An engineering material intelligent inquiry matching method and system based on natural language processing

Abstract

The application discloses an engineering material intelligent inquiry matching method and system based on natural language processing. The method comprises the following steps: obtaining an inquiry request text input by a user; performing named entity recognition by using a pre-trained language model in the field of engineering cost to extract key entities such as material name and specification model; performing semantic normalization processing through a domain ontology knowledge base; encoding the normalized conditions into vectors and combining structured fields to perform hybrid retrieval; performing adaptive statistical outlier filtering on candidate prices; calculating a confidence score according to the data volume, timeliness, semantic matching degree and data source label by using Bayesian confidence estimation, and outputting a recommended result in the form of a price interval; and receiving user feedback to continuously optimize the system. The application realizes semantic-level understanding through field customization NLP technology, overcomes the lexical gap problem, provides multi-dimensional reliable price reference, supports batch automatic processing, and significantly improves the efficiency and accuracy of engineering material inquiry.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and engineering cost information technology, specifically relating to an intelligent price inquiry and matching method and system for engineering materials based on natural language processing (NLP) technology, which is particularly suitable for material price inquiry, intelligent recommendation and batch quotation scenarios in engineering cost software. Background Technology

[0002] Construction cost management is a crucial aspect of building construction project implementation. Material costs typically account for 60% to 70% or more of the total project cost, and the accurate and efficient acquisition of material prices has a decisive impact on the precision of construction cost estimates.

[0003] Currently, engineering cost professionals mainly rely on the following methods when inquiring about material prices: (1) querying the material information prices published by the government. This type of data has a long update cycle (usually monthly or quarterly) and cannot reflect real-time market price fluctuations; (2) manually inquiring with material suppliers one by one, which is inefficient and costly; (3) relying on the historical quotation data accumulated within the company. However, this type of data is usually stored in a non-standard format, and manual judgment and matching are required when querying, which cannot form a large-scale intelligent utilization.

[0004] The material price query function in existing engineering cost software has the following technical defects: First, it generally uses keyword exact matching or simple fuzzy matching, requiring users to input standardized material names and specifications. It cannot understand colloquial or non-standardized input, resulting in a low query hit rate. Second, it lacks semantic understanding capabilities, resulting in a serious vocabulary gap; different names for the same material cannot be recognized by the system as the same query intent. Third, it lacks multi-dimensional joint recommendation capabilities, failing to comprehensively consider factors such as regional price differences, time price trends, and supplier differences to provide statistically significant confidence price ranges. Fourth, it does not support automated intelligent processing of batch table data; a large amount of repetitive batch price query work still requires manual operation.

[0005] While existing technologies include methods for applying natural language processing to information retrieval and machine learning to price prediction, there is currently no complete technical solution specifically designed for the engineering cost field that combines engineering materials domain knowledge (ontology knowledge base, domain pre-trained model) with semantic vector retrieval and multi-dimensional price recommendation. In the specialized field of engineering materials, the following deep-seated technical obstacles exist that cannot be directly addressed by general NLP techniques: First, the material naming system is extremely chaotic. The same material has numerous parallel synonyms in different regions, eras, and trades, and their semantic equivalence cannot be directly inferred using general word vectors or language models. A specially constructed engineering materials ontology knowledge base is necessary for mapping. Second, specification descriptions are chaotic and have strict engineering technical meanings. Similar specifications are extremely close in semantic vector space, but belong to different sub-items in engineering cost estimation, resulting in significant price differences. General semantic retrieval is prone to specification mismatches. Third, engineering material prices are affected by regional quota policies, tax rate differences, transportation costs, and supply chain structures. Inter-provincial price differences for the same material can reach 30% to 50% or even higher, making inquiry and matching methods from other fields unsuitable for the multi-dimensional regional price difference structure of engineering costs. Fourth, the table structure for batch inquiries of bills of quantities is highly non-standard, with diverse column descriptions and a lack of unified field naming conventions, making structured table parsing solutions from other industries unsuitable for direct application.

