A search word center word recognition method based on a multi-source voting mechanism

Abstract

This invention discloses a method for identifying search term centers based on a multi-source voting mechanism. After standardizing the search term, multiple heterogeneous information sources are invoked in parallel to parse the search term from different semantic analysis perspectives, extracting candidate centers. The search term itself then filters these candidates, and a weighted voting mechanism based on historical performance is used to determine the most representative center. Corresponding confidence information is output to ensure that the center term results do not deviate from the user's search intent, achieving accurate and stable identification of search term centers. By integrating multiple heterogeneous information sources and introducing a quantitative evaluation mechanism, this invention significantly improves the accuracy, stability, and interpretability of center term identification results, effectively reducing the interference of manually preset rules and noise signals on the center term identification results. It is applicable to search engines, information retrieval, and information matching applications.

Description

Technical Field

[0002] This invention relates to the fields of information retrieval and natural language processing, specifically to a method for identifying the center word of search terms based on a multi-source voting mechanism. Background Technology

[0004] With the continuous growth of internet information volume, search engines and information retrieval systems have become important tools for users to obtain information. Users typically express their search intent by entering search terms, and search systems need to perform semantic understanding of these terms in order to return search results that highly match the user's intent. Accurately identifying the central words in user search terms is a key technical step in improving the relevance of search results and achieving precise matching. However, facing the current diverse, colloquial, and noisy real-world search scenarios with varying user input, existing technical solutions still face severe challenges in terms of the accuracy, stability, and interpretability of central word identification, specifically manifested in four prominent problems:

[0005] (1) Inaccurate identification of the central word: Methods that rely on a single information source or rule are prone to failure due to the complex grammatical structure of search words, emerging network terms or specific domain terms. For example, relying solely on part-of-speech rules or static word matching can lead to the omission or misjudgment of core semantic entities.

[0006] (2) Unstable search results: Whether it is a system based on fixed rules or a solution based on a single deep learning model, it is limited by the blind spots of the rule or model. In search scenarios where the model training data is not fully covered, words that are not related to the actual semantics of the search terms are easily introduced, resulting in large fluctuations in the output results.

[0007] (3) Lack of reliability assessment of identification results: Existing solutions usually only output one or more candidate core words, but cannot inform users of the reliability of the results. This brings decision-making difficulties to downstream applications such as search ranking, ad triggering, and manual review, and makes it impossible to distinguish between highly reliable results and speculative results.

[0008] (4) Poor adaptability to complex semantic combinations: Search terms often contain marketing terms, spelling errors, abbreviations and complex modification relationships. Existing methods are unable to accurately extract the core semantic words that truly represent the user's intent while filtering out noise, resulting in deviations from the user's true search intent.

[0009] To address these challenges, the industry has proposed various improvement approaches. For example, patent CN110807333A proposes a method for voice command recognition using a pre-trained semantic model, emphasizing semantic understanding in specific in-vehicle scenarios. However, its core still relies on a single pre-trained large language model as the foundation for semantic understanding. Patent CN121168467A integrates multiple technologies such as semantic relevance calculation, duplicate content detection, and multimodal feature filtering to process complex data, but its ultimate goal is noise reduction and filtering, making it unsuitable for search scenarios that require determining a single, most reliable central word.

[0010] While existing technologies have attempted to address the noise and uncertainty issues in semantic understanding from different perspectives, they either lack effective multi-source parallel fusion mechanisms or lack reliability metrics for the results. This makes it difficult to guarantee the accuracy, stability, and interpretability of central word identification in complex search scenarios, and they cannot simultaneously and effectively solve the above four problems. Therefore, how to design a central word identification technology architecture that can integrate multiple heterogeneous and independent semantic analysis information sources and quantify the confidence assessment of the final fusion output is an urgent technical problem to be solved in search scenarios. Summary of the Invention

[0012] To address the shortcomings of existing technologies, this invention provides a method for identifying search term center words based on a multi-source voting mechanism. After standardizing the search term, multiple heterogeneous information sources are invoked in parallel to parse the search term from different semantic analysis perspectives, extract candidate center words, and use the search term to filter candidate words. Furthermore, a voting mechanism based on historical performance is used to determine the most representative center word, while outputting the corresponding confidence information to ensure that the center word results do not deviate from the user's search intent, thereby achieving accurate and stable identification of search term center words.

[0013] To achieve the above objectives, the present invention adopts the following technical solution:

[0014] A method for identifying the center word of a search term based on a multi-source voting mechanism includes the following steps:

[0015] Step 1: Obtain the search terms to be processed: Obtain the search terms entered by the user through any of the following systems: search engine, information retrieval system, and information matching system. The search terms can be in the form of a single word, a combination of multiple words, or a phrase containing special characters.

[0016] Step 2: Preprocess the search terms by converting them into a standardized set of search terms consisting of several words;

[0017] Step 3: Semantically analyze the standardized search term set using multiple heterogeneous information sources to generate corresponding candidate center term sets. The heterogeneous information sources refer to semantic analysis units that are independent of each other in terms of underlying data foundation, semantic analysis algorithm, and output logic, and include at least two of the following: information sources based on vector semantic similarity, information sources based on word library analysis, and semantic recognition information sources based on model reasoning. Each heterogeneous information source has its own unique semantic judgment rules, which include the algorithm logic, matching conditions, and judgment thresholds used within the information source to quantitatively evaluate and determine whether a word can serve as the core semantic representative of the search term.

