A tender information retrieval method and device based on a multi-path hybrid recall mechanism

Abstract

The application provides a kind of based on multi-path mixed recall mechanism's tender information retrieval method and device, method includes: constructing tender information knowledge base;Receive the query sentence input by user;Query sentence is parsed, and the matched knowledge slice in tender information knowledge base is retrieved as first recall result set;Query sentence is input to Sentence-BERT model to generate query vector, and the matched knowledge slice is retrieved according to query vector as second recall result set;Each knowledge slice and query sentence are respectively formed into query condition and input to ColBERT model for analysis, and the knowledge slice associated with context is output as third recall result set;First recall result set, second recall result set and third recall result set are integrated to generate mixed recall result set, and query sentence and mixed recall result set are input to ERNIE model, and output retrieval result.The method can improve the accuracy of tender information retrieval.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information retrieval, in particular to a bidding information retrieval method and device based on a multi-path hybrid recall mechanism. BACKGROUND

[0002] With the expansion of the scale of electric power material bidding, the technical parameters (such as "transformer insulation level") contained in the bidding documents, the qualification clauses (such as "supplier test report requirements") and the associated specifications (such as "material acceptance standard") present multi-modal and fragmented characteristics. The traditional retrieval system relies on a single algorithm (such as keyword matching or semantic model), which is difficult to balance precise parameter query and implicit semantic association, resulting in low efficiency of bidding review and poor policy adaptability.

[0003] Patent document CN118733732A discloses a multi-path recall retrieval electric power knowledge service method and device, which includes retrieving a retrieval vector by a retriever to obtain retrieval information; adopting a preset recall strategy to match the retrieval information in a vector database to obtain a matching result, and reordering the matching result by hybrid retrieval; inputting the reordered matching result into an electric power knowledge service large model to generate an answer result, and inputting the query information and the answer result into a security protection model for judgment. In this method, "sparse + dense" dual-channel recall is adopted, but fine-grained interactive modeling is not introduced, which cannot solve the context binding problem of technical clauses and acceptance standards in electric power material bidding, resulting in a high rate of missing key context information and difficulty in balancing retrieval efficiency and business coverage. SUMMARY

[0004] The present application provides a bidding information retrieval method and device based on a multi-path hybrid recall mechanism, which can improve the accuracy and reliability of information retrieval.

[0005] A bidding information retrieval method based on a multi-path hybrid recall mechanism, comprising:

[0006] Constructing a bidding information knowledge base, the attributes of the knowledge slices in the bidding information knowledge base including standard text segments, TF-IDF index vectors and dense semantic vectors;

[0007] Receiving a query sentence input by a user;

[0008] Parsing the query sentence, and retrieving, according to the parsing result, a knowledge slice matched and having a TF-IDF index vector as an attribute in the bidding information knowledge base as a first recall result set;

[0009] Inputting the query sentence into a Sentence-BERT model to generate a query vector, and retrieving, according to the query vector, a knowledge slice matched and having a dense semantic vector as an attribute in the bidding information knowledge base as a second recall result set;

[0010] Traverse the tender information knowledge base, input each knowledge slice and the query statement into the ColBERT model respectively to form a query condition for analysis, and output context-related knowledge slices as a third recall result set;

[0011] Integrate the first, second, and third recall result sets to generate a hybrid recall result set, and input the query statement and the hybrid recall result set into an ERNIE model trained in advance for fine rearrangement, and output a search result.

[0012] Further, the tender information knowledge base is constructed, including:

[0013] The tender document is structurally parsed to extract technical parameter fields, qualification clause fields, and associated specification fields and to standardize professional terms, thereby obtaining standard text segments about technical parameters, qualification clauses, and associated specifications;

[0014] A TF-IDF index vector is generated based on the standard text segments;

[0015] The standard text segments are input into a Sentence-BERT model to generate a dense semantic vector;

[0016] The standard text segments, the TF-IDF index vector, and the dense semantic vector are stored as knowledge slices.

[0017] Further, the query statement is parsed, and based on the parsing result, a knowledge slice that matches and has a TF-IDF index vector attribute in the tender information knowledge base is retrieved as a first recall result set, including:

[0018] It is identified whether there is an operator in the query statement, if there is an operator, a triple containing a field name, an operator, and a numerical value is generated based on the query statement, and if there is no operator, a keyword is extracted from the query statement;

[0019] A first matching degree of the triple with each knowledge slice having a TF-IDF index vector attribute in the tender information knowledge base is calculated based on a BM25 algorithm, or a second matching degree of the keyword with each knowledge slice having a TF-IDF index vector attribute in the tender information knowledge base is calculated based on the BM25 algorithm;

[0020] The knowledge slice corresponding to the triple with a first matching degree greater than a preset matching degree threshold, or the knowledge slice with a second matching degree greater than a preset matching threshold is taken as the first recall result set.

[0021] Further, retrieving, as a second recall result set, knowledge slices that match the query vector and have a dense semantic attribute in the tender information knowledge base according to the query vector, comprising:

[0022] dimensionally reducing the query vector;

[0023] clustering and dividing the knowledge slices with a dense semantic attribute in the tender information knowledge base to obtain a plurality of clustering units;

[0024] calculating the distance between the dimensionally reduced query vector and the clustering center of each clustering unit, and selecting a clustering unit with a distance less than a preset distance value as a target clustering unit;

[0025] PQ encoding the knowledge slices in the target clustering unit to obtain a PQ encoding vector;

[0026] calculating the first cosine similarity between the dimensionally reduced query vector and each PQ encoding vector, and selecting a knowledge slice corresponding to a PQ encoding vector with a first cosine similarity greater than a preset value as the second recall result set.

[0027] Further, inputting each knowledge slice and the query statement into a ColBERT model as a query condition for analysis, respectively, to output a context-related knowledge slice as a third recall result set, comprising:

[0028] The ColBERT model performs word segmentation processing on the query statement in the query condition to generate a first token sequence, and performs word segmentation processing on the knowledge slice to generate a second token sequence;

[0029] Converting the first token sequence and the second token sequence in the same query condition to generate a first word embedding vector sequence and a second word embedding vector sequence;

[0030] Calculating the second cosine similarity between each first word embedding vector in the first word embedding vector sequence and each second word embedding vector in the second word embedding vector sequence, respectively, and taking the maximum value of the second cosine similarity as the association contribution value of the first word embedding vector and the corresponding knowledge slice;

[0031] Summing all the association contribution values obtained in the same query condition to obtain the total association degree of the query statement corresponding to the knowledge slice;

[0032] Contextually binding a knowledge slice with a total association degree greater than a preset association degree threshold as the third recall result set.

