NLP-driven natural language price query intelligent matching method and system

Abstract

The application relates to the technical field of natural language processing, and discloses an NLP-driven natural language price query intelligent matching method and system, wherein the method comprises the following steps: receiving a voice recognition output price inquiry request text and a word-by-word confidence score, and recognizing suspected acoustic confusion words; performing acoustic pre-repair based on a phonological feature similarity, to generate a pre-repaired price inquiry text sequence; performing named entity recognition and multi-category sub-item grouping cutting on the pre-repaired text; verifying and correcting the pre-repaired result based on category semantic context; performing multi-granularity retrieval matching according to a category-specific granularity configuration scheme; performing multi-granularity fusion scoring and sorting on a fusion candidate set; and structurally integrating all sub-item matching results and outputting batch price inquiry results.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and more specifically, to an NLP-driven intelligent matching method and system for price queries. Background Technology

[0002] In the scenario of checking material prices in engineering cost estimation, users input price requests via natural language. The natural language processing system performs entity extraction, semantic parsing, and standardization mapping on the price request text, generates structured query conditions, performs retrieval and matching in the material price database, and outputs candidate material records and price data. When users input batch price request requests containing multiple different categories of materials via voice at the construction site, the speech recognition engine converts the speech into price request text and outputs word-by-word confidence scores.

[0003] Existing natural language processing price matching methods set up an acoustic obfuscation repair module and a multi-category sub-item separation module. First, entity extraction and sub-item separation are performed on the speech recognition text. Then, acoustic obfuscation repair is performed on abnormal words in each sub-item. Finally, standardized mapping and multi-granularity retrieval matching are performed based on the repaired text.

[0004] However, acoustic obfuscation errors generated by speech recognition disrupt the glyphs of material name terms, making it impossible for named entity recognition to correctly locate the boundaries and categories of material name entities in the obfuscated text. With entity recognition failing, multi-category grouping lacks effective entity location anchors, leading to the incorrect merging or splitting of multiple material sub-items. Furthermore, in cases of incorrect sub-item boundaries, words belonging to different material sub-items are mixed into the same context for semantic consistency verification in acoustic repair, generating cross-sub-item semantic interference. This reduces the accuracy of acoustic repair mapping, ultimately resulting in a significant deviation between multi-granularity retrieval matching results and the price data of the target material. Existing technologies suffer from the technical problem of acoustic obfuscation errors propagating step-by-step to entity recognition and grouping stages, causing cascading failures. Summary of the Invention

[0005] This invention provides an NLP-driven intelligent matching method and system for natural language price queries, which solves the technical problems in related technologies where speech recognition engines suffer from errors in material terminology recognition due to acoustic confusion and insufficient accuracy in batch price matching for multiple product categories in engineering cost query scenarios.

[0006] This invention discloses an NLP-driven intelligent matching method for price queries, comprising: receiving a price query request text containing multiple material sub-items converted by a speech recognition engine and the corresponding word-by-word confidence score sequence; performing word segmentation on the price query request text to obtain a word sequence and the confidence score corresponding to each word; and classifying words with confidence scores lower than a preset confidence threshold into a set of suspected acoustic confusion words. For each word in the suspected acoustic confusion word set, extract the phonological feature vector. Based on the phonological feature vector, query the pre-stored phonological approximation word list to obtain the acoustic bridging candidate standard term set. Calculate the phonological feature similarity score between each suspected acoustic confusion word and each candidate standard term. Select the candidate standard term with the highest phonological feature similarity score as the pre-repair replacement word and backfill it into the word sequence to generate the acoustically pre-repaired price query text sequence. Named entity recognition and multi-category sub-item grouping and segmentation are performed on the acoustically pre-repaired price query text sequence to obtain multiple query sub-item units, and the pre-repaired replacement words in each query sub-item unit are marked with a status pending verification. For each query sub-item unit, perform standardized mapping to obtain the material category identifier, obtain the category semantic context feature vector based on the material category identifier, calculate the semantic consistency score between the pre-repaired replacement word of each state to be verified and the semantic context of its category, and perform remapping replacement on the replacement word whose semantic consistency score is lower than the preset verification threshold from its candidate standard term set based on the comprehensive selection score of phonetic feature similarity and semantic consistency. Based on the granularity level configuration scheme of each sub-item's category, the standardized query results of each sub-item are decomposed into multiple granularities. Multiple rounds of semantic retrieval matching are performed independently in order of granularity level from coarse to fine. The matching results of each round are fused, scored, and sorted to output batch price query results.

[0007] Furthermore, the phonological feature similarity score is calculated as follows: the initial consonant sequence similarity, final vowel sequence similarity, and tone sequence similarity between the suspected acoustically confused words and the candidate standard terms are multiplied by their respective weight coefficients and then summed. The initial consonant sequence similarity, final vowel sequence similarity, and tone sequence similarity are all obtained using the normalized edit distance calculation method, and the sum of the three weight coefficients is 1. The encoding process of the phonological feature vector is as follows: each character in the word is converted into a pinyin representation, the pinyin is decomposed into three components: initial consonant, final vowel, and tone, and one-hot encoding is performed on each component and then concatenated to form the phonological feature vector.

[0008] Furthermore, before classifying a word into the suspected acoustic confusion word set, the method further includes: matching words with confidence scores lower than the confidence threshold with a pre-stored whitelist of terms in the field of engineering cost materials; if the word matches an entry in the term whitelist, the word is removed from the suspected acoustic confusion word set and classified into the trusted word set.

[0009] Furthermore, while selecting the candidate standard terms with the highest phonological feature similarity scores as pre-repair replacement words, the candidate standard terms with the top K phonological feature similarity scores are also retained as a set of alternative replacement words. Each alternative replacement word and its corresponding phonological feature similarity score are stored together, so that an alternative can be selected from the set of alternative replacement words when the semantic consistency verification fails, where K is the preset number of alternative candidates.

[0010] Furthermore, the multi-category sub-item grouping and cutting is performed as follows: grouping and cutting are performed based on the location information of the material name entity and the separator mark in the price query text. The separator mark position between two adjacent material name entities is used as the cutting point, and the entities on both sides of the separator mark are respectively assigned to different query sub-item units. When there is no explicit separator mark between adjacent material name entities, the starting position of the material name entity is used as the boundary starting point of the new sub-item for cutting. After grouping and splitting, integrity verification is performed on the entity combination within each query sub-item unit. If a query sub-item unit is missing a material name entity, the entity in the query sub-item unit is merged into the nearest adjacent query sub-item unit. If a query sub-item unit contains two or more material name entities, a second split is performed on the query sub-item unit with the position of each material name entity as the anchor point.

[0011] Furthermore, the calculation method for the comprehensive selection score is as follows: multiply the phonological similarity score between the original confused words and the candidate standard terms by the balance coefficient, and add the complement value of the semantic consistency score between the candidate standard terms and the category semantic context multiplied by the balance coefficient. The value of the balance coefficient is between zero and one. Before the calculation, the semantic consistency score is normalized to mean based on the range, mapping the semantic consistency score to the interval between zero and one, so that the phonological feature similarity score and the semantic consistency score have the same numerical range before the weighted summation operation is performed.

