Construction engineering name standardized classification model construction method and classification method

Abstract

This invention relates to the field of building information classification technology, specifically providing a method for constructing and classifying a standardized classification model for building project names. The method includes: performing regularization expansion and perturbation on original material samples to generate a candidate enhanced sample set; selecting candidate data with semantic consistency with the original material samples as first training data; inputting the first training data into a first generative large language model, where a linear mapping layer outputs the corresponding predicted category and corresponding confidence level based on the bias term; calculating a first loss value using a cross-entropy loss function, and updating the hidden layer parameters, linear mapping layer, and bias term of the first generative large language model until the first loss value meets the convergence requirement, thus obtaining a standardized classification model. This addresses the problem that standardized classification models in related technologies that introduce machine learning or deep learning only classify a small number of categories based on simple text classification, failing to meet the needs of large-scale multi-class problems in real-world engineering scenarios.

Description

Technical Field

[0001] This invention relates to the field of building information classification technology, and in particular to a method for constructing a standardized classification model for building project names and a classification method thereof. Background Technology

[0002] Currently, in the field of construction engineering, the standardization of naming for material samples and bills of quantities largely relies on manual comparison against industry standard documents, or semi-automated methods based on keyword matching, rule base retrieval, and traditional text standardization classification models. Due to the numerous non-standardized expressions for building materials and bill of quantities items across different regions, companies, and stages (design, bidding, construction, settlement), such as abbreviations, colloquial names, mixed model names, prefixes or suffixes of specifications, mixed unit names, misspellings, and historically inherited names, the same material or bill of quantities item often corresponds to multiple forms of expression, making the alignment process for various construction project names complex and prone to errors. In recent years, although methods have been introduced to automatically classify using machine learning or deep learning standardization classification models, these methods remain largely confined to small-scale category scenarios or simple text classification frameworks, lacking a systematic standard mapping mechanism and verifiable processes. This makes it difficult to meet the needs of large-scale, multi-category, and high-precision standardization in actual engineering scenarios.

[0003] The drawbacks of this type of method are mainly reflected in two aspects: data and model. On the data side, training a name standardization model typically requires high-quality, well-covered labeled samples to establish the correspondence between "non-standard names" and "standard names." However, in real-world engineering environments, there is a lack of uniformly organized, high-quality labeled data; historical data is scattered across different project systems with inconsistent formats and varying labeling quality. Furthermore, architectural project names often contain multi-dimensional attributes such as specifications, performance levels, and construction techniques, resulting in highly non-standard expressions, including typos, abbreviations, and combined descriptions, increasing the difficulty of classification and mapping. In addition, there is currently a lack of effective data augmentation mechanisms to simulate naming variations in real-world scenarios, leading to insufficient generalization ability of the model when faced with unfamiliar new expressions.

[0004] In terms of model training, the number of material and bill of quantities categories in the construction engineering field is enormous, involving hundreds or even thousands of standard items, making it a typical large-scale multi-class classification problem. Existing standardized classification models perform well in scenarios with a small number of categories, but when the category scale grows rapidly, they often suffer from severe category confusion and weak long-tail category recognition capabilities. Furthermore, most methods only output single-category prediction results, lacking confidence assessment and multi-candidate verification mechanisms, making it difficult to form auditable and verifiable decision-making basis in actual engineering management systems. Misjudgments can affect subsequent pricing, procurement, and statistical analysis results. In addition, while some methods based on large-scale pre-trained models perform excellently in semantic matching tasks, their large number of model parameters and high inference costs make them unsuitable for deployment in localized systems of engineering enterprises, making it difficult to balance accuracy and operational efficiency.

[0005] Standardized classification models that incorporate machine learning or deep learning in related technologies, which only classify a small number of categories based on simple text classification, cannot meet the needs of large-scale multi-class problems in real-world engineering scenarios, and no effective solution has yet been proposed. Summary of the Invention

[0006] The present invention provides a method for constructing and classifying a standardized classification model for building project names, which at least solves the problem that standardized classification models that introduce machine learning or deep learning in related technologies only classify small-scale categories based on simple text classification and cannot meet the problem of large-scale multi-category classification in actual engineering scenarios.

[0007] According to one aspect of the present invention, a method for constructing a standardized classification model for building project names is provided. The method includes: performing regularization expansion and perturbation on original material samples to generate a candidate augmented sample set; selecting candidate data consistent with the original semantics of the original material samples as first training data from the candidate augmented sample set through semantic consistency verification; constructing a linear mapping layer and a bias term on a first generative large language model; inputting the first training data into the first generative large language model, and having the linear mapping layer output the corresponding predicted category and corresponding confidence level according to the bias term; calculating a first loss value for the predicted category and the confidence level using a cross-entropy loss function, and updating the hidden layer parameters of the first generative large language model, the weights of the linear mapping layer, and the bias term according to the first loss value until the first loss value meets the convergence requirement, thereby obtaining a standardized classification model.

[0008] As an optional approach, the original material sample is augmented and perturbed according to rules to generate a candidate enhanced sample set. This includes: perturbing the original material sample according to preset rules to obtain rule enhancement operators corresponding to each rule perturbation, and incorporating them into a set of rule enhancement operators. The rule perturbation includes at least one of the following: character insertion, character deletion, character replacement, fragment order rearrangement, noise fragment injection, symbol or format specification transformation; sampling multiple sampling operators from the set of rule enhancement operators, and applying them jointly to the original material sample to obtain multiple candidate enhanced samples, which are then incorporated into the candidate enhanced sample set.

[0009] As an optional approach, semantic consistency verification is used to select candidate data whose original semantics are consistent with the original material sample from the candidate augmentation sample set as the first training data. This includes: calculating the semantic similarity between the candidate augmentation sample and the original material sample according to the text encoding function; inputting the candidate augmentation sample into the teacher model, and having the teacher model output the corresponding predicted category; using the category consistency verification function to calculate the category consistency between the predicted category of the candidate augmentation sample and the category of the original material sample; and confirming that the semantic consistency verification is passed when both the semantic similarity and category consistency between the candidate augmentation sample and the original material sample meet the corresponding preset requirements, and including the candidate augmentation sample in the first training data. The first training data includes the input building project name and the output standard name category.

[0010] As an optional approach, the first training data is input into the first generative large language model, and the linear mapping layer outputs the corresponding predicted category and corresponding confidence score based on the bias term. This includes: inputting the first training data into the first generative large language model of the standardized classification model, and having the first generative large language model output the corresponding global semantic representation vector; inputting the global semantic representation vector into the linear mapping layer of the standardized classification model, mapping it to a predetermined standard category space of thousands, calculating a one-dimensional standard category vector using the bias term, wherein each element in the one-dimensional standard category vector corresponds to the probability of a standard category; normalizing the one-dimensional standard category vector using a normalization function, and calculating the predicted category and corresponding confidence score corresponding to the first training data.

[0011] As an optional approach, before inputting the first training data into the first generative large language model and having the first generative large language model output the corresponding global semantic representation vector, the following steps are taken: fine-tuning the first generative large language model using domain adjustment instructions and classification examples; inputting the first training data into the first generative large language model and having the first generative large language model output the corresponding global semantic representation vector, including: encoding the first training data to generate a word sequence; inputting the word sequence into the first generative large language model to obtain a hidden state sequence; and taking the hidden vector of the last word in the word sequence as the global semantic representation vector.

