Information retrieval method and device, storage medium and electronic equipment

Abstract

The invention provides an information retrieval method and device, a storage medium and electronic equipment, and is applied to the technical field of information processing. The method comprises the following steps: receiving a user query containing a product name, and respectively carrying out mixed search, enhanced search and rollback search to obtain a corresponding candidate document set; and then performing correlation filtering on the three groups of candidate document sets to obtain a target document set. And after performing data source classification on each document in the target document set, outputting a classified retrieval result. According to the method, multiple document sets are comprehensively screened and filtered through the multi-stage retrieval strategy, the recall rate and accuracy of information retrieval in the diagnosis and maintenance field are effectively improved, and the problem that retrieval reliability is insufficient is solved.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, and in particular to an information retrieval method, apparatus, storage medium, and electronic device. Background Technology

[0002] Existing vector-based retrieval systems mostly employ fixed thresholds and single-stage retrieval strategies, combined with exact match product name filtering, in order to improve retrieval accuracy.

[0003] However, these methods have significant shortcomings in diagnostic and repair scenarios. Strict filtering conditions can easily lead to the false exclusion of relevant documents, reducing the recall rate; a single search strategy is difficult to adapt to diverse and incomplete query needs, limiting the system's flexibility and practicality; in addition, the product name matching method is relatively rigid and fails to effectively cover various variations of the name, further affecting the search results.

[0004] Therefore, how to improve the reliability of information retrieval in the field of diagnosis and maintenance has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, the present invention provides an information retrieval method, apparatus, storage medium, and electronic device that overcomes or at least partially solves the above problems, the technical solution of which is as follows:

[0006] An information retrieval method, comprising:

[0007] Receive user queries containing product names;

[0008] A first candidate document set is obtained by performing a mixed search using the product name;

[0009] An enhanced search was performed using the product name to obtain a second set of candidate documents.

[0010] Using the user query, a backward search is performed to obtain a third candidate document set;

[0011] The first candidate document set, the second candidate document set, and the third candidate document set are subjected to relevance filtering to obtain a target document set, wherein the target document set includes at least one document;

[0012] The data source is classified for each document in the target document set, and the classified information retrieval results are output.

[0013] Optionally, the step of performing a mixed search using the product name to obtain a first candidate document set includes:

[0014] Multiple name variations are generated based on the product name;

[0015] The document library is searched using the product name and the name variant to obtain a first candidate document containing the product name or the name variant;

[0016] Calculate the semantic similarity score between each of the first candidate documents and the user query;

[0017] Select the first candidate documents whose semantic similarity scores meet the first threshold condition, and construct the first candidate document set.

[0018] Optionally, the generation of multiple name variations based on the product name includes:

[0019] Convert the product name to lowercase to obtain the first name variant;

[0020] And / or, replace the connectors in the product name with spaces to obtain a second name variant;

[0021] And / or, remove the connector from the product name to obtain a third name variant.

[0022] Optionally, the enhanced search using the product name to obtain a second candidate document set includes:

[0023] Combine the product name with the context in the user query to generate expanded query content;

[0024] The extended query content is converted into a vector representation to obtain the query vector;

[0025] Using the query vector, a semantic vector search is performed in the document library to filter out second candidate documents whose semantic similarity scores meet the second threshold condition, and a second candidate document set is constructed.

[0026] Optionally, the step of using the user query to perform a backward search to obtain a third candidate document set includes:

[0027] Calculate the semantic similarity score between the user query and all documents;

[0028] Select third candidate documents whose semantic similarity scores meet the third threshold condition, and construct a third candidate document set;

[0029] When the number of third candidate documents that meet the third threshold condition is less than the preset minimum recall number, the threshold in the third threshold condition is lowered until the number of the selected third candidate documents reaches the preset minimum recall number.

[0030] Optionally, the step of performing relevance filtering on the first candidate document set, the second candidate document set, and the third candidate document set to obtain the target document set includes:

[0031] Fuzzy matching is performed on each candidate document in the first candidate document set, the second candidate document set, and the third candidate document set to obtain a fuzzy matching score;

[0032] Semantic analysis is performed on each candidate document in the first candidate document set, the second candidate document set, and the third candidate document set to obtain a semantic analysis score;

[0033] Candidate documents whose fuzzy matching scores are equal to or greater than a preset fuzzy matching threshold, or whose semantic analysis scores are equal to or greater than a preset semantic analysis threshold, are selected to construct a target document set.

[0034] Optionally, the documents include product manuals, user manuals, technical documents, and application cases.