[0033] Further, integrating the first recall result set, the second recall result set, and the third recall result set to generate a hybrid recall result set, comprising:

[0034] The first, second and third recall result sets are subjected to a set operation and a deduplication process to generate the mixed recall result set.

[0035] Further, the ERNIE model comprises an input layer, a Transformer encoder, an interaction aggregation module and an output layer.

[0036] The query statement and the mixed recall result set are input into a pre-trained ERNIE model for fine rearrangement, and a retrieval result is output, comprising:

[0037] The input layer receives the query statement and the mixed recall result set, and concatenates, encodes and vector converts the query statement and each knowledge slice in the mixed recall result set respectively to generate a concatenated vector.

[0038] The Transformer encoder perceives the interaction of the concatenated vector based on a multi-head attention mechanism to generate a context perception vector.

[0039] The interaction aggregation module aggregates global semantic information of the context perception vector to generate a global semantic vector.

[0040] The output layer classifies tasks based on the global semantic vector to output a matching probability of the query statement and the corresponding knowledge slice.

[0041] The knowledge slices in the mixed recall result set are arranged in descending order of the matching probability as the retrieval result.

[0042] Further, in the process of pre-training the ERNIE model, the loss function is the sum of a cross-entropy loss function and a business weight constraint term.

[0043] The business weight constraint term is obtained by summing the absolute values of the differences between all predicted business category probabilities and corresponding preset business category weight values and multiplying by a hyperparameter.

[0044] Further, the full connection layer in the output layer of the ERNIE model introduces LoRA parameters.

[0045] The method further comprises:

[0046] Detecting whether the same query statement is input by the user more than a preset number of times, and if the same query statement is input by the user more than a preset number of times, generating a first incremental learning sample according to the retrieval result corresponding to the same query statement.

[0047] Freezing the main parameters of the ERNIE model, inputting the first incremental learning sample into the ERNIE model for training, and fine-tuning the LoRA parameters.

[0048] Or,

[0049] Based on the newly added tender document, a second incremental learning sample is constructed, the main parameters of the ERNIE model are frozen, the second incremental learning sample is input into the ERNIE model for training, and the LoRA parameters are fine-tuned.

[0050] A tender information retrieval device based on a multi-path hybrid recall mechanism, comprising:

[0051] A knowledge base construction module for constructing a tender information knowledge base, the attributes of the knowledge slices in the tender information knowledge base including standard text segments, TF-IDF index vectors and dense semantic vectors;

[0052] A receiving module for receiving a query sentence input by a user;

[0053] A first retrieval module for parsing the query sentence and retrieving matching knowledge slices with TF-IDF index vectors as the first recall result set from the tender information knowledge base according to the parsing result;

[0054] A second retrieval module for inputting the query sentence into a Sentence-BERT model to generate a query vector and retrieving matching knowledge slices with dense semantic vectors as the second recall result set from the tender information knowledge base according to the query vector;

[0055] A third retrieval module for traversing the tender information knowledge base, inputting each knowledge slice and the query sentence into a ColBERT model respectively to analyze the query conditions, and outputting context-related knowledge slices as the third recall result set;

[0056] A comprehensive retrieval module for integrating the first, second and third recall result sets to generate a hybrid recall result set, inputting the query sentence and the hybrid recall result set into a pre-trained ERNIE model for fine rearrangement, and outputting a retrieval result.

[0057] Further, the knowledge base construction module constructs a tender information knowledge base, comprising:

[0058] Structurally analyzing the tender document, extracting technical parameter fields, qualification clause fields and associated specification fields, and performing professional term standardization to obtain standard text segments about technical parameters, qualification clauses and associated specifications;

[0059] Generating a TF-IDF index vector based on the standard text segments;

[0060] inputting the standard text segment into a Sentence-BERT model to generate a dense semantic vector;

[0061] storing the standard text segment, the TF-IDF index vector, and the dense semantic vector as a knowledge slice.

[0062] Further, the first retrieval module parses the query statement, and retrieves, according to a parsing result, a knowledge slice matched and having a TF-IDF index vector as a property from the tender information knowledge base as a first recall result set, including:

[0063] identifying whether there is an operator in the query statement, and if there is an operator, generating a triple containing a field name, an operator, and a numerical value according to the query statement, and if there is no operator, extracting a keyword from the query statement;

[0064] calculating, based on a BM25 algorithm, a first matching degree of the triple and each knowledge slice having a TF-IDF index vector as a property in the tender information knowledge base, or calculating, based on the BM25 algorithm, a second matching degree of the keyword and each knowledge slice having a TF-IDF index vector as a property in the tender information knowledge base;

[0065] taking the knowledge slice corresponding to the triple with a first matching degree greater than a preset matching degree threshold, or the knowledge slice with a second matching degree greater than a preset matching threshold, as the first recall result set.

[0066] Further, the second retrieval module retrieves, according to the query vector, a knowledge slice matched and having a dense semantic vector as a property from the tender information knowledge base as a second recall result set, including:

[0067] performing dimension reduction processing on the query vector;

[0068] clustering and dividing the knowledge slices having a dense semantic vector as a property in the tender information knowledge base to obtain a plurality of clustering units;

[0069] calculating distances between the query vector after the dimension reduction processing and clustering centers of the clustering units, and selecting a clustering unit with a distance less than a preset distance value as a target clustering unit;

[0070] performing PQ encoding on the knowledge slices in the target clustering unit to obtain a PQ encoding vector;

[0071] calculating a first cosine similarity between the query vector after the dimension reduction processing and each PQ encoding vector, and taking a knowledge slice corresponding to a PQ encoding vector with a first cosine similarity greater than a preset value as the second recall result set.

[0072] Further, the third retrieval module inputs each knowledge slice and the query statement into the ColBERT model respectively as query conditions for analysis, and outputs context-related knowledge slices as a third recall result set, including:

[0073] The ColBERT model performs word segmentation processing on the query statement in the query condition to generate a first token sequence, and performs word segmentation processing on the knowledge slice to generate a second token sequence;

[0074] Based on the first token sequence and the second token sequence in the same query condition, a first word embedding vector sequence and a second word embedding vector sequence are generated by conversion;

[0075] The second cosine similarity between each first word embedding vector in the first word embedding vector sequence and each second word embedding vector in the second word embedding vector sequence is calculated respectively, and the maximum value of the second cosine similarity is taken as the association contribution value of the first word embedding vector and the corresponding knowledge slice;

[0076] All association contribution values obtained in the same query condition are summed to obtain the total association degree of the query statement corresponding to the knowledge slice;

[0077] The knowledge slice with a total association degree greater than a preset association degree threshold is context-bound as the third recall result set.

