Intelligent engineering material price retrieval method based on fine-grained field perception fusion

Abstract

The invention discloses an engineering material price intelligent retrieval method based on fine-grained field perception fusion. The method comprises the steps that an original query set is received, and a material library is provided; performing semantic alignment and structure perception enhancement processing on the original queries in the original query set to generate standardized enhanced queries; calculating the global similarity of normalized enhanced query and records in a material library through a dense semantic channel and a sparse lexical item channel; performing field-level analysis on each record in the standardized enhanced query and material library, and calculating field-level fine-grained similarity of the dense semantic channel and the sparse lexical item channel on a predefined key field set; carrying out normalization processing on the global similarity and the field-level fine-grained similarity, and carrying out weighted fusion to generate a final fusion similarity; and sorting the records in the material library according to the final fusion similarity, and outputting the record with the highest rank as target material price information. The method effectively relieves the influence of noise and fuzzy query on the matching precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of engineering construction, and particularly relates to an intelligent engineering material price retrieval method based on fine-grained field perception fusion. BACKGROUND

[0002] With the complication of industrial scenarios and the rapid growth of data scale, information retrieval and knowledge alignment play an increasingly important role in industrial applications. Unlike traditional structured databases, equipment data in the engineering field often exists in the form of unstructured text, making it difficult for traditional methods relying on structured query language (SQL) to effectively cope with complex query requirements. This problem is particularly prominent in engineering material price retrieval: query texts often contain noise, abbreviations and ambiguous expressions, directly leading to a decline in retrieval accuracy, which seriously hinders the efficiency and reliability of project cost estimation. In the application of project cost estimation in the construction industry, the equipment specification information table often contains complex semantic information. How to extract key features from it and achieve accurate matching is a core challenge of the retrieval task. Existing research attempts to solve this problem from the perspective of multi-granularity semantic retrieval, for example: an information intelligent retrieval method based on multi-granularity semantic fusion, which improves matching accuracy by fusing semantic information of different granularities such as words, phrases and texts, or a partial multi-modal hashing method based on fine-grained feature fusion, which uses fine-grained feature fusion and self-attention mechanism to alleviate the influence of multi-modal feature missing. However, these methods still have deficiencies when facing real queries containing noise, abbreviations and ambiguous expressions. On the other hand, research on price knowledge alignment shows that the introduction of quota intelligent binding and knowledge graph can improve the matching efficiency of equipment specification information and quota price, but it is difficult to guarantee the stability and robustness of the matching results in complex semantic scenarios.

[0003] In recent years, information retrieval (IR) has made significant progress in search engines, question and answer systems, and personalized recommendations. Some research attempts to optimize the index structure, for example, the "interactive application research of a retrieval enhancement generation technology based on index structure optimization in the insurance field" proposed by Cheng Yinning, which introduces document segmentation, keyword extraction and semantic alignment mechanism in the indexing process, thereby improving the ability to capture key information. This method has advantages in improving retrieval coverage, but its processing ability for semantic ambiguity and cross-field expression is still insufficient.

[0004] In the aspect of semantic representation learning, the "Research on Academic Literature Knowledge Entity Extraction Method Fusing Semantic Enhancement and Knowledge Distillation" proposed by Wang Yulong et al. combines global semantics and dynamic word meaning information by embedding representation method of fusing SciBERT and ELMo model, effectively improving the modeling ability of complex context. This method performs well in deep semantic capture, but has limited ability to distinguish explicit elements such as numerical fields and unit specifications, making it difficult to be directly applied to the engineering material scene.

[0005] To fuse the advantages of sparse and dense models, the "Domain Question Answering Alignment Optimization Method Based on Keyword Extraction and Hybrid Retrieval" proposed by Yang Dongju and Yang Kun uses large language models to extract user query keywords and combines sparse retrieval and dense retrieval to achieve hybrid recall, thereby enhancing coverage and robustness. However, this method still relies on empirical settings in weight allocation and dynamic adaptability, making it difficult to achieve stable optimal fusion in complex scenarios.

[0006] In summary, although existing research has made positive progress in index optimization, sparse and dense fusion, semantic representation, and multi-task learning, there are still obvious shortcomings in the engineering material price retrieval scene. In the face of professional engineering fields containing noise, abbreviations, and ambiguous expressions, retrieval is prone to semantic drift; single modeling method is difficult to balance deep semantic understanding and explicit field matching; different channel modeling features are not fully integrated, resulting in difficulty in balancing recall coverage and discrimination accuracy in candidate sorting. SUMMARY

[0007] Therefore, the purpose of the present application is to propose an intelligent retrieval method for engineering material prices based on fine-grained field perception fusion, effectively alleviating the impact of noise and ambiguous queries on matching accuracy; design a dense-sparse dual-channel dynamic fusion candidate generation mechanism to ensure semantic discrimination ability while improving recall coverage; build a low-cost, scalable retrieval system with potential for rapid application in the engineering field.

[0008] To achieve the above technical purposes, the technical solution adopted by the present application is:

[0009] The present application provides an intelligent retrieval method for engineering material prices based on fine-grained field perception fusion, comprising the following steps:

[0010] Step 1, receive an unstructured original query set and provide a semi-structured material library;

[0011] Step 2, perform semantic alignment and structure perception enhancement processing on the original queries in the original query set to generate normalized enhanced queries;

[0012] Step 3, calculate the global similarity between the normalized enhanced query and each record in the material library through the dense semantic channel and the sparse term channel respectively;

[0013] Step 4, field-level resolution is performed on the normalized enhanced query and each record in the material library, and field-level fine-grained similarity of the dense semantic channel and the sparse term channel on a set of pre-defined key fields is calculated respectively;

[0014] Step 5, normalize the global similarity and the field-level fine-grained similarity, obtain global fusion weight and field-level fusion weight, perform weighted fusion, and generate final fusion similarity;

[0015] Step 6, sort all records in the material library according to the final fusion similarity, and output the record with the highest ranking as the target material price information.