[0006] Patent document CN112488834A discloses a knowledge graph-based method for querying engineering cost information. It achieves structured information retrieval by constructing a knowledge graph in the engineering cost domain. However, this method only supports standardized structured queries and does not address the understanding of non-standardized natural language input. Patent document CN113239133A discloses a method for predicting building material prices, using a time-series model to model historical price data. However, this method does not integrate named entity recognition and semantic matching capabilities, and cannot handle users' natural language inquiry requests. Patent document CN114820071A discloses a deep learning-based intelligent engineering cost auditing method, utilizing a pre-trained language model for semantic understanding of cost text. However, it does not involve material inquiry matching or multi-dimensional price recommendation functions. None of the above-mentioned existing technologies organically combine domain pre-trained models, domain ontology knowledge bases, hybrid semantic retrieval, and multi-dimensional price recommendation to form a complete technical solution for engineering material inquiry scenarios. Summary of the Invention

[0007] This invention aims to overcome the shortcomings of existing engineering material inquiry systems in terms of natural language understanding, semantic matching, and multi-dimensional price recommendation. It provides an intelligent inquiry and matching method and system for engineering materials based on natural language processing, which can achieve accurate semantic understanding of non-standardized user input, intelligent retrieval and recommendation of multi-dimensional price data, and automated processing of batch inquiry tasks. Furthermore, it continuously improves system performance through a continuous feedback learning mechanism.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] Step 1: Obtaining an Inquiry Request

[0010] The system retrieves user-inputted inquiry text, which can be either a single text input or a batch table import. Text input allows users to describe the materials to be inquired about in free, natural language, without adhering to strict formatting guidelines. Batch table import allows users to directly import table files containing multiple material information entries into the system; the system automatically parses the column names using a table structure semantic recognition algorithm, eliminating the need for manual field mapping configuration by the user.

[0011]

Step Two: Named Entity Recognition

[0012] Named entity recognition (NER) is performed on the obtained inquiry request text to extract key entities such as material name (MAT), specification (SPEC), unit of measurement (UNIT), region (LOC), and time (TIME). The NER employs a sequence labeling method based on a pre-trained language model in the engineering cost domain. Using a general pre-trained language model with a Transformer architecture as a foundation, domain-adaptive pre-training is performed on an engineering cost professional corpus with a total size of no less than 120 million characters (approximately 240 million tokens). This pre-training uses a masked language model (MLM) task with a mask probability of 15%, a learning rate of 2e-5 (linearly decaying after 1000 linear warm-up steps), a batch size of 32, and training for 3 epochs. The AdamW optimizer (β1=0.9, β2=0.999, weights...) is used. (Attenuation 0.01); then supervised fine-tuning is performed on an engineering materials named entity dataset that uses the BIOES annotation system, is independently annotated by two people and verified by arbitration (the consistency Kappa coefficient between annotators is not less than 0.90), contains no less than 8,000 annotated sentences and 35,000 annotated entity instances. The supervised fine-tuning learning rate is 1e-5, the batch size is 16, the training is performed for 5 epochs, and an early stopping strategy is adopted (stopping when the validation set loss does not decrease for 2 consecutive epochs); the professional corpus includes at least professional texts such as engineering cost standards and specifications, material information price announcements, supplier quotations and bills of quantities.

[0013]

Step 3: Semantic Normalization

[0014] The extracted non-standard entities are semantically normalized using a built-in ontology knowledge base in the engineering materials domain to obtain standardized query conditions. This ontology knowledge base stores material alias mapping relationships in the form of hash tables and includes a specification normalization rule base based on a set of regular expressions (e.g., mapping non-standard specification descriptions to standard formats) and a unit conversion table, mapping colloquial and non-standard material descriptions to system standard material codes. Based on the material classification system of the "Construction Engineering Quantity List Pricing Specification," the ontology knowledge base supports various standardized conversions, such as mapping "Grade III rebar" to "HRB400" and "20 mm" to "φ20 mm," and is continuously expanded and maintained through a version management mechanism.

[0015] Step 3: Complete the query criteria (optional)

[0016] When the entities extracted in step two are missing a region entity or a time entity, the missing query parameters are automatically inferred and completed based on the context and user historical behavior logs. The user historical behavior logs are stored indexed by user identifier.