[0078] Further, the comprehensive retrieval module integrates the first recall result set, the second recall result set, and the third recall result set to generate a hybrid recall result set, including:

[0079] The first recall result set, the second recall result set, and the third recall result set are subjected to set operation and de-duplication processing to generate the hybrid recall result set.

[0080] Further, the ERNIE model includes an input layer, a Transformer encoder, an interaction aggregation module, and an output layer;

[0081] The comprehensive retrieval module inputs the query statement and the hybrid recall result set into the pre-trained ERNIE model for fine rearrangement, and outputs a retrieval result, including:

[0082] The input layer receives the query statement and the hybrid recall result set, and respectively concatenates, encodes, and vector converts the query statement and each knowledge slice in the hybrid recall result set to generate a concatenated vector;

[0083] The Transformer encoder perceives and interacts with the concatenated vector based on a multi-head attention mechanism to generate a context-aware vector;

[0084] The interaction aggregation module performs global semantic information aggregation on the context awareness vector to generate a global semantic vector;

[0085] The output layer performs task classification based on the global semantic vector to output a matching probability of the query statement and the corresponding knowledge slice;

[0086] The knowledge slices in the mixed recall result set are arranged in descending order of the matching probability as the retrieval result.

[0087] Further, in the process of pre-training the ERNIE model, the loss function is the sum of the cross-entropy loss function and a business weight constraint term;

[0088] The business weight constraint term is obtained by summing the absolute values of the differences between all predicted business category probabilities and corresponding preset business category weight values and multiplying by a hyperparameter.

[0089] Further, a full connection layer in the output layer of the ERNIE model introduces LoRA parameters.

[0090] The device further comprises an incremental learning module for:

[0091] Detecting whether the same query statement input by the user exceeds a preset number of times, and if the same query statement input by the user exceeds the preset number of times, generating a first incremental learning sample according to the retrieval result corresponding to the same query statement;

[0092] Freezing the main parameters of the ERNIE model, inputting the first incremental learning sample into the ERNIE model for training, and fine-tuning the LoRA parameters;

[0093] Alternatively,

[0094] Based on the newly added tender document, a second incremental learning sample is constructed, the main parameters of the ERNIE model are frozen, the second incremental learning sample is input into the ERNIE model for training, and the LoRA parameters are fine-tuned.

[0095] The method and device for retrieving tender information based on the multi-path mixed recall mechanism provided by the application have at least the following beneficial effects:

[0096] (1) By constructing a tender information knowledge base containing standardized knowledge slices, using a multi-path mixed recall mechanism of sparse channels, dense channels and mixed channels, combining ERNIE model dynamic rearrangement and LoRA incremental learning, simultaneously meeting the three requirements of precise parameter matching, term semantic expansion and clause context association in power material tendering, and realizing real-time self-adaptation of policy changes, thereby breaking through the bottleneck of retrieval accuracy and dynamic business coverage rate, improving the efficiency of tender review, and providing efficient and accurate knowledge retrieval support for power material tendering;

[0097] (2) Through the ERNIE model dynamic rearrangement optimization, the practicability of the retrieval result is improved, and combined with the LoRA incremental learning technology, when the user multiple click behavior or policy update is detected, the model parameter update is completed, the system can dynamically adapt to the business change, maintains high retrieval precision, effectively improves the power bidding auditing efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0098] Figure 1 The flow chart of one embodiment of the bidding information retrieval method provided by the application based on the multi-path hybrid recall mechanism.

[0099] Figure 2 The flow chart of one embodiment of the bidding information knowledge base construction in the bidding information retrieval method provided by the application based on the multi-path hybrid recall mechanism.

[0100] Figure 3 The flow chart of one embodiment of the sparse channel recall in the bidding information retrieval method provided by the application based on the multi-path hybrid recall mechanism.

[0101] Figure 4 The flow chart of one embodiment of the dense channel recall in the bidding information retrieval method provided by the application based on the multi-path hybrid recall mechanism.

[0102] Figure 5 The flow chart of one embodiment of the mixed channel recall in the bidding information retrieval method provided by the application based on the multi-path hybrid recall mechanism.

[0103] Figure 6 The flow chart of one embodiment of the fine rearrangement in the bidding information retrieval method provided by the application based on the multi-path hybrid recall mechanism.

[0104] Figure 7 The structural schematic diagram of one embodiment of the bidding information retrieval device provided by the application based on the multi-path hybrid recall mechanism. DETAILED DESCRIPTION

[0105] In order to better understand the above technical solutions, the above technical solutions will be described in detail in combination with the drawings in the specification and specific embodiments.

[0106] REFERENCE Figure 1 In some embodiments, a bidding information retrieval method based on a multi-path hybrid recall mechanism is provided, comprising:

[0107] S1, a bidding information knowledge base is constructed, the attributes of the knowledge slices in the bidding information knowledge base include standard text segments, TF-IDF index vectors and dense semantic vectors;

[0108] S2, receiving a query sentence input by a user;

[0109] S3, parsing the query sentence, retrieving a matched knowledge slice with a TF-IDF index vector as a first recall result set from the tender information knowledge base according to a parsing result;

[0110] S4, inputting the query sentence to a Sentence-BERT model to generate a query vector, and retrieving a matched knowledge slice with a dense semantic vector as a second recall result set from the tender information knowledge base according to the query vector;

[0111] S5, traversing the tender information knowledge base, inputting each knowledge slice and the query sentence into a ColBERT model respectively to analyze, and outputting a context-related knowledge slice as a third recall result set;

[0112] S6, integrating the first, second and third recall result sets to generate a hybrid recall result set, inputting the query sentence and the hybrid recall result set into an ERNIE model trained in advance to perform fine rearrangement, and outputting a retrieval result.

