Product defect information crawling method based on large model refinement

Abstract

The invention discloses a large model refinement-based product defect information crawling method, which comprises the following steps of: firstly, collecting original text data related to product defects from various network channels through a crawler method; secondly, performing fine adjustment on the pre-trained large language model on the collected data set, so that the semantic content related to the defect can be better understood and analyzed; and finally, carrying out deep analysis on original defect data by utilizing the fine-tuned model, and accurately extracting products, defective parts, damage events and various valuable defect feature information. The method gives full play to the advantages of a large language model in semantic comprehension and text analysis, core content related to product defects can be accurately recognized, irrelevant noise information can be effectively filtered out, the accuracy of information extraction is greatly improved, potential defect information hidden in user feedback is mined, and the method is suitable for popularization and application. And an enterprise is helped to find problems in time and carry out targeted improvement.

Description

Technical Field

[0001] The present invention relates to the technical field of information crawling, and in particular to a product defect information crawling method based on large model refinement. Background Art

[0002] In recent years, with the rapid development of the Internet, web information crawling technology has been widely used in the field of product defect information collection. This technology can quickly and efficiently extract data related to product defect information from massive amounts of web information, providing comprehensive and objective data support for risk assessment. With the advent of the Web 2.0 era, more and more information is scattered on dynamic interactive platforms such as social media, forums, and blogs, bringing new challenges to crawling.

[0003] Nowadays, there are many different types of products, and traditional crawling methods cannot effectively identify and filter these redundant information. With the continuous development of machine learning technology, people can use large language models to classify and analyze crawled information data in a more detailed and accurate manner. In recent years, the combination of fine-tuned large models and information crawling technology has achieved good results in product defect information crawling scenarios, but there are also some problems.

[0004] On the one hand, companies usually use traditional web crawling technology to collect complaints and comments related to product defects from various channels. However, with the continuous enrichment of product types and the increasing diversification of information channels, these raw data are filled with a large amount of noise information that is irrelevant to actual defects. Traditional crawling methods are difficult to effectively identify and filter these irrelevant contents, resulting in low efficiency in subsequent information processing and analysis.

[0005] Even if companies are able to collect a large amount of product defect data, they still need to invest a lot of manpower in manual feature extraction and screening. The feature engineering process is time-consuming and labor-intensive, and cannot fully capture the complex patterns and potential correlations hidden in the data. It is impossible to efficiently extract key information that is valuable for product improvement from massive data.

[0006] On the other hand, even if existing technologies can preliminarily identify the main types of defects through methods such as manual recognition, such analysis is usually limited to surface phenomena, and it is difficult to deeply understand the potential causes behind each category. Lacking in-depth semantic analysis and knowledge reasoning capabilities of defect information, it is difficult to summarize targeted product improvement measures. The analysis process is too subjective and empirical, making it difficult to replicate and promote the method to other product areas. Summary of the invention

[0007] The problem to be solved by the present invention is to provide a product defect information crawling method based on large model refinement, to conduct in-depth and refined analysis on the massive product defect information data collected by the enterprise, to discover the key to the product defect information, and to accurately extract valuable defect-related content.

[0008] The present invention adopts the following technical solution: a product defect information crawling method based on large model refinement, comprising the following steps:

[0009] S1. Data collection and annotation: Use web crawler methods to collect text data containing product defect information, annotate the collected text data, and create annotated data sets;

[0010] S2. Data preprocessing: Clean the labeled data set, extract the data content related to product defects, and obtain the product defect data set;

[0011] S3. Model selection and model training: Select a suitable large language model as the baseline model, and use the labeled product defect dataset to train the baseline model;

[0012] S4. Fine-tune the large language model and perform model evaluation and optimization: refine the crawled information, use the trained large language model to extract features from the collected product defect data, and optimize the performance of the large language model in processing product defect-related semantics;

[0013] S5. Deploy the large language model: Deploy the fine-tuned large language model to the application to extract product defect information in real time;

[0014] S6. Monitoring and iteration: Continuously monitor the performance of the large language model, and continuously iterate and optimize the large language model based on feedback and new data from actual applications.