[0012] Furthermore, the independent execution of multiple rounds of semantic retrieval matching in order of granularity from coarse to fine includes: in the coarse-grained round, performing semantic retrieval in the material price database using material category and product form as search conditions to obtain a coarse-grained candidate set; and in the medium-grained round, performing semantic filtering on the coarse-grained candidate set using series-level information as filtering conditions to obtain a medium-grained candidate set. In the fine-grained rounds, parameterized precise matching is performed on the medium-grained candidate set using parameter-level information as the matching condition to obtain the fine-grained candidate set; the filtered records whose matching scores are within the retrieval interval in each round are included in the retrieval set. The lower bound of the retrieval interval is the filtering threshold of the current granular round minus the preset retrieval tolerance value, and the upper bound is the filtering threshold. The fine-grained candidate set and the retrieval set are merged to obtain the fused candidate set.

[0013] Furthermore, the calculation method of the multi-granularity fusion score is as follows: for each candidate material record in the fusion candidate set, calculate its matching degree score in the coarse-grained, medium-grained and fine-grained dimensions respectively, and obtain the multi-granularity fusion matching score by weighted summation according to the granularity weight coefficient of the category to which the query sub-item unit belongs. The sum of the three granularity weight coefficients is 1. A decay factor is applied to the candidate records in the retrieved set, and the multi-granularity fusion matching score of the candidate records in the retrieved set is multiplied by the decay factor to obtain the adjusted fusion matching score. The decay factor is greater than zero and less than one.

[0014] Furthermore, the output of batch price lookup results includes: structurally integrating the sorting and matching results of each query sub-item unit according to the original appearance order of each query sub-item unit in the price lookup request text; calculating the product of the unit price of the first-ranked matching result in each query sub-item unit and the corresponding quantity of that query sub-item unit; and summing the product results of all query sub-item units to obtain the estimated total combined price. Output the independent matching results of each query sub-item, the estimated total price of the combination, and the acoustic repair annotation information. The acoustic repair annotation information includes the original confused words of the words that have undergone acoustic repair in each query sub-item, the speech recognition confidence score, the pre-repaired replacement words, the final replacement words after category perception correction, and the semantic consistency score between the final replacement words and the category semantic context.

[0015] This invention provides an NLP-driven intelligent matching system for price queries, comprising: The confusion identification module is used to receive the price inquiry request text containing multiple material sub-items after being converted by the speech recognition engine and the corresponding word-by-word confidence score sequence, perform word segmentation processing on the price inquiry request text to obtain the word sequence and the confidence score corresponding to each word, and classify words with confidence scores lower than a preset confidence threshold into a suspected acoustic confusion word set; The acoustic pre-repair module is used to extract phonological feature vectors for each word in the suspected acoustic confusion word set, query a pre-stored phonological approximation word list based on the phonological feature vectors to obtain a set of acoustic bridging candidate standard terms, calculate the phonological feature similarity score and select the candidate standard term with the highest score as the pre-repair replacement word to backfill and generate the acoustically pre-repaired price query text sequence. The entity recognition and grouping module is used to perform named entity recognition and multi-category sub-item grouping and segmentation on the acoustically pre-repaired price query text sequence, obtain multiple query sub-item units, and mark the pre-repaired replacement words in each query sub-item unit with a status pending verification. The category awareness correction module is used to perform standardized mapping to obtain material category identifiers for each query sub-item unit, calculate the semantic consistency score of each replacement word to be verified based on the category semantic context feature vector, and perform remapping and replacement on replacement words with semantic consistency scores lower than the preset verification threshold based on the comprehensive selection score; The multi-granularity retrieval and matching module is used to perform multi-granularity decomposition on the standardized query results of each sub-item based on the granularity level configuration scheme of the category to which each sub-item belongs, and independently perform multiple rounds of semantic retrieval and matching in order of granularity level from coarse to fine. The fusion scoring and output module is used to perform multi-granularity fusion scoring and sorting on the matching results of each round, and to structurally integrate the sorted matching results of each query sub-item unit according to the original order of appearance and output the batch price query results.

[0016] This invention addresses the technical problem of cascading failures caused by acoustic confusion errors propagating to entity recognition and grouping stages in multi-material mixed price inquiry scenarios via voice input by performing acoustic pre-repair based on word-by-word confidence scores before named entity recognition. This achieves the technical effect of enabling named entity recognition to correctly locate material name entity boundaries based on the pre-repaired text, providing effective entity location anchors for grouping, and preventing the cascading propagation of acoustic confusion errors. Furthermore, by introducing independent category semantic contexts for each query sub-item after grouping to verify and remap the pre-repair results, it avoids cross-sub-item semantic interference and improves the accuracy of acoustic repair mapping. Finally, by independently loading granularity level configuration schemes by category and performing independent multi-granularity retrieval matching, it adapts the retrieval matching process of each query sub-item to its category granularity structure, obtaining candidate material records and price data appropriate to the price inquiry intent. Attached Figure Description

[0017] Figure 1 This is a flowchart of the NLP-driven intelligent matching method for price queries provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the distribution of confidence scores for word segmentation results provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the similarity of phonetic features of suspected confusing words provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the ranking of the comprehensive similarity of the phonological features of candidate standard terms provided in an embodiment of the present invention; Figure 5This is a schematic diagram of the semantic consistency verification results of the pre-repaired replacement words provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the multi-granularity retrieval and filtering funnel for each sub-item provided in the embodiments of the present invention; Figure 7 This is a schematic diagram illustrating the multi-granularity fusion score of the first-ranked matching record for each sub-item provided in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the composition of the subtotal of the batch price lookup items and the combined total price provided in an embodiment of the present invention; Figure 9 This is a schematic diagram comparing the confidence level and verification score before and after acoustic repair according to an embodiment of the present invention. Detailed Implementation

[0018] In the scenario of checking material prices in engineering cost estimation, users input price requests via natural language. The natural language processing system performs entity extraction, semantic parsing, and standardization mapping on the price request text, generates structured query conditions, performs retrieval and matching in the material price database, and outputs candidate material records and price data. When users input batch price request requests containing multiple different categories of materials via voice at the construction site, the speech recognition engine converts the speech into price request text and outputs word-by-word confidence scores.

[0019] Existing natural language processing (NLP) price matching methods typically employ acoustic obfuscation repair and multi-category sub-item separation modules. These modules first extract entities and separate sub-items from the speech-recognized text, then repair abnormal words within each sub-item using acoustic obfuscation repair, and finally perform standardized mapping and multi-granularity retrieval matching based on the repaired text. However, acoustic obfuscation errors generated by speech recognition disrupt the glyphs of material name terms, making it impossible for the NLP system's named entity recognition to correctly locate the boundaries and categories of material name entities within the obfuscated text. Given the failure of entity recognition, multi-category grouping based on entity location lacks effective entity location anchors to determine the boundaries of each sub-item, leading to the incorrect merging or splitting of multiple material sub-items. Furthermore, when sub-item boundaries are incorrect, subsequent independent acoustic repair of each sub-item results in words belonging to different material sub-items being mixed into the same context for semantic consistency verification, causing cross-sub-item semantic interference. This reduces the accuracy of the acoustic repair mapping, ultimately causing the structured query conditions to deviate from the user's true price-checking intent, and resulting in a significant deviation between the multi-granularity retrieval matching results and the target material's price data.