[0113] Further, referring to Figure 2 , in step S1, the tender information knowledge base is constructed, including:

[0114] S11, structurally analyzing the tender document, extracting technical parameter fields, qualification clause fields and associated specification fields, and performing professional term standardization to obtain standard text segments about technical parameters, qualification clauses and associated specifications;

[0115] S12, generating a TF-IDF index vector based on the standard text segments;

[0116] S13, inputting the standard text segments into a Sentence-BERT model to generate a dense semantic vector;

[0117] S14, storing the standard text segments, TF-IDF index vectors and dense semantic vectors as knowledge slices.

[0118] Specifically, constructing the tender information knowledge base is the basis of the entire retrieval method, and its quality directly affects the accuracy and efficiency of subsequent retrieval. This step mainly performs in-depth processing on the power material tender document to generate standardized knowledge slices.

[0119] The power material bidding document is structured and analyzed. The power material bidding document usually contains a large amount of information, covering technical parameters, qualification clauses and related specifications. The technical parameter field may involve specific indicators such as the insulation level of the transformer, the rated capacity, etc.; the qualification clause field includes the test report requirements that the supplier needs to have, the qualification certificate type, etc.; the related specification field involves the material acceptance standard, the industry implementation specification, etc. Through a special analysis program, these key fields are extracted from the document, and the unstructured document is converted into structured data, laying the foundation for subsequent processing.

[0120] Further, the standardization of power professional terms is carried out. There are a large number of professional terms in the power field, and some terms have aliases, abbreviations, etc., such as "transformer" which may be called "transformer" in different documents, which will affect the accuracy of knowledge retrieval. Therefore, it is necessary to establish a power equipment term mapping table, which includes standard terms, alias terms and associated parameter groups. For example, the standard term "high-voltage circuit breaker" may have an alias such as "high-voltage switch", and the associated parameter group includes rated voltage, rated current, etc. Through the term library, the extracted terms are matched with the standard terms to realize the standardization of the terms and ensure that the same terms with different expressions can be correctly identified in the subsequent retrieval process.

[0121] Specifically, in step S12, the standard text segment is segmented, a word table is constructed, TF (term frequency) and IFD (inverse document frequency) are calculated, TD and IDF are multiplied to obtain TF-IDF value, and the TF-IDF value is converted into vector form according to the order of each word in the word table to obtain the TF-IDF index vector.

[0122] Further, in step S13, the Sentence-BERT model embeds the standard text segment to generate a dense semantic vector.

[0123] Further, in step S14, the final obtained bidding information knowledge base contains the original standard text segment, the corresponding TF-IDF index vector and the dense semantic vector. The original standard text segment is the key information segment extracted from the document, which retains the original form of the information; the TF-IDF index vector is used for subsequent sparse channel recall, which realizes precise matching by calculating the importance of the word in the document; the dense semantic vector generated by Sentence-BERT is used for dense channel recall, which can better capture the semantic information of the text and realize semantic-level retrieval.

[0124] Further, reference is made to Figure 3, step S3 is sparse channel recall, the query statement is parsed, and according to the analysis result, the knowledge slice matched and the attribute is the TF-IDF index vector in the tender information knowledge base is retrieved as the first recall result set, including:

[0125] S31, identify whether there is an operator in the query statement, if there is an operator, generate a triple containing field name, operator and numerical value according to the query statement, if there is no operator, extract the keywords from the query statement;

[0126] S32, based on the BM25 algorithm, calculate the first matching degree of the triple and each knowledge slice with the attribute of TF-IDF index vector in the tender information knowledge base, or based on the BM25 algorithm, calculate the second matching degree of the keywords and each knowledge slice with the attribute of TF-IDF index vector in the tender information knowledge base;

[0127] S33, the triple corresponding to the knowledge slice with the first matching degree greater than the preset matching degree threshold, or the knowledge slice with the second matching degree greater than the preset matching threshold is taken as the first recall result set.

[0128] Specifically, in step S31, the query statement input by the user may contain an operator, for example: "transformer insulation level ≥ F level", and there may be no operator, for example: transformer insulation level. If there is an operator, generate a triple containing field name, operator and numerical value according to the query statement, the triple contains field name, operator and numerical value; For example: transformer insulation level, ≥, F level. If there is no operator, directly extract the keywords.

[0129] Further, in step S32, based on the BM25 algorithm, calculate the first matching degree of the triple and each knowledge slice with the attribute of TF-IDF index vector in the tender information knowledge base, or based on the BM25 algorithm, calculate the second matching degree of the keywords and each knowledge slice with the attribute of TF-IDF index vector in the tender information knowledge base, the calculation formula is:

[0130] ; (1)

[0131] Wherein, score(D, Q) is the matching degree of knowledge slice D and query statement Q, n represents the number of word segmentation items in the query statement, which can be the number of elements in the triple, or the number of keywords, IDF(q i ) represents the i-th word segmentation item q iThe inverse document frequency is given by f(qi,D), which represents the word frequency factor. The higher the word frequency, the more important the word segment is in the document. k1 represents the adjustment parameter, b represents the length penalty parameter, |D| represents the total number of words in knowledge slice D, and avqdl represents the average number of words in all knowledge slices.

[0132] Inverse document frequency (IVF) is calculated using the following formula:

[0133] (2)

[0134] Where N represents the total number of knowledge slices, n(q i ) indicates that it contains the word segment q i The number of knowledge slices.

[0135] Inverse document frequency (IVF) is used to measure the general importance of a term. The fewer documents a term appears in, the higher its IVF and the greater its importance.

[0136] Length normalization term This is used to balance the impact of documents of different lengths on search results, preventing longer documents from receiving higher scores due to containing more terms.

[0137] Set adjustment parameters Controlling word frequency saturation means that once word frequency reaches a certain level, its contribution to the score increases at a slower rate. The length penalty is controlled; a higher value results in a more significant penalty for longer documents. These parameters allow the BM25 algorithm to more accurately calculate the match between documents and queries, thereby retrieving knowledge slices that meet the precise parameter conditions.

[0138] Further, in step S33, for query statements containing operators, the knowledge slices corresponding to triples with a first matching degree greater than a preset matching degree threshold are used as the first recall result set. For query statements not containing operators, the knowledge slices with a second matching degree greater than a preset matching threshold are used as the first recall result set.