[0035] An information retrieval device includes: a user query receiving unit, a hybrid search unit, an enhanced search unit, a fallback search unit, a relevance filtering unit, and a data source classification unit;

[0036] The user query receiving unit is used to receive user queries containing product names;

[0037] The hybrid search unit is used to perform a hybrid search using the product name to obtain a first candidate document set;

[0038] The enhanced search unit is used to perform an enhanced search using the product name to obtain a second candidate document set;

[0039] The backtracking search unit is used to perform a backtracking search based on the user query to obtain a third candidate document set;

[0040] The relevance filtering unit is used to perform relevance filtering on the first candidate document set, the second candidate document set, and the third candidate document set to obtain a target document set, wherein the target document set includes at least one document;

[0041] The data source classification unit is used to classify the data sources of each document in the target document set and output the classified information retrieval results.

[0042] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the information retrieval method.

[0043] An electronic device includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other via the bus; the processor is used to call program instructions in the memory to execute the information retrieval method.

[0044] By employing the above technical solutions, this invention provides an information retrieval method, apparatus, storage medium, and electronic device that receives user queries containing product names; performs a hybrid search using the product names to obtain a first candidate document set; performs an enhanced search using the product names to obtain a second candidate document set; performs a fallback search using the user query to obtain a third candidate document set; performs relevance filtering on the first, second, and third candidate document sets to obtain a target document set, wherein the target document set includes at least one document; classifies the data sources of each document in the target document set, and outputs the classified information retrieval results. This invention, by receiving user queries containing product names and employing a multi-stage retrieval strategy involving hybrid search, enhanced search, and fallback search, comprehensively filters and relevance-based multiple candidate document sets, and finally classifies and outputs the target document set, effectively improves the recall and accuracy of information retrieval in the diagnostic and repair field, and solves the problem of insufficient retrieval reliability.

[0045] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0046] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0047] Figure 1 A flowchart illustrating one embodiment of the information retrieval method provided by this invention is shown.

[0048] Figure 2 The diagram shows a specific implementation of step S110 in the information retrieval method provided in this embodiment of the invention.

[0049] Figure 3 The diagram shows a specific implementation of step S120 in the information retrieval method provided in this embodiment of the invention.

[0050] Figure 4 The diagram shows a specific implementation of step S130 in the information retrieval method provided in this embodiment of the invention.

[0051] Figure 5The diagram shows a specific implementation of step S140 in the information retrieval method provided in this embodiment of the invention.

[0052] Figure 6 A schematic diagram of the structure of the information retrieval device provided in an embodiment of the present invention is shown;

[0053] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present invention is shown. Detailed Implementation

[0054] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0055] With the development of information retrieval technology, vector representation-based retrieval systems are increasingly being applied in various scenarios, especially in the field of product-related diagnosis and repair, where they hold significant value. Existing technologies typically employ fixed thresholds and single-stage retrieval strategies, combined with exact-matching product name filtering mechanisms, in order to improve retrieval accuracy and relevance.

[0056] However, such traditional solutions have significant shortcomings. First, strict filtering conditions often lead to the incorrect exclusion of some relevant documents, reducing the recall rate of the retrieval system, especially in diagnostic and repair scenarios. Second, a single retrieval strategy lacks adaptability to diverse query needs, making it difficult to effectively handle incomplete or complex user queries, thus limiting the system's usability and flexibility. Finally, the product name matching method is relatively rigid, mainly relying on exact matching, which cannot fully cover various variations and expressions of product names, making it difficult to comprehensively recall relevant information.

[0057] Based on this, the present invention provides an information retrieval method. By receiving user queries containing product names, it performs hybrid search, enhanced search, and fallback search to obtain corresponding candidate document sets. Subsequently, relevance filtering is applied to the three candidate document sets to obtain the target document set. After classifying the documents in the target document set according to their data sources, the categorized retrieval results are output. It is evident that the present invention, through a multi-stage retrieval strategy, comprehensively filters and selects multiple document sets, effectively improving the recall and accuracy of information retrieval in the diagnostic and maintenance field, and solving the problem of insufficient retrieval reliability.

[0058] like Figure 1The diagram shows a flowchart of one embodiment of the information retrieval method provided by this invention. The method may include:

[0059] S100: Receive user queries containing product names.

[0060] The product name refers to the specific product identifier included in the user's query, such as "ABC-Smart Lock," which is used to limit the search scope and improve relevance.

[0061] User queries refer to information entered by users, including product names and related requirements, which guides the retrieval system to find matching documents.

[0062] Specifically, in this embodiment of the invention, by receiving a user query containing a product name, the user first obtains the query content submitted by the user, extracts the product name, and prepares for subsequent multi-stage retrieval.

[0063] As examples, embodiments of the present invention can receive query strings input by users through the provided API interface or user interface, perform basic preprocessing on the received query strings including removing extra spaces and uniform character encoding, and identify and extract the core "product name" entity from them.

[0064] S110. Use product names to perform a mixed search to obtain the first candidate document set.

[0065] Hybrid search refers to combining multiple variations of product names and semantic vector technology to rigorously filter documents and obtain a highly accurate first candidate document set.

[0066] The first candidate document set refers to the document set obtained through hybrid search, which includes relevant documents with a high degree of semantic matching with the product name.