[0012] As an optional approach, the method further includes: constructing second training data based on the original material samples, wherein the second training data includes input data and output data, the input data being a list of building project names and candidate standard name categories, and the output data being the nth candidate standard name category or rejection in the list; the second training data being divided into multiple training samples based on whether the candidate standard name category list includes the correct category and the number of interference items in the correct category; fine-tuning the second generative large language model using domain adjustment instructions and classification examples; performing staged training on the second generative large language model based on the proportion of different types of training samples in the second training data; calculating the second loss value output by each model based on the sequence-level cross-entropy loss function, and updating the hidden layer parameters of the second generative large language model according to the second loss value until the second loss value meets the convergence requirement, thus obtaining the verification model.

[0013] As an optional approach, the second training data includes four types of training samples: positive samples, difficult-to-positive samples, negative samples, and difficult-to-negative samples. Specifically, the candidate standard name category list for the positive samples contains the correct category, and the number of interference items similar to the correct category is less than a preset number; the candidate standard name category list for the difficult-to-positive samples contains the correct category, and the number of interference items similar to the correct category is greater than a preset number; the candidate standard name category list for the negative samples does not contain the correct category, and the number of interference items similar to the correct category is less than a preset number; the candidate standard name category list for the difficult-to-negative samples does not contain the correct category, and the number of interference items similar to the correct category is greater than a preset number.

[0014] As an optional approach, the second generative large language model is trained in stages based on the proportion of different types of training samples in the second training data. This includes: selecting different types of training samples according to the proportion of the corresponding stages; generating prompt words based on the input data of the training samples and a prompt word template, wherein the prompt word template includes a given sample name, a list of candidate standard categories, finding the standard category in the list that is most relevant to the given sample name, and rejecting the sample if none of them match; inputting the prompt words into the second generative large language model, and having the second generative large language model output the corresponding verification result and verification confidence.

[0015] As an optional approach, the staged training includes three stages, in sequence: a first stage, a second stage, and a third stage; the positive samples in the first stage are the main class; the second stage introduces difficult positive samples; and the third stage introduces negative samples and difficult negative samples, with the ratio of positive samples, difficult positive samples, negative samples, and difficult negative samples being 1:1:1:1.

[0016] As an optional approach, the method further includes: when the confidence level of the output of the standardized classification model is within a second confidence level interval, obtaining multiple standard categories with the highest confidence levels based on the standardized classification model; constructing input data for the verification model based on the multiple standard categories, and inputting it into the verification model to obtain verification categories and verification confidence levels; and determining the final category identification result based on the verification categories and the predicted categories.

[0017] According to another aspect of the present invention, a standardized classification method for building project names is provided, comprising: inputting the building project name to be classified into a standardized classification model to obtain the corresponding predicted category and the corresponding confidence level; wherein, the standardized classification model is obtained by the above-described method for constructing a standardized classification model for building project names; when the confidence level is within a first confidence level interval, taking the output predicted category as the final category identification result; when the confidence level is within a second confidence level interval, performing a verification through a verification model to determine the final category identification result, wherein the confidence level of the second confidence level interval is less than that of the first confidence level interval, and the verification model is obtained by the above-described method for constructing a standardized classification model for building project names; and when the confidence level is within a third confidence level interval, manually verifying the predicted category and the corresponding confidence level.

[0018] According to another aspect of the present invention, an electronic device is provided, comprising: a processor, and a memory storing a program, the program including instructions that, when executed by the processor, cause the processor to perform the methods described above.

[0019] According to another aspect of the present invention, a computer program product is provided, comprising a computer program / instructions that, when executed by a processor, implement the method described above.

[0020] The method for constructing a standardized classification model for building project names provided in this invention generates a candidate augmented sample set by regularizing and perturbing the original material samples. Through semantic consistency verification, candidate data with the original semantic consistency of the original material samples is selected from the candidate augmented sample set as the first training data. This innovative data augmentation mechanism, employing both rule-driven and semantic consistency verification, simulates non-standard writing styles such as typos, abbreviations, and inconsistent formats in engineering projects. It expands the training samples without requiring extensive new annotations, improving the model's robustness to noise and long-tail representations.

[0021] A linear mapping layer and bias terms are constructed on a first generative large language model. The first training data is input into the first generative large language model, and the linear mapping layer outputs the corresponding predicted category and confidence level based on the bias terms. A first loss value for the predicted category and confidence level is calculated using the cross-entropy loss function, and the hidden layer parameters, linear mapping layer, and bias terms of the first generative large language model are updated based on the first loss value until the first loss value meets the convergence requirement, resulting in a standardized classification model. Based on the generative large language model, combined with the linear mapping layer and bias terms, the semantic vector representation is projected onto a predefined standard name category space through the linear mapping layer, achieving a unified score for all candidate categories and outputting the confidence level of the corresponding category. This approach retains the semantic understanding capability of a large model while also considering the computational efficiency of large-scale category mapping, effectively integrating generative and discriminative capabilities. It solves the problem that standardized classification models that introduce machine learning or deep learning in related technologies only classify a small number of categories based on simple text classification, which cannot meet the needs of large-scale multi-class problems in practical engineering scenarios. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating a method for constructing a standardized classification model for building project names, according to an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the standardized classification model architecture of an embodiment of the present invention.

[0025] Figure 3 This is a flowchart of a standardized classification method for building project names according to an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of the electronic device of the present invention. Detailed Implementation

[0027] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the invention. It should be understood that the accompanying drawings and embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the invention.

[0028] To address the aforementioned technical challenges, this embodiment provides a standardization method for material and bill of quantities names in the construction engineering field. Based on a unified data preprocessing and sample augmentation mechanism, it introduces a lightweight semantic matching and standardized classification model suitable for large-scale category scenarios. Combined with confidence assessment and multi-candidate verification strategies, it achieves high-precision mapping and automatic verification of non-standard names. By establishing a stable and scalable name standardization framework, the consistency and manageability of engineering data can be improved, laying a reliable data foundation for subsequent cost analysis, intelligent auditing, and digital management.

[0029] Embodiments of the present invention provide a method for constructing a standardized classification model for building project names, comprising:

[0030] Step S101: Regularize and perturb the original material sample to generate a candidate enhancement sample set;

[0031] Step S102: Through semantic consistency verification, select candidate data that is consistent with the original semantics of the original material sample from the candidate enhancement sample set as the first training data.

[0032] Step S103: Construct a linear mapping layer and bias terms on the first generative large language model;

[0033] Step S104: Input the first training data into the first generative large language model, and the linear mapping layer outputs the corresponding predicted category and the corresponding confidence level according to the bias term;

[0034] Step S105: Calculate the first loss value of the predicted category and confidence using the cross-entropy loss function, and update the hidden layer parameters, weights of the linear mapping layer and bias terms of the first generative large language model according to the first loss value until the first loss value meets the convergence requirement, thus obtaining the standardized classification model.

[0035] The above steps can be performed by computers, processors, servers, etc., and can interface with devices storing or retrieving original material samples. This allows for amplification and training based on the original material samples to obtain a standardized classification model. This standardized classification model can then be used to determine standardized classification names for the names of building projects to be classified.

[0036] The above method generates a candidate augmentation sample set by regularizing and perturbing the original material samples. Through semantic consistency verification, candidate data with the original semantics consistent with the original material samples is selected from the candidate augmentation sample set as the first training data. This innovative data augmentation mechanism, which combines rule-driven and semantic consistency verification, simulates non-standard writing styles such as typos, abbreviations, and formatting errors in engineering. It expands the training samples without requiring a large number of new annotations, improving the model's robustness to noise and long-tailed representations.