[0018] Step 4: Filter candidate center words based on the search terms themselves: Compare all candidate center words output by each heterogeneous information source in Step 3 with the standardized search term set obtained in Step 2, and retain only the candidate words that exist in the standardized search term set to form an effective candidate center word set;

[0019] Step 5: Weighted voting determination and confidence calculation: Statistically analyze the occurrence of each valid candidate center word in different heterogeneous information sources in Step 4. According to the pre-configured voting weight of each information source, sum the voting scores of each valid candidate center word by weight. Take the valid candidate center word with the highest voting score as the final center word, and calculate the confidence of the final center word based on the distribution of the voting results.

[0020] In step 1, obtaining the search terms to be processed includes obtaining the search terms entered by the user in real time and obtaining the search terms from the historical search log data stored in the system.

[0021] The preprocessing in step 2 specifically includes: performing word segmentation, case normalization, special character cleaning, word standardization, and filtering on the search terms obtained in step 1 in sequence; the word standardization includes word form regularization, spelling correction, singular / plural regularization, and synonym expression regularization; the filtering refers to filtering words with no substantial semantic contribution from the search terms based on a preset thesaurus, including banned words, stop words, and words with low semantic weight; the words with low semantic weight are words included in the preset low semantic contribution thesaurus.

[0022] In step 3, each heterogeneous information source runs independently in either parallel or asynchronous mode and outputs a set of candidate center words.

[0023] In step 3, the information source based on vector semantic similarity vectorizes the search terms and information items, calculates the semantic similarity between them, selects the top K information items with the highest similarity, and extracts the keywords of the information items as candidate center words; the keywords refer to words extracted from the information items that can characterize the name and core attributes of the information items;

[0024] In step 3, the information source based on the thesaurus analysis uses a constructed semantic knowledge base and rule base to match search terms with words in the knowledge base. When a search term matches a word in the semantic knowledge base, the word is determined as a candidate center word. The semantic knowledge base includes at least a general language thesaurus, an industry domain thesaurus, a category thesaurus, and an attribute thesaurus.

[0025] In step 3, the semantic recognition information source based on model reasoning uses a computational model with semantic understanding capabilities to identify core words by analyzing the contribution value of each word in the search term to expressing the overall search intent; specifically, it obtains a quantitative score representing the importance of each word in the search term generated by the model, and selects the most important words as candidate core words according to a preset score threshold or ranking rule.

[0026] In step 3, when extracting the candidate center words based on the information source of vector semantic similarity, the frequency of each candidate center word appearing in the first K information items is counted.

[0027] In step 4, the comparison is an exact match judgment, which specifically includes: performing an exact match on each word in the candidate center word set in step 3, and judging whether the word is exactly the same as a word in the standardized search word set. If they are the same, they are retained; otherwise, they are discarded.

[0028] In step 5, the voting weight is calculated and configured based on the historical recognition accuracy of each heterogeneous information source;

[0029] Step 5 specifically includes:

[0030] Step 501: Record the output state of each valid candidate center word in each heterogeneous information source, and calculate the basic voting value, where:

[0031] For an information source based on vector semantic similarity, its basic voting value is equal to the ratio of the frequency of the effective candidate center word in the first K information items to K;

[0032] For other heterogeneous information sources, if the valid candidate center word is output by the information source, the basic voting value is 1; otherwise, it is 0.

[0033] Step 502: Based on the voting weights configured for each heterogeneous information source, the basic voting values ​​obtained by the effective candidate center words in Step 501 are weighted, and the voting score of each effective candidate center word is calculated.

[0034] Step 503: Calculate the confidence level of the final center word based on the distribution of the voting results. The confidence level is related to the weighted voting score of the final center word, the weighted voting score of the second-ranked valid candidate center word, and the number of heterogeneous information sources from which the final center word is output.

[0035] In step 503, the confidence level of the center words is calculated using the following weighted formula:

[0036]

[0037] Where Conf represents the confidence level of the final center word; Vc represents the weighted voting score of the final center word; ΣVi represents the sum of the weighted voting scores of all valid candidate center words; V2 represents the weighted voting score of the valid candidate center word that ranks second in votes; and when hour, When there is no valid candidate center word ranked second, Nc represents the number of heterogeneous information sources that output the final center word; N represents the total number of heterogeneous information sources that participated in the voting.

[0038] α, β, and γ are all weighting coefficients used to measure the share of votes received. Voting advantage and multi-source consistency rate The degree of influence of the three indicators on the final confidence level, and α+β+γ=1.

[0039] Compared with the prior art, the present invention has at least the following beneficial effects:

[0040] (1) This invention introduces and integrates multiple heterogeneous information sources based on vector semantic similarity, lexicon analysis and model reasoning. Each information source analyzes the search terms from different semantic analysis dimensions, which effectively avoids the error in identifying the central word caused by data bias, model misjudgment or incomplete rule coverage of a single information source, and significantly improves the accuracy and anti-interference ability of the central word identification results.

[0041] (2) This invention filters multi-source candidate words by using the preprocessed search terms themselves as constraints, and forces the final candidate central word to be anchored within the user's original query range, thereby greatly reducing the absolute dependence on external manually configured word libraries or rules and reducing the risk of introducing irrelevant semantics due to incomplete or incorrect configuration.

[0042] (3) The present invention determines the final central word based on the weighted voting mechanism, and calculates the confidence of the central word comprehensively from three dimensions: the share of votes, leading advantage, and multi-source consensus, in combination with the distribution of central word votes. This enables quantitative evaluation of the representativeness, uniqueness and stability of the central word, and provides a reliable quantitative basis for search matching, ranking and governance decisions.