[0067] Specifically, in this embodiment of the invention, preliminary document screening can be performed using product names and their variant combinations to obtain preliminary candidate documents; semantic similarity calculation and threshold screening are performed on the preliminary candidate documents to construct a first candidate document set that meets strict relevance standards, so as to achieve dual protection of accurate name matching and semantic relevance, and ensure a high degree of consistency between search results and query intent.

[0068] S120. Use product names to perform enhanced search and obtain a second set of candidate documents.

[0069] Among them, enhanced search refers to integrating product names into the context of user queries and using pure vector search with a more lenient threshold to obtain a more semantically relevant set of second candidate documents.

[0070] The second candidate document set refers to the set of documents obtained by enhanced search, which includes a broad range of documents that are semantically related.

[0071] Specifically, embodiments of the present invention can semantically fuse product names with query context to generate extended queries, obtain semantically relevant documents through vectorized retrieval, and construct a second candidate document set based on relaxed threshold criteria. This approach enhances the coverage of complex query intents while ensuring semantic relevance.

[0072] S130. Use the user query to perform a backtracking search and obtain the third candidate document set.

[0073] Among them, backtracking search refers to performing vector similarity search based solely on the semantics of the user query, relaxing product name restrictions, ensuring the recall of potentially relevant documents, and forming a third candidate document set.

[0074] The third candidate document set refers to the set of documents obtained by backtracking the search, which is used to ensure the basic number of search results.

[0075] Specifically, embodiments of the present invention can perform preliminary screening by directly calculating the semantic matching degree between the query and the document, and use a dynamic threshold adjustment mechanism to ensure the minimum recall size, thereby constructing a third candidate document set with basic relevance, so as to provide a fallback guarantee when strict matching fails and maintain the basic availability of the system's retrieval capabilities.

[0076] S140. Perform relevance filtering on the first candidate document set, the second candidate document set, and the third candidate document set to obtain the target document set, wherein the target document set includes at least one document.

[0077] The target document set refers to the high-quality document set retained after relevance filtering of the candidate document sets at each stage, which serves as the basis for the final retrieval results.

[0078] Specifically, embodiments of the present invention can implement dual scoring and screening, including fuzzy matching and semantic analysis, on multiple candidate document sets. By determining thresholds, a target document set that meets at least one relevance condition is constructed. The surface feature matching and deep semantic association are comprehensively considered to ensure that the final result has both retrieval breadth and semantic accuracy.

[0079] S150. Classify the data sources of each document in the target document set and output the classified information retrieval results.

[0080] Among them, information retrieval results refer to the clearly categorized and highly relevant retrieval content output to users after the target document set has been classified and organized.

[0081] Specifically, embodiments of the present invention can classify and organize target document sets according to attributes such as document source and type to form structured search results, making it easier for users to quickly locate the information they need.

[0082] As examples, embodiments of the present invention can traverse each document in the target document set and classify them into preset categories (such as "official user manuals", "community Q&A posts", "technical blog articles", and "video tutorials") based on their metadata (such as source URLs, tags, and document types) or content characteristics. After classification, embodiments of the present invention can organize and format the search results by category, and finally present them to the user in a structured form (such as a grouped list), making it convenient for the user to quickly locate the required content according to the information type.

[0083] This invention provides an information retrieval method, comprising: receiving a user query containing a product name; performing a hybrid search using the product name to obtain a first candidate document set; performing an enhanced search using the product name to obtain a second candidate document set; performing a fallback search using the user query to obtain a third candidate document set; performing relevance filtering on the first, second, and third candidate document sets to obtain a target document set, wherein the target document set includes at least one document; classifying the documents in the target document set according to their data sources; and outputting the classified information retrieval results. This invention, by receiving a user query containing a product name and employing a multi-stage retrieval strategy involving hybrid search, enhanced search, and fallback search, comprehensively filters and relevance-based multiple candidate document sets, and finally classifies and outputs the target document set, effectively improves the recall and accuracy of information retrieval in the diagnostic and repair field, and solves the problem of insufficient retrieval reliability.

[0084] Optional, based on Figure 1 The method shown is as follows: Figure 2 The diagram shows a specific implementation of step S110 in the information retrieval method provided in this embodiment of the invention. Step S110 may specifically include:

[0085] S200: Generate multiple name variations based on the product name.

[0086] Name variants refer to various written forms generated from the original product name through a series of rules, which are used to expand the matching scope and improve the coverage and accuracy of retrieval.

[0087] Specifically, embodiments of the present invention can generate a set of name variants by applying various rules to the original product name and adding pinyin, English or abbreviation forms according to the actual situation, in order to expand the matching range and ensure that various possible naming methods can be covered in the subsequent retrieval stage.