[0017]

Step 4: Hybrid Semantic Search

[0018] The normalized query conditions are encoded into high-dimensional query vectors using a pre-trained sentence vector model. Cosine similarity is used as the metric, and a near nearest neighbor algorithm (such as HNSW) is employed to recall candidate price records with similarity exceeding a preset threshold in the vector index of the material price database. The preset threshold is 0.80 by default (0.75 for general materials such as cement and steel, and 0.85 for special materials), overcoming the lexical gap problem of traditional keyword matching. At the same time, the normalized regional code, time range, and other structured fields are used as filtering conditions to accurately filter the candidate set, forming a hybrid retrieval strategy that combines vector semantic retrieval with structured filtering, balancing recall and retrieval accuracy.

[0019]

Step 5: Filtering Outliers

[0020] Outlier filtering based on a statistical model is performed on the filtered candidate price record set, employing an adaptive dual-model mechanism: when the effective sample size N for similar materials is greater than or equal to 30, a normal distribution model is established, using the historical mean μ and standard deviation σ of similar material prices as parameters, and price data exceeding the interval [μ-2σ, μ+2σ] are marked as outliers; when the effective sample size N is less than 30, a robust statistical method based on the interquartile range (IQR) is switched, calculating the difference IQR between the first quartile Q1 and the third quartile Q3, and price data below Q1-1.5×IQR or above Q3+1.5×IQR are marked as outliers. Price data marked as outliers have reduced weight in the subsequent confidence calculation in step six to retain their value as a statistical reference.

[0021] [Step Six: Confidence Calculation and Result Output]

[0022] Based on the data volume, timeliness, and corresponding data source labels of each record in the effective price record set, the confidence score is calculated using the Bayesian confidence estimation method. The Bayesian confidence estimation integrates the following four types of factors: (1) Reference data volume factor: the more effective price records there are, the higher the confidence score; (2) Data timeliness factor: price records closer to the current date are given higher weights, and historical data exceeding the preset timeliness threshold reduces their contribution to the confidence score; (3) Semantic matching factor: the higher the mean cosine similarity of the recalled candidate set in step four, the higher the confidence score; (4) Data source reliability factor: records with the data source label of government information price announcement are given the highest reliability weight (prior probability 0.9), records with the data source label of quotations confirmed by at least two suppliers are given the second highest weight (prior probability 0.7), and records with the data source label of unilateral quotation are given the lowest weight (prior probability 0.4). The recommendation results are output in the form of lowest price, reference price, and highest price range, along with the data source and confidence score.

[0023]

Step Seven: Output the Results

[0024] The final recommendation results will be output in a standardized format, including material standard name, specifications, region, time range, recommended price range, confidence score, and reference data source information. It supports both single-item display and batch backfilling.

[0025] [Step 8: Feedback and Learning]

[0026] The system receives user feedback on recommendation results, including price confirmation, price correction, or rating. It writes the user-confirmed price data into the material price database as a reference data source for subsequent recommendations. At the same time, it continuously optimizes and updates the named entity recognition model and semantic normalization rules based on accumulated feedback data to achieve continuous improvement in system performance.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] (1) Break through the limitations of keyword matching and achieve semantic-level understanding. By introducing customized NLP technology in the field of engineering cost, the system can accurately understand the user's non-standardized and colloquial expressions, greatly improve the query hit rate, effectively solve the vocabulary gap problem in existing technologies, and reduce the user's operating threshold.

[0029] (2) Provides multi-dimensional, confidence-based intelligent price recommendations. By combining multiple factors such as regional price differences, data timeliness, supplier differences, and outlier filtering with Bayesian confidence estimation, it provides a price range that is more valuable than a single price query.

[0030] (3) Supports batch automated processing, significantly improving efficiency. The automatic parsing and parallel matching function of batch tables frees cost engineers from a large number of repetitive inquiry tasks, and the parallel computing mechanism greatly shortens the processing time of batch tasks.

[0031] (4) Continuous optimization is achieved through a feedback learning mechanism. By receiving user feedback, the system continuously accumulates confirmed price data into the historical database and drives the continuous updating of the model, so that the system can continuously improve its performance as the usage increases, forming a positive data flywheel.