[0020] Therefore, the technical problem to be solved in this implementation is how to solve the cascading failure problem caused by the propagation of acoustic confusion errors to the entity recognition and grouping segmentation stages in the multi-material mixed price query scenario of voice input.

[0021] According to an embodiment of this invention, an NLP-driven intelligent matching method for natural language price queries is provided. It should be understood that the execution entity of the method described in this embodiment is a computing device equipped with natural language processing capabilities. This computing device pre-stores a material price database and a phonological approximation word list. The material price database contains standard names, specifications, price data, and category classification information of engineering cost materials. The phonological approximation word list records the correspondence between standard terms in the material field and their acoustically confounded variants.

[0022] At least one embodiment of the present invention discloses an NLP-driven intelligent matching method for natural language price queries, such as... Figure 1 As shown, it includes the following steps: Step 1: Receive the price inquiry request text and word-by-word confidence scores, perform word segmentation processing, and identify suspected acoustically confused words; The system receives a natural language price inquiry request text containing multiple material sub-items, converted by a speech recognition engine, along with a word-by-word confidence score sequence corresponding to the price inquiry request text. It then performs word segmentation processing on the price inquiry request text using a word segmentation model to obtain a word sequence. and the confidence scores for each word. ,in For the first One word, For the first The confidence score corresponding to each word The total number of words, This is the position index within the word sequence. For each word... Confidence score Compared with the preset credibility threshold Comparisons were made, and those with confidence scores below the confidence threshold were... The words are classified into the suspected acoustic confusion word set. The confidence score should not be lower than the confidence threshold. The words are included in the set of credible words. .

[0023] It should be noted that the word-by-word confidence score mentioned above is the posterior probability value calculated by the speech recognition engine for each output word during the acoustic decoding process. This posterior probability value reflects the degree of confidence of the speech recognition engine in the correctness of the current word recognition result, and its value ranges from [value missing]. The closer the score is to 1, the higher the reliability of the recognition result.

[0024] It should be noted that the above credibility thresholds The confidence threshold is an empirical value pre-determined based on historical data of voice recognition for engineering cost and material price inquiries. The value of should be chosen to ensure a high recall rate for acoustically confusing error words; for example, the confidence threshold can be set. Set to a value between 0.7 and 0.85.

[0025] In this embodiment of the application, in order to reduce the probability of correctly identified low-frequency technical terms being misclassified as suspected confusing words, the words are classified into the suspected acoustic confusing word set. Previously, the process also included the following steps: assigning confidence scores below a confidence threshold. The word is matched against a pre-stored whitelist of terms in the field of engineering cost and materials. If the word matches an entry in the whitelist, it is removed from the set of suspected acoustic confusion words. Removed from and added to the set of credible words .

[0026] Step 2: Perform acoustic pre-repair on suspected acoustically confused words based on phonological feature similarity to generate a pre-repaired price lookup text sequence; Set of suspected acoustically confusing words Each word in Extracting phonological feature vectors Based on phonological feature vectors Query the pre-stored phonological similar word list to obtain words that are similar to... The acoustic bridging candidate standard term set whose phonological features are within the preset search radius ,in For the first One candidate standard terminology, The total number of candidate terms. This serves as an index for candidate standard terms. It also calculates suspected acoustic confusion terms. Terminology of each candidate standard Phonological feature similarity score between :

[0027] in, For words With candidate standard terminology The similarity of the initial consonant sequences between them For the similarity of vowel sequences, For tone sequence similarity, , , The weight coefficients of each component are given and satisfy the following conditions: Among them, the similarity of initial consonant sequences similarity of vowel sequences Harmony and tone sequence similarity All values ​​were obtained using the normalized edit distance calculation method. Specifically, the edit distance between the word and the corresponding syllable component sequence of the candidate standard term was divided by the maximum value of the lengths of the two sequences, resulting in a value range of [value missing]. The similarity score is calculated by adjusting the edit distance; the smaller the edit distance, the higher the similarity score.

[0028] It should be noted that the above-mentioned phonological feature vectors The encoding process is as follows: each character in the word is converted into a Pinyin representation; the Pinyin is decomposed into three components: initial consonant, final vowel, and tone; each component is then encoded using one-hot encoding and concatenated to form the phonological feature vector of the word. Since the initial consonant, final vowel, and tone components are all converted into equal-length numerical vectors through one-hot encoding, the numerical ranges of each component are consistent. , , All three items are within the range of values. The dimensionless similarity scores can be directly weighted and summed.

[0029] Furthermore, the aforementioned preset search radius serves to limit the scope of the phonologically similar word list search: focusing on words suspected of acoustic confusion. Phonological feature vectors For the query center, only the phonological feature vector and the phonological feature vector are considered. Terms whose distance between them does not exceed the preset search radius are included in the acoustic bridging candidate standard terminology set. Entries whose distance exceeds the preset search radius are not included in the subsequent phonological feature similarity calculation, thereby controlling the size of the candidate set while excluding irrelevant terms with excessive phonological differences.

[0030] The candidate standard terms with the highest phonological feature similarity scores were selected as the pre-repair replacement words. , replace the pre-repaired words Replace the original suspected acoustically confusing words in the corresponding positions of the word sequence to generate an acoustically pre-corrected price lookup text sequence. ,in For the first time after acoustic pre-repair Words in each position.

[0031] It should be noted that the above-mentioned phonological approximation lexicon is organized in an inverted structure with the phonological feature vector of the standard term as the index key. Each record contains the standard term text, the phonological feature vector of the standard term, and the material category identifier to which the standard term belongs. The phonological approximation lexicon is pre-built based on the standard term lexicon in the field of engineering cost materials, generating acoustically confounded variants for each standard term and establishing corresponding relationships before storing them.

[0032] In this embodiment of the application, in order to provide alternative repair solutions in the subsequent category perception verification stage, the candidate standard terms with the highest phonetic feature similarity scores are selected as pre-repair replacement words. At the same time, it also retains the similarity scores of phonological features in the order of ranking. The candidate standard terms for the position are used as a set of alternative replacements. ,in The preset number of alternative candidates. For the first The first alternative word for a suspected acoustically confusing term. For the first The second alternative word to a suspected acoustically confusing term. For the first The first suspected acoustic misleading word Each candidate replacement word is stored along with its corresponding phonological feature similarity score. This is used in subsequent steps to retrieve replacement words from the candidate replacement word set if semantic consistency verification fails. An alternative solution was selected.