[0088] Furthermore, embodiments of the present invention can convert the product name to lowercase to obtain a first name variant. Replacing hyphens in the product name with spaces yields a second name variant. Removing hyphens from the product name yields a third name variant.

[0089] As examples, embodiments of the present invention can take a specific product name (such as "ABC-Smart Lock") as input. A predefined set of rules is used to programmatically generate various common textual variations of that name. These rules may include: converting the original text to all lowercase (such as "abc-Smart Lock"); replacing hyphens (such as hyphens and underscores) with spaces (such as "abc Smart Lock"); or directly removing hyphens (such as "abc Smart Lock"). Furthermore, depending on actual needs and product characteristics, the rule set can be extended to generate pinyin, common English translations, or abbreviations, outputting a list containing the original name and all its generated variations.

[0090] S210. Search the document library using the product name and name variants to obtain the first candidate document containing the product name or name variant.

[0091] Specifically, in this embodiment of the invention, a hybrid vector search and text matching can be performed on the document library. First, candidate documents containing any product name or its variants are selected, and obviously irrelevant documents are initially filtered out to obtain the first candidate document containing the product name or its variants.

[0092] As examples, embodiments of the present invention use product names and a list of their variations as a set of keywords for rapid searching within the document library's index. The retrieval logic can be as follows: if a document contains any word or phrase from this set of keywords precisely in a specific field (such as title, body summary, or product tag), the document is initially selected, thereby efficiently filtering out all documents that explicitly mention the product from a massive amount of documents, forming a rough but highly relevant initial candidate document pool.

[0093] S220. Calculate the semantic similarity score between each first candidate document and the user query.

[0094] The semantic similarity score refers to the distance or relevance between the vector representations of the query and the document, reflecting the degree of semantic similarity between them. The higher the score, the closer the semantics, which is used to help filter documents that match the query intent.

[0095] Specifically, in this embodiment of the invention, the user query can be semantically matched with each first candidate document, and the vector distance and relevance score between the two can be calculated to quantify the semantic similarity between the document and the query.

[0096] As examples, embodiments of the present invention can convert the content or key content fragments of all first candidate documents and the original user query statement into high-dimensional vectors respectively. Then, by calculating the cosine similarity between the vector of each document and the query vector, a quantified semantic similarity score is calculated for each document, thereby reflecting the degree of deep semantic association between the document content and the user's query intent.

[0097] S230. Select the first candidate documents whose semantic similarity scores meet the first threshold condition, and construct the first candidate document set.

[0098] The first threshold condition refers to the semantic similarity score standard set in the hybrid search filter (such as distance < 1.5 and relevance > 0.05). Only documents that meet this threshold can be retained in the first candidate document set to ensure high relevance and accuracy of the search results.

[0099] Specifically, in embodiments of the present invention, documents with high semantic relevance can be filtered out based on a preset threshold in the first threshold condition, and noisy documents below the threshold can be eliminated to form a high-quality first candidate document set.

[0100] As examples, embodiments of the present invention may pre-set a high first threshold condition to ensure the accuracy of the final result. All candidate documents with calculated scores are traversed, and only those documents with semantic similarity scores reaching or exceeding this threshold are retained. These rigorously selected documents contain the product name or its variations in their surface text and are highly relevant to the user's query in their deeper semantics, collectively forming the first candidate document set.

[0101] This invention, through generating multiple name variations based on product names, effectively expands the matching scope, avoids omissions due to naming differences, and performs precise searches in the document library, initially filtering relevant documents and improving search coverage and accuracy. Next, semantic similarity calculations are performed on candidate documents and user queries to ensure that the selected documents are highly semantically aligned with the query. Then, by setting strict thresholds, a first candidate document set is constructed, thereby minimizing irrelevant information, improving the relevance and accuracy of search results, and enhancing the overall search performance.

[0102] Optional, based on Figure 1 The method shown is as follows: Figure 3 The diagram shows a specific implementation of step S120 in the information retrieval method provided in this embodiment of the invention. Step S120 may specifically include:

[0103] S300: Combine the product name with the context in the user query to generate extended query content.

[0104] Extended query content refers to a richer and more complete textual expression formed by organically combining the product name with contextual information in the user's query (including the user's specific needs, usage scenarios, and descriptive details). Extended query content not only includes core product keywords but also covers multi-dimensional information about the query intent, enhancing the semantic expressiveness of the query and enabling the retrieval system to more accurately understand user needs, thereby improving the relevance and recall rate of the match.

[0105] Specifically, embodiments of the present invention can extract the user's original query text and its contextual information, and then integrate the pre-generated product name and its variations into the context to construct richer and more specific extended query content.

[0106] As examples, embodiments of the present invention can generate extended content such as "How to install ABC-Smart Lock" and "Installation steps and precautions for ABC-Smart Lock" for queries such as "How to install?" and product name "ABC-Smart Lock", thereby making up for the ambiguity or lack of information in the original query and constructing a query expression that better represents the user's real needs.