[0016] Further, the step 1 specifically comprises:

[0017] Step 11, receive a set of unstructured original queries oriented to the engineering material price retrieval scene Wherein, q i represents the i-th original query described in natural language, i represents the index of the original query, m represents the total number of original queries, and the original query contains noise, industry abbreviations and colloquial expressions;

[0018] Step 12, provide a semi-structured material library T=(S,R), wherein S is a structure pattern, is a record set, r j represents the j-th record, j represents the index of the record, n represents the total number of records, and each record r j j1 j2 jk is composed of multiple heterogeneous fields, wherein f j1 represents the first heterogeneous field under the j-th record, f j2 represents the second heterogeneous field under the j-th record, and f jk represents the k-th heterogeneous field under the j-th record, k represents the total number of heterogeneous fields; the heterogeneous fields include structured numerical / enumeration fields and free text fields.

[0019] Further, the step 2 specifically comprises:

[0020] Step 21, perform semantic alignment processing on each original query in the original query set to generate a standardized query semantically equivalent;

[0021] Step 22, perform structure-aware enhancement processing on the standardized query to form a structured query field representation;​​​

[0022] Step 23, splicing the standardized query and the query field representation to generate a normalized enhanced query.

[0023] Further, the step 21 specifically includes:

[0024] The semantic alignment processing includes at least one of synonym normalization, abbreviation expansion and unit normalization, and the noise, abbreviation and ambiguous expression in the original query are converted into a standard form through synonym normalization mapping, abbreviation expansion rule set and unit normalization conversion function to generate a semantically equivalent standardized query.

[0025] The step 22 specifically includes:

[0026] Step 221, defining a field set F of a domain:

[0027] F={Material, Spec, Unit, …}

[0028] Wherein, the field set of the domain is divided into material Material, specification Spec and unit Unit according to the engineering field;

[0029] Step 222, extracting a pre-defined key field and its corresponding value from the standardized query by rule matching, dictionary mapping or numerical recognition to obtain a structured query field representation

[0030]

[0031] Wherein, f represents a key field, V f is the value domain of the key field f, and z f is the value of the key field f.

[0032] The step 23 specifically includes:

[0033] Splicing the standardized query and the structured query field representation to obtain a normalized enhanced query

[0034]

[0035] Wherein, C(·,·) is a splicing function.

[0036] Further, the synonym normalization, abbreviation expansion and unit normalization specifically include:

[0037] The synonym normalization specifically refers to mapping an abbreviation or alias w in the domain vocabulary V to a standard word The normalization obtains a vocabulary

[0038] The abbreviation expansion is to replace an abbreviation a with a standard form b by using a rule set R'=(a,b);

[0039] The unit normalization is to convert a numerical-unit pair (v,u) into a unified system of units by using a conversion function .

[0040] Further, the step 3 specifically comprises:

[0041] Step 31, constructing a dense-sparse dual-channel framework, including a dense semantic channel and a sparse term channel;

[0042] Step 32, in the dense semantic channel, a pre-trained semantic encoder φ is used to encode the normalized enhanced query and each record r in the material library j into a query semantic vector and a record vector .

[0043]

[0044] wherein Norm(φ) represents an L2 normalization operation; x generally refers to a result after encoding by a pre-trained semantic encoder φ; represents a real number field; represents a set of d-dimensional real number vectors;

[0045] The similarity between the normalized enhanced query and each record r in the material library j is calculated by a cosine inner product between vectors, to obtain a global dense semantic similarity

[0046]

[0047] wherein 〈·,·> represents an inner product operation of the query semantic vector and the record vector v(r j );

[0048] Step 33, in the sparse term channel, a term set is obtained based on a segmentation function τ(·), and a BM25 model is used to calculate the term matching similarity between the normalized enhanced query and each record r in the material library j , to obtain a global sparse term similarity

[0049]

[0050] Where t represents a word, tf(t,r) j ) represents the word t in the record r j In the context of word frequency, k1 represents a hyperparameter that adjusts word frequency saturation, and its function is to control the word frequency tf(t,r). j The degree of influence of ) on the final correlation score; b∈[0,1] is used to control the strength of normalization; |r j | indicates a record r j Length, avg dl The average record length is denoted by idf(t); idf(t) represents the inverse record frequency, defined as:

[0051]

[0052] Where N is the total number of records, n t This represents the number of records containing the word t.

[0053] Furthermore, step 4 specifically includes:

[0054] Step 41: Parse the corresponding content of the same key field from each record in the normalized enhanced query and the material library respectively;

[0055] Step 42: For each key field f, use the semantic encoder of the dense semantic channel to calculate the similarity between its field content vectors, and obtain the field-level dense similarity.

[0056] Step 43: For each key field f, use the BM25 model of the sparse term channel to calculate the similarity between its field content terms, and obtain the field-level sparse term similarity.

[0057] Furthermore, step 5 includes:

[0058] Step 51: Apply the Min-Max normalization operator to the global dense semantic similarity. Global sparse term similarity Field-level dense similarity Similarity to field-level sparse terms Normalize it to unify its scale to the [0,1] interval:

[0059]

[0060] in, express Normalization results express Normalization results express Normalization results express Normalization results;

[0061] Step 52: Input the semantic vector of the normalized enhanced query into the trained multilayer perceptron model to predict the global fusion weight and multiple field-level fusion weights for the current query;

[0062] Step 53: Use the global fusion weight to linearly weight the normalized global dense semantic similarity and the normalized global sparse term similarity to obtain the dual-channel fusion similarity of global semantics.

[0063]

[0064] Among them, a i ∈[0,1] represents the global fusion weight;

[0065] Step 54: Apply the field-level fusion weights to linearly weight the normalized field-level dense similarity and normalized field-level sparse similarity for each field to obtain the dual-channel fusion similarity of field-level semantics.