[0139] Further, refer to Figure 4 Step S4 is dense channel recall, which involves retrieving matching knowledge slices with dense semantic vector attributes from the bidding information knowledge base based on the query vector, as the second recall result set, including:

[0140] S41. Dimensionally reduce the query vector;

[0141] S42. Cluster the knowledge slices with dense semantic vector attributes in the bidding information knowledge base to obtain multiple clustering units;

[0142] S43, calculate the distance between the query vector after dimensionality reduction processing and the cluster center of each cluster unit, and select the cluster unit with a distance less than a preset distance value as a target cluster unit;

[0143] S44, PQ encode the knowledge slices in the target cluster unit to obtain a PQ encoding vector;

[0144] S45, calculate the first cosine similarity between the query vector after dimensionality reduction processing and each PQ encoding vector, and select the knowledge slice corresponding to the PQ encoding vector with a first cosine similarity greater than a preset value as the second recall result set.

[0145] Specifically, the query statement is input into the Sentence-BERT model to generate a query vector. Sentence-BERT is a BERT-based sentence embedding model that can convert sentences into dense semantic vectors and better capture the semantic information of sentences.

[0146] Further, in step S41, the query vector is subjected to dimensionality reduction processing, specifically, the 768-dimensional dense vector generated by Sentence-BERT is reduced to 64 dimensions by PCA. High-dimensional vectors increase the cost of calculation and storage, and PCA dimensionality reduction can reduce the dimensionality while preserving the main information, thereby improving retrieval efficiency.

[0147] Further, in step S42, the IVF algorithm is used to cluster similar vectors in the same unit by clustering the knowledge slices with dense semantic vectors in the tender information knowledge base, thereby obtaining multiple cluster units. Subsequent retrieval only needs to be performed in part of the relevant units, thereby reducing the amount of calculation.

[0148] Further, in step S43, the distance between the query vector after dimensionality reduction processing and the cluster center of each cluster unit is calculated, and the cluster unit with a distance less than a preset distance value is selected as a target cluster unit.

[0149] Further, in steps S44 and S45, each vector is quantized to an 8-byte index code using PQ (Product Quantization) encoding, the first cosine similarity between the query vector after dimensionality reduction processing and each PQ encoding vector is calculated, and the knowledge slice corresponding to the PQ encoding vector with a first cosine similarity greater than a preset value is selected as the second recall result set.

[0150] Further, with reference to Figure 5 , step S5 is a hybrid channel recall, each knowledge slice and the query statement are input into the ColBERT model for analysis to output context-related knowledge slices as a third recall result set, including:

[0151] S51, the ColBERT model performs word segmentation processing on the query statement in the query condition to generate a first word sequence, and performs word segmentation processing on the knowledge slice to generate a second word sequence;

[0152] S52, based on the first word sequence and the second word sequence in the same query condition, a first word embedding vector sequence and a second word embedding vector sequence are generated by conversion;

[0153] S53, the second cosine similarity between each first word embedding vector in the first word embedding vector sequence and each second word embedding vector in the second word embedding vector sequence is calculated respectively, and the maximum value of the second cosine similarity is taken as the association contribution value of the first word embedding vector and the corresponding knowledge slice;

[0154] S54, sum all the association contribution values obtained in the same query condition to obtain the total association degree of the query statement corresponding to the knowledge slice;

[0155] S55, the knowledge slice with a total association degree greater than a preset association degree threshold is contextually bound as a third recall result set.

[0156] Specifically, all knowledge slices in the bidding information knowledge base are traversed, each knowledge slice and the query statement are composed into a query condition respectively, assuming that there are M knowledge slices in the bidding information knowledge base, then M query conditions are generated.

[0157] In step S51, the ColBERT model performs word segmentation processing on the query statement Q in the query condition to generate a first word sequence: {q1, q2, q3…q n}。

[0158] The knowledge slice D is processed by word segmentation to generate a second word sequence: {d1, d2, d3…d m}。

[0159] Further, in step S52, the ColBERT model converts the first word sequence and the second word sequence in the same query condition to generate a first word embedding vector sequence and a second word embedding vector sequence .

[0160] In step S53, the cosine similarity is calculated by the following formula:

[0161] ; (3)

[0162] wherein E qi represents the i-th word embedding vector in the first word embedding vector sequence E dj represents the second word embedding vector sequence the jth word embedding vector.

[0163] The calculated cosine similarity is taken as the association contribution value of the word embedding vector in the corresponding query condition.

[0164] Further, in steps S54 and S55, all the association contribution values calculated in the same query condition are summed to obtain the total association degree of the query statement corresponding knowledge slice in the query condition. The corresponding knowledge slice in the query condition with a total association degree greater than a preset association degree threshold is contextually bound as the third recall result set.

[0165] Further, in step S6, the first, second, and third recall result sets are integrated to generate a hybrid recall result set, including:

[0166] The first, second, and third recall result sets are subjected to a set operation and a deduplication process to generate the hybrid recall result set.

[0167] Through the above processing, the retrieval results of the three channels can be integrated to avoid missing important information, while removing duplicates and reducing redundancy in subsequent processing.

[0168] Further, the ERNIE model includes an input layer, a Transformer encoder, an interaction aggregation module, and an output layer.

[0169] Reference Figure 6 The query statement and the hybrid recall result set are input into a pre-trained ERNIE model for fine rearrangement, and a retrieval result is output, including:

[0170] S61, the input layer receives the query statement and the hybrid recall result set, and concatenates, encodes, and vector converts the query statement and each knowledge slice in the hybrid recall result set respectively to generate a concatenated vector;

[0171] S62, the Transformer encoder perceives and interacts with the concatenated vector based on a multi-head attention mechanism to generate a context perception vector;

[0172] S63, the interaction aggregation module aggregates global semantic information of the context perception vector to generate a global semantic vector;

[0173] S64, the output layer performs task classification based on the global semantic vector to output a matching probability of the query statement and the corresponding knowledge slice;

[0174] S65, the knowledge slices in the hybrid recall result set are arranged in descending order of matching probability as the retrieval result.

[0175] Specifically, the ERNIE model is a cross-encoder based on ERNIE that can directly calculate the matching degree between query statements and knowledge slices by encoding them, resulting in good ranking performance.

[0176] After the above processing, knowledge slices are output to the power bidding system in descending order of matching probability. The matching probability is the result of comprehensive consideration of multi-path hybrid recall and dynamic reordering optimization. The higher the matching probability of a knowledge slice, the higher the degree of matching with the user query, and the more valuable it is to the user.

[0177] When deploying the power bidding platform, the user query input interface is coupled to the State Grid ECP2.0 bidding system, making it convenient for users to perform query operations in the familiar bidding system; the knowledge slice output format follows the CIM model of the IEC61970 standard, ensuring that the output knowledge is compatible and interactive with other systems in the power industry, thereby improving the integration and usability of the entire system.