[0107] S310. Convert the extended query content into a vector representation to obtain the query vector.

[0108] In this context, the query vector refers to the mathematical vector representation of the expanded query content, obtained through vectorization techniques in natural language processing (such as word embedding and sentence vector encoding). The query vector captures the semantic features of the query text in a multi-dimensional space and is used to calculate similarity with pre-computed document vectors in the document library, thereby achieving efficient semantic-based retrieval.

[0109] Specifically, embodiments of the present invention can utilize a pre-trained natural language processing model to extend the input encoder of the query text, converting the text into a fixed-dimensional vector representation, i.e., a query vector. The query vector captures the semantic features of the text in the semantic space, facilitating subsequent similarity calculation with document vectors and achieving efficient semantic-based matching.

[0110] S320. Using the query vector, perform semantic vector search in the document library, filter out second candidate documents whose semantic similarity scores meet the second threshold condition, and construct the second candidate document set.

[0111] The second threshold condition refers to the semantic similarity score standard set in the enhanced search filtering (such as distance < 2.0 and relevance > 0.01). Only documents that meet this threshold can be retained in the first candidate document set to ensure the recall of more semantically related documents that may not strictly match the product name.

[0112] Specifically, in this embodiment of the invention, the query vector can be used as input to perform vector retrieval in a pre-calculated and stored document vector index, calculating the distance and relevance score between the query vector and each document vector. Documents that meet the set conditions are then selected and collected into a second candidate document set.

[0113] As examples, embodiments of the present invention can use query vectors to perform nearest neighbor searches across a pre-vectorized document library. The cosine similarity between the query vector and each document vector is calculated, generating a semantic similarity score for each document. A relatively lenient second threshold condition is then applied for document filtering. All documents with scores exceeding this threshold, even if their content does not precisely mention the product name, are included in the second candidate document set as long as they are semantically highly relevant to the extended query.

[0114] This invention combines product names with contextual information in user queries to generate extended query content, which is then converted into query vectors for semantic vector search. This enhances the search's ability to more comprehensively understand and express the user's complex needs, overcomes the limitations of single keyword matching, and improves recall and semantic relevance. As a result, a richer and more accurate set of second candidate documents is obtained, which not only improves the coverage and accuracy of the retrieval but also effectively captures the user's potential intent, ensuring better search results in diverse query scenarios.

[0115] Optional, based on Figure 1 The method shown is as follows: Figure 4 The diagram shows a specific implementation of step S130 in the information retrieval method provided in this embodiment of the invention. Step S130 may specifically include:

[0116] S400. Calculate the semantic similarity score between the user query and all documents.

[0117] Specifically, in this embodiment of the invention, user queries can be vectorized using a natural language processing model to obtain semantic vectors for the queries. Then, for each document, a semantic similarity score is calculated between it and the user query vector using cosine similarity or Euclidean distance, so that each document and the user query are assigned a semantic similarity score.

[0118] S410. Select the third candidate documents whose semantic similarity scores meet the third threshold condition, and construct the third candidate document set.

[0119] The third threshold condition refers to the semantic similarity score standard set in the fallback search filter (such as distance < 3 and relevance > 0.001). Only documents that meet this threshold can be retained in the third candidate document set. The purpose is to ensure that even if a document does not explicitly contain a product name, it can still be recalled based on semantic relevance, preventing the omission of potentially valuable content.

[0120] Specifically, embodiments of the present invention can preset an extremely lenient third threshold condition (such as semantic similarity score > 0.001 or vector distance < 3.0). The semantic similarity scores of all documents are traversed, and all documents with scores reaching or exceeding this initial threshold are filtered out to form a preliminary third candidate document set. The goal of this stage is to recall as broadly as possible any documents that may have a weak semantic connection to the user's query, ensuring that no potentially useful information is overlooked.

[0121] S420. When the number of third candidate documents that meet the third threshold condition is less than the preset minimum recall number, the threshold in the third threshold condition is lowered until the number of selected third candidate documents reaches the preset minimum recall number.

[0122] The preset minimum recall number refers to the minimum threshold number of candidate documents (such as 10 or 20) set to ensure basic coverage of the search results. When the actual number of candidate documents filtered through the third threshold is less than this minimum value, the third threshold conditions will be automatically relaxed step by step (such as increasing the distance threshold or decreasing the relevance threshold) to expand the recall range until the preset minimum recall number is met, thereby ensuring the sufficiency and diversity of the search results.

[0123] Specifically, this embodiment of the invention sets a preset minimum recall number as a baseline. By checking the number of preliminary third candidate documents, if the number is insufficient, a dynamic threshold adjustment mechanism is triggered: the screening criteria for the third threshold condition can be automatically reduced in predetermined steps (e.g., reducing the similarity score threshold from 0.001 to 0.0005, or relaxing the vector distance threshold from 3.0 to 3.5), and the screening is performed again. This process is iterative, gradually relaxing the semantic similarity requirements until the number of recalled documents meets or exceeds the preset minimum recall number, thereby ensuring that the final output third candidate document set can at least reach the required basic size and maintain the availability of the retrieval service.