[0032] (5) Make full use of the enterprise's private data assets to form a differentiated technological barrier. The system transforms the enterprise's historical price data into the core capability of intelligent query. The richer the accumulated data assets, the higher the system's recommendation accuracy, forming a data and model barrier that is difficult for competitors to replicate in the short term. Attached Figure Description

[0033] Figure 1 This is a flowchart of the overall process of the method described in this invention, showing the complete process from obtaining the inquiry request to the result output and feedback learning (including steps one to eight).

[0034] Figure 2 This is a schematic diagram of the internal structure of the NLP preprocessing module of the present invention, illustrating the collaborative relationship between the three sub-units: named entity recognition, semantic normalization, and query condition completion.

[0035] Figure 3 This is a flowchart of the hybrid semantic retrieval module of the present invention, illustrating the retrieval strategy that combines vector semantic retrieval with structured filtering.

[0036] Figure 4 This is a flowchart of the multi-dimensional price recommendation module of the present invention, showing the processing steps and their order of outlier filtering (step five) and confidence calculation (step six).

[0037] Figure 5 This is a flowchart of the batch table inquiry processing of the present invention, showing the entire process of table import, structure recognition, parallel processing and result backfilling.

[0038] Figure 6 This is an overall architecture diagram of the system described in this invention. Detailed Implementation

[0039] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0040] Example 1: Intelligent price matching for single text input

[0041] This embodiment uses the example of an engineering cost estimator checking the current year's price of 20mm grade III rebar in Beijing to illustrate the specific execution process of the method of the present invention.

[0042] Step 1: Obtaining the Request for Quotation. The user enters the following text into the quotation function interface of the engineering cost software: "Price of 20mm Grade III rebar in Beijing this year." The system receives this text as a quotation request and initiates the processing flow.

[0043] Step 2, Named Entity Recognition. The system performs named entity recognition on the input text, identifying: Region entity LOC: [Beijing region], Material name entity MAT: [Grade III rebar], Specification entity SPEC: [20 mm], and Time entity TIME: [This year].

[0044] Step 3, Semantic Normalization. The semantic normalization unit completes the mapping through the ontology knowledge base: region [Beijing] is mapped to the standard administrative division code [110000], material [Grade III rebar] is mapped to the standard code [HRB400], specification [20 mm] is mapped to the standard specification [φ20 mm], and time [this year] is mapped to the current year. All necessary conditions for this query have been extracted; the query condition completion step does not need to be executed.

[0045] Step 4: Hybrid Semantic Retrieval. The system encodes the normalized material information (HRB400 φ20mm) into a high-dimensional query vector using a sentence vector model. Using cosine similarity as a metric, it recalls candidate material price records with a similarity exceeding 0.80 in the HNSW vector index of the material price database. Simultaneously, using the region code [110000] and the current year's time range as structured filtering conditions, it obtains a set of relevant price records for the Beijing area for the current year (a total of 45 candidate records).

[0046] Step 5: Statistical outlier filtering. With a sample size N=45≥30 from the 45 candidate records, a normal distribution model is established. The price mean μ and standard deviation σ are calculated. Outlier price records exceeding the range [μ-2σ, μ+2σ] are filtered out, resulting in 38 valid price records. The 7 filtered outlier records have their weight reduced in the confidence calculation in step 6.

[0047] Step Six: Confidence Calculation and Result Output. The confidence calculation unit, based on the amount of data from 38 valid records (relatively large, positive contribution), timeliness (high proportion of data from the last two years, positive contribution), mean cosine similarity (0.91, positive contribution), and data source labels (5 government information prices and 33 supplier confirmed quotations), uses Bayesian confidence estimation to calculate a comprehensive confidence score of 0.92 (out of 1.0). A recommended price range is generated: minimum price 3800 yuan / ton, reference price 3980 yuan / ton, and maximum price 4200 yuan / ton.

[0048] Step 7, Results Output. The system displays the following on the inquiry interface: Material Name [Hot-rolled ribbed steel bar HRB400 φ20mm], Region [Beijing], Time [Current Year], Recommended Price [3800 to 4200 RMB / ton, Reference Price 3980 RMB / ton], Confidence Level [92%], Data Source [Internal historical quotation database, 38 valid records referenced; 5 government information prices, 33 supplier confirmed quotations].