[0178] Furthermore, during the pre-training of the ERNIE model, the loss function is the sum of the cross-entropy loss function and the business weight constraint term;

[0179] The business weight constraint term is obtained by summing the absolute values ​​of the differences between the probabilities of all predicted business categories and the corresponding preset business category weight values, and then multiplying by a hyperparameter.

[0180] According to the power business rules, business category weight values ​​are preset, and the business weight constraints follow the rule that qualification certificate weight value > price quotation weight value > delivery option weight value. This is because in power material bidding, the supplier's qualification certificate is the foundation for ensuring material quality and supply capacity, and has the highest priority; the price quotation directly affects the bidding cost and has the second highest priority; the delivery date relates to the timeliness of material supply and has a relatively lower priority.

[0181] In ERNIE model training, the loss function incorporates business weight constraints, as shown below:

[0182] (4)

[0183] in, This represents the loss function of the ERNIE model. Let λ represent the cross-entropy loss function, and let λ represent the hyperparameter. w is a weight constraint term. i Let i be the probability of the predicted business category. Let s be the weight value of the i-th preset business category, and s be the total number of business categories.

[0184] For example, business categories include certifications, quotations, and delivery dates; business weight constraints can be set. wherein, is a qualification certificate business category weight value, is a business category weight value, is a delivery time business category weight value.

[0185] hyperparameters for adjusting the strength of the business constraints, The greater the value, the greater the impact of the business constraints on the loss function, and the more the model tends to follow the preset weight rules;

[0186] Through the action of the ERNIE model and the loss function, the mixed recall result set is sorted according to the weight priority, improving the practicality of the retrieval result.

[0187] Further, the full connection layer in the output layer of the ERNIE model introduces LoRA parameters;

[0188] The method further comprises:

[0189] detecting whether the same query statement input by the user exceeds a preset number of times, and if the same query statement input by the user exceeds the preset number of times, generating a first incremental learning sample according to the retrieval result corresponding to the same query statement;

[0190] Freezing the main parameters of the ERNIE model, inputting the first incremental learning sample into the ERNIE model for training, and fine-tuning the LoRA parameters;

[0191] Alternatively,

[0192] Based on the newly added bidding documents, a second incremental learning sample is constructed, the main parameters of the ERNIE model are frozen, the second incremental learning sample is input into the ERNIE model for training, and the LoRA parameters are fine-tuned.

[0193] Specifically, incremental learning adaptation is used to enable the system to adapt to changes in user behavior and policy updates, maintaining the stability and timeliness of retrieval performance.

[0194] When detecting that the user retrieves the same behavior data or policy update instruction, incremental learning is activated. The triggering conditions of the incremental learning module include that the same query statement input by the user exceeds a preset number of times and that new bidding documents are added. The same query statement is input multiple times, indicating that the query statement may be more in line with the user's needs, and the system needs to learn this preference; the addition of new bidding documents means that the business rules may change, and the system needs to adapt to new policy requirements in a timely manner.

[0195] The incremental learning module adopts the LoRA technology of RoPE position encoding;

[0196] The full connection layer in the output layer of the ERNI model is injected with a rank decomposition matrix, and the parameters of the rank decomposition matrix are fine-tuned based on incremental learning samples. The LoRA technology can realize the rapid adaptation of the model while reducing the amount of calculation by freezing most of the parameters of the pre-trained model and only training the low-rank decomposition matrix.

[0197] The construction of the LoRA bypass matrix is based on the formula , wherein W is the original 1 weight matrix, W' is the constructed rank decomposition matrix, A and B are low-rank adaptation matrices. The operation process is as follows:

[0198] Decompose the original weight matrix , wherein d=1024 is the input dimension, and k=4096 is the output dimension. The original weight matrix is the parameter of the pre-trained model, which contains a large amount of knowledge.

[0199] Construct a low-rank adaptation matrix and , r=128 and r<0.05k. The rank r of the low-rank matrix is much smaller than the dimension of the original weight matrix. By training the two low-rank matrices, new knowledge can be captured, the model can be updated, and the number of parameters and the cost of calculation can be reduced.

[0200] Reference Figure 7 In some embodiments, a tender information retrieval device based on a multi-path mixed recall mechanism is provided, comprising:

[0201] A knowledge base construction module 201 is configured to construct a tender information knowledge base, wherein the attributes of the knowledge slices in the tender information knowledge base include standard text segments, TF-IDF index vectors, and dense semantic vectors.

[0202] A receiving module 202 is configured to receive a query sentence input by a user.

[0203] A first retrieval module 203 is configured to parse the query sentence, retrieve matching knowledge slices with the attribute of TF-IDF index vectors in the tender information knowledge base as a first recall result set according to the parsing result.

[0204] A second retrieval module 204 is configured to input the query sentence into a Sentence-BERT model to generate a query vector, and retrieve matching knowledge slices with the attribute of dense semantic vectors in the tender information knowledge base as a second recall result set according to the query vector.

[0205] A third retrieval module 205 is configured to traverse the tender information knowledge base, input each knowledge slice and the query sentence into a ColBERT model respectively to analyze the query conditions, and output context-related knowledge slices as a third recall result set.

[0206] The integrated retrieval module 206 is configured to integrate the first, second and third recall result sets to generate a mixed recall result set, input the query statement and the mixed recall result set into an ERNIE model trained in advance to perform fine rearrangement, and output a retrieval result.

[0207] Further, the knowledge base construction module 201 constructs a bidding information knowledge base, including:

[0208] The bidding document is structurally parsed to extract technical parameter fields, qualification clause fields and associated specification fields and perform professional term standardization to obtain standard text segments about technical parameters, qualification clauses and associated specifications;

[0209] The standard text segments are used to generate TF-IDF index vectors;

[0210] The standard text segments are input into a Sentence-BERT model to generate dense semantic vectors;

[0211] The standard text segments, TF-IDF index vectors and dense semantic vectors are stored as knowledge slices.