[0033] Step 3: Perform named entity recognition and multi-category sub-item grouping and segmentation on the acoustically pre-repaired price query text sequence to obtain multiple query sub-item units; Price lookup text sequence after acoustic pre-restoration Perform named entity recognition in the materials domain to extract material name, specification, quantity, and unit of measurement entities from the price query text, and obtain the positional information of each entity in the price query text sequence. Based on the positional information of the material name entities and the delimiters in the price query text, perform grouping and segmentation on the identified entity sequence, aggregating entities belonging to the same material sub-item into a query sub-item unit, and obtaining a set of multiple query sub-item units. ,in For the first Each query sub-item unit, This serves as the index for the query sub-item unit. Simultaneously, the pre-repaired replacement words contained in each query sub-item unit are marked with a pending verification status, and each pending verification replacement word is associated and stored with its corresponding original obfuscated words and the set of acoustic bridging candidate standard terms.

[0034] It should be noted that the aforementioned separators include conjunctions and punctuation marks appearing in the price query text. Conjunctions, such as "plus," "and," and "as well as," indicate a parallel relationship. The grouping and segmentation are performed as follows: the separator position between two adjacent material name entities is used as the segmentation point, and the entities on both sides of the separator are assigned to different query sub-item units; when there is no explicit separator between adjacent material name entities, the starting position of the material name entity is used as the boundary starting point of the new sub-item for segmentation.

[0035] It should be noted that the data structure of each of the above query sub-items includes: a material name entity field, a specification entity field, a quantity entity field, a unit of measurement entity field, an index field for each entity location, and a list of replacement words to be verified field. Each record in the list of replacement words to be verified field contains pre-repaired replacement words, original obfuscated words, and a set of alternative replacement words. And its corresponding phonological feature similarity score.

[0036] In this embodiment of the application, in order to improve the accuracy of grouping and segmentation, after performing grouping and segmentation based on the location of the material name entity and the separator, the following steps are also included: performing integrity verification on the entity combination within each query sub-item unit to determine whether each query sub-item unit contains both a material name entity and a quantity entity; if a query sub-item unit lacks a material name entity, then the entities in the query sub-item unit are merged into the adjacent query sub-item unit with the nearest location to the query sub-item unit; if a query sub-item unit contains two or more material name entities, then a secondary segmentation is performed on the query sub-item unit with the location of each material name entity as the anchor point.

[0037] Step 4: Perform standardized mapping on each query sub-item unit and verify and correct the pre-repair results based on the category semantic context to obtain the standardized query results after category-aware correction; For each query sub-item unit Perform standardized mapping processing independently to convert query sub-item units The material name entity in the database is matched with the standard material names in the material price database to obtain the standard material subject and material category identifier for each sub-item. Based on material category identification Obtain the semantic context feature vector of frequently co-occurring terms in this category from the material price database. Semantic context feature vector The acquisition method is: identified by material category. As a retrieval criterion, the name text of all material records belonging to this category is extracted from the material price database. Word vector encoding is performed on the name text of each material in this category, and the mean vector is calculated. This mean vector is then used as the semantic context feature vector for this category. .

[0038] For query sub-item cells Pre-repair replacement words for each pending verification status Extracting semantic feature vectors ,in For querying sub-item units Inner The pre-repaired replacement words in the unverified state are calculated using cosine similarity. Semantic consistency score with the semantic context of this product category:

[0039] in, Pre-repair replacement words semantic feature vectors, For querying sub-item units The semantic context feature vector of the product category, semantic feature vector The model, Semantic context feature vector The modulus. It should be noted that semantic feature vectors... With semantic context feature vector All are generated by the same word vector encoding model, have the same vector space dimension, and the calculated cosine similarity value ranges from [value range missing]. The dimensionless score can be directly used for threshold comparison.

[0040] semantic consistency score With the preset verification threshold Comparison, among which , verification threshold The value is pre-determined based on historical data statistics of semantic verification in the field of engineering cost materials, and is used to distinguish whether the semantic consistency between the pre-repaired replacement word and the semantic context of its category is sufficient: the semantic consistency score is not lower than the verification threshold. Replacement words, confirm their pre-repair results, and update the status of the replacement words from pending verification to confirmed; for semantic consistency scores below the verification threshold Replacement terms, from its set of candidate standard terms for acoustic bridging. The candidate standard term with the highest comprehensive selection score is selected for remapping and replacement. The comprehensive selection score is calculated as follows:

[0041] in, This is a balance coefficient between phonological similarity and semantic consistency. Original confusing words With candidate standard terminology Phonological similarity scores between them Candidate standard terms With category Semantic consistency scores between semantic contexts. It should be noted that phonological feature similarity scores... The range of values ​​is Semantic consistency score The range of values ​​is Since the two have different dimensions, the semantic consistency score needs to be determined before substituting them into the formula. Perform mean normalization based on the range to adjust the semantic consistency score. Mapped to The interval is used to make two items have the same numerical range before performing a weighted summation operation to obtain the standardized query results of each sub-item after category perception correction.

[0042] Furthermore, the aforementioned balance coefficients The range of values ​​is This is used to control the overall score selection based on phonological feature similarity and semantic consistency. Relative weights in: balance coefficient The larger the value, the greater the influence of phonological feature similarity on the selection of candidate standard terms; the balance coefficient... The smaller the value, the greater the impact of category semantic consistency on the selection of candidate standard terms; the balance coefficient... The specific values ​​are determined in advance based on historical verification data of acoustic repair of engineering cost materials.

[0043] Step 5: Based on the category-specific granularity level configuration scheme, perform multi-granularity decomposition on the standardized query results of each sub-item to generate a set of multi-granularity search units specific to each sub-item; Based on the category perception correction, the material category identifiers corresponding to the main terms of each sub-item standard material are as follows: Each granularity level configuration scheme is loaded from a pre-stored granularity level configuration library. The granularity level configuration scheme defines the mapping relationship between each entity type under the material category and the coarse-grained, medium-grained, and fine-grained levels. Coarse-grained corresponds to category-level information such as material category and product form, medium-grained corresponds to series-level information such as material grade and series model, and fine-grained corresponds to parameter-level information such as specific specifications and dimensional values.

[0044] Each query sub-item unit Standardized terms in the query sub-item unit The granularity level configuration scheme of the category is decomposed to obtain the query sub-item unit. Dedicated coarse-grained search unit Medium-granularity retrieval unit and fine-grained retrieval unit Composed of query sub-item units Multi-granularity retrieval unit set .

[0045] It should be noted that the above particle size distribution schemes vary depending on the material type. For example, for steel, coarse particle size corresponds to steel type (e.g., section steel) and product form (e.g., channel steel), medium particle size corresponds to material grade (e.g., Q235B), and fine particle size corresponds to cross-sectional specifications (e.g., specific values ​​for height and web thickness). For concrete, coarse particle size corresponds to concrete type (e.g., ready-mixed concrete), medium particle size corresponds to strength grade (e.g., C30), and fine particle size corresponds to parameters such as supply method and transportation distance.