[0043] (4) This invention integrates multi-source heterogeneous information sources, search term self-constraints and voting judgment methods to achieve refined analysis of the semantics of different words in multi-word combinations and phrases of search terms, so as to accurately capture the core search intent of users. It can effectively support the identification of central words in search engines, information retrieval systems and information matching systems, and has good versatility and scalability. Attached Figure Description

[0045] Figure 1 This is a flowchart of a search term center word identification method based on a multi-source voting mechanism in an embodiment of the present invention;

[0046] Figure 2 This is a flowchart of the candidate center word identification method based on multiple heterogeneous information sources in an embodiment of the present invention;

[0047] Figure 3 This is a flowchart of the voting determination process based on dynamic weights in an embodiment of the present invention;

[0048] Figure 4 This is a structural diagram of a search term center word recognition system based on a multi-source voting mechanism in an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] This patented method can be applied to scenarios such as search engines, information retrieval systems, and information matching systems. It is used to perform semantic analysis on user-input search terms and identify their core semantic terms. In search engines and information retrieval systems, users typically express their search intent by inputting search terms. Search terms often contain multiple terms or modifiers. Accurately identifying the central term representing the user's core search intent from the search terms is a key factor affecting the accuracy and stability of search result matching. Existing methods for identifying the central term often rely on the analysis results of a single rule or model, which are easily affected by noisy data or human configuration biases, leading to inaccurate central term identification.

[0052] This invention introduces multiple heterogeneous information sources to analyze search terms from different semantic analysis perspectives. Under the constraints of the search terms themselves, candidate center words are screened, and a voting mechanism is used to determine the most representative center word. Simultaneously, confidence information is output, thereby achieving accurate and stable identification of the search term center word and reducing the cost of manual intervention. The heterogeneous information sources described in this invention refer to multiple semantic analysis units that are based on different data foundations, employ different semantic judgment algorithms, and are capable of independently performing semantic analysis on the same input and outputting results. Their heterogeneity lies in the fundamental differences in analysis dimensions and technical paths, rather than merely differences in data sources.

[0053] In one exemplary embodiment of the present invention, the following steps are included:

[0054] Step 1: Obtain search terms to be processed: Obtain search terms entered by the user through any of the following systems: search engine, information retrieval system, and information matching system. The search terms can be in the form of a single word, a combination of multiple words, or a phrase containing special characters. Obtaining search terms to be processed also includes obtaining search terms entered by the user in real time and obtaining search terms from the historical search log data stored in the system.

[0055] Step 2: Preprocess the search terms by converting them into a standardized set of search terms consisting of several words. The preprocessing specifically includes: performing word segmentation, case normalization, special character cleaning, word standardization, and filtering on the search terms obtained in Step 1. Word standardization includes word form regularization, spelling correction, singular / plural regularization, and synonym regularization. The filtering process involves filtering words with no substantial semantic contribution from the search terms based on a pre-defined thesaurus. These words include banned words, stop words, and words with low semantic weight. The words with low semantic weight are those included in the pre-defined low-semantic-contribution thesaurus.

[0056] Step 3: Semantically analyze the standardized search term set using multiple heterogeneous information sources to generate corresponding candidate center term sets. The heterogeneous information sources refer to semantic analysis units that are independent of each other in terms of underlying data foundation, semantic analysis algorithm, and output logic, and include at least two of the following: information sources based on vector semantic similarity, information sources based on word library analysis, and semantic recognition information sources based on model reasoning. Each heterogeneous information source has its own unique semantic judgment rules, which include the algorithm logic, matching conditions, and judgment thresholds used within the information source to quantitatively evaluate and determine whether a word can serve as the core semantic representative of the search term.

[0057] Each heterogeneous information source can run independently in either parallel or asynchronous manner and output a set of candidate center words;

[0058] The information source based on vector semantic similarity vectorizes the search terms and information items, calculates the semantic similarity between them, selects the top K information items with the highest similarity, and extracts the keywords of the information items as candidate center words; the keywords refer to words extracted from the information items that can characterize the name and core attributes of the information items;

[0059] The information source based on the lexicon analysis utilizes a constructed semantic knowledge base and rule base to match search terms with words in the knowledge base. When a search term matches a word in the semantic knowledge base, that word is determined as a candidate center word. The semantic knowledge base includes at least a general language lexicon, an industry domain lexicon, a category lexicon, and an attribute lexicon.

[0060] The semantic recognition information source based on model reasoning utilizes a computational model with semantic understanding capabilities to identify core words by analyzing the contribution value of each word in the search term to expressing the overall search intent. Specifically, it obtains a quantitative score representing the importance of each word in the search term generated by the model, and selects the most important words as candidate core words based on a preset score threshold or ranking rule.

[0061] When extracting the candidate center words from the information source based on vector semantic similarity, the frequency of each candidate center word appearing in the first K information items is counted.

[0062] Step 4: Filter candidate center words based on the search terms themselves: Compare all candidate center words output by each heterogeneous information source in Step 3 with the standardized search term set obtained in Step 2, and retain only the candidate words that exist in the standardized search term set to form an effective candidate center word set;

[0063] The comparison is a precise matching judgment, which specifically includes: performing precise matching on each word in the candidate center word set in step 3, and judging whether the word is exactly the same as a word in the standardized search word set. If they are the same, they are retained; otherwise, they are discarded.