[0015] Preferably, in step S1, a web crawler method is used to collect text data containing product defect information, and the method is as follows:

[0016] S1.1. Determine the target data source: According to the product type and enterprise needs, determine the network channels that need to be crawled, including online customer reviews, complaint platforms, recall notices, and industry reports;

[0017] S1.2. Determine the crawling strategy: Build a crawler program based on the network channels that need to be crawled, determine the crawling rules and parameters, obtain text data containing product defect information, and cover different types of products and defects.

[0018] Furthermore, in step S1, the collected text data is annotated, and the keywords and corresponding categories of each text data containing product defect information are annotated, including: company name, product name, defect problem, injury incident, cause of injury time, product main category, and product sub-category.

[0019] Preferably, the data preprocessing in step S2 includes the following sub-steps:

[0020] S2.1. Clean the labeled data set, remove HTML tags and irrelevant character noise information; convert the text data into a format suitable for model input;

[0021] S2.2, segment the text, divide the text data into sentences and words, extract basic units, and combine each text data and its corresponding annotation information into training samples;

[0022] S2.3. Use the set prompt format to organize the text data according to the keywords and corresponding categories, so that after the text data is input, the product defect information is output according to the set prompt format.

[0023] Among them, the tools for text segmentation are Tokenizer, NLTK or SpaCy.

[0024] Preferably, in step S3, a suitable large language model is selected as the baseline model, and the large language model is GPT-3 or higher version, which can handle text generation and extraction tasks.

[0025] Furthermore, the baseline model is trained using the labeled product defect dataset, and the method includes the following sub-steps:

[0026] S3.1. Set training parameters, including learning rate, batch size, and number of training rounds;

[0027] S3.2. Training model: Use the labeled product defect dataset to train a large language model, allowing the large language model to learn how to extract key information from the text;

[0028] S3.3. Verify the effect: monitor the change in loss during training.

[0029] Preferably, in step S4, the trained large language model is used to perform feature extraction on the collected product defect data to obtain a fine-tuned large language model, and the method includes the following sub-steps:

[0030] S4.1. Conduct context modeling, use the self-attention mechanism of the large language model to analyze the relationship between sentences and words, identify context information, and identify key phrases in sentences through context information;

[0031] S4.2. Perform semantic embedding, convert text data into vector representation, use the embedding layer of the large language model to capture word semantic information, calculate the similarity between the embedding vectors of different words, compare the vector distances of different words, and identify synonyms;

[0032] S4.3, keyword identification: perform feature extraction, identify important words in text data, and use part-of-speech tagging methods to help extract specific types of keywords;

[0033] S4.4. Perform syntactic analysis, use dependency syntactic analysis to determine the grammatical relationship of words, identify key phrases and functions; identify subject-verb-object structures to help extract core information;

[0034] S4.5. Model prompt design: Construct effective prompts to explicitly require the large language model to extract specific types of information; conduct example learning to enable the large language model to learn how to extract key information in similar contexts by providing multiple examples, and let the large language model get used to the extraction format through input-output comparison.

[0035] Further, in step S4, the model evaluation tests the performance of a large language model on a validation set, improves the reliability of the evaluation through a cross-validation method, analyzes the accuracy and consistency of the extraction, evaluates the accuracy and recall of the extraction, monitors the change in loss during the validation process, and adjusts the prompts and training samples according to the evaluation results to optimize the model performance;

[0036] Based on the fine-tuned large language model, we integrate and output the results: generate structured output, output the extracted key information in a structured format, including JSON or CSV format; integrate information, combine multiple extraction results, generate a comprehensive report or summary, and help users obtain key information;

[0037] Preferably, in step S6, the performance of the large language model is continuously monitored, accurate and valuable defect feature information is extracted from the product defect data, support is provided for subsequent product improvement analysis, and user feedback on the extraction results is collected to identify the strengths and weaknesses of the large language model. New product defect data and feedback are integrated into the training set, continuous learning is performed, and the model is regularly updated to adapt to new product requirements.