[0046] In this embodiment of the application, in order to handle the situation where some granular information is missing in a user's price query request, the following steps are also included when performing multi-granularity decomposition: performing information integrity checks on each granularity level retrieval unit; if the entity information corresponding to a certain granularity level is missing in the query sub-item unit... If the retrieval unit does not exist, the retrieval unit at that granularity level is marked as the default state, and the filtering operation at that granularity level is skipped in subsequent retrieval and matching steps. The candidate set of the previous granularity level is directly passed to the next granularity level.

[0047] Step 6: Perform multiple rounds of semantic retrieval matching independently for each query sub-item unit in order of granularity from coarse to fine, and obtain the fusion candidate set for each sub-item; For each query sub-item unit According to query sub-item unit Multiple rounds of semantic retrieval operations are executed independently at a dedicated granularity level, from coarse to fine.

[0048] Step 601: In the coarse-grained round, query sub-item cells coarse-grained retrieval unit Using material categories and product forms as search criteria, a semantic search is performed in the material price database to obtain a coarse-grained candidate set. .

[0049] Step 602: In the medium-granularity round, use the medium-granularity retrieval unit The series-level information in the data serves as the filtering condition in the coarse-grained candidate set. Perform semantic filtering to obtain a medium-granularity candidate set. .

[0050] Step 603: In the fine-grained rounds, use fine-grained retrieval units. The parameter-level information in the data serves as the matching condition, within the medium-granularity candidate set. Perform parameterized exact matching to obtain a fine-grained candidate set. .

[0051] Step 604: Perform boundary retrieval processing on the records that were filtered out in each round, ensuring that the matching scores in each granularity round fall within the retrieval range. The filtered records are included in the retrieval set. ,in This is the filtering threshold for the current granularity round. For the preset retrieval tolerance value, the fine-grained candidate set is... With the collection Merge to retrieve query sub-item cells Fusion candidate set .

[0052] Furthermore, the filtering threshold The filtering threshold is configured independently for each granularity round. The range of values ​​is This is used to distinguish whether candidate records meet the filtering conditions in the current granularity round, with a matching score not lower than the filtering threshold. Records that match below the filtering threshold will be retained for the next round. Records were filtered out; retrieval tolerance value This is a preset retrieval tolerance value used to determine the lower bound of the retrieval range. The value of satisfies Make the retrieval range It remains within the valid range.

[0053] Furthermore, the semantic retrieval and semantic filtering operations used in each granularity round in steps 601 to 603 are performed as follows: the text information in the retrieval conditions is converted into semantic vectors through a word vector encoding model, and the cosine similarity is calculated with the semantic vectors of the corresponding fields of each material record in the candidate set. The cosine similarity is used as the matching score for that round, and the matching score is not lower than the filtering threshold. Records with matching scores below the filtering threshold will be included in the next round of the candidate set. The records are filtered out and then enter the retrieval judgment process.

[0054] In this embodiment of the application, in order to address the situation where precise matching is missed due to differences in specification parameter descriptions during the fine-grained matching stage, the following steps are included before performing parameterized precise matching in step 603: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] The specifications and parameters in the database undergo numerical normalization, converting parameter values ​​expressed in different units of measurement into the standard units of measurement used in the material price database. The criteria for precise parameter matching are relaxed from strict equality to numerical differences within the preset parameter tolerance range.

[0055] Step 7: Perform multi-granularity fusion scoring and ranking on the fusion candidate set of each sub-item to obtain the ranking matching results of each sub-item; For each query sub-item unit Fusion candidate set Each candidate material record in Calculate the candidate material records separately Matching scores on coarse-grained, medium-grained, and fine-grained dimensions. , and According to the query sub-item unit The granularity weight coefficients of the product category are weighted and fused to obtain a multi-granularity fusion matching score:

[0056] in, , , These are the weight coefficients for the coarse-grained, medium-grained, and fine-grained dimensions, respectively, and satisfy the following conditions: Each weight coefficient is determined by the query sub-item unit. The granularity level configuration scheme for the product category is predefined. Among these, the coarse-grained dimension matching score... Based on candidate material records Material categories and query sub-items coarse-grained retrieval unit Cosine similarity calculation between material categories, and matching score at medium granularity. Based on candidate material records Series model information and medium-granularity retrieval unit The normalized edit distance between the text strings is calculated by dividing the edit distance between the two strings by the maximum of their lengths and then taking the complement. The resulting value ranges from [value missing]. Match score; fine-grained dimension match score Based on candidate material records Specifications and fine-grained search unit The calculation of the numerical similarity between the intermediate parameter values ​​is specifically performed by dividing the absolute value of the difference between the two values ​​by the fine-grained retrieval unit. Take the complement of the parameter value and truncate it to the nearest integer. The interval is used to obtain the parameter similarity score. It should be noted that the coarse-grained dimension matching score... Medium-granularity dimension matching score Match score with fine-grained dimensions The calculation results for all three items are within the range of [value range missing]. The dimensionless scores can be directly weighted and summed.

[0057] According to multi-granularity fusion matching score Sort query sub-items in descending order Fusion candidate set Retrieve query sub-item units from each candidate material record in the database. The sorting and matching results.

[0058] In this embodiment of the application, in order to distinguish the reliability differences of candidate records from different granularity rounds in the fusion score, the multi-granularity fusion matching score is calculated. At the same time, the collection and retrieval were also carried out. Apply a decay factor to candidate records ( ), will retrieve the collection The multi-granularity fusion matching score of candidate records is adjusted to , to make the return collection Candidate records are ranked after normal candidate records with the same matching degree in the sorting results.

[0059] Step 8: Structure and integrate the sorting and matching results of each sub-item according to their original order of appearance, and output the batch price lookup results; The sorted matching results of each query sub-item are structured and integrated according to the original order of appearance of each query sub-item in the price query request text. For each query sub-item, the sorted candidate material information, corresponding price data, and quantity information are output. The product of the unit price of the first-ranked matching result in each query sub-item and the corresponding quantity is calculated. The product results of all query sub-items are then summarized to obtain the estimated total price of the batch price query. :

[0060] in, To query the total number of sub-item cells, For querying sub-item units The unit price corresponding to the top-ranked match result. For querying sub-item units The numerical value corresponding to the medium number of entities. The index variable for summation takes values ​​from 1 to... This iterates through all the query sub-items sequentially. It should be noted that the unit price... The unit is yuan per unit of measurement, and the quantity is numerical. The unit is the quantity value of the corresponding unit of measurement. The product of the two is in yuan. The units of the products of each query sub-item are consistent, and the summation operation can be performed directly to estimate the total price. The unit is yuan.

[0061] Output the independent matching results of each query sub-item and the estimated total price of the combination. The data includes acoustic repair annotation information for each query sub-item, which includes words that have undergone acoustic pre-repair and category perception correction, their original confusion forms, and the final correction results. The structured integration output data is in a multi-level structure organized by sub-item number. The output record for each query sub-item unit contains the following fields: sub-item number, original text fragment of the sub-item, standardized material main term, sorted candidate material record list (each record includes material name, specifications, unit price, data source, and multi-granularity fusion matching score), sub-item quantity and unit of measurement, subtotal of sub-item amount, and acoustic repair annotation list.