[0064] Step 5: Weighted voting determination and confidence calculation: Statistically analyze the occurrence of each valid candidate center word in different heterogeneous information sources in Step 4. According to the pre-configured voting weight of each information source, sum the voting scores of each valid candidate center word by weight. Take the valid candidate center word with the highest voting score as the final center word, and calculate the confidence of the final center word based on the distribution of the voting results.

[0065] The voting weights are calculated and configured based on the historical recognition accuracy of each heterogeneous information source.

[0066] Step 5 specifically includes:

[0067] Step 5-1: Record the output state of each valid candidate center word in each heterogeneous information source, and calculate the basic voting value, where:

[0068] For an information source based on vector semantic similarity, its basic voting value is equal to the ratio of the frequency of the effective candidate center word in the first K information items to K;

[0069] For other heterogeneous information sources, if the valid candidate center word is output by the information source, the basic voting value is 1; otherwise, it is 0.

[0070] Step 5-2: Based on the voting weights configured for each heterogeneous information source, the basic voting values ​​obtained by the effective candidate center words in Step 5-1 are weighted, and the voting score of each effective candidate center word is calculated.

[0071] Step 5-3: Calculate the confidence level of the final center word based on the distribution of the voting results. The confidence level is related to the weighted voting score of the final center word, the weighted voting score of the second-ranked valid candidate center word, and the number of heterogeneous information sources from which the final center word is output.

[0072] In step 5-3, the confidence level of the center words is calculated using the following weighted formula:

[0073]

[0074] Where Conf represents the confidence level of the final center word; Vc represents the weighted voting score of the final center word; ΣVi represents the sum of the weighted voting scores of all valid candidate center words; V2 represents the weighted voting score of the valid candidate center word that ranks second in votes; Nc represents the number of heterogeneous information sources for outputting the final center word; and N represents the total number of heterogeneous information sources participating in the voting.

[0075] α, β, and γ are all weighting coefficients used to measure the share of votes received. Voting advantage and multi-source consistency rate The degree of influence of the three indicators on the final confidence level, and α+β+γ=1.

[0076] like Figure 1 The flowchart shown in this embodiment of the invention illustrates a method for identifying the center word of a search term based on a multi-source voting mechanism, and the specific steps include:

[0077] Step 101: Obtain search terms to be processed: Obtain search terms entered by the user through any of the following systems: search engine, information retrieval system, and information matching system. The search terms can be in the form of a single word, a combination of multiple words, or a phrase containing special characters. Obtaining search terms to be processed includes obtaining search terms entered by the user in real time and obtaining search terms from historical search log data stored in the system.

[0078] Step 102: Preprocess the search terms to convert them into a standardized search term set: Perform word segmentation, case normalization, special character cleaning, word standardization, and filtering on the search terms from Step 101 in sequence to finally obtain a standardized search term set consisting of several words.

[0079] The word segmentation process breaks down continuous text into word units with independent semantic meaning;

[0080] The case normalization process unifies the case of letters, eliminating case differences.

[0081] The special character cleaning process involves removing symbols that do not participate in semantic expression;

[0082] The term standardization process involves standardizing the word form, correcting spelling, standardizing singular and plural forms, and standardizing synonyms of search terms to eliminate differences caused by different word forms or synonymous expressions. For example, it standardizes non-standard expressions commonly found in the Internet environment, such as spelling errors, abbreviations, and internet slang.

[0083] The filtering process refers to filtering prohibited words, stop words, and low-semantic-weight words from the search terms based on a preset thesaurus. The low-semantic-weight words are those included in the preset low-semantic-contribution thesaurus, which have a low semantic contribution to representing the core intent of the search terms, including but not limited to marketing terms, generalized descriptive terms, or modifiers.

[0084] Through step 102, the present invention converts the original search terms into standardized words, providing a standard comparison basis for subsequent center word identification based on multi-source voting, effectively reducing the interference of noisy data, expression differences, or words with no actual semantic contribution on the final identification results.

[0085] This invention takes the complex search term "creative-galaxy & E-dimable ceilling-fan-lights product for sale" from a real-world product search scenario as an example, and performs the following preprocessing on the search term:

[0086] (1) Word segmentation: The connectors in the search terms are segmented to obtain the initial search term sequence as ["creative", "galaxy", "&", "E", "dimable", "ceilling", "fan", "lights", "product", "for", "sale"].

[0087] (2) Case normalization: Convert all letters to lowercase to obtain ["creative", "galaxy", "&", "e", "dimable", "ceilling", "fan", "lights", "product", "for", "sale"]

[0088] (3) Special character cleaning: Remove special symbols that do not participate in semantic expression, such as "&" and "-", and retain only the terms, resulting in the term sequence ["creative", "galaxy", "e", "dimable", "ceilling", "fan", "lights", "product", "for", "sale"].

[0089] (4) Variant word processing:

[0090] Lexical correction: "dimable" has been corrected to the more commonly used "dimmable";

[0091] Spell correction: Correct the misspelling "ceilling" to "ceiling";

[0092] Singular-plural normalization: restore the plural form "lights" to the singular form "light";

[0093] Synonym conventions: Map the abbreviation or variant of "e" to "electric";

[0094] After processing with variant words, the following set is obtained: ["creative", "galaxy", "electric", "dimmable", "ceiling", "fan", "light", "product", "for", "sale"].