[0049] Step 8, Feedback and Learning. Once the user confirms that the reference price of 3980 yuan / ton is the actual purchase price, the system writes this record (data source tag: user-confirmed price) into the material price database. At the same time, the entity recognition results of this query are updated to the model training sample pool, triggering incremental fine-tuning updates of the model periodically.

[0050] Example 2: Batch Table Intelligent Price Inquiry

[0051] This example illustrates the system's batch processing flow when a user imports a table containing multiple material information entries.

[0052] Table parsing stage. The user imports an Excel spreadsheet containing 50 rows of material information, with column names such as Material Description, Quantity, Project Item, and Remarks. The system's table structure semantic recognition submodule performs semantic classification inference on each column name, identifying the "Material Description" column as the inquiry field (confidence level 0.97), without requiring manual specification from the user; it automatically extracts the 50 rows of material descriptions into a batch inquiry task list.

[0053] Parallel processing phase. The system submits 50 price inquiry tasks to the parallel processing queue, and executes steps two through seven independently for each task. In an 8-core parallel environment, the total processing time for the 50 price inquiry tasks is reduced by approximately 80% compared to sequential processing.

[0054] Result backfilling stage. After batch processing, the system automatically adds recommended price range, reference price, confidence level, and data source columns to the original table, backfilling the recommendation results for each material row by row to generate a complete intelligent quotation table. Recommendation results with a confidence level below the set threshold (default 0.6) are highlighted to prompt the user for manual review.

[0055] Example 3: System Architecture and Implementation of Each Module

[0056] The system described in this invention adopts a modular architecture in which the memory and processor work together.

[0057] Material Price Database (105): Stores standardized, multi-dimensional material price data. Each price record includes fields for material code, material name, specifications, region code, price, tax-inclusive label, supplier information, data date, and data source label. It also maintains a vector index (based on the HNSW algorithm for semantic retrieval) and a B+ tree structured index (for conditional filtering). The values ​​for the data source label field are as follows: 0 - Government information price announcement (prior reliability weight 0.9); 1 - Quotations confirmed by multiple suppliers (prior reliability weight 0.7); 2 - Quotations from a single supplier (prior reliability weight 0.4); 3 - Historical prices confirmed by the user (prior reliability weight 0.85).

[0058] NLP preprocessing module (102): includes named entity recognition unit (201), semantic normalization unit (202) and query condition completion unit (203). The Named Entity Recognition (NER) unit is based on a pre-trained language model using the Transformer architecture. It is further pre-trained on a corpus of engineering cost estimation linguistics containing at least 120 million characters (MLM task, mask probability 15%, learning rate 2e-5, AdamW optimizer, 3 epochs). Then, it undergoes supervised fine-tuning on a NER dataset containing at least 8,000 annotated sentences and 35,000 annotated entity instances (Kappa ≥ 0.90) (learning rate 1e-5, 5 epochs, early stopping). It supports the recognition of five entity classes (MAT, SPEC, UNIT, LOC, TIME). The Semantic Normalization (SNU) unit incorporates an ontology knowledge base in the engineering materials domain, storing alias mapping relationships in a hash table and maintaining a specification normalization rule base based on a set of regular expressions, providing efficient search and conversion services. The Query Completion (CCM) unit intelligently infers missing query parameters based on context and user history logs.

[0059] The semantic retrieval module (103) includes a vector encoding unit (301), an ANN retrieval unit (302), and a structured filtering unit (303). The vector encoding unit uses a pre-trained sentence vector model finely tuned on the engineering cost corpus to encode standardized material information into high-dimensional vectors. The ANN retrieval unit is based on a vector database and uses the HNSW approximate nearest neighbor algorithm to construct a vector index. It uses cosine similarity as a metric and performs Top-K similarity retrieval on records with a default threshold of 0.80 or higher. The structured filtering unit performs precise filtering on the vector retrieval results according to structured conditions such as regional coding and time range.