[0037] A linear mapping layer and a bias term are constructed on the first generative large language model. The first training data is input into the first generative large language model, and the linear mapping layer outputs the corresponding predicted category and the corresponding confidence level according to the bias term. The first loss value of the predicted category and confidence level is calculated by the cross-entropy loss function, and the hidden layer parameters, linear mapping layer and bias term of the first generative large language model are updated according to the first loss value until the first loss value meets the convergence requirement, and a standardized classification model is obtained.

[0038] Based on a generative large language model, this method combines a linear mapping layer and a bias term to project semantic vector representations onto a predefined standard name category space. This enables a unified score for all candidate categories and outputs the confidence score for each category. It retains the semantic understanding capabilities of a large model while also considering the computational efficiency of large-scale category mapping, effectively integrating generative and discriminative capabilities. This addresses the shortcomings of standardized classification models that incorporate machine learning or deep learning, which only classify a small number of categories based on simple text classification and cannot meet the needs of large-scale multi-class problems in real-world engineering scenarios.

[0039] The aforementioned original material samples may have manually labeled, non-standard names, and corresponding standard name data pairs. For example, "C30 concrete" and its corresponding "C30 concrete".

[0040] Regularization expansion and perturbation, also known as simulating non-standard writing such as typos, abbreviations, and random ordering, perturbs and expands the original material sample. For example, after perturbing "C30 concrete", we can get "C30 concrete soil", "C30 concrete", "C30 cement", etc., all of which correspond to the same standard name "C30 concrete". In this way, we can obtain multiple candidate enhanced samples.

[0041] These candidate augmented samples, after perturbation, may undergo semantic changes, leading to significant deviations in standard name category recognition. This results in errors or inconsistencies in the correspondence between non-standard and standard names, reducing sample accuracy and consequently affecting subsequent model training. Therefore, this embodiment requires semantic consistency judgment to select suitable candidate augmented samples as the first training data, ensuring the accuracy of the first training data.

[0042] At the output of the first generative large language model, a linear mapping layer and its corresponding bias term are constructed. The input dimension of the linear mapping layer is consistent with the dimension of the global semantic representation vector output by the first generative large language model, and the output dimension is consistent with the preset number of standard categories. The global semantic representation vector is input into the linear mapping layer, and after weighted operation and superposition with the bias term, a score vector in the standard category space is generated, realizing the mapping from semantic representation to standard categories. The semantic vector is then projected onto a standard space of thousands of categories.

[0043] The confidence level mentioned above is obtained by normalizing the score vector, which is the probability value of the predicted category and can characterize the credibility of the corresponding predicted category.

[0044] The aforementioned cross-entropy loss is a standard loss for polyphenol-type tasks. It calculates the first loss value for the predicted class and confidence level using the cross-entropy loss function. This first loss value characterizes the difference between the predicted class output by the standardized classification model and the true standard class. Through backpropagation, the first loss value simultaneously updates the hidden layer parameters, weights of the linear mapping layer, and bias terms of the first generative large language model, allowing the model to gradually reduce classification error during iterations until the loss converges, completing the training of the standardized classification model.

[0045] By regularizing and perturbing the original material samples, and combining semantic consistency verification to screen training data, the model can effectively expand the expressive diversity of training samples while keeping the original categories unchanged, and improve the robustness of the model to non-standard naming. By constructing a linear mapping layer and bias term on the basis of the generative large language model, and using cross-entropy loss for supervised training, the model can achieve high-precision classification prediction in a large-scale standard class space, while ensuring stable model training and fast convergence.

[0046] As an alternative approach, the original material samples are subjected to regularization augmentation and perturbation to generate a candidate enhanced sample set, including:

[0047] The original material sample is perturbed according to the preset rules, and the rule enhancement operators corresponding to each rule perturbation are obtained and included in the rule enhancement operator set. The rule perturbation includes at least one of the following: character insertion, character deletion, character replacement, fragment order rearrangement, noise fragment injection, symbol or format specification transformation.

[0048] Multiple sampling operators are sampled from the set of rule-enhancing operators and applied jointly to the original material sample to obtain multiple candidate enhancement samples, which are then included in the candidate enhancement sample set.

[0049] In the field of construction engineering, construction project names, such as material names and bill of quantities names, typically include specifications, units of measurement, brand / model, process descriptions, and synonymous abbreviations. Furthermore, real-world business data commonly exhibits non-standard practices such as typos, missing characters, disordered order, mixed use of symbols, and noisy prefixes and suffixes. To reduce overfitting of the model to "standard writing forms" and improve robustness to noisy naming, this embodiment introduces a rule-driven data augmentation component during the training phase: each original name undergoes multi-strategy perturbation and expansion to generate several semantically equivalent or nearly equivalent variant samples. A semantic consistency verification mechanism is then used to filter out augmented results with semantic shifts, thereby constructing training data covering a broader expression space.

[0050] The overall process can be summarized as follows: name normalization → rule perturbation generation → semantic consistency verification → data storage. Rule perturbation includes character-level perturbation (addition, deletion, and modification), word / fragment rearrangement, symbol and format replacement, and controlled noise fragment injection; semantic consistency verification ensures that the augmented samples and the original samples remain consistent in the sense of "standard category," avoiding the introduction of label noise.

[0051] Specifically, let the original training dataset be... , where x i Indicates non-standard name text, y i This represents the corresponding standard name category, and N is the number of data points in the dataset. For each sample x... i Construct a set of rule-enhancing operators:

[0052] ,

[0053] Inserting enhancement operators for characters Remove enhancement operators for characters Enhanced operators for character replacement Enhancement operator for fragment order rearrangement Inject enhancement operators into noisy segments, Enhance operators for symbol or format specification transformations.

[0054] For any input The enhancement process samples k operators from the set of enhancement operators and then applies them jointly to the data x. Typically, k=1, denoted as... , For T1-T k The j-th operator in the middle, its For T j The sampling distribution of the operator can be set empirically with different perturbation intensities and probabilities, and the combination length can be adjusted accordingly. Controlling the reinforcement strength, in the formula This is the function composition operator, pronounced "circle". It indicates that multiple functions / operators are applied sequentially from right to left. The complete meaning of the above formula is: first, apply the first enhancement operator T1 to the original data x to obtain T1(x); then apply the second enhancement operator T2 to the result of T1(x) to obtain T2(T1(x)); and so on, until the k-th enhancement operator T is applied. k Ultimately, the enhanced data was obtained. .

[0055] For each sample Generate a set of candidate augmentation samples:

[0056]

[0057] By employing various rule perturbations such as character insertion, deletion, replacement, fragment rearrangement, noise injection, and format transformation, the model can realistically simulate common non-standard expressions in engineering scenarios, such as typos, misuse of abbreviations, disordered specifications, and inconsistent formats. This makes the enhanced samples more closely resemble actual business data and significantly improves the model's generalization ability to noisy text.

[0058] As an optional approach, semantic consistency verification is used to select candidate data from the candidate augmentation sample set that are semantically consistent with the original material sample as the first training data, including:

[0059] The semantic similarity between the candidate enhanced sample and the original material sample is calculated based on the text encoding function;

[0060] The candidate augmentation samples are input into the teacher model, and the teacher model outputs the corresponding predicted category.

[0061] The class consistency verification function is used to calculate the class consistency between the predicted class of the candidate augmented sample and the class of the original material sample;

[0062] If the semantic similarity and category consistency between the candidate augmented sample and the original material sample both meet the corresponding preset requirements, the semantic consistency verification is confirmed, and the candidate augmented sample is included in the first training data. The first training data includes the input building project name and the output standard name category.