[0095] (5) Filtering: Based on the preset list of prohibited words, stop words, and low semantic weight words, delete prohibited words, stop words, and low semantic weight words from the search terms. Low semantic weight words refer to words preset in the low semantic contribution word library within the product search domain that only have marketing significance, generalized descriptions, or modifying functions, and do not characterize the core category, functional attributes, or specifications of the product, such as "product," "on sale," and "new." In this embodiment, the words "product" and "for sale" that do not characterize specific product functional attributes are deleted. This is because in the product library, all product entries are referred to as "product," and this word cannot distinguish any specific product; "for sale" indicates a state of the product, not its product attribute; the appearance of these two words will interfere with the search system's recognition of the core product entity "ceiling fan light."

[0096] After preprocessing in step 102, the final standardized search term set is: ["creative", "galaxy", "electric", "dimmable", "ceiling", "fan", "light"]. This standardized search term set will serve as the unified input basis for candidate center word selection and multi-source voting in subsequent steps. If a candidate center word is a multi-word phrase, it is determined whether the phrase completely matches the continuous word sequence in the standardized search term set; if it matches, the phrase is retained as a whole.

[0097] Step 103: Utilize multiple heterogeneous information sources to perform semantic parsing on the standardized search term set, generating corresponding candidate center word sets: Invoke preset heterogeneous information sources to independently perform semantic analysis on the standardized search term set, and output the candidate center words identified by each source; the heterogeneous information sources refer to semantic analysis units that are independent of each other in terms of underlying data foundation, semantic analysis algorithm, and output logic. The heterogeneous information sources include at least two of the following: information sources based on vector semantic similarity, information sources based on lexicon analysis, and semantic recognition information sources based on model reasoning; each type of heterogeneous information source has its own built-in semantic judgment rules, which are the algorithmic logic, matching conditions, and judgment thresholds used within the information source to quantitatively evaluate and determine whether a word can serve as the core semantic representative of a search term.

[0098] To clearly demonstrate the "differences" among the various heterogeneous information sources, the table below compares the three information sources used in this embodiment:

[0099] Differences between three heterogeneous information sources

[0100] Heterogeneous information source types The underlying data it depends on Core analysis algorithms and logic Rules for outputting candidate words Information sources based on vector semantic similarity Unstructured product information database, word vector model Similarity Calculation and Related Retrieval Statistical analysis of highly similar products Information sources based on lexicon analysis Structured, tagged semantic knowledge base and rule base Exact matching and rule-based judgment Symbol matching based on a pre-set knowledge base Semantic recognition information source based on model reasoning Semantic analysis model Context-aware representation learning and weight quantization Based on the model's overall understanding and importance assessment of the context.

[0101] The vector-based semantic similarity information source vectorizes the standardized search terms and information items (such as product information), calculates the semantic similarity between them, selects the top K products (information items) with the highest similarity, and extracts the keywords of these products as candidate center words. The keywords refer to words extracted from the product information that can represent the product name and core attributes of the product. The frequency of each candidate center word appearing in the keywords of the K products is counted and denoted as freq(word). This frequency information will be used as the basis for the voting strength of the vector information source for the candidate center words in the subsequent voting calculation. If a word appears repeatedly in the keywords of multiple similar products, it indicates that the word has a stronger correlation with this group of highly similar products and is more representative of the core semantics than a word that appears only once in a few products.

[0102] Following the example in step 102, the preprocessed search term "creative galaxy electric dimmable ceiling fan light" is vectorized with the product titles and attribute descriptions in the product information database, and the semantic similarity between the two is calculated, such as using cosine similarity. Products are then sorted in descending order based on their similarity scores, and the top 50 products with the highest similarity are selected and compiled. Keywords associated with the product are extracted from the product keyword database. In this example, the keywords are a pre-generated thesaurus containing keywords representing product names and attributes, such as "ceiling fan," "fan," and "light" for product A. The extracted keywords are deduplicated to form a candidate core keyword set. In this case, the final extracted keywords are: "ceiling," "fan," "light," "lamp," and "led," with frequencies of 28, 35, 42, 40, and 30, respectively.

[0103] The information source based on the thesaurus analysis uses a constructed semantic knowledge base to perform multi-dimensional semantic matching on the standardized search term set, and uses the matched words as candidate center words; the semantic knowledge base is a composite structure containing multi-level semantic knowledge and judgment logic, and includes at least a general language thesaurus, an industry domain thesaurus, a product category thesaurus, an attribute thesaurus, and a rule base.

[0104] The general language lexicon includes basic vocabulary and their basic semantic relationships, which are used to confirm the validity of terms and their basic parts of speech, such as nouns, verbs, and adjectives, to ensure that the search terms are legitimate natural language words.

[0105] The industry-specific thesaurus includes high-frequency core words from specific business areas, used to limit search terms to specific business categories. For example, in the lighting field, words such as "fan", "light", "lamp", "led", and "panel" will be included here.

[0106] The product category thesaurus contains terms that can stand alone as core product categories, which is key to identifying what products users want to buy. The terms in this thesaurus can be directly mapped to core categories in the product classification system, such as "fan" and "light".

[0107] The attribute terminology database includes descriptive and limiting terms used to describe product functions, specifications, style, materials, and other dimensions. For example, "dimmable" (functional attribute), "creative" (style attribute), and "ceiling" (installation limiting attribute). This type of terminology helps the system understand the user's specific requirements for core product categories.

[0108] The rule base stores a series of judgment logics and priority strategies for comprehensively processing the matching results of the aforementioned thesaurus. The rule base defines logics such as "product category words have higher priority than attribute words", "when there are multiple category words, take the word with the lowest semantic level", and "a specific combination structure "attribute word + category word" is regarded as a composite central word" to achieve accurate parsing of semantic relationships between multiple terms.