[0060] The price recommendation module (104) includes a multi-dimensional joint retrieval unit (401), an outlier filtering unit (402), and a confidence calculation unit (403). The outlier filtering unit implements the above-mentioned adaptive dual-model detection mechanism (using the 2σ normal distribution model when N≥30, and switching to the IQR method when N<30); the confidence calculation unit adopts the Bayesian confidence estimation framework, and calculates the comprehensive confidence based on four factors: data volume, timeliness, semantic matching degree, and data source reliability. The prior probability of each source category is set according to the enumerated value of the data source label field.

[0061] Feedback Learning Module (107): Receives user feedback on recommendation results, writes confirmed price data (data source label set to 3) into the material price database, and periodically uses accumulated feedback data for incremental fine-tuning of the NLP model and rule expansion of the ontology knowledge base. For every 500 user-confirmed labeled data entries accumulated, an incremental fine-tuning of the named entity recognition model is triggered, with the fine-tuning learning rate set to one-tenth of the initial training rate.

[0062] Data access module (108): Supports connection to government-released material price data sources and external supplier quotation interfaces. Through standardized data access adapters, it realizes continuous updating and expansion of material price data. New access records are automatically assigned corresponding data source label field values ​​according to the source type.

Claims

1. A method for intelligent price inquiry and matching of engineering materials based on natural language processing, characterized in that, Includes the following steps: S1. Obtain the text of the inquiry request entered by the user; S2. Perform named entity recognition on the inquiry request text to extract at least one material entity and at least one specification entity; S3. Perform semantic normalization processing on the extracted material entities and specification entities, and map non-standard entities to preset material standard codes; S4. The normalized query conditions are matched in the material price database using a hybrid retrieval strategy. The hybrid retrieval strategy includes: encoding the normalized query conditions into a query vector, using cosine similarity as a metric to recall candidate price record sets with similarity exceeding a preset threshold; and using structured fields as filtering conditions to precisely filter the candidate price record sets. S5. Perform outlier filtering based on a statistical model on the filtered candidate price record set to obtain the effective price record set; S6. Based on the data volume, timeliness, and corresponding data source tags of each record in the effective price record set, calculate the confidence score, and output the recommendation result containing the reference price and the confidence score in the form of a price range.

2. The method according to claim 1, characterized in that, The named entity recognition in step S2 adopts a sequence labeling method based on a pre-trained language model in the field of engineering cost. The pre-trained language model is based on a general Transformer architecture pre-trained model and is further pre-trained in a domain-adaptive manner on an engineering cost professional corpus with a total scale of no less than 120 million characters. The professional corpus includes at least professional texts such as engineering cost standards and specifications, material information price announcements, supplier quotations, and bills of quantities. Then, it is fine-tuned under supervision on an engineering material named entity dataset that adopts the BIOES annotation system, is independently annotated by two people and verified by arbitration (the consistency Kappa coefficient between annotators is no less than 0.90), and contains no less than 8,000 annotated sentences and 35,000 annotated entity instances. The extracted categories of named entity recognition include at least material name entities, specification and model entities, unit of measurement entities, region entities, and time entities.

3. The method according to claim 1, characterized in that, The semantic normalization process described in step S3 is implemented through a built-in ontology knowledge base in the field of engineering materials. The ontology knowledge base stores the material alias mapping relationship in the form of a hash table and includes a specification normalization rule base based on a set of regular expressions and a unit conversion table to map colloquial or non-standardized material descriptions to the system's standard material codes. The ontology knowledge base is based on the material classification system of the "Construction Engineering Quantity List Pricing Specification" and supports dynamic iteration and expansion through a version management mechanism.

4. The method according to claim 1, characterized in that, The query vector in step S4 is obtained by encoding the normalized query conditions through a pre-trained sentence vector model; the candidate price record set is retrieved by using an approximate nearest neighbor algorithm in the vector index of the material price database; the structured field filtering conditions include regional codes and time ranges.

5. The method according to claim 1, characterized in that, The outlier filtering based on the statistical model described in step S5 employs an adaptive dual-model mechanism: (a) When the effective sample size N of the same type of material is greater than or equal to 30, a normal distribution model is established, with the historical price mean μ and standard deviation σ of the same type of material as parameters, and price data that exceed the interval [μ-2σ, μ+2σ] are marked as outliers; (b) When the effective sample size N is less than 30, switch to a robust statistical method based on interquartile range (IQR), calculate the difference IQR between the first quartile Q1 and the third quartile Q3, and mark price data that are lower than Q1-1.5×IQR or higher than Q3+1.5×IQR as outliers. (c) Price data marked as outliers are weighted less in the confidence calculation in step S6.