[0038] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0039] 1. The product defect information crawling method of the present invention utilizes the powerful semantic understanding and text generation capabilities of a large language model to effectively identify and filter noise information that is irrelevant to actual defects, and accurately extract valuable defect-related content from massive data.

[0040] 2. The product defect information crawling method of the present invention can avoid the information redundancy and irrelevant problems existing in traditional crawling methods, greatly improve the efficiency of subsequent analysis and processing, and greatly reduce the cost and time expenditure of manual feature engineering, thereby improving the comprehensiveness and objectivity of information extraction.

[0041] 3. The method of the present invention can be flexibly applied to different types of network channels and supports the collection and analysis of product defect data in multiple languages, greatly enhancing the ability of enterprises to obtain comprehensive defect information. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of the product defect information crawling method of the present invention;

[0043] Figure 2 This is an information crawling flow chart of the product defect information crawling method of the present invention;

[0044] Figure 3 This is a detailed flow chart of the large model of the product defect information crawling method of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the application is further elaborated in detail below in conjunction with the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in the field on this embodiment belong to the protection scope of the present invention. At the same time, for the step numbering in the embodiment of the present invention, it is only set for the convenience of explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0046] In one embodiment of the present invention, the crawled text data of product defect information, such as recall data, feedback information, etc., are effectively refined and analyzed to remove redundant information in the information, providing a basis for subsequent optimization of product design.

[0047] like Figure 1 As shown, the following steps are included:

[0048] S1. Data collection and annotation: determine the target information URL, crawl the information using a web crawler, collect text data containing product defect information, annotate the collected text data, and create annotated data sets;

[0049] S2. Data preprocessing: Clean the labeled data set, extract the data content related to product defects, and obtain the product defect data set;

[0050] S3. Model selection and model training: Select a suitable large language model as the baseline model, and train the baseline model based on the existing annotated product defect dataset;

[0051] S4. Fine-tune the large language model and perform model evaluation and optimization: refine the crawled information, use the trained large language model to extract features from the collected product defect data, and optimize the performance of the large language model in processing product defect-related semantics;

[0052] S5. Deploy the large language model: Deploy the fine-tuned large language model to the application, extract product defect information in real time, and obtain refined product defect data;

[0053] S6. Monitoring and iteration: Continuously monitor the performance of the large language model, and continuously iterate and optimize the large language model based on feedback and new data from actual applications.

[0054] The product defect information crawling method based on the large model refinement in this embodiment is as follows:

[0055] 1. Data Collection

[0056] First, if Figure 2 As shown, according to the needs, the target data source is determined, based on the web crawler method, the crawler strategy is designed and the crawler program is written to collect text data containing product defect information from various channels.

[0057] This data can come from product recall notices, consumer complaints, user feedback, industry reports, etc., ensuring data diversity to cover different types of products and defects.

[0058] 2. Data Annotation

[0059] Label the collected data, identify key information, and create a labeled data set. Each text needs to be labeled with keywords and corresponding categories. Key fields include: company name, product name, defect, injury incident, cause of injury incident caused by defect, product main category, product sub-category.

[0060] For example:

[0061] Company Name: Kunshan XX Lighting Technology Co., Ltd.

[0062] Product name: One-pole XX brand (XGMZ-020) fixed general purpose lamp (LED surface mounted)

[0063] Defect: Metal solder joints

[0064] Injury incident: Electric shock / electric shock

[0065] Cause of injury caused by defect: The metal solder joints of the connector pins can be touched in lamps marked as non-replaceable light sources

[0066] Product main category: furniture and building decoration building materials

[0067] Product categories: Outdoor lighting fixtures and devices

[0068] 3. Data Preprocessing

[0069] Preprocess the labeled data, including: Text cleaning: removing irrelevant information, punctuation marks, and extra spaces.