[0109] Following the example in step 102, the standardized search term set [“creative”, “galaxy”, “electric”, “dimmable”, “ceiling”, “fan”, “light”] is input into the information source based on the lexicon analysis. This information source calls the constructed semantic knowledge base to perform the following parsing and matching:

[0110] The general-purpose language lexicon confirms that all terms are valid semantic terms;

[0111] The industry-specific thesaurus limits the search to the "home lighting / home appliances" sector.

[0112] The product category thesaurus matched the core category terms "fan" and "light" in the "home lighting / home appliances" category.

[0113] The attribute dictionary matched "dimmable" (functional attribute), "electric" (energy attribute), "ceiling" (installation structure attribute), "creative" (style attribute), and "galaxy" (theme attribute).

[0114] Based on the preset rule that "product category words have higher priority than attribute words", the rule base determines "fan" and "light" as the core candidate central words. At the same time, based on the rule that "if installation structure attributes co-occur with core category words, they can be used as supplementary candidate words", "ceiling" is also listed as a candidate central word.

[0115] Therefore, the candidate headwords output by this information source are "fan", "light", and "ceiling".

[0116] The semantic recognition information source based on model reasoning utilizes a computational model with semantic understanding capabilities to identify core terms by analyzing the contribution of each word in the search term to expressing the overall search intent. Specifically, it obtains a quantitative score representing the importance of each word in the search term generated by the model, and selects the words with the highest importance as candidate center words based on a preset score threshold or ranking rule. In practice, this score reflects the necessity of the word in constituting the overall semantics. The score can be obtained through various technical means, such as, but not limited to, attention weight distribution, using the average attention weight received by the word as its score; or using the gradient information of the model to quantify the importance of each word by calculating a saliency map; the system filters words according to preset rules, such as "selecting the two words with the highest scores" or "all words with scores greater than 0.7", and outputs the final set of candidate center words.

[0117] In this embodiment, the model information source calculates and analyzes each standardized search term, generating an importance score for each term. The assumed example data is: fan: 0.88, light: 0.85, ceiling: 0.65, electric: 0.45, dimmable: 0.40. According to the preset judgment rule "select the two terms with the highest importance scores", the terms "fan" (0.88) and "light" (0.85) with the highest scores are selected as candidate center terms.

[0118] The results from the three heterogeneous information sources are deduplicated and summarized to obtain a set of candidate center words, including: “ceiling”, “fan”, “light”, “lamp”, and “led”.

[0119] Each heterogeneous information source can run independently in parallel or asynchronous mode and output a set of candidate center words. Refer to the appendix for the workflow. Figure 2 The flowchart shows a candidate center word identification method based on multiple heterogeneous information sources. Each heterogeneous information source independently analyzes the standardized search term set based on its own judgment rules, and finally outputs a candidate center word set.

[0120] Step 104: Filter candidate center words based on the search terms themselves: Compare each candidate word in the candidate center word set in Step 103 with the standardized search term set obtained after preprocessing in Step 102, and retain only those candidate words that exist in the standardized search term set to form a valid candidate center word set for voting.

[0121] Specifically, for each word in the candidate center word set obtained in step 103, an exact match is performed and a judgment is made. If the word is exactly the same as a word in the standardized search term set, it is retained; otherwise, it is discarded. The exact match method includes string matching and term matching. This step uses the search terms themselves as a natural constraint to ensure that all candidate words entering the final decision come from the user's original input, effectively filtering out external noise words that may be introduced in multi-source analysis.

[0122] Following the example from step 102, using the filtering rules from step 104, we determine whether each word in the candidate core word set exists in the standardized search term set. The determination and results are as follows:

[0123] "ceiling" is retained in the standardized search term set.

[0124] The term "fan" is retained in the standardized search term set.

[0125] The term "light" is retained in the standardized search term set.

[0126] If “lamp” is not in the standardized search term set, it will be removed.

[0127] If “led” is not in the standardized search term set, it will be removed.

[0128] The set of valid candidate center words is: {"ceiling", "fan", "light"}, which will participate in the subsequent voting decision. Step 104 limits the multi-source analysis to the user's original expression, avoiding noise that may be introduced by multi-source fusion. For example, words like "lamp" and "led" are not part of the user's original expression and are therefore deleted.

[0129] Step 105: Voting Determination and Confidence Calculation Based on Dynamic Weights: The occurrence of each valid candidate center word in different heterogeneous information sources from Step 104 is statistically analyzed. The voting scores of each valid candidate center word are obtained by weighted summation according to the pre-configured voting weights of each information source. The valid candidate center word with the highest voting score is selected as the final center word. The confidence of the final center word is calculated based on the distribution of the voting results, at least according to its share of votes, leading advantage, and multi-source consistency rate. The voting weights are calculated and configured based on the historical recognition accuracy of each heterogeneous information source; when the voting weights of each heterogeneous information source are not configured, the voting weights of each information source are the same.