[0124] This invention, through calculating the semantic similarity score between a user query and all documents, enables the broad and flexible mining of potentially relevant content without relying on strict matching of product names, ensuring that the recalled results are semantically relevant. Simultaneously, when the number of recalled documents is insufficient to the preset minimum recall number, the threshold is automatically and dynamically adjusted to effectively expand the recall scope, avoid missing important information, guarantee the sufficiency and diversity of search results, and improve the robustness and user experience of the overall retrieval system.

[0125] Optional, based on Figure 1 The method shown is as follows: Figure 5 The diagram shows a specific implementation of step S140 in the information retrieval method provided in this embodiment of the invention. Step S140 may specifically include:

[0126] S500. Perform fuzzy matching on each candidate document in the first, second, and third candidate document sets to obtain a fuzzy matching score.

[0127] The fuzzy matching score is a numerical score calculated by performing fuzzy text matching (such as based on edit distance, word overlap, and keyword matching degree) on candidate document text and user queries and product names. It measures the degree of formal similarity between the document content and the query terms. The fuzzy matching score reflects the literal closeness of the match between the document and the user query and product name.

[0128] Specifically, embodiments of the present invention can calculate the surface text similarity between the key content (such as title or summary) of each candidate document and the user query text, product name and its variations, based on character or word-level similarity algorithms; this is known as the fuzzy matching score. A higher fuzzy matching score indicates a higher degree of direct matching between the document and the query in terms of keyword spelling and phrase composition. The fuzzy matching score is primarily used to capture documents that, while not precisely matching the product name, use highly relevant synonyms or near-synonyms.

[0129] S510. Perform semantic analysis on each candidate document in the first, second, and third candidate document sets to obtain semantic analysis scores.

[0130] The semantic analysis score refers to the score obtained by calculating the semantic similarity between candidate documents and user queries and product names using natural language processing technology. The semantic analysis score reflects the semantic relevance between documents and queries, and can capture information with similar meanings but different expressions.

[0131] Specifically, embodiments of the present invention can use a deep learning semantic model to perform deep semantic analysis on each candidate document. The model understands the contextual meaning, intent, and conceptual relationships between the document and the query, and outputs a semantic analysis score. The semantic analysis score incorporates more complex features, enabling the identification of documents that, even if the surface text does not match, are highly related in deep semantics, such as documents that answer similar questions but target different product models.

[0132] S520. Filter out candidate documents whose fuzzy matching score is equal to or greater than the preset fuzzy matching threshold, or whose semantic analysis score is equal to or greater than the preset semantic analysis threshold, and construct the target document set.

[0133] The preset fuzzy matching threshold refers to the minimum standard for fuzzy matching score set in advance. Only when the fuzzy matching score of a candidate document reaches or exceeds the threshold is the document considered to be sufficiently relevant to the query in terms of textual conformity, and thus enters the subsequent filtering or display stage.

[0134] The preset semantic analysis threshold refers to the minimum standard for semantic analysis score set in advance. Only when the semantic analysis score of a candidate document reaches or exceeds the threshold is the document considered to be sufficiently matched with the query in terms of semantic relevance and meets the user's information needs, and thus it is retained for the final result output.

[0135] Specifically, this embodiment of the invention presets two independent thresholds: one for fuzzy matching score (e.g., score >= 15) and the other for semantic analysis score (e.g., score >= 0.05). This embodiment of the invention examines each candidate document; if either its fuzzy matching score or semantic analysis score reaches or exceeds the corresponding threshold, the document is determined to be highly relevant and retained. All documents meeting this condition constitute the final target document set, thus ensuring that the search results include both highly accurate documents with high keyword matching and supplementary documents that are highly relevant in semantic understanding, balancing accuracy and recall.

[0136] This invention employs a dual-scoring system to filter documents by performing fuzzy matching and semantic analysis on each candidate document in multiple candidate document sets, thereby comprehensively covering diverse document types that may be associated with user queries. The dual-threshold mechanism of fuzzy matching and semantic analysis scores not only effectively ensures that documents are highly relevant to the product name and query format, but also accurately captures content with different expressions but high semantic relevance. This significantly improves the comprehensiveness and accuracy of the recall results, reducing the risk of missing closely related information and optimizing the quality of the final target document set, thus significantly enhancing the intelligence level of information retrieval and the user experience.

[0137] Optionally, in the above Figure 1Based on one or more corresponding embodiments, in another optional embodiment provided by the present invention, the documents provided by the present invention may include product manuals, user manuals, technical documents and application cases.

[0138] The product manual can be written by the product manufacturer or supplier and is a document that details the product's specifications, features, usage methods, safety precautions, and maintenance information.