[0070] Word segmentation and encoding: Convert the labeled text data into a format suitable for model input, usually using Tokenizer for word segmentation and organizing it using a specific prompt format.

[0071] For example: Input: [crawled text]

[0072] Output: Company name: [Company name], Product name: [Product name], Defect: [Defect], Injury incident: [Injury incident], Reason why defect led to injury incident: [Reason], Product main category: [Main category], Product sub-category: [Category].

[0073] 4. Model selection

[0074] Based on the existing refined data, select a suitable large language model as the baseline model to ensure that the model can handle text generation and extraction tasks.

[0075] In particular, this embodiment selects GPT-3 as the large language model, and higher versions are also applicable to the method of the present invention.

[0076] 5. Fine-tune the model

[0077] Use annotated datasets to fine-tune large language models, such as Figure 3 As shown, this embodiment uses the Transformers library of Hugging Face to fine-tune the GPT-3 model. The fine-tuning steps include:

[0078] Load the dataset, load the GPT-3 model and tokenizer tool, perform data preprocessing, and set training parameters; training parameters include: selecting an appropriate learning rate, batch size, number of training rounds, etc.

[0079] Training model: Use labeled target data to train the model so that the model can learn how to extract key information from the text and obtain detailed information about product defects.

[0080] Validation effect: Test the model performance on the validation set and evaluate the extracted accuracy and recall.

[0081] 6. Model Evaluation

[0082] This embodiment uses indicators such as F1-score, accuracy and recall rate to evaluate the model, and the reliability of the evaluation can be improved through the cross-validation method.

[0083] 7. Deploy the model

[0084] Deploy the trained model to the application to extract product defect information in real time. Ensure that the deployment environment can support the operation of the model.

[0085] 8. Monitor and Iterate

[0086] Continuously monitor the performance of the model, and continuously iterate and optimize the model based on feedback and new data from actual applications.

[0087] Preferably, this embodiment refines the crawled information and uses the trained large language model to extract features from the collected product defect data. The specific method is as follows:

[0088] 1.1 Text Preprocessing

[0089] Clean data: remove irrelevant characters, punctuation, and extra spaces to ensure that the text is clean; convert the text to a unified format (such as lowercase) to reduce the impact of uppercase and lowercase on the extraction results.

[0090] Sentence and word segmentation: Split text information into sentences and words for subsequent analysis; use word segmentation tools (such as NLTK or SpaCy) to process the text and extract basic semantic units.

[0091] 1.2. Understanding the context

[0092] Context modeling: Utilize the self-attention mechanism to analyze the relationships between sentences, words, and words in text information, identify important contextual information, and identify the relationship between key phrases in a sentence and their context, which helps to understand their importance; identify the semantic roles of keywords, such as topics, modifiers, etc., through contextual information.

[0093] Semantic embedding: Convert text into vector representation and use the model’s embedding layer to capture the semantic information of words; calculate the similarity between the embedding vectors of different words and compare the vector distances of different words to help identify synonyms and related concepts.

[0094] 1.3 Keyword Identification

[0095] Feature extraction: Identify important words such as nouns, verbs, adjectives, etc. in the text; use part-of-speech tagging (POStagging) to identify nouns, verbs, adjectives, etc. to determine which words are more semantically important.

[0096] For example, nouns are often candidates for features because they often refer to entities or concepts.

[0097] Syntactic analysis: Use dependency parsing to determine the grammatical relationship of words, identify key phrases and their functions; identify subject-verb-object structures to help understand the core information of the text.

[0098] 1.4 Model prompt design

[0099] Construct effective prompts: Design targeted prompts that explicitly require the model to extract specific types of information and guide the model to extract specific information.

[0100] For example: Extract keywords from the following text: {text}

[0101] Output: company name, product name, defects, injury incidents, etc.