[0212] Further, the first retrieval module 203 parses the query statement and retrieves matching knowledge slices with TF-IDF index vectors as the first recall result set from the bidding information knowledge base according to the parsing result, including:

[0213] Identify whether there is an operator in the query statement, if there is an operator, generate a triple containing a field name, an operator and a numerical value according to the query statement, and if there is no operator, extract keywords from the query statement;

[0214] Calculate the first matching degree of the triple and each knowledge slice with a TF-IDF index vector in the bidding information knowledge base based on the BM25 algorithm, or calculate the second matching degree of the keywords and each knowledge slice with a TF-IDF index vector in the bidding information knowledge base based on the BM25 algorithm;

[0215] The knowledge slice corresponding to the triple with a first matching degree greater than a preset matching degree threshold, or the knowledge slice with a second matching degree greater than a preset matching threshold, is used as the first recall result set.

[0216] Further, the second retrieval module 204 retrieves matching knowledge slices with dense semantic vectors as the second recall result set from the bidding information knowledge base according to the query vector, including:

[0217] The query vector is processed by dimension reduction;

[0218] cluster the knowledge slices with the attribute of dense semantic vectors in the tender information knowledge base to obtain a plurality of clustering units;

[0219] calculate the distance between the query vector after dimension reduction processing and the clustering center of each clustering unit, and select the clustering unit with a distance less than a preset distance value as a target clustering unit;

[0220] PQ encode the knowledge slices in the target clustering unit to obtain a PQ encoding vector;

[0221] calculate the first cosine similarity between the query vector after dimension reduction processing and each PQ encoding vector, and select the knowledge slice corresponding to the PQ encoding vector with a first cosine similarity greater than a preset value as the second recall result set.

[0222] Further, the third retrieval module 205 inputs each knowledge slice and the query statement into the ColBERT model as a query condition for analysis, and outputs the context-related knowledge slice as the third recall result set, including:

[0223] The ColBERT model performs word segmentation processing on the query statement in the query condition to generate a first token sequence, and performs word segmentation processing on the knowledge slice to generate a second token sequence;

[0224] convert the first token sequence and the second token sequence in the same query condition to generate a first word embedding vector sequence and a second word embedding vector sequence;

[0225] calculate the second cosine similarity between each first word embedding vector in the first word embedding vector sequence and each second word embedding vector in the second word embedding vector sequence, and take the maximum value of the second cosine similarity as the association contribution value of the first word embedding vector and the corresponding knowledge slice;

[0226] Sum all the association contribution values obtained in the same query condition to obtain the total association degree of the knowledge slice corresponding to the query statement;

[0227] Contextually bind the knowledge slice with a total association degree greater than a preset association degree threshold as the third recall result set.

[0228] Further, the comprehensive retrieval module 206 integrates the first recall result set, the second recall result set, and the third recall result set to generate a hybrid recall result set, including:

[0229] Perform a set operation on the first recall result set, the second recall result set, and the third recall result set and perform a de-duplication process to generate the hybrid recall result set.

[0230] Further, the ERNIE model comprises an input layer, a Transformer encoder, an interaction aggregation module, and an output layer.

[0231] The comprehensive retrieval module inputs the query statement and the mixed recall result set into a pre-trained ERNIE model for fine rearrangement, and outputs a retrieval result, comprising:

[0232] The input layer receives the query statement and the mixed recall result set, and respectively concatenates, encodes, and vector converts the query statement and each knowledge slice in the mixed recall result set to generate a concatenated vector;

[0233] The Transformer encoder perceives and interacts with the concatenated vector based on a multi-head attention mechanism to generate a context perception vector;

[0234] The interaction aggregation module aggregates global semantic information of the context perception vector to generate a global semantic vector;

[0235] The output layer classifies tasks based on the global semantic vector to output a matching probability of the query statement and the corresponding knowledge slice;

[0236] The knowledge slices in the mixed recall result set are arranged in descending order of the matching probability as the retrieval result.

[0237] Further, in the process of pre-training the ERNIE model, the loss function is the sum of a cross-entropy loss function and a business weight constraint term;

[0238] The business weight constraint term is obtained by summing the absolute values of the differences between all predicted business category probabilities and corresponding preset business category weight values, and multiplying by a hyperparameter.

[0239] Further, the full connection layer in the output layer of the ERNIE model introduces LoRA parameters;

[0240] The device further comprises an incremental learning module for:

[0241] Detecting whether the same query statement input by the user exceeds a preset number of times, and if the same query statement input by the user exceeds the preset number of times, generating a first incremental learning sample according to the retrieval result corresponding to the same query statement;

[0242] Freezing the main parameters of the ERNIE model, inputting the first incremental learning sample into the ERNIE model for training, and fine-tuning the LoRA parameters;

[0243] Or,

[0244] Based on the newly added tender document, a second incremental learning sample is constructed, the main parameters of the ERNIE model are frozen, the second incremental learning sample is input into the ERNIE model for training, and the LoRA parameters are fine-tuned.

[0245] The tender information retrieval method and device based on the multi-path hybrid recall mechanism provided by the above embodiments at least have the following beneficial effects:

[0246] (1) By constructing a tender information knowledge base containing standardized knowledge slices, using a multi-path hybrid recall mechanism of sparse channels, dense channels and mixed channels, combining ERNIE model dynamic rearrangement and LoRA incremental learning, the three requirements of accurate parameter matching, term semantic expansion and clause context association in power material bidding are met, and real-time self-adaptation of policy changes is realized, thereby breaking through the bottleneck of retrieval accuracy and dynamic business coverage rate, improving the efficiency of tender review, and providing efficient and accurate knowledge retrieval support for power material bidding;

[0247] (2) By optimizing the ERNIE model dynamic rearrangement, the practicability of the retrieval result is improved, and combined with the LoRA incremental learning technology, when detecting multiple click behaviors or policy updates of the user, the model parameter update is completed, ensuring that the system can dynamically adapt to business changes, maintain high retrieval accuracy, and effectively improve the efficiency of power bidding review.

[0248] Although the preferred embodiments of the present application have been described, those skilled in the art can make further changes and modifications to these embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application. Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application also intends to include these modifications and variations.