[0130] like Figure 3 The flowchart of the voting determination based on dynamic weights in this embodiment of the invention specifically includes:

[0131] Step 105-1: Dynamically calculate the voting weight of each heterogeneous information source based on its historical accuracy. For example, select a historical search term set containing 1000 labeled true center words as a test set, and run each heterogeneous information source to process this test set. If the candidate center word set output by the heterogeneous information source contains the true center word of the search term, then this is counted as a "correct identification". Count the number of correct identifications Ni for each heterogeneous information source, and calculate its initial accuracy Ai = Ni / 1000. Assuming the calculated vector source Avector source = 0.85, word source Alexicon source = 0.78, and model source Amodel source = 0.80, then the weights can be obtained through normalization:

[0132] The vector source weight Wvector source = 0.85 / (0.85+0.78+0.80) ≈ 0.35;

[0133] W-thesaurus source = 0.78 / (0.85+0.78+0.80) ≈ 0.32;

[0134] W_model_source = 0.80 / (0.85+0.78+0.80) ≈ 0.33

[0135] In order to demonstrate the impact of weight differences on voting results, this embodiment of the invention configures a specific set of weights for demonstration, namely: the weight of the vector source voting is configured as 0.5, the weight of the lexicon source voting is configured as 0.3, and the weight of the model source voting is configured as 0.2. This set of weight values ​​reflects the relative importance relationship calculated based on historical accuracy (vector source > lexicon source ≈ model source), and the values ​​are rounded for clarity.

[0136] Step 105-2: Use binary labeling to count the occurrence of each valid candidate center word in different heterogeneous information sources. If a valid candidate center word is output by a heterogeneous information source, it is determined that it appears in that heterogeneous information source, which means that the information source votes for the valid candidate center word, and the vote count is 1. If a valid candidate center word is not output by a heterogeneous information source, it is determined that it does not appear in that heterogeneous information source, which means that the information source does not vote for the valid candidate center word, and the vote count is 0. Among them, the vote count of the information source based on vector semantic similarity is calculated according to the frequency of the valid candidate center word in the first K products, specifically freq(word) / K. The higher the frequency, the closer the association between the word and the highly similar product, and the higher its credibility as a core semantic representative.

[0137] Step 105-3: Calculate the voting score by weighting the votes from each information source and selecting the word with the highest score as the final center word. The voting scores are shown in the table below:

[0138] Candidate central words Vector source (0.5) Thesaurus source (0.3) Model source (0.2) Total Score ceiling 0.56（28 / 50） 1 0 0.5×0.56+0.3×1+0.2×0=0.58 fan 0.7（35 / 50） 1 1 0.5×0.7+0.3×1+0.2×1=0.85 light 0.84（42 / 50） 1 1 0.5×0.84+0.3×1+0.2×1=0.92

[0139] Based on the voting results, the word with the highest score was "light", and "light" was determined as the final word.

[0140] As another optional implementation, if the system configures each information source to have the same weight, then the votes of each information source are directly summed arithmetically to obtain the voting score of the candidate center word.

[0141] Step 105-4: Calculate the confidence score of the center word based on the distribution of the voting results, specifically including:

[0142] Step 105-4-1: Calculate Vc = 0.92, ΣVi = 2.35, V2 = 0.85, Nc = 3, N = 3; calculate the final vote share of the central word Vc / ΣVi, the vote advantage (Vc-V2) / Vc, and the multi-source vote consistency rate Nc / N, which are 0.39, 0.07, and 1, respectively.

[0143] The vote share of the central word refers to the proportion of the votes of the central word to the total score of all valid candidate central words. The higher the value, the higher the total support the central word receives among all candidate words, and the stronger its certainty as a representative of the core semantics of the search term.

[0144] The central word's voting advantage refers to the score difference between the central word and the second-highest-scoring valid candidate central word. A higher value indicates a more significant leading advantage of the central word relative to other candidate words, a less controversial voting result, and a higher certainty that it represents the sole core semantic representative.

[0145] The multi-source consensus rate of the central word refers to the proportion of the number of information sources for the central word to the total number of information sources. The higher this value, the higher the consensus among different heterogeneous information sources on the central word, the lower the risk of the result being affected by noise and bias from a single information source, and the stronger the reliability of the central word.

[0146] Step 105-4-2: Calculate the confidence score of the central term using the following weighted formula. The confidence score calculation of this invention comprehensively considers a composite index of three dimensions: vote share, leading advantage, and multi-source consensus. It quantitatively evaluates the representativeness, uniqueness, and stability of the central term, providing a refined and configurable decision-making basis for subsequent search strategies.

[0147]

[0148] Where: Conf represents the confidence level of the final center word; a higher value indicates a more reliable recognition result. And when... hour, When there is no valid candidate center word ranked second, .

[0149] Vc: Represents the weighted voting score of the final center word, which is the final score of the word calculated in step 105-3.

[0150] ΣVi: Represents the sum of weighted voting scores for all valid candidate center words.

[0151] V2: Represents the weighted voting score of the second-ranked valid candidate headword.

[0152] Nc: Indicates the number of heterogeneous information sources that output the final headword.

[0153] N: Represents the total number of heterogeneous information sources participating in the vote.

[0154] α, β, γ: These are weighting coefficients used to measure the influence of three indicators—vote share (Vc / ΣVi), vote advantage ((Vc-V2) / Vc), and multi-source consistency rate (Nc / N)—on the final confidence score, respectively, and satisfying α + β + γ = 1. Their values ​​need to be configured according to the emphasis placed on the reliability of the recognition results in the business scenario. For example, when the confidence score is required to better reflect the absolute representativeness of the central word, the α value is set to the maximum; when the confidence score is required to better reflect whether the final central word's winning advantage is obvious, the β value is set to the maximum. In this embodiment of the invention, the weighting coefficients of the confidence formula are set to α=0.5, β=0.3, and γ=0.2. This configuration reflects a business strategy that prioritizes the overall vote share of the central keyword (α=0.5), i.e., the high representativeness of the central keyword; while also considering the degree of its leading advantage (β=0.3) and consistency across information sources (γ=0.2). This is suitable for general search scenarios that seek both representativeness and stability for the central keyword. Substituting into the formula for calculation:

[0155] Conf = 0.5*(0.92 / 2.35) + 0.2*((0.92-0.85) / 0.92) + 0.3*(3 / 3) = 0.6

[0156] Therefore, the output center word is "light", with a confidence level of approximately 0.6.