[0062] In this embodiment of the application, in order to provide users with traceable information of the repair process so that users can confirm the accuracy of the price query results, when outputting the acoustic repair annotation information, the following is also included: for the words that have undergone acoustic repair in each query sub-item unit, output their original confused words, speech recognition confidence scores, pre-repaired replacement words, final replacement words after category perception correction, and semantic consistency scores between the final replacement words and the category semantic context, so that users can judge the reliability of the repair results.

[0063] This implementation addresses the problem of acoustic obfuscation causing the failure of natural language processing entity recognition cascade in voice input multi-material mixed price query scenarios. By adjusting the processing order through acoustic pre-repair based on word-by-word confidence scores before named entity recognition, the material names and terms corrupted by acoustic obfuscation are restored to a text form close to the standard terminology before entering the entity recognition stage. This is because the text processed by named entity recognition is a price query text sequence that has undergone acoustic pre-repair. Instead of the original obfuscated text, named entity recognition can correctly locate the boundary positions of material name entities based on the recovered term glyphs, thus providing effective entity location anchors for grouping and segmentation. Therefore, acoustic obfuscation errors are blocked in the entity recognition stage and no longer propagate to the grouping and segmentation stage, solving the cascading failure problem caused by acoustic obfuscation errors in existing methods.

[0064] Furthermore, this implementation utilizes the independent category semantic context of each query sub-item unit to perform semantic consistency verification and remapping on the pre-repair results after grouping and segmentation. Because the acoustic pre-repair stage selects replacement words solely based on phonological similarity, lacking semantic constraints related to material categories, it may map potentially acoustically confusing words to standard terms that are phonologically similar but do not match the category. By introducing category semantic context verification after grouping and segmentation, the pre-repair replacement words within each query sub-item unit undergo consistency verification under the semantic constraints of their respective categories. For replacement words with insufficient semantic consistency, a candidate standard term set from the acoustic bridging is used. In the process, candidate standard terms with higher consistency with the current category semantic context are reselected for remapping. Because the verification and correction processes of each query sub-item are independent, words belonging to different material sub-items will not be mixed into the same category semantic context for verification, thus avoiding the impact of cross-sub-item semantic interference on the accuracy of acoustic restoration.

[0065] Furthermore, in this embodiment, each query sub-item unit independently loads a category-specific granularity level configuration scheme and performs independent multi-granularity search matching. This is because materials of different categories have significant differences in granularity structure, and a uniform granularity level configuration scheme cannot adapt to the unique granularity levels of each category. By configuring the granularity level mapping relationship independently by category, the search matching process for each material sub-item can filter and match layer by layer according to the granularity structure of its category. Therefore, each query sub-item unit can obtain candidate material records and price data that are appropriate to its price search intent.

[0066] The following is an example of an application of the present invention, such as... Figures 2-9 As shown, the implementation process is as follows: A cost engineer for a construction project (Project No.: PROJ-HB-2024) requested bulk material price quotes via voice input on a mobile terminal at the construction site. On the morning of March 15, 20XX, the engineer needed to simultaneously query the current market prices of three structural materials: channel steel (Q235B, specification 16), ready-mixed concrete (C30, pumped), and threaded steel bars (HRB400, diameter 20mm). Due to construction noise interference at the site, the voice recognition engine encountered acoustic confusion errors in recognizing some material terms. The system needed to perform acoustic repair and multi-category search matching on the recognized text, ultimately outputting the price data for the three materials and an estimated total combined price.

[0067] The engineer's voice input request for a price quote is converted by a speech recognition engine into the text: "I need to check the price of No. 16 Caogang Q235B concrete, plus 50 cubic meters of C30 pumped ready-mixed concrete, and the price of HRB400 20mm diameter threaded steel bars." The system receives this text and a sequence of word-by-word confidence scores, performs word segmentation using a word segmentation model, and obtains the word sequence and the corresponding confidence score for each word. The confidence score of each word is then compared with a confidence threshold. By comparison, a set of suspected acoustically confusing words was identified. Among them, the confidence score for "Cao Gang" was 0.52, and the confidence score for "Jing Jin" was 0.61, both below the threshold of 0.75, and were initially classified as... Subsequently, a matching test was performed between the two words and the terminology whitelist, but no whitelist entries were found. This confirmed the entry into the whitelist. All other words had a confidence level of at least 0.75 and were therefore included in the set of reliable words. 。

[0068] Table 1 Word Segmentation Results, Confidence Scores, and Classifications

[0069] For the two suspected confusing words "Caogang" and "Jingjin" in , the phonetic feature vectors are extracted respectively, the phonetic approximation word list is queried, the respective acoustic bridging candidate standard term sets are obtained, and the phonetic feature similarity scores are calculated (initial consonants), [[ID=十六]] (final consonants), (tones).

[0070] Taking "Caogang" as an example, its pinyin is cáogāng, and the pinyin of the candidate term "channel steel" is cáogāng. The similarity of the initial consonant sequence , the similarity of the final consonant sequence , the similarity of the tone sequence , and the calculation of the phonetic feature similarity is as follows:

[0071] Taking "Jingjin" as an example, its pinyin is jīngjīn, and the pinyin of the candidate term "ribbed steel bar" is luówéngāngjīn. Due to the different number of characters, after calculating the normalized edit distance of the initial consonant sequence , the final consonant sequence , the tone sequence ; the pinyin of the candidate term "steel bar" is gāngjīn, , , .

[0072]

[0073] Select the candidate standard term with the highest phonetic feature similarity for each word as the pre-repair replacement word, and at the same time retain the top 3 alternative replacement words. "Caogang" is pre-repaired to "channel steel", "Jingjin" is pre-repaired to "steel bar", and after backfilling, a pre-repaired price query text sequence is generated .

[0074] Table 2 Calculation of Phonetic Similarity of Suspected Confusing Words and Pre-repair Results

[0075] The pre-repaired price query text sequence is: "Need to check the price of No. 16 channel steel Q235B, add 50 cubic meters of C30 pumped commercial concrete, and the quotation of HRB400 steel bar with a diameter of 20".

[0076] For the pre-repaired price query text sequence Named entity recognition was successfully performed, locating three categories of material name entities and their associated specifications, quantities, and units of measurement. Based on the delimiters "plus" and "and", the entity sequence was grouped into three query sub-items by determining the cut points. And mark the words from the pre-repaired items in each sub-item as pending verification status.

[0077] because It contains the material name entity "channel steel" (from pre-repair) and the specification entities "Q235B" and "16". The quantity entity is not explicitly given. After the system performs an integrity check, it is marked as the default state, and the quantity is set to be added later. The material name entity "ready-mixed concrete", the specification entity "C30" and "pumped", the quantity entity "50", and the unit of measurement "cubic meters" are included, and the integrity verification has passed. It includes the material name entity "Rebar" (from pre-repair), the specification entity "HRB400" and "Diameter 20", and the quantity entity is default. Each sub-item contains only one material name entity and does not require secondary cutting.