[0066]

[0067] Among them, a i,f ∈[0,1] represents the field-level fusion weights;

[0068] Step 55: Fuse the dual-channel similarity of global semantics. Dual-channel fusion similarity with semantics at each field level Summing and dividing by the number of fields plus one yields the final fused similarity.

[0069]

[0070] in, The similarity is a dual-channel fusion of global semantics; This represents the dual-channel fusion similarity for field-level semantics.

[0071] Furthermore, the training process of the multilayer perceptron model includes:

[0072] 1) Constructing the training samples:

[0073] For the i-th normalized enhanced query Enumerate weight values ​​'a' in the discrete weight set A = {0..1, 0.2, ..., 0.9}, and select those that result in the true target record. Enter based on Optimal weights of the retrieved Top-K ranking candidate set As a reference value;

[0074]

[0075] in, For the indicator function, r i Represents the records being retrieved; Top-K indicates the normalization-enhanced query. The top K records with the highest similarity;

[0076] Similarly, for the i-th normalized enhanced query Enumerate weight values ​​'a' in the discrete weight set A = {0.1, 0.2, ..., 0.9}, and select those that allow the true target record to be recorded. Enter based on Optimal weights of the retrieved Top-K ranking candidate set As a reference value;

[0077]

[0078] The adaptive weight training sample set T is obtained as follows:

[0079]

[0080] Among them, Q train To query the training set;

[0081] 3) Using the query semantic vector For input, a multilayer perceptron is used. Output fusion weights As a predicted value:

[0082]

[0083] Among them, g Θ The neural network representing the multilayer perceptron maps d-dimensional vectors to prediction weights; Represents a set of d-dimensional real vectors;

[0084] Similarly, with the query semantic vector For input, a multilayer perceptron is used. Output field-level fusion weights As a predicted value:

[0085]

[0086] 3) During the training phase, the optimization objective is to minimize the mean squared error between the predicted and reference values, while introducing... The regularization term is used to train the multilayer perceptron model using the AdamW optimizer:

[0087]

[0088] Where Θ represents the learning parameters of the weighted scoring model, This represents the loss function that guides the training of the neural network; λ is the regularization coefficient. express Regular terms.

[0089] Furthermore, during the inference phase, the fusion weights predicted by the multilayer perceptron model are... Perform range clipping operation:

[0090]

[0091] Here, clip(·) indicates that the weights are truncated to a closed interval. Inside, The value is selected as [0.7, 0.9].

[0092] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0093] This invention proposes an intelligent retrieval method for engineering material prices based on fine-grained field-aware fusion, effectively alleviating the shortcomings in retrieval accuracy and stability caused by low retrieval efficiency, long processing time, and subjective human intervention. Through semantic alignment and structure-aware enhancement, it achieves standardized expression of non-standardized queries; by combining dense semantic channels and sparse term channels, it effectively balances semantic understanding and precise term matching. Simultaneously, through a dynamic weight learning mechanism, it achieves stable identification of target records under unique solution constraints. Attached Figure Description

[0094] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0095] Figure 1 This is an execution flowchart of an intelligent retrieval method for engineering material prices based on fine-grained field perception fusion provided in an embodiment of the present invention.

[0096] Figure 2 This is a graph showing the impact of different weight parameters on the model retrieval performance, provided in an embodiment of the present invention. Detailed Implementation

[0097] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0098] Please see Figure 1 The present invention provides an intelligent retrieval method for engineering material prices based on fine-grained field-aware fusion, comprising the following steps:

[0099] Step 1: Receive the unstructured raw query set and provide a semi-structured material library;

[0100] In this embodiment, step 1 specifically includes:

[0101] Step 11: Receive the unstructured raw query set for the engineering material price retrieval scenario. Where, q i This represents the i-th original query described in natural language, where i represents the index of the original query, and m represents the total number of original queries. The original query contains noise, industry abbreviations, and colloquial expressions.

[0102] Step 12: Provide a semi-structured material library T = (S, R), where S is the structural pattern. For a record set, r j Let represent the j-th record, where j represents the index of the record, n represents the total number of records, and r represents the number of records per record. j =f j1 ,f j2 ,…,f jk It consists of multiple heterogeneous fields, where f j1 f represents the first heterogeneous field in the j-th record. j2 This indicates the second heterogeneous field in the j-th record, f. jk This indicates that the j-th record contains the k-th heterogeneous field, where k represents the total number of heterogeneous fields. The heterogeneous fields include structured numerical / enumerated fields (such as units, specifications, years, etc.) and free text fields (such as project name, project characteristics, work content description, etc.).

[0103] Step 2: Perform semantic alignment and structure-aware enhancement processing on the original queries in the original query set to generate normalized enhanced queries;

[0104] In the scenario of searching for engineering material prices, the original query q i∈Q often contains noise, abbreviations, and domain-specific ambiguities. Directly using these for comparison can easily lead to semantic drift and make field alignment difficult. Therefore, this invention, while maintaining semantic equivalence, introduces query semantic enhancement and structure-aware modeling to standardize the input query, making core elements such as name, specifications, parameters, and units explicitly presented, thereby reducing semantic ambiguity and enhancing its comparability with structured fields. Specifically, semantic alignment and structure-aware enhancement operators are introduced:

[0105]

[0106] Where Q is the original query set. This is the normalized enhanced query set.

[0107] The original query q is enhanced by semantic enhancement operators. i Convert to normalized expression This generates representations that are both semantically faithful and structurally explicit, laying a solid foundation for subsequent similarity modeling.

[0108] In this embodiment, step 2 specifically includes:

[0109] Step 21: Perform semantic alignment processing on each original query in the original query set to generate semantically equivalent standardized queries; the goal of semantic alignment is to unify noise, abbreviations, and diverse colloquial expressions into a standard form, and this process yields standardized queries. Its semantic expression is equivalent to the original query, but its form is more standardized.