[0063] To avoid introducing semantic shifts or incorrect labels during the enhancement process, this embodiment introduces a semantic consistency verification mechanism after enhancement, retaining only samples that maintain consistency with the original semantics.

[0064] First, through the text encoding function Calculate the semantic similarity between the augmented sample and the original sample. When the semantic similarity is not lower than a preset threshold First, it is assumed that the semantics remain stable; second, a category consistency verification function is introduced. It is required to enhance the category of the sample after it has been predicted by the teacher model. satisfy Or its confidence level is higher than the set threshold , A threshold is preset for class consistency; when the confidence level of the teacher's model prediction is higher than this value, the class is considered consistent.

[0065] The teacher model is a mature and standardized classification model that has been trained and whose classification results are reliable enough. It is specifically used to verify whether the sample categories after data augmentation are correct.

[0066] Based on the above dual constraints, a verification function is defined. I[] is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. A threshold is preset for semantic similarity. Augmented samples are only included in the augmented dataset when the validation function is 1, resulting in the augmented training set.

[0067]

[0068] The original labeled training dataset, This is the enhanced training set.

[0069] By combining rule perturbation with semantic verification, this embodiment significantly expands the sample expression space while ensuring label consistency, effectively improving the model's ability to identify misspellings, deformations, and long-tailed expressions, and providing stable and high-quality data support for subsequent large-scale standardized classification models.

[0070] By employing a dual verification mechanism of semantic similarity and category consistency, we ensure that augmented samples and original samples are consistent in semantics and standard categories, avoiding label noise introduced by data augmentation, guaranteeing the accuracy and reliability of training data, and providing a high-quality, highly consistent training foundation for standardized classification models.

[0071] As an optional approach, the first training data is input into the first generative large language model, and the linear mapping layer outputs the corresponding predicted category and the corresponding confidence level based on the bias term. This includes: inputting the first training data into the first generative large language model of the standardized classification model, and the first generative large language model outputting the corresponding global semantic representation vector; inputting the global semantic representation vector into the linear mapping layer of the standardized classification model, mapping it to a predetermined standard category space of thousands, and calculating a one-dimensional standard category vector using the bias term, wherein each element in the one-dimensional standard category vector corresponds to the probability of a standard category; normalizing the one-dimensional standard category vector using a normalization function, and calculating the predicted category and the corresponding confidence level corresponding to the first training data.

[0072] The standardized classification model is a unified semantic understanding and classification mapping framework built upon the first generative large language model. Specifically, the input non-standard names are first segmented and encoded, and then fed into the pre-trained first generative large language model to obtain their corresponding global semantic representation vectors. Subsequently, this semantic representation is projected onto a predefined standard category space through a linear mapping layer to achieve large-scale classification prediction for thousands of categories.

[0073] The first generative large language model outputs the probability distribution of each candidate category and selects the category with the highest probability as the prediction result. This probability value is also output as the confidence score for subsequent hierarchical decision-making and verification. During the training phase, supervised learning is used to optimize the model, enabling it to maintain stable discriminative ability in complex expressions, long-tailed categories, and high-similarity category scenarios. By combining the semantic understanding capability of the generative large language model with a discriminative classification structure, this normalization classification model can achieve high-precision and robust name normalization mapping in a large-scale category space.

[0074] Specifically, let the input non-standard name be a string ( First, it is converted into a sequence of word tokens through word segmentation and encoder. The data is then input into a pre-trained generative language model to obtain the corresponding hidden state sequence. ,in The latent vector representation of the last token serves as the global semantic representation vector for the name. Let be a d-dimensional real vector space.

[0075] Then a linear mapping layer is constructed. With bias term , for The 3D real vector space maps semantic vectors to a predefined standard category space, with a total of 3D categories. Typically numbering in the thousands, this yields a category logits vector, specifically a one-dimensional vector where each element represents the probability of a category. These probabilities are unnormalized and are normalized using the softmax normalization function below to become true probabilities. :

[0076]

[0077] Calculate the predicted probability distribution for each category using the Softmax function:

[0078]

[0079] And select the category corresponding to the highest probability. As a standardized prediction result, its probability value is also used. As the classification confidence output, zc and zk refer to the data at the c-th and k-th positions in z, respectively, which are actually a certain category. For example, in z=(0.1,0.1,0.2,0.2,0.3,0.1), if c=1, zc=0.1; if c=2, zc=0.1; if c=3, zc=0.2, and so on for k.

[0080] To improve the discrimination ability in large-scale category scenarios, cross-entropy loss function is used for supervised optimization during the training phase:

[0081]

[0082] in, For cross-entropy loss, This provides true standard category labels. By leveraging the powerful semantic understanding capabilities of generative large models and the efficient category discrimination capabilities of discriminative mapping structures, stable and accurate standard name pre-determination is achieved in a space of thousands of categories, providing a reliable probabilistic basis for subsequent confidence-based hierarchical verification mechanisms.

[0083] By mapping the global semantic representation vector to a standard class space of thousands and normalizing the output class probabilities through Softmax, we can achieve efficient and stable large-scale multi-class classification while maintaining the strong semantic understanding ability of generative models. The output results have clear confidence, which is convenient for subsequent hierarchical decision-making and verification.

[0084] As an optional approach, before inputting the first training data into the first generative large language model and having the first generative large language model output the corresponding global semantic representation vector, the following steps are taken: fine-tuning the first generative large language model using domain adjustment instructions and classification examples; inputting the first training data into the first generative large language model and having the first generative large language model output the corresponding global semantic representation vector, including: encoding the first training data to generate a word sequence; inputting the word sequence into the first generative large language model to obtain a hidden state sequence; and taking the hidden vector of the last word in the word sequence as the global semantic representation vector.

[0085] Instruction fine-tuning involves using natural language instructions combined with task examples to enable the first generative large language model to learn to perform specific vertical domain tasks as needed.

[0086] For example, a natural language instruction could be: "This is the name of a building material; please map it to a standard name." Input a structured instruction template into the model, clearly defining the task objective: Input: Non-standard project name; Output: Standard name + confidence level.

[0087] By using labeled data from the construction engineering field for supervised fine-tuning, the large language model can be made to understand industry terminology, abbreviations, colloquialisms, typos, specifications, models, processes, and units, and can distinguish highly similar material / list items, thereby transforming the general large model into a vertical domain-specific model.

[0088] The role of instruction fine-tuning is to improve domain semantic understanding, enhance the ability to distinguish large-scale categories, reduce misjudgment of similar categories, and make the model output stable, controllable, and in line with business specifications.

[0089] The first generative large language model processes the input training data into a global semantic representation vector as follows: First, it segments and encodes the non-standard input names, for example: "C30 concrete pump" is segmented into "C30 / concrete / pump". Encoding: The text is converted into a numerical vector so that the large language model can process it.

[0090] The input is fed into the pre-trained first generative large language model to obtain its corresponding global semantic representation vector. That is, the processed text is input into the large model Qwen-2.5-3B-Instruct. After reading it, the large model will output a global semantic representation vector, which contains the complete meaning of the sentence: what material, what specifications, what process, and what purpose.

[0091] The global semantic representation vector is then projected onto a predefined standard class space through a linear mapping layer. This linear mapping can be understood as a classifier or a neural network layer, with the standard class space containing a list / database of standard names (thousands of classes). By inputting the global semantic representation vector into the linear mapping layer, the layer maps it to a specific class from the standard names. This enables large-scale classification prediction for thousands of classes.