6. The method according to claim 1, characterized in that, The confidence score calculation in step S6 takes into account the following factors: (a) Reference data quantity factor: The more records in the effective price record set, the higher the confidence level; (b) Data timeliness factor: Price records closer to the current date are given higher weight, and historical data that exceeds the preset timeliness threshold has a reduced contribution to confidence. (c) Semantic matching factor: The higher the mean cosine similarity of the candidate price record set in step S4, the higher the confidence level; (d) Data source reliability factor: The records with the data source label of government information price announcement are given the highest reliability weight, the records with the data source label of quotation confirmed by multiple suppliers are given the second highest weight, and the records with the data source label of unilateral quotation are given the lowest weight.

7. The method according to claim 1, characterized in that, The inquiry request also includes a batch table import format, the processing steps of which include: (a) The column name text of the imported table is classified and inferred by multi-label semantic recognition algorithm through table structure. The semantic categories of various columns such as inquiry field column, quantity column, project column and remarks column are automatically identified. The classification and inference is based on the column name classification model trained in the engineering cost business scenario, without the need for users to manually configure field mapping. (b) Submit the parsed multi-line inquiry data to the batch processing queue through a parallel computing mechanism, and each line inquiry task executes steps S2 to S6 independently; (c) Automatically populate the batch matching results back into the original table, and add a recommended price range column, a confidence level column, and a data source column to the original table to generate a complete quotation table with recommended prices.

8. The method according to claim 1, characterized in that, It also includes step S7: receiving user feedback on the recommendation results, the feedback including price confirmation, price correction or result rating; writing the user-confirmed price data into the material price database, and continuously optimizing and updating the named entity recognition model in step S2 and the semantic normalization rules in step S3 based on the accumulated feedback data.

9. The method according to claim 1, characterized in that, Between step S3 and step S4, there is also a query condition completion step: when the entities extracted in step S2 are missing a region entity or a time entity, the missing query parameters are automatically inferred and completed based on the context and user historical behavior logs, and the user historical behavior logs are stored according to the user identifier index.

10. An intelligent price inquiry and matching system for engineering materials based on natural language processing, characterized in that, include: The memory is used to store the material price database. Each price record in the material price database contains at least a material code field, a material name field, a specification field, a region field, a price field, a data source label field, and a data date field, and each price record is associated with a pre-generated vector representation. A processor, coupled to the memory, is configured to perform the following operations: Obtain the inquiry request text; Named entity recognition is performed on the inquiry request text to extract at least one material entity and at least one specification entity; The material and specification entities are semantically normalized using an ontology knowledge base to obtain normalized query conditions. The normalized query conditions are encoded into query vectors, and cosine similarity is used as a metric to recall candidate price record sets with similarity exceeding a preset threshold in the vector index of the memory. Structured fields are used as filtering conditions for precise filtering. Perform statistical outlier filtering on the filtered candidate price record set to obtain the valid price record set; Based on the data volume, timeliness, and data source tag field of each record in the effective price record set, a confidence score is calculated, and the recommendation result is output in the form of a price range.

11. The system according to claim 10, characterized in that, When performing named entity recognition, the processor calls a sequence labeler based on a pre-trained language model in the engineering cost domain. This pre-trained language model is based on a general pre-trained model with a Transformer architecture, and is obtained through domain-adaptive pre-training on an engineering cost professional corpus with a total size of no less than 120 million characters and supervised fine-tuning on an engineering materials named entity recognition dataset. When performing semantic normalization, the processor queries a built-in engineering materials domain ontology knowledge base, which includes material alias mapping relationships and a specification normalization rule base. The memory also maintains vector indexes and multi-dimensional structured indexes to support hybrid retrieval.

12. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the intelligent inquiry and matching method for engineering materials based on natural language processing as described in any one of claims 1 to 9.

Images