[0110] Preferably, step 21 specifically includes:

[0111] The semantic alignment process includes at least one of the following operations: synonym normalization, abbreviation expansion, and unit normalization. Through synonym normalization mapping, abbreviation expansion rule set, and unit normalization conversion function, noise, abbreviations, and fuzzy expressions in the original query are converted into standard forms, generating semantically equivalent standardized queries.

[0112] Specifically, the synonym normalization, abbreviation expansion, and unit normalization include:

[0113] Synonym normalization specifically involves defining the synonym normalization mapping as follows: Map abbreviations or aliases w in domain vocabulary V to standard terms. Normalization yields vocabulary

[0114] Abbreviation expansion uses the rule set R' = (a, b) to replace the abbreviation 'a' with the standard form 'b'.

[0115] Unit normalization is the process of using conversion functions to normalize a value-unit pair (v, u) (where v is the value and u is the unit). Convert it to a unified measurement system (e.g., “2cm” → “20mm”).

[0116] Step 22: Perform structure-aware enhancement processing on the standardized query to form a structured query field representation; after obtaining the standardized query... Next, key fields need to be further made explicit to align with the database's structured schema.

[0117] Preferably, step 22 specifically includes:

[0118] Step 221, Define the domain field set F:

[0119] F = {Material, Spec, Unit, ...}

[0120] The domain field set is divided into Material, Spec, and Unit according to the engineering domain.

[0121] Step 222: Using rule matching, dictionary mapping, or numerical recognition methods, extract data from standardized queries. Extract predefined key fields and their corresponding values ​​to obtain a structured query field representation.

[0122]

[0123] Where f represents the key field, V f For the range of values ​​of the key field f, z f The value of the key field f;

[0124] Step 23: Concatenate the standardized query with the query field representation to generate a normalized enhanced query.

[0125] Preferably, step 23 specifically includes:

[0126] Standardized query With structured query field representation Concatenating and serializing results in normalized enhanced queries.

[0127]

[0128] Where C(·,·) is the concatenation function.

[0129] Step 3: Calculate the global similarity between the normalized enhanced query and each record in the material database through both dense semantic channels and sparse term channels;

[0130] In the candidate generation stage, single-channel similarity modeling often has limitations. While dense vectors generated through dense semantic channels can capture deep semantic relationships, they are insufficient when dealing with key fields related to engineering materials, such as specifications, numerical values, and units. Sparse term channels, on the other hand, use term-based models (such as BM25) to provide accurate matching, but struggle to cover semantic synonyms and contextual dependencies. Therefore, to better adapt to the coexistence of semantic and numerical elements in engineering material retrieval, a field-aware fine-grained fusion mechanism is introduced on top of the traditional dense-sparse dual-channel framework.

[0131] In this embodiment, step 3 specifically includes:

[0132] Step 31: Construct a dense-sparse dual-channel framework, including a dense semantic channel and a sparse term channel;

[0133] Step 32: In dense semantic channels, a pre-trained semantic encoder is used. Normalization Enhanced Queries and each record r in the material library j Mapping to query semantic vector and recording vector v(r) j ):

[0134] and

[0135] Where Norm(φ) represents the L2 normalization operation; x generally refers to the result encoded by the pre-trained semantic encoder φ; Represents the real number field; Represents a set of d-dimensional real vectors;

[0136] Query enhancement through normalization using the cosine inner product of vectors. and each record r in the material library j The similarity is used to obtain the global dense semantic similarity.

[0137]

[0138] Where <·,·> represent query semantic vectors. and recording vector v(r) j Inner product operation;

[0139] Step 33: In the sparse term channel, obtain the term set based on the word segmentation function τ(·), and use the BM25 model to calculate the normalized enhanced query. With each record r in the material library jThe term similarity between terms is used to obtain the global sparse term similarity.

[0140]

[0141] Where t represents a word, tf(t,r) j ) represents the word t in the record r j In the context of word frequency, k1 represents a hyperparameter that adjusts word frequency saturation, and its function is to control the word frequency tf(t,r). j The degree of influence of ) on the final correlation score; b∈[0,1] is used to control the strength of normalization; |r j | indicates a record r j Length, avg dl The average record length is denoted by idf(t); idf(t) represents the inverse record frequency, defined as:

[0142]

[0143] Where N is the total number of records, n t This represents the number of records containing the word t.

[0144] Step 4: Perform field-level parsing on the normalized enhanced query and each record in the material library, and calculate the field-level fine-grained similarity of the dense semantic channel and the sparse term channel on the predefined key field set respectively;

[0145] To further enhance adaptability to engineering material retrieval scenarios, a field-aware fine-grained fusion mechanism is introduced based on global dual-channel modeling. Traditional global weighted fusion is prone to the problem of "semantic matching correct but field mismatch" when dealing with engineering records containing multiple heterogeneous elements: for example, querying "Φ..."

[0146] The similarity between "20 steel bars" and "Φ25 steel bars, 20 tons" may still be very high, but the inconsistency in the specification field is ignored. Therefore, this invention proposes to perform dense-sparse dual-channel modeling and fusion at the field level to ensure accurate alignment of each key element.

[0147] In this embodiment, step 4 specifically includes:

[0148] Step 41: Parse the corresponding content of the same key field from each record in the normalized enhanced query and the material library respectively;

[0149] Step 42: For each key field f, use the semantic encoder of the dense semantic channel to calculate the similarity between its field content vectors, and obtain the field-level dense similarity.

[0150] Step 43: For each key field f, use the BM25 model of the sparse term channel to calculate the similarity between its field content terms, and obtain the field-level sparse term similarity.

[0151] Step 5: Normalize the global similarity and field-level fine-grained similarity, obtain the global fusion weight and field-level fusion weight, perform weighted fusion, and generate the final fusion similarity.