[0092] Traditional models can only classify dozens of categories, but by combining large model semantic representation and linear mapping layers, we can easily classify into thousands of standard names with relatively high accuracy.

[0093] There are many types of generative large language models, and they can be selected through experimental evaluation. For example, the classification module was evaluated on general text classification datasets (IMDB, SST-2, Yelp, DBpedia, TREC, MultiNLI, SNLI) and self-made domain datasets (QiandanCD, CailiaoCD), and the experimental results are shown in Table 1.

[0094] Table 1. Evaluation results of various base models on different general text classification datasets.

[0095]

[0096] Based on the results in Table 1 and experimental observations, the following conclusions can be drawn:

[0097] The baseline model (Qwen2.5 series Instruct models) without task fine-tuning has already demonstrated strong performance on general text classification datasets. For example, on standard datasets such as IMDB, SST-2, Yelp, DBpedia, and TREC, the classification accuracy steadily improves as the model size increases from 3B to 32B, reaching a maximum of 97.3% and 96.2% on tasks such as DBpedia and TREC, respectively. This indicates that large-scale general-purpose language models already possess good semantic understanding and strong generalization capabilities in common text classification tasks.

[0098] However, on the self-made large-scale category classification datasets QiandanCD and CailiaoCD, its performance is significantly lower than that of general datasets. For example, the accuracy of the 3B model on these two datasets is only 75.2% and 73.3%, respectively. Even when expanded to the 32B model, the performance only improves to 83.5% and 84.9%. This indicates that although the general pre-trained model has good general semantic understanding capabilities, its domain knowledge and discriminative ability are still insufficient when facing classification tasks with larger category scales, finer semantic boundaries, and obvious domain characteristics. It is difficult to directly adapt to large-scale category classification scenarios in vertical domains.

[0099] Building upon this foundation, this embodiment selects the Qwen2.5-3B-Instruct baseline model as the training base and fine-tunes it for classification tasks. Experimental results show that the fine-tuned model (GC) achieves significant improvements across all datasets. On general text classification datasets, the model performance is further enhanced, reaching 96.7%, 97.4%, 98.1%, 99.0%, and 97.5% on IMDB, SST-2, Yelp, DBpedia, and TREC datasets, respectively, all exceeding the original baseline model. Simultaneously, it also improves to 92.3% and 93.4% on the NLI tasks MultiNLI and SNLI, respectively. More significantly, on the self-made domain datasets QiandanCD and CailiaoCD, the model accuracy reaches 91.3% and 94.7%, respectively, representing improvements of over 16% and 21% compared to the un-fine-tuned 3B baseline model. This demonstrates that the fine-tuning strategy significantly enhances the model's discriminative ability in complex category spaces within vertical domains.

[0100] By fine-tuning the instructions, the generative large language model is adapted to the naming characteristics of the construction engineering field, thereby improving the domain semantic understanding ability. By using the last token latent vector as the global semantic representation, the overall semantic information of the name can be fully captured, enabling the model to maintain stable representation ability when faced with non-standard names such as abbreviations, variants, and mixed specifications.

[0101] As an optional approach, the method further includes: constructing second training data based on the original material samples, wherein the second training data includes input data and output data, the input data being a list of building project names and candidate standard name categories, and the output data being the nth candidate standard name category or rejection in the list; the second training data being divided into multiple training samples based on whether the candidate standard name category list includes the correct category and the number of interference items in the correct category; fine-tuning the second generative large language model using domain adjustment instructions and classification examples; performing staged training on the second generative large language model based on the proportion of different types of training samples in the second training data; calculating the second loss value of each model output based on the sequence-level cross-entropy loss function, and updating the hidden layer parameters of the second generative large language model according to the second loss value until the second loss value meets the convergence requirement, thus obtaining the verification model.

[0102] The verification model uses a second generative large language model as its base and is independent of the standardized classification model. It can use an open-source generative model that is isomorphic to the standardized classification model (such as Qwen-2.5-3B-Instruct) or other generative language models with semantic reasoning capabilities.

[0103] The verification model is based on a generative decoder-only architecture. It learns by fine-tuning instructions to complete fine-grained semantic comparison, optimal category selection and rejection judgment within a limited candidate set.

[0104] During training, a second set of training data is first constructed, assuming the standard class set is... Each category Corresponding to a standard name text The original labeled dataset is represented as follows: ,in For the name to be standardized, This step aims to unify the training samples of the validation model into a "candidate constraint decision" form, that is, to construct each sample as an input. With output The input consists of "name to be standardized + list of candidate standard names", and the output is the candidate labels to be selected, thus obtaining the training set for validating the model. M represents the total number of all constructed "candidate constraint judgment" samples, which is also the total number of the second training data. The training set mentioned above... To form second training data for the training samples.

[0105] Specifically, for any sample data The fixed number of candidates is (For example Construct candidate set And form a verification model input. :

[0106]

[0107] Where j = 1 - K. Supervision labels use selective labeling: if... Then let Indicates the position of the correct category in the candidate list; if Then let This indicates that there is no correct answer in the candidate set, thus explicitly training the model's ability to reject answers.

[0108] The method of fine-tuning instructions is similar to that of the first generative large language model, with the main purpose of improving the general model's ability to better handle specific architectural domains.

[0109] Training is conducted in stages, from easy to difficult, to make the model training process more stable and converge more efficiently, avoiding model training oscillations or discrimination failures caused by large differences in sample difficulty.

[0110] Cross-entropy loss optimizes generative output sequences, directly supervises the model to output correct candidate numbers or rejection results, maximizes the likelihood probability of the model for the target output sequence, and makes the model output more in line with task constraints; at the same time, it can effectively suppress irrelevant outputs, improve the consistency, controllability and accuracy of the verification results, and ensure the reliability of the two-stage verification decision.

[0111] Staged training ensures a reasonable learning path and strong model robustness; cross-entropy loss ensures accurate output and stable decision-making. Together, these two techniques enable the verification model to accurately identify optimal candidates and reliably reject unmatched items in complex engineering naming scenarios, significantly reducing the overall system's classification error rate.

[0112] By constructing a second training dataset that includes candidate constraints and rejection capabilities, and by employing staged training and sequence-level cross-entropy loss to optimize the verification model, the model is equipped with the ability to perform fine-grained semantic comparison, optimal selection, and rejection when there is no match in the limited candidate list, which significantly improves the verification accuracy of low-confidence samples and the overall reliability of the system.

[0113] Specifically, given the name to be standardized With the set of candidate standard names The model needs to output either the best-matching candidate class or the rejection label "NONE". Therefore, an input cue is constructed for each piece of second training data. With target output Model learning conditional probabilities:

[0114]

[0115] Supervised learning is achieved by maximizing the log-likelihood of the true output sequence, thereby enabling the model to perform semantic alignment, contrastive reasoning, and rejection judgment within a limited candidate space.

[0116] In terms of specific training data organization, this embodiment utilizes four constructed sample sets: positive sample set... Difficult Correction Sample Set Negative Sample Set With the difficult sample set The training difficulty and capability coverage are controlled using a stratified sample + dynamic allocation method. The mixed training distribution is defined as follows:

[0117]

[0118] in and This invention examines training effectiveness by setting different ratios, including: using only positive samples (…). ); Only use positive + negative ( (), Only positive + difficult positive; positive + negative + difficult positive; positive + negative + difficult negative; and positive + negative + difficult positive + difficult negative combinations are used.