[0139] The user manual can be an operation guide for end users, containing practical information such as product installation steps, operation procedures, troubleshooting, and daily maintenance.

[0140] Technical documents can be materials covering technical aspects such as product design, development, testing, architecture, and implementation details.

[0141] Application cases can be documents that showcase specific application examples and successful experiences of the product in real-world usage environments.

[0142] This invention, through information retrieval covering multiple document categories, not only meets users' diverse needs for product functions, operation guides, technical details, and practical application scenarios, but also enhances the professionalism and practicality of the retrieval results, thereby improving the user experience of information retrieval.

[0143] Although the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous.

[0144] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0145] Corresponding to the above method embodiments, this invention also provides an information retrieval device, the structure of which is as follows: Figure 6 As shown, it may include: a user query receiving unit 10, a hybrid search unit 20, an enhanced search unit 30, a fallback search unit 40, a relevance filtering unit 50, and a data source classification unit 60.

[0146] User query receiving unit 10 is used to receive user queries containing product names.

[0147] Hybrid search unit 20 is used to perform a hybrid search using product names to obtain the first candidate document set.

[0148] Enhanced search unit 30 is used to perform enhanced search using product names to obtain a second set of candidate documents.

[0149] The fallback search unit 40 is used to perform a fallback search based on the user query to obtain a third candidate document set.

[0150] The relevance filtering unit 50 is used to perform relevance filtering on the first candidate document set, the second candidate document set, and the third candidate document set to obtain the target document set, wherein the target document set includes at least one document.

[0151] The data source classification unit 60 is used to classify the data sources of each document in the target document set and output the classified information retrieval results.

[0152] Optionally, the hybrid search unit 20 can be used to generate multiple name variations based on the product name; to search the document library using the product name and name variations to obtain first candidate documents containing the product name or name variations; to calculate the semantic similarity score between each first candidate document and the user query; and to filter out the first candidate documents whose semantic similarity scores meet the first threshold condition to construct a first candidate document set.

[0153] Optionally, the hybrid search unit 20 can be further used to convert the product name to lowercase to obtain a first name variant; and / or, replace the hyphens in the product name with spaces to obtain a second name variant; and / or, remove the hyphens in the product name to obtain a third name variant.

[0154] Optionally, the enhanced search unit 30 can be used to combine the product name with the context in the user query to generate extended query content; convert the extended query content into a vector representation to obtain a query vector; and use the query vector to perform a semantic vector search in the document library to filter out second candidate documents whose semantic similarity scores meet the second threshold condition, thus constructing a second candidate document set.

[0155] Optionally, the fallback search unit 40 can be used to calculate the semantic similarity score between the user query and all documents; filter out third candidate documents whose semantic similarity scores meet the third threshold condition, and construct a third candidate document set; when the number of third candidate documents that meet the third threshold condition is less than the preset minimum recall number, lower the threshold in the third threshold condition until the number of filtered third candidate documents reaches the preset minimum recall number.

[0156] Optionally, the relevance filtering unit 50 can be used to perform fuzzy matching on each candidate document in the first candidate document set, the second candidate document set, and the third candidate document set to obtain a fuzzy matching score; perform semantic analysis on each candidate document in the first candidate document set, the second candidate document set, and the third candidate document set to obtain a semantic analysis score; and filter out candidate documents whose fuzzy matching score is equal to or greater than a preset fuzzy matching threshold, or whose semantic analysis score is equal to or greater than a preset semantic analysis threshold, to construct a target document set.

[0157] Optional documents include product manuals, user manuals, technical documents, and application examples.

[0158] This invention provides an information retrieval device for: receiving user queries containing product names; performing a hybrid search using the product names to obtain a first candidate document set; performing an enhanced search using the product names to obtain a second candidate document set; performing a fallback search using the user query to obtain a third candidate document set; performing relevance filtering on the first, second, and third candidate document sets to obtain a target document set, wherein the target document set includes at least one document; classifying the data sources of each document in the target document set; and outputting the classified information retrieval results. This invention, by receiving user queries containing product names and employing a multi-stage retrieval strategy involving hybrid search, enhanced search, and fallback search, comprehensively filters and relevance-based multiple candidate document sets, and finally classifies and outputs the target document set, effectively improves the recall and accuracy of information retrieval in the diagnostic and repair field, and solves the problem of insufficient retrieval reliability.

[0159] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0160] The information retrieval device includes a processor and a memory. The user query receiving unit 10, hybrid search unit 20, enhanced search unit 30, backtracking search unit 40, relevance filtering unit 50, and data source classification unit 60 are all stored in the memory as program units. The processor executes the program units stored in the memory to realize the corresponding functions.

[0161] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured. By adjusting kernel parameters, a multi-stage retrieval strategy can be used to comprehensively screen and filter multiple document sets, effectively improving the recall and accuracy of information retrieval in the diagnostic and maintenance field, and solving the problem of insufficient retrieval reliability.