[0152] In this embodiment, step 5 includes:

[0153] Step 51: To eliminate the difference in numerical scale between dense semantic similarity and sparse term similarity, the global dense semantic similarity is normalized using the Min-Max normalization operator. Global sparse term similarity Field-level dense similarity Similarity to field-level sparse terms Normalize it to unify its scale to the [0,1] interval:

[0154]

[0155] in, express Normalization results express Normalization results express Normalization results express Normalization results;

[0156] Step 52: Input the semantic vector of the normalized enhanced query into the trained multilayer perceptron model to predict the global fusion weight and multiple field-level fusion weights for the current query;

[0157] To balance the relative importance of semantic matching and term matching across different queries, a query-based dual-channel adaptive weighting mechanism is introduced. For each query... Learn a fusion weight α∈[0,1], and then perform linear fusion on the normalized scores of the two channels.

[0158] Step 53: Use the global fusion weight to linearly weight the normalized global dense semantic similarity and the normalized global sparse term similarity to obtain the dual-channel fusion similarity of global semantics.

[0159]

[0160] Among them, a i∈[0,1] represents the global fusion weight;

[0161] Step 54: Apply the field-level fusion weights to linearly weight the normalized field-level dense similarity and normalized field-level sparse similarity for each field to obtain the dual-channel fusion similarity of field-level semantics.

[0162]

[0163] Among them, a i,f ∈[0,1] represents the field-level fusion weights;

[0164] Step 55: Fuse the dual-channel similarity of global semantics. Dual-channel fusion similarity with semantics at each field level Summing and dividing by the number of fields plus one yields the final fused similarity.

[0165]

[0166] in, The similarity is a dual-channel fusion of global semantics; This represents the dual-channel fusion similarity for field-level semantics.

[0167] The training process of the multilayer perceptron model includes:

[0168] 1) Constructing the training samples:

[0169] For the i-th normalized enhanced query Enumerate weight values ​​'a' in the discrete weight set A = {0..1, 0.2, ..., 0.9}, and select those that result in the true target record. Enter based on Optimal weights of the retrieved Top-K ranking candidate set As a reference value;

[0170]

[0171] in, For the indicator function, r i Represents the records being retrieved; Top-K indicates the normalization-enhanced query. The top K records with the highest similarity;

[0172] Similarly, for the i-th normalized enhanced query Enumerate weight values ​​'a' in the discrete weight set A = {0.1, 0.2, ..., 0.9}, and select those that allow the true target record to be recorded. Enter based on Optimal weights of the retrieved Top-K ranking candidate set As a reference value;

[0173]

[0174] If none of the 'a' values ​​are in the Top-K, the implementation default value of min(A) is used to avoid empty labels. Therefore, the adaptive weight training sample set T is obtained as follows:

[0175]

[0176] Among them, Q train To query the training set;

[0177] 2) Using the query semantic vector For input, a multilayer perceptron is used. Output fusion weights As a predicted value:

[0178]

[0179] Among them, g Θ The neural network representing the multilayer perceptron maps d-dimensional vectors to prediction weights; Represents a set of d-dimensional real vectors;

[0180] Similarly, with the query semantic vector For input, a multilayer perceptron is used. Output field-level fusion weights As a predicted value:

[0181]

[0182] In actual reasoning, if the continuous weights output by the model are used directly... Extreme value biases may occur, leading to over-reliance on a single channel. This can amplify the error of a particular channel, resulting in unstable sorting results. To suppress fluctuations caused by extreme values, the fusion weights predicted by the multilayer perceptron model are adjusted during the inference phase. Perform range clipping operation:

[0183]

[0184] Here, clip(·) indicates that the weights are truncated to a closed interval. Inside, The value is selected as [0.7, 0.9].

[0185] 3) To make the predicted value It can approximate the optimal reference value obtained based on discrete search. Predicted value It can approximate the reference value This invention models the weight prediction problem as a regression task. During the training phase, the optimization objective is to minimize the mean squared error between the predicted and reference values, while simultaneously introducing... The regularization term is used to train the multilayer perceptron model using the AdamW optimizer:

[0186]

[0187] Where Θ represents the learning parameters of the weighted scoring model, This represents the loss function that guides the training of the neural network; λ is the regularization coefficient. express Regular terms.

[0188] Step 6: Sort all records in the material library according to the final fusion similarity, and output the record with the highest ranking as the target material price information.

[0189] The objective of this invention is to target a specific original query q. i Select a unique target record r from the record set R. i :

[0190]

[0191] in, Indicates for all; q i Let ∈ represent the i-th original query, ∈ denotes "belongs to", and Q denotes the original query set. This indicates that there exists a unique one, R represents the set of records, st represents such that, arg max represents returning the record that maximizes the similarity, and Sim(·,·) represents the semantic similarity function.

[0192] 1. Experimental data:

[0193] (1) Price database statistics

[0194] The data in this experiment comes from actual engineering pricing scenarios to ensure the authenticity and practicality of the experiment. The price database serves as the target dataset for retrieval and contains 30,000 engineering pricing records, covering fields such as serial number, project code, project name, project characteristics, work content description, unit of measurement, 24-year labor guidance price, materials supplied by the client, main material loss rate, and remarks.

[0195] (2) Query set:

[0196] The query set, based on a real-world pricing example, contains 500 search requests to simulate actual search scenarios. Fields include sequence number, project code, project name, project feature description, unit of measurement, labor cost, and brand. During the query process, matching items are retrieved from the price database based on the project name and feature description in the query set, which are then used to populate the project code and labor cost fields. The query set retains textual noise present in real-world applications (such as non-standard expressions), effectively validating the model's ability to handle complex inputs.