[0078] Table 3 Results of Grouping and Segmenting Multi-Category Sub-Items

[0079] Perform normalization mapping on each of the three query sub-item units to obtain the material category identifier for each sub-item: Mapped to the "Steel Profiles" category (CAT-XG). Mapped to the "ready-mixed concrete" class (CAT-HNT). Mapped to the "Reinforcing Steel" class (CAT-GJ).

[0080] Semantic context feature vectors for each category are extracted from the material price database, and semantic consistency scores are calculated for the replacement words to be verified within each sub-item. , verification threshold .

[0081] The semantic consistency score of "channel steel" in the semantic context of the steel category.

[0082] Semantic consistency score of "reinforcing steel" in the semantic context of the reinforcing steel category

[0083] Since both pre-repaired replacement words passed semantic consistency verification, there is no need to trigger the comprehensive selection score remapping process. The final standardized query results after category-aware correction are shown in the table below.

[0084] Table 4. Category Perception Verification Results and Standardization Query Results

[0085] Based on the material category identifier of each sub-item, the granularity level configuration scheme of the corresponding category is loaded from the granularity level configuration library, and multi-granularity decomposition is performed on the standardized terms of each sub-item to generate a set of multi-granularity retrieval units exclusive to each sub-item.

[0086] (Steel profiles / channel steel categories): Coarse-grained steel is "steel profiles, channel steel", medium-grained steel is "Q235B", and fine-grained steel is "No. 16 (web height 160mm)"; (Commercial concrete categories): Coarse particle size is "Commercial concrete, pumped", medium particle size is "C30", and fine particle size is "Pumped, locally supplied"; (Reinforcing bar categories): Coarse-grained is "reinforcing bar, threaded reinforcing bar", medium-grained is "HRB400", and fine-grained is "20mm in diameter".

[0087] Table 5. Decomposition results of multi-granularity retrieval units for each sub-item

[0088] For each of the three query sub-items, multiple rounds of semantic retrieval matching are performed independently in a coarse-to-medium-to-fine granularity order, with each granularity's filtering threshold set to [value missing]. retrieval tolerance value The retrieval range is .

[0089] by (Channel steel) as an example: For the coarse-grained round, semantic search was performed in the material price database using "section steel, channel steel" to obtain a coarse-grained candidate set (a total of 18 section steel records were retrieved); for the medium-grained round, "Q235B" was used to filter the coarse-grained candidate set, retaining 11 records and filtering 7, of which 2 records had matching scores falling within the range of... The results were included in the re-fishing set; the fine-grained rounds were precisely matched with the medium-grained candidate set using parameterization with the parameter "16 (belly height 160mm)". Before matching, the specification parameters were normalized and 4 results were retained; the fine-grained candidate set and the re-fishing set were merged to obtain a fusion candidate set of 6 results. , Perform the same process to obtain their respective fusion candidate sets.

[0090] Calculate the multi-granularity fusion matching score for each candidate record in the candidate set of each sub-item fusion. The particle size weighting coefficients for each category are predefined by the configuration scheme: For steel profiles, coarse-grained weight is 0.20, medium-grained weight is 0.35, and fine-grained weight is 0.45; for ready-mixed concrete, coarse-grained weight is 0.25, medium-grained weight is 0.45, and fine-grained weight is 0.30; for reinforcing steel, coarse-grained weight is 0.20, medium-grained weight is 0.30, and fine-grained weight is 0.50. A decay factor is applied to candidate records in the retrieved set. .

[0091] by Taking the record ranked first in the fusion candidate set (material number MAT-XG-016-Q235B) as an example, its three-granularity matching scores are 0.92 for coarse granularity, 0.95 for medium granularity, and 0.88 for fine granularity. The fusion matching score is calculated as follows:

[0092] according to After sorting in descending order, the matching results of the first-ranked sub-item are shown in the table below.

[0093] Table 6 shows the top-ranked matching results for each sub-item.

[0094] The sorting and matching results of the three query sub-items are structured and integrated according to their original order of appearance, and the subtotal of each sub-item and the estimated total price of the combination are calculated. In this price search... The quantity is 50 cubic meters. and If the quantity is defaulted, the system will record the default quantity as 1 unit of measurement for estimation and indicate the default status in the output to prompt the user to supplement.

[0095]

[0096] Table 7. Structured Integration Output of Batch Price Lookup

[0097] The data flow throughout the implementation process reflects strict logical connections between steps: the suspected confusing words "Cao Gang" and "Jing Jin" identified in step 1 directly drive the phonetic similarity calculation in step 2, and the pre-repaired text sequence generated in step 2... As input for named entity recognition in step 3, the entity boundaries are correctly located and the three-item grouping is completed. The sub-item units and verification labels output in step 3 are passed to step 4. The category semantic context verifies both pre-repair replacement words. The standardized query results directly enter the category-specific granular decomposition in step 5. The three sets of multi-granularity retrieval units output in step 5 drive the three independent multi-round semantic retrieval links in step 6. After the candidate set is merged and weighted by the fusion score in step 7, it is finally integrated and output in step 8 according to the original sub-item order to form a complete batch price query result containing acoustic repair traceability information. The combined total price estimate is 33,220 yuan.

[0098] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent matching of price queries driven by NLP, characterized in that, Includes the following steps: The system receives a price inquiry request text containing multiple material sub-items, converted by a speech recognition engine, and a corresponding word-by-word confidence score sequence. It then performs word segmentation on the price inquiry request text to obtain a word sequence and a confidence score for each word. Words with confidence scores lower than a preset confidence threshold are classified into a set of suspected acoustic confusion words. For each word in the suspected acoustic confusion word set, extract the phonological feature vector. Based on the phonological feature vector, query the pre-stored phonological approximation word list to obtain the acoustic bridging candidate standard term set. Calculate the phonological feature similarity score between each suspected acoustic confusion word and each candidate standard term. Select the candidate standard term with the highest phonological feature similarity score as the pre-repair replacement word and backfill it into the word sequence to generate the acoustically pre-repaired price query text sequence. Named entity recognition and multi-category sub-item grouping and segmentation are performed on the acoustically pre-repaired price query text sequence to obtain multiple query sub-item units, and the pre-repaired replacement words in each query sub-item unit are marked with a status pending verification. For each query sub-item unit, perform standardized mapping to obtain the material category identifier, obtain the category semantic context feature vector based on the material category identifier, calculate the semantic consistency score between the pre-repaired replacement word of each state to be verified and the semantic context of its category, and perform remapping replacement on the replacement word whose semantic consistency score is lower than the preset verification threshold from its candidate standard term set based on the comprehensive selection score of phonetic feature similarity and semantic consistency. Based on the granularity level configuration scheme of each sub-item's category, the standardized query results of each sub-item are decomposed into multiple granularities. Multiple rounds of semantic retrieval matching are performed independently in order of granularity level from coarse to fine. The matching results of each round are fused, scored, and sorted to output batch price query results.