[0102] Conduct example learning: Provide multiple examples to let the model learn how to extract key information in similar contexts, and help the model understand the extraction format and requirements; let the model get used to the extraction format through "input-output" comparison.

[0103] For example:

[0104] Input: Kunshan XX Lighting Technology Co., Ltd. recalls a number of XX brand XGMZ-020 lamps.

[0105] Output: Company name: Kunshan XX Lighting Technology Co., Ltd., Product name: Yiji XX brand XGMZ-020 lamp.

[0106] 1.5. Result integration and output

[0107] Generate structured output: Output the extracted key information in a structured format (such as JSON or CSV) to facilitate subsequent analysis and processing.

[0108] Perform information integration: Combine multiple extraction results to generate comprehensive reports or summaries to help users quickly obtain key information.

[0109] 1.6 Training and fine-tuning

[0110] Instance-based fine-tuning: Fine-tune the model using labeled example data to allow the model to learn how to extract key information from new text; monitor the loss changes during training to ensure that the model is learning effectively.

[0111] Evaluation and optimization: Test model performance on the validation set and analyze the accuracy and consistency of extraction; adjust prompts and training samples based on the evaluation results to optimize model performance.

[0112] 1.7 Feedback Mechanism

[0113] User feedback: Collect user feedback on extraction results to identify the strengths and weaknesses of the model; dynamically adjust prompts and models based on feedback.

[0114] Continuous learning: Incorporate new data and feedback into the training set and regularly update the model to adapt to new information.

[0115] Preferably, this embodiment performs prompt fine-tuning on the large model GPT3.5. The following is an example of rule setting and input and output.

[0116] Preset detailed rules, as follows:

[0117] {"role":"system","content":"Assume that you are a named entity recognition model. Now I will give you a text. 1. Please identify the entity names in the sentence according to the examples and requirements, and output them in the form of a list. 2. The output formats must have only three types: [\"defective component i\",\"defective component name\"], [\"consequence i\",\"consequences caused by the defect\".], [\"cause i\",\"consequences caused by the defect\".], none of the three entities can be missing, and cause i must be the cause of consequence i. You need to identify the defective component, the consequences caused, and the reasons for the consequences caused by the component, and each product may have more than one A defect. 3. The entity type generated in each sentence cannot be repeated! 4. If the defects and consequences are not mentioned in the text, just fill in the content as not mentioned. 5. In the answer, only ',' can be used for commas, and the Chinese version of commas ',' cannot be used. Not only should ',' be used between entities, but also within entities. Remember this! 6. If no information about the defects, consequences and causes of small parts is found in the text you provided, the output will be: [\"Defective Part 1\",\"Not Mentioned\"],[\"Consequence 1\",\"Not Mentioned\".],[\"Reason 1\",\"Not Mentioned\".]. 7. Only output the format I requested, without other unnecessary modifications and words! 8. The answer should be short and the summary should be accurate."}.

[0118] The input is as follows:

[0119] {"role":"user","content":"Defect: The braking force of the front brake of the bicycle is too large. When the front brake applies a force of 50N, the braking force is 172N, which exceeds the maximum limit of 120N in the standard GB 14746-2006 "Safety Requirements for Children's Bicycles". Consequence: When the handbrake is used to brake the bicycle in an emergency, the braking force is too large, which can easily cause the vehicle to lose control and cause children to fall and get injured."}.

[0120] The output is as follows:

[0121] {"role":"assistant","content":"[\"Defective component 1\",\"Bicycle front brake\"],[\"Consequence 1\",\"Child fell\"],[\"Cause 1\",\"Excessive braking force during emergency braking of the bicycle\"]"}.

[0122] The input is as follows:

[0123] {"role":"user","content":"Defect: The toe clearance of this children's bicycle (for installing mudguards) is less than 89mm, which does not meet the standard GB 14746-2006 "Safety Requirements for Children's Bicycles". Consequence: When turning, the rider's feet are likely to miss or hit obstacles, which can easily cause falls."}.

