentiment quantization value containing the textile quality risk index, E = -5 × Ea × Eb, Ea is the sentiment score, Ea = -1, Eb is the degree adverb score, Eb = the sum of the degree adverb scores in the negative text containing the textile quality risk index / the total number of negative texts;

[0070] Through the calculation formula of the above textile quality risk score, the quality risk of the short pants garment in step (3.1) is evaluated, wherein the total negative quality text data amount is 511, the negative text data amount of fiber content is 39, the negative text data amount of color fastness is 384, the negative text data amount of odor is 76, and the negative text data amount of bursting strength is 12, and the results are as follows:

[0071] The degree adverb score Eb of fiber content is 1.1, and the risk score P is 4.62;

[0072] The degree adverb score Eb of color fastness is 1.411, and the risk score P is 135.19;

[0073] The degree adverb score Eb of odor is 0.875, and the risk score P is 37.02;

[0074] The degree adverb score Eb of bursting strength is 1.77, and the risk score P is 1.37;

[0075] (3.3) According to the risk score, the risk level is divided, and the risk assessment result is output; wherein the division of risk level is: P≥180 belongs to high risk, 64≤P<180 belongs to higher risk, P<64 belongs to general risk;

[0076] According to the above risk level division, it can be known from the risk level division of the woven shorts in step (3.2) that the color fastness of the shorts garment belongs to higher risk, and the fiber content, odor and bursting strength belong to general risk, so the supervision department needs to pay special attention to the detection of the "color fastness" index of the textile product;

[0077] And according to the report on product quality supervision and spot check in 2024 released by Zhejiang Market Supervision Administration in 2025, the detection result of 714street shorts garment in e-commerce platform sampling is that the water fastness is unqualified. Comparing the above measured results with the sampling detection results proves that the design of the method is meaningful and effective for the identification and evaluation of textile quality risk.

Claims

1. A textile quality risk identification and evaluation method based on natural language evaluation, characterized by: Firstly, a knowledge base of textile quality risk identification field is established; then, a large language model is combined with the knowledge base of textile quality risk identification field to extract features of textile quality risk evaluation text data, and feature indexes are obtained, including quality risk features, sentiment words and degree adverbs; finally, the weights of textile quality risk indexes are determined, the quality risk score of textiles is calculated according to the extracted feature indexes, the risk level is divided according to the risk score, and the risk evaluation result is output; The knowledge base of the textile quality risk identification field includes a textile quality risk feature dictionary having a mapping relationship between textile quality risk indexes and textile quality risk feature words, and an emotion word dictionary and a degree adverb dictionary; The calculation formula of the textile quality risk score is as follows: P=S×O×E; Wherein, S represents the risk degree score of the textile quality risk index, S=10×s, s is the weight value of the textile quality risk index; O represents the occurrence frequency of the textile quality risk index, O=10×(ia / in), ia is the number of negative text data containing the textile quality risk index, in is the number of total quality risk text data; E value is the negative text emotion quantization value containing the textile quality risk index, E=-5×Ea×Eb, Ea is the emotion score, Ea=-1, Eb is the degree adverb score, Eb=the sum of the degree adverb scores in the negative text containing the textile quality risk index / the total number of negative texts.

2. The natural language evaluation-based textile quality risk identification and evaluation method according to claim 1, characterized in that, The textile quality risk index refers to the detection index for detecting textiles in the textile detection standards GB 18401-2010, GB 31701-2015, GB 21500-2008, GB 18383-2007, GB 20400-2006, GB 30585-2014, GB 25038-2010 and GB 18587-2001.

3. The natural language evaluation-based textile quality risk identification and evaluation method according to claim 2, characterized in that, The extraction process of the textile quality risk feature word is as follows: firstly, the textile quality risk evaluation text data is collected, then the SimHash algorithm is used to remove duplicate texts, the regular expression of the re module in Python is used to filter special symbols, and short texts with a string length less than 5 are deleted, then the large language model is used to assist in identifying invalid comments, and only the sentences containing textile quality descriptions are reserved to form a corpus; then the text data in the corpus is input into the large language model, candidate feature words are generated through instructions, potential feature words are found out by using the mask prediction function of BERT, and the correlation degrees of all words in the corpus and the candidate feature words are calculated by combining the Word2Vec embedding vectors; finally, the n candidate feature words with the highest correlation degrees are selected as extended feature words, and the repeated words are removed, then the extended feature words obtained by screening are manually filtered to remove unreasonable extended words, and the textile quality risk feature words are obtained.

4. The natural language evaluation-based textile quality risk identification and evaluation method according to claim 3, characterized in that, In combination with the Word2Vec embedding vector, the relevance of all words in the corpus to the candidate feature words is calculated. The specific process is as follows: first, use the corpus and the gensim library in python to train the Word2Vec model, second, use the trained Word2Vec model to obtain the vector representation of the candidate feature words, and finally traverse all the words in the corpus, calculate the cosine similarity with the candidate feature words, and sort them in descending order of relevance.

5. The natural language evaluation-based textile quality risk identification and evaluation method according to claim 4, characterized in that, The textile quality risk feature dictionary with the mapping relationship between the textile quality risk indicators and the textile quality risk feature words is The feature word table with the membership relationship between the textile quality risk feature words and the textile quality risk indicators is established by matching the textile quality risk feature words and the textile quality risk indicators through artificial screening.

6. The natural language evaluation-based textile quality risk identification and evaluation method according to claim 5, characterized in that, The textile quality risk evaluation text data is extracted by using a large language model combined with a textile quality risk identification domain knowledge base. The specific process is as follows: (1) Use Langchain to read the text in the textile quality risk identification domain knowledge base, and divide the text into chunks; (2) Use the text2vec-large-chinese Chinese embedding model to perform embedding vectorization processing on the text divided into chunks to obtain vector data; then use the FAISS vector database tool to store the vector data in the vector database; (3) Select a large language model, inject the textile quality risk identification domain knowledge base, and use the instruction fine-tuning technology to build a multi-task prompt template containing textile quality feature word extraction, sentiment analysis and degree discrimination. The template content includes extracting textile quality risk feature words, textile quality risk indicators, sentiment polarity, and degree adverbs and degree adverb scores from textile quality risk evaluation text data, and requires the large language model to read, match and process the text data from the vector database; (4) Use the multi-task prompt template containing textile quality feature word extraction, sentiment analysis and degree discrimination built in step (3) to extract features from the textile quality risk evaluation text data.

7. The natural language evaluation-based textile quality risk identification and evaluation method according to claim 1, characterized in that, According to the risk score, the risk level is divided into: P≥x2 belongs to high risk, x1≤P<x2 belongs to higher risk, P<x1 belongs to general risk, where x2>x1.

8. The natural language evaluation-based textile quality risk identification and evaluation method according to claim 7, characterized in that, x1=64, x2=180.

Images