Claims

1. A method for searching tender information based on a multi-path hybrid recall mechanism, characterized in that, The method comprises the following steps: constructing a tender information knowledge base, wherein the attributes of knowledge slices in the tender information knowledge base include standard text segments, TF-IDF index vectors, and dense semantic vectors; receiving a query sentence input by a user; parsing the query sentence, retrieving, according to a parsing result, a knowledge slice that matches and has an attribute of a TF-IDF index vector in the tender information knowledge base as a first recall result set: identifying whether there is an operator in the query sentence, and if there is an operator, generating a triple containing a field name, an operator, and a numerical value according to the query sentence; if there is no operator, extracting a keyword from the query sentence; calculating a first matching degree of the triple and each knowledge slice with an attribute of a TF-IDF index vector in the tender information knowledge base based on a BM25 algorithm, and adding the knowledge slice corresponding to the triple with a first matching degree greater than a preset matching degree threshold to the first recall result set; inputting the query sentence into a Sentence-BERT model to generate a query vector, and retrieving, according to the query vector, a knowledge slice that matches and has an attribute of a dense semantic vector in the tender information knowledge base as a second recall result set: performing dimension reduction processing on the query vector; clustering and dividing the knowledge slices with an attribute of a dense semantic vector in the tender information knowledge base to obtain a plurality of clustering units; calculating the distance between the query vector after dimension reduction processing and the clustering centers of each clustering unit, and selecting a clustering unit with a distance less than a preset distance value as a target clustering unit; performing PQ encoding on the knowledge slices in the target clustering unit to obtain a PQ encoding vector; calculating a first cosine similarity between the query vector after dimension reduction processing and each PQ encoding vector, and adding the knowledge slice corresponding to the PQ encoding vector with a first cosine similarity greater than a preset value to the second recall result set; traversing the tender information knowledge base, inputting each knowledge slice and the query sentence into a ColBERT model respectively to analyze a query condition, and outputting a knowledge slice with context association as a third recall result set: the ColBERT model performs word segmentation processing on the query sentence in the query condition to generate a first word token sequence, and performs word segmentation processing on the knowledge slice to generate a second word token sequence; converting the first word token sequence and the second word token sequence in the same query condition to generate a first word embedding vector sequence and a second word embedding vector sequence; calculating a second cosine similarity between each first word embedding vector in the first word embedding vector sequence and each second word embedding vector in the second word embedding vector sequence, and taking the maximum value of the second cosine similarity as the association contribution value of the first word embedding vector and the corresponding knowledge slice; summing all the association contribution values obtained in the same query condition to obtain a total association degree of the query sentence corresponding to the knowledge slice; contextually binding the knowledge slice with a total association degree greater than a preset association degree threshold as the third recall result set. The first, second and third recall result sets are integrated to generate a mixed recall result set, and the query statement and the mixed recall result set are input into a pre-trained ERNIE model for fine rearrangement, and a retrieval result is output.

2. The method of claim 1, wherein, The tender information knowledge base is constructed, including: The tender document is structurally parsed to extract technical parameter fields, qualification clause fields and associated specification fields and to perform professional term standardization, and standard text segments about technical parameters, qualification clauses and associated specifications are obtained; A TF-IDF index vector is generated based on the standard text segments; The standard text segments are input into a Sentence-BERT model to generate dense semantic vectors; The standard text segments, the TF-IDF index vector and the dense semantic vector are stored as knowledge slices.

3. The method of claim 1, wherein, The query statement is parsed, and matching knowledge slices with the TF-IDF index vector are retrieved from the tender information knowledge base as a first recall result set according to the parsing result, including: The second matching degrees of the keywords and each knowledge slice with the TF-IDF index vector in the tender information knowledge base are calculated based on a BM25 algorithm; and the knowledge slice with a second matching degree greater than a preset matching threshold is taken as the first recall result set.

4. The method of claim 1, wherein, The first, second and third recall result sets are integrated to generate a mixed recall result set, including: The first, second and third recall result sets are subjected to a set operation and a deduplication process to generate the mixed recall result set.

5. The method of claim 1, wherein, The ERNIE model includes an input layer, a Transformer encoder, an interaction aggregation module and an output layer; The query statement and the mixed recall result set are input into a pre-trained ERNIE model for fine rearrangement, and a retrieval result is output, including: The input layer receives the query statement and the mixed recall result set, and concatenates, encodes and vector converts the query statement and each knowledge slice in the mixed recall result set respectively to generate a concatenated vector; The Transformer encoder perceives the interaction of the concatenated vector based on a multi-head attention mechanism to generate a context-aware vector; The interaction aggregation module aggregates global semantic information of the context-aware vector to generate a global semantic vector; The output layer classifies tasks based on the global semantic vector to output a matching probability of the query statement and the corresponding knowledge slice; The knowledge slices in the mixed recall result set are arranged in descending order of the matching probability as the retrieval result.

6. The method according to claim 1 or 5, characterized in that, During pre-training of the ERNIE model, a loss function is a sum of a cross-entropy loss function and a business weight constraint term; The business weight constraint term is obtained by summing the absolute values of the differences between all predicted business category probabilities and corresponding preset business category weight values and multiplying by a hyperparameter.

7. The method of claim 5, wherein, A full connection layer in the output layer of the ERNIE model introduces a LoRA parameter; The method further includes: whether the same query sentence is input more than a preset number of times, if the same query sentence is input more than the preset number of times, generating a first incremental learning sample according to a search result corresponding to the same query sentence; freezing the main parameters of the ERNIE model, inputting the first incremental learning sample into the ERNIE model for training, and fine-tuning the LoRA parameters; or, based on the newly added bidding document to construct a second incremental learning sample, freeze the main parameters of the ERNIE model, input the second incremental learning sample into the ERNIE model for training, and fine-tune the LoRA parameters.

8. A tender information search apparatus based on a multi-path hybrid recall mechanism applied to the method according to any one of claims 1 to 7, characterized in that, including: a knowledge base construction module for constructing a bidding information knowledge base, the attributes of the knowledge slices in the bidding information knowledge base including standard text segments, TF-IDF index vectors and dense semantic vectors; a receiving module for receiving a query sentence input by a user; a first retrieval module for parsing the query sentence, retrieving a matching knowledge slice with a TF-IDF index vector attribute in the bidding information knowledge base as a first recall result set according to the parsing result; a second retrieval module for inputting the query sentence into a Sentence-BERT model to generate a query vector, and retrieving a matching knowledge slice with a dense semantic vector attribute in the bidding information knowledge base as a second recall result set according to the query vector; a third retrieval module for traversing the bidding information knowledge base, inputting each knowledge slice and the query sentence into a ColBERT model for analysis respectively to output a context-related knowledge slice as a third recall result set; a comprehensive retrieval module for integrating the first, second and third recall result sets to generate a hybrid recall result set, inputting the query sentence and the hybrid recall result set into a pre-trained ERNIE model for fine rearrangement, and outputting a search result.

Images