[0124] The output is as follows:

[0125] {"role":"assistant","content":"[\"Defective component 1\",\"Toe clearance (for mudguard installation)\"],[\"Consequence 1\",\"Easily causes falls\"],[\"Reason 1\",\"Feet are prone to miss steps or hit obstacles when turning\"]"}.

[0126] The input is as follows:

[0127] {"role":"user","content":"Defect: The bicycle wheel clamping force is insufficient. The minimum disassembly torque of the rear wheel right axle nut is 19.5N·m, which fails to reach 70% of the manufacturer's nominal tightening torque required by the standard GB 14746-2006 "Safety Requirements for Children's Bicycles". Consequence: The wheel clamping force is insufficient. The locking nut of the rear wheel is easy to loosen or fall off when children are riding, which may cause the wheel to loosen or separate and cause children to fall."}.

[0128] The output is as follows:

[0129] {"role":"assistant","content":"[\"Defective component 1\",\"Wheel\"],[\"Consequence 1\",\"Child falls\"],[\"Reason 1\",\"Insufficient wheel clamping force\"],[\"Defective component 2\",\"Rear wheel right axle nut\"],[\"Consequence 2\",\"Wheel is loose\"],[\"Reason 2\",\"The locking nut of the rear wheel is easy to loosen or fall off\"]"}.

[0130] The input is as follows:

[0131] {"role":"user","content":"Defects: After durability testing, the rivet under the right armrest of this stroller fell off, which does not meet the mechanical and physical performance requirements of the standard GB14748-2006 "Safety Requirements for Child Strollers". In addition, the "Warning" information word is not marked on the product manual, which does not meet the safety warning requirements of the standard GB14748-2006 "Safety Requirements for Child Strollers". Consequences: If the rivet under the right armrest of the stroller falls off, it will affect the safe use of the stroller; the lack of warning information words in the instruction manual may cause improper adult supervision and cause danger to children."}.

[0132] The output is as follows:

[0133] {"role":"assistant","content":"[\"Defective component 1\",\"The rivet under the right armrest fell off\"],[\"Consequence 1\",\"Affects the safe use of the stroller\"],[\"Reason 1\",\"Rivet fell off\"],[\"Defective component 2\",\"Instruction manual\"],[\"Consequence 2\",\"Causes danger to children\"],[\"Reason 2\",\"No "Warning" information word is marked, causing improper adult supervision\"]"},

[0134] {"role":"user","content":"1. Dangerous gap (roof) 2. Applicable age for strollers 3. Safety warnings"},

[0135] {"role":"assistant","content":"[\"Defective component 1\",\"Dangerous gap (at the roof)\"],[\"Consequence 1\",\"Not mentioned\"],[\"Reason 1\",\"Not mentioned\"]"},

[0136] {"role":"user","content":""},

[0137] {"role":"assistant","content":"[\"Defective Part 1\",\"Not mentioned\"],[\"Consequence 1\",\"Not mentioned\"],[\"Reason 1\",\"Not mentioned\"]"},

[0138] {"role":"user","content":"Defects: 1. The front fork connection of the product broke after testing; 2. The protective cover cap fell off under a tensile force of 30N, and an exposed protrusion was formed after the screws were tightened; 3. There was no permanent mark on the stem tube indicating the minimum insertion depth; 4. The right handle cover fell off under a tensile force of 65N; 5. The product's logo and instructions for use did not meet the standard requirements. Consequences: 1. The product has insufficient mechanical strength, which may cause the risk of children falling out or falling and getting injured; 2. The protective cover cap is easily pulled off to form an exposed protrusion, which may cause the risk of stabbing or scratching the child's skin; 3. There is no permanent mark on the stem tube indicating the minimum insertion depth, which may cause the risk of children falling forward or falling out and getting injured. 4. The handle cover was pulled off during use, which may cause the exposed handlebar steel pipe wall to scratch the child or the child to fall from the car; 5. The applicable age and weight are not indicated, which may cause children to misuse and cause injuries; Failure to provide warning words and correct assembly instructions may lead to abuse or incorrect assembly, causing accidental injuries to children."},