2. The method according to claim 1, characterized in that, The phonological feature similarity score is calculated as follows: the initial consonant sequence similarity, final vowel sequence similarity, and tone sequence similarity between the suspected acoustically confused words and the candidate standard terms are multiplied by their respective weight coefficients and then summed. The initial consonant sequence similarity, final vowel sequence similarity, and tone sequence similarity are all obtained using the normalized edit distance calculation method, and the sum of the three weight coefficients is 1. The encoding process of the phonological feature vector is as follows: each character in the word is converted into a pinyin representation, the pinyin is decomposed into three components: initial consonant, final vowel, and tone, and one-hot encoding is performed on each component and then concatenated to form the phonological feature vector.

3. The method according to claim 1, characterized in that, Before classifying a word into the suspected acoustic confusion word set, the method further includes: matching words with confidence scores lower than the confidence threshold with a pre-stored whitelist of terms in the field of engineering cost materials; if the word matches an entry in the term whitelist, the word is removed from the suspected acoustic confusion word set and classified into the trusted word set.

4. The method according to claim 1, characterized in that, While selecting the candidate standard terms with the highest phonological feature similarity scores as pre-repair replacement words, the candidate standard terms with the top K phonological feature similarity scores are also retained as a set of alternative replacement words. Each alternative replacement word and its corresponding phonological feature similarity score are stored together, so that an alternative can be selected from the set of alternative replacement words when the semantic consistency verification fails, where K is the preset number of alternative candidates.

5. The method according to claim 1, characterized in that, The multi-category sub-item grouping and cutting is performed as follows: grouping and cutting are performed based on the location information of the material name entity and the separator mark in the price query text. The separator mark position between two adjacent material name entities is used as the cutting point, and the entities on both sides of the separator mark are respectively assigned to different query sub-item units. When there is no explicit separator mark between adjacent material name entities, the starting position of the material name entity is used as the boundary starting point of the new sub-item for cutting. After grouping and splitting, integrity verification is performed on the entity combination within each query sub-item unit. If a query sub-item unit is missing a material name entity, the entity in the query sub-item unit is merged into the nearest adjacent query sub-item unit. If a query sub-item unit contains two or more material name entities, a second split is performed on the query sub-item unit with the position of each material name entity as the anchor point.

6. The method according to claim 1, characterized in that, The calculation method for the comprehensive selection score is as follows: multiply the phonological similarity score between the original confused words and the candidate standard terms by the balance coefficient, and add the semantic consistency score between the candidate standard terms and the category semantic context multiplied by the balance coefficient. The balance coefficient ranges from zero to one. Before the calculation, the semantic consistency score is normalized to mean based on the range, mapping the semantic consistency score to the interval between zero and one, so that the phonological feature similarity score and the semantic consistency score have the same numerical range before the weighted summation operation is performed.

7. The method according to claim 1, characterized in that, The independent execution of multiple rounds of semantic retrieval matching in order of granularity from coarse to fine includes: In the coarse-grained round, semantic retrieval is performed in the material price database using material category and product form as search criteria to obtain a coarse-grained candidate set; In the medium-granularity rounds, semantic filtering is performed on the coarse-granular candidate set using serial-level information as the filtering condition to obtain the medium-granularity candidate set; In the fine-grained rounds, parameterized precise matching is performed on the medium-grained candidate set using parameter-level information as the matching condition to obtain the fine-grained candidate set; the filtered records whose matching scores are within the retrieval interval in each round are included in the retrieval set. The lower bound of the retrieval interval is the filtering threshold of the current granularity round minus the preset retrieval tolerance value, and the upper bound is the filtering threshold. The fine-grained candidate set and the retrieval set are merged to obtain the fused candidate set.

8. The method according to claim 7, characterized in that, The calculation method of the multi-granularity fusion score is as follows: For each candidate material record in the fusion candidate set, calculate its matching degree score in the coarse-grained, medium-grained and fine-grained dimensions respectively, and obtain the multi-granularity fusion matching score by weighted summation according to the granularity weight coefficient of the category to which the query sub-item unit belongs. The sum of the three granularity weight coefficients is 1. A decay factor is applied to the candidate records in the retrieved set, and the multi-granularity fusion matching score of the candidate records in the retrieved set is multiplied by the decay factor to obtain the adjusted fusion matching score. The decay factor is greater than zero and less than one.

9. The method according to claim 1, characterized in that, The output batch price lookup results include: structurally integrating the sorting and matching results of each query sub-item unit according to the original appearance order of each query sub-item unit in the price lookup request text; calculating the product of the unit price of the first-ranked matching result in each query sub-item unit and the corresponding quantity of that query sub-item unit; and summing the product results of all query sub-items units to obtain the estimated total combined price. Output the independent matching results of each query sub-item, the estimated total price of the combination, and the acoustic repair annotation information. The acoustic repair annotation information includes the original confused words of the words that have undergone acoustic repair in each query sub-item, the speech recognition confidence score, the pre-repaired replacement words, the final replacement words after category perception correction, and the semantic consistency score between the final replacement words and the category semantic context.

10. An NLP-driven intelligent matching system for price queries, used to execute the method described in any one of claims 1 to 9, characterized in that, include: The confusion identification module is used to receive the price inquiry request text containing multiple material sub-items after being converted by the speech recognition engine and the corresponding word-by-word confidence score sequence, perform word segmentation processing on the price inquiry request text to obtain the word sequence and the confidence score corresponding to each word, and classify words with confidence scores lower than a preset confidence threshold into a suspected acoustic confusion word set; The acoustic pre-repair module is used to extract phonological feature vectors for each word in the suspected acoustic confusion word set, query a pre-stored phonological approximation word list based on the phonological feature vectors to obtain a set of acoustic bridging candidate standard terms, calculate the phonological feature similarity score and select the candidate standard term with the highest score as the pre-repair replacement word to backfill and generate the acoustically pre-repaired price query text sequence. The entity recognition and grouping module is used to perform named entity recognition and multi-category sub-item grouping and segmentation on the acoustically pre-repaired price query text sequence, obtain multiple query sub-item units, and mark the pre-repaired replacement words in each query sub-item unit with a status pending verification. The category awareness correction module is used to perform standardized mapping to obtain material category identifiers for each query sub-item unit, calculate the semantic consistency score of each replacement word to be verified based on the category semantic context feature vector, and perform remapping and replacement on replacement words with semantic consistency scores lower than the preset verification threshold based on the comprehensive selection score. The multi-granularity retrieval and matching module is used to perform multi-granularity decomposition on the standardized query results of each sub-item based on the granularity level configuration scheme of the category to which each sub-item belongs, and independently perform multiple rounds of semantic retrieval and matching in order of granularity level from coarse to fine. The fusion scoring and output module is used to perform multi-granularity fusion scoring and sorting on the matching results of each round, and to structurally integrate the sorted matching results of each query sub-item unit according to the original order of appearance and output the batch price query results.

Images