[0119] Experimental and engineering verifications show that using positive examples alone can easily lead to the model forcing selection when candidates are missing, thus lacking the ability to reject. Adding only negative examples can improve rejection, but it is still prone to misjudgment under high similarity interference. After adding difficult positive / difficult negative examples, the model's ability to characterize semantic boundaries and fine-grained differences is significantly enhanced, and the overall robustness is higher.

[0120] Regarding the training objective, this embodiment employs the sequence-level cross-entropy loss of standard SFT for each sample. (in Minimize the negative log-likelihood for the output token sequence:

[0121]

[0122] θ represents the trainable parameters of the second generative large language model. When a correct answer exists among the candidates, the target output is the index or standard name text of the correct candidate; when no correct answer exists among the candidates, the target output is rejection (NONE), thus enabling the model to explicitly learn the decision boundary between "optional" and "non-optional" during the training phase. The corresponding prompt word template can be set as follows:

[0123] Given sample name: Candidate criterion category list: Please find the standard category in the candidate list that is most relevant to the sample name; if none match, please output "Rejection".

[0124] This suggestion constrains the task to a comparison and selection problem within the candidate set, prompting the model to form a controllable reasoning path of "multiple candidate comparison → selection / rejection".

[0125] In terms of training strategy, this embodiment adopts a phased training approach to gradually improve capabilities: the first phase mainly uses positive examples (and a small number of negative examples) to enable the model to quickly learn basic candidate matching and output formats; the second phase introduces difficult positive examples to strengthen the model's fine-grained semantic discrimination and comparative reasoning capabilities under "highly similar candidate interference"; the third phase introduces negative examples and difficult negative examples to ensure that the model can maintain stable rejection even in scenarios where "candidates are missing but the overall candidates are highly similar," avoiding mistaking similar categories as the correct answer. The above phased process can be seen as gradually increasing the proportion of difficult samples in the training set, i.e., making... The training sessions increase progressively, thus achieving a learning effect that progresses from easy to difficult.

[0126] Based on the results of comparing various ratios, the optimal fine-tuning strategy determined by this invention is to use four types of samples simultaneously and adopt a balanced ratio, namely:

[0127]

[0128] Under this strategy, the second generative large language model can accurately select the correct answer when candidates include it, correctly output rejection when candidates are missing, and maintain stronger stability and generalization ability under high similarity interference. Through the aforementioned contrast-enhanced SFT (Supervised Fine-Tuning) fine-tuning method, this embodiment significantly improves the adaptability of the validation model to complex candidate constraint judgment tasks without changing the model architecture, providing reliable semantic decision-making support for subsequent confidence level review.

[0129] As an optional approach, the second training data specifically includes four types of training samples: positive samples, hard positive samples, negative samples, and hard negative samples. Specifically, the candidate standard name category list for positive samples contains the correct category, and the number of interference items similar to the correct category is lower than a preset number; the candidate standard name category list for hard positive samples contains the correct category, and the number of interference items similar to the correct category is higher than a preset number; the candidate standard name category list for negative samples does not contain the correct category, and the number of interference items similar to the correct category is lower than a preset number; and the candidate standard name category list for hard negative samples does not contain the correct category, and the number of interference items similar to the correct category is higher than a preset number.

[0130] This embodiment constructs four categories of data—"positive / difficult positive / negative / difficult negative"—and characterizes their difficulty. It defines a name similarity function. (This can be achieved using edit distance normalization, Jaccard similarity coefficient, or vector cosine similarity, etc.), and the difficulty is measured by the maximum interference intensity in the candidate set.

[0131]

[0132] Given threshold (For example ),when When candidate interference is considered to be relatively small (normal), The candidate sample is considered to have high similarity interference (difficulty). Based on this criterion, this invention constructs four types of training samples respectively.

[0133] Firstly, positive examples require that the candidate list contains the correct category and that the remaining candidates have low similarity to the correct category. In other words, the number of distractors similar to the correct category is less than a preset number, which can be expressed as:

[0134]

[0135]

[0136]

[0137] And marked This is the basic semantic matching capability used to train the model.

[0138] Secondly, Hard Positive examples require that the candidate list contains the correct class and several distractor classes that are highly similar to the correct class; that is, the number of distractor items similar to the correct class is higher than a preset number. ,For example:

[0139]

[0140] in Used for injecting highly similar interference terms. Used to fill in the number of candidates, still marked This is to enhance the model's fine-grained semantic analysis and comparative reasoning capabilities.

[0141] Third, negative examples require that the candidate list does not contain the correct category and that the candidates are significantly different from the correct category. In other words, the number of distractors similar to the correct category is less than a preset number. This can be represented as:

[0142]

[0143] And marked This is used to train the model to avoid "forced wrong selection" when candidates are missing.

[0144] Fourth, Hard Negative examples require that the candidate list does not contain the correct category but is highly similar to the correct category overall. In other words, the number of distractors similar to the correct category is higher than a preset number. This can be represented as:

[0145]

[0146] Same label This is used to improve the model's ability to stably reject and maintain output reliability even in environments with high similarity interference.

[0147] To better fit the online "Top-(K)" candidate review scenario, this invention sorts and perturbs the candidate list: first, the candidates are sorted in descending order according to the similarity between the name to be standardized and the candidate standard name.

[0148]

[0149] This makes the training input distribution closer to the candidate set generated by the first-stage classifier in the actual system; secondly, it can be based on probability. Slightly swapping adjacent candidates avoids the model relying on candidate position bias rather than semantic judgment; at the same time, price-based perturbations such as symbol unification and space / bracket normalization can be applied to the candidate standard names to improve the model's robustness to format differences.

[0150] Finally, by setting a reasonable ratio for the four types of samples (such as...) Furthermore, the sampling weights for long-tail categories are increased to obtain results with sufficient coverage and clear difficulty stratification. This system enhances the verification model's ability to perform fine-grained semantic analysis, comparative reasoning, and rejection judgment under multiple candidate conditions, providing high-quality data support for the accurate verification of low-confidence samples.

[0151] By systematically constructing four types of samples—positive, hard positive, negative, and hard negative—the verification model can gradually learn fine-grained discrimination and stable rejection capabilities under high similarity interference from basic matching, effectively solving the problems of traditional models being prone to forced classification and confusion of highly similar categories.

[0152] As an optional approach, the second generative large language model is trained in stages by varying the proportions of different types of training samples in the second training data. This includes: selecting different types of training samples according to the proportions of the corresponding stages; generating prompt words based on the input data of the training samples and a prompt word template, wherein the prompt word template includes a given sample name, a list of candidate standard categories, identifying the standard category in the list that is most relevant to the given sample name, and rejecting the sample if none of them match; inputting the prompt words into the second generative large language model, which then outputs the corresponding verification result and verification confidence level.

[0153] By standardizing the task format through prompt word templates, the review model can perform candidate comparison and selection judgments according to standardized logic, resulting in stable, controllable, and interpretable outputs. Combined with segmented training of samples of different difficulty, the model's robustness in complex engineering name scenarios is further improved.

[0154] As an optional approach, phased training includes three phases, in sequence: Phase 1, Phase 2, and Phase 3. In Phase 1, positive samples are the main class; in Phase 2, difficult positive samples are introduced; and in Phase 3, negative samples and difficult negative samples are introduced, with the ratio of positive samples, difficult positive samples, negative samples, and difficult negative samples being 1:1:1:1.

[0155] By introducing positive, difficult positive, negative, and difficult negative samples in stages, from easy to difficult, the model first learns the basic matching rules and then gradually strengthens its fine-grained discrimination ability under high similarity interference and its rejection ability when there is no match. The training process is more stable and converges more efficiently, avoiding model training oscillations or discrimination failures caused by large differences in sample difficulty.