[0197] 2. Experimental Setup:

[0198] In its implementation, this paper's method relies on existing open-source models and classic retrieval tools. During the query semantic enhancement stage, the large language model Qwen3-32B is employed, which possesses strong semantic understanding and instruction-following capabilities, effectively mitigating noise, abbreviations, and ambiguous expressions in the original query, thereby generating a more semantically standardized enhanced query. The dense semantic channel uses Qwen3-Embedding-8B as the text encoder, mapping queries and records to a shared vector space; the sparse term channel is implemented based on BM25, leveraging its advantages in matching explicit terms and numerical parameters. The scores of both channels are Min-Max normalized along the query-by-query dimension and then fused using linear weighting to form a hybrid channel. The validation set uses a fusion weight set to [value missing]. α =0.7, achieving a better balance between semantic information and term matching. In verifying the impact of reranking on algorithm performance, the qwen3-reranker-0.6B model [Qwen3Embedding: Advancing Text Embedding and Reranking Through Foundation Models] was used.

[0199] 3. Evaluation Metrics

[0200] To comprehensively evaluate the effectiveness of the proposed method in both candidate generation and ranking stages, this paper selects Recall@K and Mean Reciprocal Rank (MRR) as core metrics to evaluate the system performance from the two dimensions of coverage and ranking quality, respectively.

[0201] First, Recall@K measures whether, during the candidate generation phase, a true target record can be included in a candidate set of size K. It is defined as follows:

[0202]

[0203] Where N represents the number of test queries, r i For the i-th query, Represents the corresponding candidate set. • This is an indicator function. A high Recall@K value indicates that the candidate generation module can effectively cover the correct answer.

[0204] Secondly, MRR is used to evaluate the overall quality of the ranking phase, particularly the relative position of the target record within the candidate set. It is defined as:

[0205]

[0206] Among them, rank i Represents the actual target record r i The position of the correct answer in the ranking results. A higher MRR value indicates that the model is able to more consistently rank the correct answer at the top, thus improving the final discrimination performance.

[0207] 4. Baselines

[0208] To verify the effectiveness of the proposed model, four representative retrieval models were selected as baselines:

[0209] (1)Qwen3-Embedding-8B[Qwen3 Embedding:Advancing Text Embedding andReranking Through Foundation Models]

[0210] The Qwen3 series features the latest proprietary text embedding model with 8 billion parameters, supports 100+ natural languages ​​and multiple programming languages, and is widely used in retrieval, clustering, and ranking tasks.

[0211] (2)SimCSE[Fengshenbang 1.0:Being the Foundation of Chinese CognitiveIntelligence]

[0212] SimCSE is an unsupervised sentence vector representation model within a contrastive learning framework, capable of capturing sentence-level semantic similarity. The Erlangshen-SimCSE-110M-Chinese version we selected is optimized for Chinese, with a parameter size of 110 million, and is commonly used for Chinese retrieval and similarity calculation tasks.

[0213] (3)BM25[The Probabilistic Relevance Framework:BM25 and Beyond]

[0214] The classic probability-based sparse retrieval model takes into account factors such as word frequency, inverse document (record) frequency, and document (record) length normalization, and has robust performance in explicit keyword and parameter matching.

[0215] (4)TF-IDF[A statistical interpretation of term specificity and its application in retrieval]

[0216] By combining term frequency (TF) and inverse document (record) frequency (IDF) to measure term importance, the discriminative power of sparse representations can be improved, making it one of the most commonly used benchmark retrieval methods.

[0217] 5. Experimental Results

[0218] To comprehensively evaluate the effectiveness of the proposed method, experiments were conducted on a real engineering material price retrieval dataset, and the analysis was carried out from two aspects: overall performance and ablation comparison.

[0219] Table 1. Analysis of the main experiment

[0220] Model Recall@20 MRR@20 Recall@30 MRR@30 BM25 78.12％ 50.97％ 81.16％ 49.13％ TF-IDF 78.67％ 51.07％ 81.16％ 49.74％ SimCSE 23.55％ 15.76％ 29.09％ 18.16％ Qwen3-Embedding 84.76％ 50.86％ 88.37％ 50.10％ Ours 86.98％ 56.82％ 90.58％ 56.43％

[0221] Table 1 shows a comparison of retrieval performance under different channel combinations. The results indicate that a single channel has limitations in both recall and ranking quality, while the dual-channel fusion scheme proposed in this invention achieves better recall and ranking.

[0222] MRR achieved optimal performance in both dimensions.

[0223] From the single-channel results, the dense semantic channel based on Qwen3-Embedding achieved recall rates of 84.76% and 88.37% for Recall@20 and Recall@30, respectively, indicating its ability to effectively capture semantic synonyms and contextual associations, demonstrating strong recall capabilities. However, its MRR index was relatively low, reflecting shortcomings in ranking accuracy. In contrast, the sparse term channel based on BM25 achieved recall rates of 78.12% and 81.16% for Recall@20 and Recall@30, respectively, significantly lower than the dense channel, but its MRR performance was similar, indicating that term matching still plays an important role in accurate ranking.

[0224] Overall, the dual-channel fusion scheme proposed in this invention achieves recall rates of 86.98% and 90.58% for Recall@20 and Recall@30, respectively, and MRR@20 and MRR@30 rates of 56.82% and 56.43%, respectively, all significantly outperforming single-channel methods. This result demonstrates that the fusion scheme can fully leverage the recall advantages of dense semantic modeling and the accuracy of sparse term matching, achieving effective complementarity between the two.

[0225] Table 2 Ablation Experiment

[0226]

[0227] Further ablation experiments were conducted to verify the roles of each module in the overall algorithm. The experiments removed or combined the query semantic enhancement, dense semantic channel, sparse term channel, and re-ranking modules, and evaluated them on the same dataset. The results are shown in Table 2.

[0228] First, when using the dense semantic channel (Qwen3 Embedding) alone, Recall@20 reaches 84.76%, indicating that the semantic vector can effectively capture the deep semantic relevance between the query and the material description. However, its MRR@20 is only 50.86%, showing shortcomings in the fine-grained ranking of candidate results. In contrast, when using the sparse term channel (BM25) alone, Recall@20 is only 78.12%, significantly lower than embedding, but its MRR performance is close. This suggests that BM25 has certain advantages in the precise matching of explicit terms, but its overall coverage is limited.