[0139] {"role":"assistant","content":"[\"Defective component 1\",\"Front fork connection\"],[\"Consequence 1\",\"Children were injured by falling out or falling\"],[\"Reason 1\",\"Insufficient mechanical strength of the product\"],[\"Defective component 2\",\"Protective cover\"],[\"Consequence 2\",\"Stab or scratch children's skin\"],[\"Reason 2\",\"Falling under 30N pulling force, forming exposed protrusions after the screws were tightened\"],[\"Defective component 3\",\"Standbar\"],[\"Consequence 3\",\"Causing children to fall forward and hit or fall out\"],[\"Reason 3\",\"No permanent mark indicating the minimum insertion depth\"],[\"Defective component 4\",\"Handle cover\"],[\"Consequence 4\",\"The exposed steel pipe wall of the handlebar scratches children or children fall from the bike\"],[\"Reason 4\",\"The handle cover is pulled off during use\"],[\"Defective component 5\",\"Product markings and instructions for use\"],[\"Consequence 5\",\"May cause harm to children due to misuse or cause accidental injury to children\"],[\"Reason 5\",\"The applicable age and weight are not marked, and warning words and correct assembly instructions are not provided\"]"}.

[0140] In specific applications, this implementation is based on the defect information data of children's bicycles of various brands collected in a certain city. After being refined based on the large model, it presents richer and more accurate results, as shown in the following table:

[0141]

[0142] In the table, the first column shows the date of announcement information for different products, the second column lists the product names in the product defect information, the third column indicates the name of the manufacturer, the fourth column is the place of origin, the fifth column is the county or county-level city place of origin, the sixth column is the product defect and the consequences caused by the defect, the seventh column is the main category of the product, and the eighth column is the sub-category of the product.

[0143] From these refined data, we can clearly see the specific components, hazards and various product information of each product defect. Compared with the traditional simple crawling method, the traditional method requires a lot of manual resources for feature engineering. The method based on the large model in this embodiment can automatically extract various valuable defect feature information, greatly reducing the manual workload, improving the accuracy of information extraction, and digging out potential defect information hidden in user feedback, which is helpful for enterprises to discover problems in time and make targeted improvements, providing a valuable basis for enterprises to make targeted product improvements, enabling enterprises to obtain insights into product defects more efficiently and economically.

[0144] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A product defect information crawling method based on large model refinement, characterized in that: The steps include: S1. Data collection and annotation: Use web crawler methods to collect text data containing product defect information, annotate the collected text data, and create annotated data sets; S2. Data preprocessing: Clean the labeled data set, extract the data content related to product defects, and obtain the product defect data set; S3. Model selection and model training: Select a suitable large language model as the baseline model, and use the labeled product defect dataset to train the baseline model; S4. Fine-tune the large language model and perform model evaluation and optimization: refine the crawled information, use the trained large language model to extract features from the collected product defect data, and optimize the performance of the large language model in processing product defect-related semantics; S5. Deploy the large language model: Deploy the fine-tuned large language model to the application to extract product defect information in real time; S6. Monitoring and iteration: Continuously monitor the performance of the large language model, and continuously iterate and optimize the large language model based on feedback and new data from actual applications.

2. The product defect information crawling method based on large model refinement according to claim 1 is characterized in that: In step S1, a web crawler method is used to collect text data containing product defect information, and the method is as follows: S1.

1. Determine the target data source: According to the product type and enterprise needs, determine the network channels that need to be crawled, including online customer reviews, complaint platforms, recall notices, and industry reports; S1.

2. Determine the crawling strategy: Build a crawler program based on the network channels that need to be crawled, determine the crawling rules and parameters, obtain text data containing product defect information, and cover different types of products and defects.