[0156] The model employs a three-stage course-based training approach with a balanced 1:1:1:1 ratio of four types of samples, gradually improving its discrimination ability from easy to difficult. While ensuring basic matching accuracy, it significantly enhances the discrimination stability and rejection reliability in highly similar interference scenarios, making the verification results more in line with actual engineering needs.

[0157] As an alternative approach, the method further includes: when the confidence level of the output of the standardized classification model is in the second confidence level interval, obtaining multiple standard categories with the highest confidence levels based on the standardized classification model; constructing input data for the verification model based on the multiple standard categories, and inputting it into the verification model to obtain the verification category and verification confidence level; and determining the final category recognition result based on the verification category and the predicted category.

[0158] Figure 2 This is a schematic diagram of the standardized classification model architecture of an embodiment of the present invention, such as... Figure 2 As shown, by triggering a review mechanism based on confidence level, while ensuring efficient output of high-confidence samples, a refined secondary verification is performed on medium-confidence samples to reduce the risk of misjudgment in large-scale classification. Based on Top-K candidates, a constraint review is performed, which significantly improves the accuracy and reliability of the standardization results while controlling the amount of computation.

[0159] Based on the output of the standardized classification model, a confidence-driven hierarchical decision-making mechanism is introduced, modeling the overall standardization process as a two-stage risk control problem. Let the first stage be the standardization classification model's response to the input name... Output category probability distribution ,in The prediction category is: .

[0160] The highest probability is used as the confidence level of the sample. The verification model is first based on Stratify samples according to a preset threshold: when It was determined to be a high-confidence sample at that time. For high confidence thresholds, the system directly outputs the standard name. To reduce overall inference overhead and ensure high throughput; when When the sample is determined to be uncertain, a two-stage semantic review is triggered. For low confidence threshold; when (Optional strategy) If a sample is identified as a low-confidence, high-risk sample, it can be directly entered into the manual review / rollback rule base, thereby avoiding unreliable output when there is insufficient information.

[0161] When the second-stage review is triggered, the review model selects a Top-K candidate class set based on the probability distribution of the first stage to constrain the search space, thereby reducing the original size of the search space. The decision problem (involving thousands of classes) is compressed into a candidate constraint selection problem. Specifically, let:

[0162]

[0163] And construct a list of candidate standard names ( Then, the generative validation model, fine-tuned with S4, is invoked. The input is standardized as "name to be standardized + list of candidate standard names", and the model outputs the index of the best matching candidate. Or, the condition for rejecting the label NONE is generated as follows:

[0164]

[0165] like For a given candidate index, the two-stage prediction category is: Output standard name ;like This indicates that if there is no sufficiently matching standard name in the candidate set, the system will enter a risk fallback path (such as manual review or expansion of the candidate set) to prevent forced output of incorrect results when there is high similarity interference or missing candidates.

[0166] To further enhance the controllability and interpretability of the two-stage decision-making process, this embodiment introduces a "verification confidence" and "consistency judgment" mechanism under candidate constraints, forming a more stringent risk control. Let the maximum probability of the output distribution of the generative validation model be the two-stage confidence:

[0167]

[0168] when and If the system accepts the two-stage result, it will either trigger a rollback (REVIEW) or expand the candidate set (e.g., by adding...). (Expand from 10 to 20) and then review. Meanwhile, to reduce unnecessary two-phase calls, an optional "model consistency" rule can be introduced: if the category selected in the second phase... Compared with the first stage prediction If they match, output and record "Reviewed and approved" directly; if they do not match, arbitrate based on two-stage confidence levels, for example, using weighted fusion scoring.

[0169]

[0170] in Indicates the two-stage approach to candidate categories The normalized support (which can be obtained by the probability mapping of the output index by the generative model) is ultimately selected as follows:

[0171]

[0172] The "conflict samples" are then marked as high-risk samples for subsequent data backflow and threshold calibration.

[0173] Through the above mechanism, the verification model achieves a two-stage verification process of "high-confidence rapid pass + medium-confidence semantic verification + low-confidence risk fallback": the first stage is responsible for efficiently generating candidate and preliminary judgment results in a large-scale category space, and the second stage performs stronger semantic comparison and rejection judgment within a small candidate set, thereby significantly reducing the risk of misclassification. This design improves system reliability while keeping overall computational overhead under control, and provides traceable evidence for each decision by outputting confidence levels, candidate sets, and verification results, enhancing the system's interpretability and engineering feasibility.

[0174] Figure 3 This is a flowchart of a standardized classification method for building project names according to an embodiment of the present invention, such as... Figure 3 As shown, another invention in this embodiment also provides a standardized classification method for building project names, including:

[0175] Step S301: Input the name of the building project to be classified into the standardized classification model to obtain the corresponding predicted category and the corresponding confidence level.

[0176] Step S302: If the confidence level is within the first confidence level interval, the output predicted category is taken as the final category recognition result.

[0177] Step S303: When the confidence level is within the second confidence level interval, the final category recognition result is determined by verifying the model. The confidence level of the second confidence level interval is lower than that of the first confidence level interval.

[0178] Step S304: If the confidence level is in the third confidence level range, the predicted category and the corresponding confidence level are manually reviewed.

[0179] By adopting a three-level decision-making mechanism of "high confidence direct output, medium confidence model verification, and low confidence manual review", the system achieves an optimal balance between automation efficiency and result reliability. This not only improves the processing speed of standardized project names, but also effectively avoids misclassification from entering the business process, thereby improving the system's usability and security.

[0180] Embodiments of the present invention also provide a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a method according to an embodiment of the present invention.

[0181] Embodiments of the present invention also provide a computer program product, including a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform the method of an embodiment of the present invention.

[0182] Embodiments of the present invention also provide an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform the method of the embodiments of the present invention.

[0183] refer to Figure 4 The present invention will now be described in the form of a structural block diagram of an electronic device that can serve as an embodiment of the present invention, which is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0184] like Figure 4 As shown, the electronic device includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 402 or a computer program loaded from a storage unit 408 into a random access memory (RAM) 403. The RAM 403 may also store various programs and data required for the operation of the electronic device. The computing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0185] Multiple components in the electronic device are connected to I / O interface 405, including: input unit 406, output unit 407, storage unit 408, and communication unit 409. Input unit 406 can be any type of device capable of inputting information into the electronic device. Input unit 406 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 407 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 408 may include, but is not limited to, disks and optical discs. Communication unit 409 allows the electronic device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, and / or wireless communication transceivers, such as Bluetooth devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0186] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, CPUs, graphics processing units (GPUs), various special-purpose artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention may be implemented as a computer program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program may be loaded and / or installed on an electronic device via ROM 402 and / or communication unit 409. In some embodiments, the computing unit 401 may be configured to perform the methods described above by any other suitable means (e.g., by means of firmware).

[0187] Computer programs for implementing the methods of embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0188] In the context of embodiments of the present invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, or infrared systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0189] It should be noted that the term "comprising" and its variations used in the embodiments of the present invention are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "a plurality" mentioned in the embodiments of the present invention are illustrative and not restrictive, and those skilled in the art should understand that unless explicitly indicated otherwise in the context, they should be understood as "one or more".

[0190] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this invention are subject to strict compliance with relevant laws, regulations, and regulatory requirements in their collection, storage, use, processing, transmission, provision, and disclosure, and adhere to the principles of legality, legitimacy, necessity, and good faith. The acquisition of relevant information and data is premised on the user's explicit consent or other legitimate reasons, and a clear and convenient authorization management approach is provided to the user, allowing the user to independently choose to consent, withdraw consent, or refuse to provide relevant information. For functions that rely on user information, if the user does not authorize or withdraws authorization, the corresponding technical function cannot be implemented, and the technical solution of this invention is not applicable in this scenario.