[0162] This invention provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the information retrieval method.

[0163] This invention provides a processor for running a program, wherein the program executes the information retrieval method during runtime.

[0164] like Figure 7 As shown, this embodiment of the invention provides an electronic device 1000, which includes at least one processor 1001, at least one memory 1002 connected to the processor 1001, and a bus 1003. The processor 1001 and the memory 1002 communicate with each other via the bus 1003. The processor 1001 is used to call program instructions in the memory 1002 to execute the aforementioned information retrieval method. The electronic device in this document can be a server, PC, PAD, mobile phone, etc.

[0165] The present invention also provides a computer program product that, when executed on an electronic device, is suitable for executing a program that initializes an information retrieval method step.

[0166] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0167] In a typical configuration, an electronic device includes one or more processors (CPUs), memory, and a bus. The electronic device may also include input / output interfaces, network interfaces, etc.

[0168] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and memory includes at least one memory chip. Memory is an example of computer-readable media.

[0169] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0170] In the description of this invention, it should be understood that if the terms "upper", "lower", "front", "rear", "left" and "right" are used to indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the position or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0171] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0172] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0173] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the present invention.

Claims

1. An information retrieval method, characterized in that, include: Receive user queries containing product names; A first candidate document set is obtained by performing a mixed search using the product name; An enhanced search was performed using the product name to obtain a second set of candidate documents. Using the user query, a backward search is performed to obtain a third candidate document set; The first candidate document set, the second candidate document set, and the third candidate document set are subjected to relevance filtering to obtain a target document set, wherein the target document set includes at least one document; The data source is classified for each document in the target document set, and the classified information retrieval results are output.

2. The method according to claim 1, characterized in that, The step of performing a mixed search using the product name to obtain a first candidate document set includes: Multiple name variations are generated based on the product name; The document library is searched using the product name and the name variant to obtain a first candidate document containing the product name or the name variant; Calculate the semantic similarity score between each of the first candidate documents and the user query; Select the first candidate documents whose semantic similarity scores meet the first threshold condition, and construct the first candidate document set.

3. The method according to claim 2, characterized in that, The generation of multiple name variations based on the product name includes: Convert the product name to lowercase to obtain the first name variant; And / or, replace the connectors in the product name with spaces to obtain a second name variant; And / or, remove the connector from the product name to obtain a third name variant.

4. The method according to claim 1, characterized in that, The enhanced search using the product name to obtain a second candidate document set includes: Combine the product name with the context in the user query to generate expanded query content; The extended query content is converted into a vector representation to obtain the query vector; Using the query vector, a semantic vector search is performed in the document library to filter out second candidate documents whose semantic similarity scores meet the second threshold condition, and a second candidate document set is constructed.

5. The method according to claim 1, characterized in that, The step of using the user query to perform a backward search to obtain a third candidate document set includes: Calculate the semantic similarity score between the user query and all documents; Select third candidate documents whose semantic similarity scores meet the third threshold condition, and construct a third candidate document set; When the number of third candidate documents that meet the third threshold condition is less than the preset minimum recall number, the threshold in the third threshold condition is lowered until the number of the selected third candidate documents reaches the preset minimum recall number.

6. The method according to claim 1, characterized in that, The step of performing relevance filtering on the first candidate document set, the second candidate document set, and the third candidate document set to obtain the target document set includes: Fuzzy matching is performed on each candidate document in the first candidate document set, the second candidate document set, and the third candidate document set to obtain a fuzzy matching score; Semantic analysis is performed on each candidate document in the first candidate document set, the second candidate document set, and the third candidate document set to obtain a semantic analysis score; Candidate documents whose fuzzy matching scores are equal to or greater than a preset fuzzy matching threshold, or whose semantic analysis scores are equal to or greater than a preset semantic analysis threshold, are selected to construct a target document set.

7. The method according to any one of claims 1 to 6, characterized in that, The documents include product manuals, user manuals, technical documents, and application cases.

8. An information retrieval device, characterized in that, include: User query receiving unit, hybrid search unit, enhanced search unit, fallback search unit, relevance filtering unit, and data source classification unit; The user query receiving unit is used to receive user queries containing product names; The hybrid search unit is used to perform a hybrid search using the product name to obtain a first candidate document set; The enhanced search unit is used to perform an enhanced search using the product name to obtain a second candidate document set; The backtracking search unit is used to perform a backtracking search based on the user query to obtain a third candidate document set; The relevance filtering unit is used to perform relevance filtering on the first candidate document set, the second candidate document set, and the third candidate document set to obtain a target document set, wherein the target document set includes at least one document; The data source classification unit is used to classify the data sources of each document in the target document set and output the classified information retrieval results.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the information retrieval method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The electronic device includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the information retrieval method as described in any one of claims 1 to 7.

Images