3. The product defect information crawling method based on large model refinement according to claim 2 is characterized in that: In step S1, the collected text data is annotated, and the keywords and corresponding categories of each text data containing product defect information are annotated, including: company name, product name, defect problem, injury incident, cause of injury time, product main category, and product sub-category.

4. The product defect information crawling method based on large model refinement according to claim 1 is characterized in that: The data preprocessing in step S2 includes the following sub-steps: S2.

1. Clean the labeled data set, remove HTML tags and irrelevant character noise information; convert the text data into a format suitable for model input; S2.2, segment the text, divide the text data into sentences and words, extract basic units, and combine each text data and its corresponding annotation information into training samples; S2.

3. Use the set prompt format to organize the text data according to the keywords and corresponding categories, so that after the text data is input, the product defect information is output according to the set prompt format.

5. The product defect information crawling method based on large model refinement according to claim 4 is characterized in that: The tools for word segmentation of text are Tokenizer, NLTK or SpaCy.

6. The product defect information crawling method based on large model refinement according to claim 1 is characterized in that: In step S3, a suitable large language model is selected as the baseline model. The large language model is GPT-3 or higher, which can handle text generation and extraction tasks.

7. The product defect information crawling method based on large model refinement according to claim 6 is characterized in that: The baseline model is trained using the labeled product defect dataset. The method includes the following sub-steps: S3.

1. Set training parameters, including learning rate, batch size, and number of training rounds; S3.

2. Training model: Use the labeled product defect dataset to train a large language model, allowing the large language model to learn how to extract key information from the text; S3.

3. Verify the effect: monitor the change in loss during training.

8. The product defect information crawling method based on large model refinement according to claim 7 is characterized in that: In step S4, the trained large language model is used to extract features from the collected product defect data to obtain a fine-tuned large language model. The method includes the following sub-steps: S4.

1. Conduct context modeling, use the self-attention mechanism of the large language model to analyze the relationship between sentences and words, identify context information, and identify key phrases in sentences through context information; S4.

2. Perform semantic embedding, convert text data into vector representation, use the embedding layer of the large language model to capture word semantic information, calculate the similarity between the embedding vectors of different words, compare the vector distances of different words, and identify synonyms; S4.3, keyword identification: perform feature extraction, identify important words in text data, and use part-of-speech tagging methods to help extract specific types of keywords; S4.

4. Perform syntactic analysis, use dependency parsing to determine the grammatical relationships of words, and identify key phrases and functions; Identify subject-verb-object structures to help extract core information; S4.

5. Model prompt design: construct effective prompts that explicitly ask the large language model to extract specific types of information; Perform example learning by providing multiple examples so that the large language model can learn how to extract key information in similar contexts, and let the large language model get used to the extraction format through input-output comparison.

9. The product defect information crawling method based on large model refinement according to claim 8 is characterized in that: The model evaluation described in step S4 tests the performance of a large language model on a validation set, improves the reliability of the evaluation through a cross-validation method, analyzes the accuracy and consistency of extraction, evaluates the accuracy and recall of extraction, monitors the change in loss during the validation process, and adjusts the prompts and training samples according to the evaluation results to optimize the model performance; Based on the fine-tuned large-scale language model, the results are integrated and output: structured output is generated, and the extracted key information is output in a structured format, including JSON format or CSV format; information is integrated, and multiple extraction results are combined to generate a comprehensive report or summary to help users obtain key information.

10. The product defect information crawling method based on large model refinement according to claim 1 is characterized in that: In step S6, the performance of the large language model is continuously monitored to extract accurate and valuable defect feature information from the product defect data to provide support for subsequent product improvement analysis, and user feedback on the extraction results is collected. By identifying the strengths and weaknesses of the large language model, new product defect data and feedback are integrated into the training set, and continuous learning is carried out to update the model to adapt to new product requirements.

Images