[0191] The steps described in the method embodiments provided by the present invention can be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of protection of the present invention is not limited in this respect.

[0192] The term "embodiment" in this specification refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of the invention. The appearance of this phrase in various places in the specification does not necessarily imply the same embodiment, nor does it imply independence or alternativeity from other embodiments. The various embodiments in this specification are described in a related manner, with reference to each other for similar or identical parts. In particular, for apparatus, device, and system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, and relevant details are referred to in the description of the method embodiments.

[0193] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for constructing a standardized classification model for building project names, characterized in that, include: The original material samples are augmented and perturbed in a regular manner to generate a set of candidate enhanced samples; Through semantic consistency verification, candidate data that is consistent with the original semantics of the original material sample is selected as the first training data from the candidate enhanced sample set. Construct a linear mapping layer and bias terms on the first generative large language model; The first training data is input into the first generative large language model, and the linear mapping layer outputs the corresponding predicted category and the corresponding confidence level according to the bias term. The first loss value of the predicted category and the confidence level is calculated using the cross-entropy loss function. The hidden layer parameters of the first generative large language model, the weights of the linear mapping layer, and the bias term are updated based on the first loss value until the first loss value meets the convergence requirement, thus obtaining a standardized classification model.

2. The method according to claim 1, characterized in that, The original material samples are augmented and perturbed to generate a candidate enhanced sample set, including: The original material sample is perturbed according to preset rules to obtain rule enhancement operators corresponding to each rule perturbation, which are then included in the rule enhancement operator set. The rule perturbation includes at least one of the following: character insertion, character deletion, character replacement, fragment order rearrangement, noise fragment injection, symbol or format specification transformation. Multiple sampling operators are sampled from the set of rule enhancement operators and applied together to the original material sample to obtain multiple candidate enhancement samples, which are then included in the candidate enhancement sample set.

3. The method according to claim 1, characterized in that, Through semantic consistency verification, candidate data that is semantically consistent with the original material sample is selected from the candidate augmentation sample set as the first training data, including: The semantic similarity between the candidate augmented sample and the original material sample is calculated based on the text encoding function; The candidate augmentation samples are input into the teacher model, and the teacher model outputs the corresponding predicted category. The class consistency between the predicted class of the candidate augmented sample and the class of the original material sample is calculated using a class consistency verification function. If the semantic similarity and category consistency between the candidate augmented sample and the original material sample both meet the corresponding preset requirements, the semantic consistency verification is confirmed, and the candidate augmented sample is included in the first training data. The first training data includes the input building project name and the output standard name category.

4. The method according to claim 1, characterized in that, The first training data is input into the first generative large language model, and the linear mapping layer outputs the corresponding predicted category and the corresponding confidence level based on the bias term, including: The first training data is input into the first generative large language model of the standardized classification model, and the first generative large language model outputs the corresponding global semantic representation vector. The global semantic representation vector is input into the linear mapping layer of the standardized classification model and mapped to a predetermined standard category space of thousands. A one-dimensional standard category vector is calculated using the bias term, wherein each element in the one-dimensional standard category vector corresponds to the probability of a standard category. The one-dimensional standard category vector is normalized using a normalization function, and the predicted category and corresponding confidence level of the first training data are calculated.

5. The method according to claim 4, characterized in that, Before inputting the first training data into the first generative large language model, and before the first generative large language model outputs the corresponding global semantic representation vector, the following steps are included: The first generative large language model is fine-tuned using domain-adjusting instructions and classification examples. The first training data is input into the first generative large language model, and the first generative large language model outputs the corresponding global semantic representation vector, including: The first training data is encoded to generate a word sequence; The word sequence is input into the first generative large language model to obtain the hidden state sequence; The latent vector of the last word in the word sequence is taken as the global semantic representation vector.

6. The method according to claim 1, characterized in that, The method further includes: Based on the original material samples, a second training data is constructed, wherein the second training data includes input data and output data. The input data is the name of the building project and a list of candidate standard name categories, and the output data is the nth candidate standard name category or rejection in the list. The second training data is divided into multiple training samples according to whether the candidate standard name category list includes the correct category and the number of interference items in the correct category. By using domain-adjusting instructions and classification examples, the second generative large language model is fine-tuned. The second generative large language model is trained in stages by using the proportion of different types of training samples in the second training data. Based on the sequence-level cross-entropy loss function, the second loss value of each model output is calculated, and the hidden layer parameters of the second generative large language model are updated according to the second loss value until the second loss value meets the convergence requirement, thus obtaining the kernel model.

7. The method according to claim 6, characterized in that, The second training data includes four types of training samples: positive samples, hard positive samples, negative samples, and hard negative samples. The candidate standard name category list of the positive sample contains the correct category, and the number of interference items similar to the correct category is less than a preset number. The candidate standard name category list of the difficult-to-correct sample contains the correct category, and the number of interference items similar to the correct category is higher than the preset number; The candidate standard name category list of the negative sample does not contain the correct category, and the number of interference items similar to the correct category is lower than the preset number; The candidate standard name category list for the difficult-to-bear samples does not contain the correct category, and the number of interference items similar to the correct category is higher than the preset number.

8. The method according to claim 6, characterized in that, The second generative large language model is trained in stages based on the proportion of different types of training samples in the second training data, including: Different types of training samples are selected based on the proportion of the corresponding stage. Based on the input data of the training samples, and combined with the prompt word template, prompt words are generated. The prompt word template includes a given sample name, a list of candidate standard categories, and the selection of the standard category most relevant to the given sample name from the list of candidate standard categories. If none of them match, the sample is rejected. The prompt word is input into the second generative large language model, which then outputs the corresponding verification result and verification confidence level.

9. The method according to claim 8, characterized in that, The phased training includes three phases, which are: Phase 1, Phase 2, and Phase 3 in sequence. The positive examples in the first phase are the primary class; The second stage introduces difficult-to-correct examples; The third stage introduces negative samples and hard negative samples, and the ratio of positive samples, hard positive samples, negative samples, and hard negative samples is 1:1:1:

1.

10. The method according to claim 6, characterized in that, The method further includes: When the confidence level of the output of the standardized classification model is within the second confidence level interval, the highest confidence level standard categories are obtained based on the standardized classification model. The input data for constructing the verification model is based on the multiple standard categories, and then input into the verification model to obtain the verification category and verification confidence level; The final category identification result is determined based on the verification category and the predicted category.

11. A standardized classification method for the names of construction projects, characterized in that, include: The name of the building project to be classified is input into the standardized classification model to obtain the corresponding predicted category and the corresponding confidence level; wherein, the standardized classification model is obtained according to the construction method of the standardized classification model of the building project name as described in any one of claims 1 to 5; If the confidence level is within the first confidence level range, the output predicted category will be used as the final category identification result. When the confidence level is within the second confidence level interval, a verification model is used to verify the final category identification result, wherein the confidence level of the second confidence level interval is less than that of the first confidence level interval, and the verification model is obtained according to the construction method of the standardized classification model of building project names as described in any one of claims 6 to 10; If the confidence level is within the third confidence level range, the predicted category and the corresponding confidence level will be manually reviewed.

12. An electronic device, comprising: A processor and a memory storing a program, characterized in that the program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 11.

13. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the method of any one of claims 1 to 11.

Images