[0229] Furthermore, when query semantic enhancement is introduced, the algorithm's Recall@20 increases from 82.83% to 86.98%, MRR@20 increases from 52.07% to 56.82%, and it also improves by approximately 4 percentage points in both Recall@30 and MRR@30. This indicates that query semantic enhancement can effectively alleviate the semantic drift problem by making key elements such as specification parameters and units explicit, thereby simultaneously improving candidate coverage and ranking accuracy.

[0230] It is worth noting that while introducing a reranking module into single-channel retrieval results can bring some improvement, it failed to provide further gains under the dual-channel fusion framework, and even declined in some metrics. This may be because the fusion strategy itself effectively takes into account both semantics and term matching, resulting in a more balanced candidate result. However, the general reranking model, lacking domain adaptation, weakens this balance. Based on this, the final method abandons the reranker and directly adopts the result of dual-channel fusion and dynamic weight learning.

[0231] Figure 2The impact of different weight parameters α on the model's retrieval performance is demonstrated. The curves show that as α increases, Recall@10, Recall@15, and Recall@20 all exhibit a monotonically increasing trend, indicating that semantic channels contribute significantly to the overall performance. When α is low, the model over-relies on sparse channels, resulting in significantly insufficient performance. As α gradually increases, the Recall value increases rapidly, indicating that dense semantic representation has an advantage in capturing complex query semantics. However, when α exceeds 0.7, the increase in the curve gradually slows down, stabilizing in the range of 0.7 to 0.9, indicating that excessive reliance on semantic channels no longer yields significant benefits.

[0232] Therefore, we can conclude that, on the one hand, semantic channels contribute more to the task than sparse channels, but complete reliance is still not robust enough; on the other hand, the fusion strategy works best with medium to high semantic weights, and its performance is relatively stable in the range of 0.7 to 0.9. Therefore, during the interval pruning process in the inference phase, we set the lower and upper bounds of the intervals to...

[0233] This invention proposes an intelligent retrieval method for engineering material prices based on fine-grained field-aware fusion, effectively alleviating the shortcomings in retrieval accuracy and stability caused by low retrieval efficiency, long processing time, and subjective human intervention. This method achieves standardized expression of non-standardized queries through semantic alignment and structure-aware enhancement; simultaneously, it combines the synergistic fusion of dense semantic channels and sparse term channels, and introduces a dynamic weight learning mechanism to improve term matching accuracy while ensuring semantic understanding. Experimental results show that this method outperforms traditional baselines in metrics such as Recall@K and MRR, achieving a balance between candidate coverage and ranking accuracy.

[0234] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An intelligent search method for engineering material price based on fine-grained field-aware fusion, characterized in that, The method comprises the following steps: Step 1, receiving a set of unstructured original queries and providing a semi-structured material library; Step 2, performing semantic alignment and structure-aware enhancement processing on the original queries in the set of original queries to generate normalized enhanced queries; Step 3, calculating global similarity between the normalized enhanced queries and each record in the material library through a dense semantic channel and a sparse term channel respectively; Step 4, performing field-level resolution on the normalized enhanced queries and each record in the material library, and calculating field-level fine-grained similarity of the dense semantic channel and the sparse term channel on a set of pre-defined key fields respectively; Step 5, performing normalization processing on the global similarity and the field-level fine-grained similarity, obtaining global fusion weights and field-level fusion weights, and performing weighted fusion to generate final fusion similarity; Step 6, ranking all records in the material library according to the final fusion similarity, and outputting the record with the highest ranking as target material price information.

2. The fine-grained field-aware fusion-based engineered material price intelligent retrieval method of claim 1, wherein, The step 1 specifically comprises: Step 11, receiving a set of unstructured raw queries oriented to an engineering material price retrieval scenario where q i denotes the ith raw query described in natural language, i denotes the index of the raw query, and m denotes the total number of raw queries, the raw query containing noise, industry abbreviations, and colloquial expressions; Step 12, providing a semi-structured material library T = (S, R), wherein S is a structure pattern, is a record set, r j represents the jth record, j represents the index of the record, n represents the total number of records, each record r j = f j1 , f j2 , …, f jk is composed of a plurality of heterogeneous fields, wherein f j1 represents the first heterogeneous field under the jth record, f j2 represents the second heterogeneous field under the jth record, f jk represents the kth heterogeneous field under the jth record, k represents the total number of heterogeneous fields; the heterogeneous fields include structured numerical / enumeration fields and free text fields.

3. The method for intelligent search of engineering material price based on fine-grained field-aware fusion according to claim 1, characterized in that, The step 2 specifically comprises: Step 21, performing semantic alignment processing on each original query in the set of original queries to generate a standardized query that is semantically equivalent; Step 22, performing structure-aware enhancement processing on the standardized query to form a structured query field representation; Step 23, splicing the standardized query and the query field representation to generate a normalized enhanced query.

4. The method for intelligent search of engineering material price based on fine-grained field-aware fusion according to claim 3, characterized in that, The step 21 specifically comprises: The semantic alignment processing includes at least one of the following operations: synonym normalization, abbreviation expansion and unit normalization, through synonym normalization mapping, abbreviation expansion rule set and unit normalization conversion function, the noise, abbreviation and ambiguous expression in the original query are converted into a standard form to generate a semantic equivalent standardized query The step 22 specifically comprises: Step 221, defining a field set F: F={Material,Spec,Unit,…} Wherein, the field set is divided into material Material, specification Spec and unit Unit according to the engineering field; Step 222, extracting pre-defined key fields and their corresponding values from the standardized query by rule matching, dictionary mapping or numerical recognition, to obtain a structured query field representation :​ wherein f represents a key field, V f is a value domain of the key field f, z f is a value of the key field f. The step 23 specifically comprises: standardized query structured query field representation concatenating the serialized, resulting normalized enhanced query : Wherein, C(·,·) is a splicing function.

5. The fine-grained field-aware fusion-based engineered material price intelligent retrieval method of claim 4, wherein, The synonym normalization, abbreviation expansion and unit normalization specifically comprise: The synonym normalization specifically is: mapping an abbreviation or alias w in the domain vocabulary V to a standard word The normalization results in a vocabulary Abbreviation expansion is to replace the abbreviation a with the standard form b by using a rule set R′=(a,b); The unit normalization is a conversion function for a numerical-unit pair (v, u) to convert it into a system of units.

6. The fine-grained field-aware fusion-based engineered material price intelligent retrieval method of claim 1, wherein, The step 3 specifically comprises: Step 31, constructing a dense-sparse dual-channel framework, including a dense semantic channel and a sparse term channel; Step 32, in the dense semantic channel, a pre-trained semantic encoder φ is employed: respectively map the normalized augmented query and each record r j in the material base to query semantic vector and record vector v(r j ): and where Norm(φ) represents an L2 normalization operation; x generally refers to the result of the pre-trained semantic encoder φ; denotes the real number field; denotes a set of d-dimensional real number vectors; The normalized boosted query is calculated by the cosine inner product between vectors and the similarity of each record r in the material base j , obtaining the global dense semantic similarity : where <·,·> denotes the inner product of query semantic vector and record vector v(r j ) of the record r Step 33, in the sparse term channel, based on the word segmentation function τ(·) to get the term set, and using the BM25 model to calculate the normalized enhanced query The similarity between the word items in each record r j in the material library, get the global sparse term similarity : where t denotes a word, tf(t, r j ) is the term frequency of word t in record r j , k1 is a hyper-parameter to adjust the saturation of term frequency, which controls the degree of influence of term frequency tf(t, r k ) on the final relevance score; b ∈ [0, 1] is used to control the strength of normalization; |r j | denotes the length of record r j , avg dl is the average record length; idf(t) denotes the inverse record frequency, defined as: where N is the total number of records, n t is the number of records containing the word t.

7. The fine-grained field-aware fusion-based engineered material price intelligent retrieval method of claim 1, wherein, The step 4 specifically comprises: Step 41, respectively resolving the corresponding contents of the same key fields from the normalized enhanced query and each record in the material library; Step 42, for each key field f, compute the similarity between its field content vectors using the semantic encoder of the dense semantic channel, obtaining field-level dense similarity Step 43, for each key field f, compute the similarity between its field content terms using the BM25 model of the sparse term channel, to get the field-level sparse term similarity 8. The fine-grained field-aware fusion-based engineered material price intelligent retrieval method of claim 1, wherein, The step 5 has the following steps: Step 51, normalize the global dense semantic similarity and the global sparse term similarity by Min-Max normalization operator respectively Global sparse term similarity Field-level dense similarity and field-level sparse term similarity to the scale of [0, 1] interval. wherein represents the normalized results, represents the normalized results, represents the normalized results, represents the normalized results; Step 52, inputting the semantic vector of the normalized enhanced query into the trained multi-layer perception model to predict the global fusion weight and the plurality of field-level fusion weights for the current query; Step 53, linearly weighting the normalized global dense semantic similarity and the normalized global sparse term similarity using the global fusion weight to obtain a dual-channel fusion similarity of the global semantic : wherein a i ∈ [0,1] is a global fusion weight; Step 54, linearly weighting the normalized field-level dense similarity and the normalized field-level sparse similarity of the corresponding field using the respective field-level fusion weight, to obtain the dual-channel fusion similarity of the field-level semantics : wherein a i,f ∈ [0,1] is a fusion weight for each field level; Step 55, summing the global semantic bi-channel fusion similarity and the individual field-level semantic bi-channel fusion similarity and dividing by the number of fields plus one to obtain the final fusion similarity : wherein, is a global semantic dual-channel fusion similarity; is a field-level semantic dual-channel fusion similarity.

9. The fine-grained field-aware fusion-based engineered material price intelligent retrieval method of claim 8, wherein, The training process of the multi-layer perception model comprises: 1) constructing training samples: For the i-th normalized enhanced query Enumerate weight values ​​'a' in the discrete weight set A = {0..1, 0.2, ..., 0.9}, and select those that allow the true target record to be recorded. Enter based on Optimal weights of the retrieved Top-K ranking candidate set As a reference value; wherein, is an indicator function, r i denotes the records being searched; Top-K denotes the top K records with the highest similarity to the normalized enhanced query . Likewise, for the i-th normalized boosted query Enumerate weight values a in the discrete weight set A = {0.1, 0.2,..., 0.9}, select the one that can make the true target record Enter the based on The optimal weight of the Top-K ranked candidate set retrieved As a reference value; The adaptive weight training sample set T is obtained as follows: wherein Q train is a query training set; 2) with the query semantic vector as input, a multi-layer perceptron g Θ : output fusion weights as prediction where g Θ a neural network representing a multi-layer perceptron that maps a d-dimensional vector to a predicted weight; denotes a set of d-dimensional real-valued vectors; Likewise, the query semantic vector is inputted to a multi-layer perceptron g Θ : output field-level fusion weights as a predicted value: 3) In the training stage, the optimization target is to minimize the mean square error between the predicted value and the reference value, and an e2 regular term is introduced, and the multi-layer perception model is trained by using the AdamW optimizer: wherein Θ represents a learning parameter of the weight assignment model, represents a Loss function guiding the training of the neural network; λ is a regularization coefficient, represents a regularization term.

10. The fine-grained field-aware fusion-based engineered material price intelligent retrieval method of claim 9, wherein, In the inference phase, performing an interval clipping operation on the fusion weight predicted by the multi-layer perceptron model In the inference phase, performing an interval clipping operation on the fusion weight predicted by the multi-layer perceptron model where clip( ) denotes clipping the weight to the closed interval [0, 1], is chosen as [0.7, 0.9].

Images