[0157] When the selected candidate center words are empty, the system directly outputs the result of no center words; when the selected candidate center words are not empty and the final confidence level does not reach the preset threshold, the system outputs the candidate word and marks it as a low-confidence center word.

[0158] In practical applications, a confidence threshold is preset based on business needs. When the confidence of the identified headword is higher than this threshold, it is used as a high-confidence headword for search result recall, relevance calculation, and ranking. When the confidence is lower than the threshold, it is used as a low-confidence headword for expanding semantic usage, such as expanding the recall scope to avoid missing potentially relevant results; participating in multi-intent comprehensive ranking as a less weighted feature; and serving as an auxiliary input to trigger more complex semantic analysis, rather than the sole decision-making basis. Through this differentiated processing, the system can ensure high relevance of search results when the headword identification is clear, and maintain the overall recall capability and robustness of the system when there is uncertainty in the identification, avoiding search failures due to headword misjudgment.

[0159] like Figure 4 This is a structural diagram of a search term center word identification system based on a multi-source voting mechanism in an embodiment of the present invention. The system includes a search term acquisition module, a search term preprocessing module, a candidate center word extraction module, a candidate center word filtering module, and a center word determination module.

[0160] The search term acquisition module is used to acquire the search terms to be processed;

[0161] The search term preprocessing module is connected to the search term acquisition module and the preset thesaurus, and is used to perform line segmentation processing, case normalization processing, special character cleaning processing, variant word processing, and no filtering processing on the search terms;

[0162] The candidate central word extraction module is connected to the search term preprocessing module and is used to call multiple heterogeneous information sources to extract candidate central words related to the search term.

[0163] The candidate central word filtering module is connected to the candidate central word extraction module and is used to filter effective candidate central words that are consistent with the search term from the candidate central words.

[0164] The central word determination module is connected to the candidate central word filtering module and the weight configuration unit for configuring voting weights. It is used to calculate the voting score of the effective candidate central words, determine the word with the highest score as the central word, and calculate the corresponding setting information.

[0165] The above modules are implemented as software modules, hardware modules, or a combination of software and hardware, and can be deployed on the same computing device or in any distributed computing environment.

[0166] The method and system of this invention can be applied to search engines, information retrieval systems, and information matching systems for scenarios such as semantic understanding of search terms, optimization of search result matching, management of search terms, and evaluation of search quality, and can effectively improve the accuracy of search result matching and the stability of system operation.

[0167] The beneficial effects achieved by this invention are as follows:

[0168] (1) Improve the accuracy and stability of central word identification. By introducing multiple heterogeneous information sources and adopting a voting statistics mechanism, the impact of misjudgment from a single information source on the central word identification results is effectively reduced, thereby improving the consistency and reliability of the identification results.

[0169] (2) Reduce the risks associated with manual configuration and subjective intervention. This invention uses the search term itself as a constraint condition for selecting candidate center words, reducing the reliance on manually configured center words or rules and lowering the possibility of introducing irrelevant words due to improper configuration.

[0170] (3) Provide reliable quantitative information on the results of the center word identification. By analyzing the voting distribution of candidate center words, the confidence information corresponding to the center words is output, providing a reliable quantitative basis for subsequent search matching, ranking and governance decisions.

[0171] (4) It has good versatility and scalability. The method and system of the present invention can be flexibly deployed in search engines, information retrieval systems or information matching systems. The heterogeneous information source types and voting weights can be expanded or adjusted according to business needs, and it has a wide range of applications.

[0172] The above embodiments do not limit the present invention in any way. Based on the above description, those skilled in the art can make various changes and modifications without departing from the technical concept of the present invention. All other improvements and applications made to the above embodiments by equivalent transformation are within the protection scope of the present invention. The technical scope of the present invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A search term center word recognition method based on a multi-source voting mechanism, characterized in that, Includes the following steps: Step 1: Obtain the search terms to be processed: Obtain the search terms entered by the user through any of the following systems: search engine, information retrieval system, and information matching system. The search terms can be in the form of a single word, a combination of multiple words, or a phrase containing special characters. Step 2: Preprocess the search terms by converting them into a standardized set of search terms consisting of several words; Step 3: Semantically analyze the standardized search term set using multiple heterogeneous information sources to generate corresponding candidate center term sets. The heterogeneous information sources refer to semantic analysis units that are independent of each other in terms of underlying data foundation, semantic analysis algorithm, and output logic, and include at least two of the following: information sources based on vector semantic similarity, information sources based on word library analysis, and semantic recognition information sources based on model reasoning. Each heterogeneous information source has its own unique semantic judgment rules, which include the algorithm logic, matching conditions, and judgment thresholds used within the information source to quantitatively evaluate and determine whether a word can serve as the core semantic representative of the search term. Step 4: Filter candidate center words based on the search terms themselves: Compare all candidate center words output by each heterogeneous information source in Step 3 with the standardized search term set obtained in Step 2, and retain only the candidate words that exist in the standardized search term set to form an effective candidate center word set; Step 5: Weighted voting determination and confidence calculation: Statistically analyze the occurrence of each valid candidate center word in different heterogeneous information sources in Step 4. According to the pre-configured voting weight of each information source, sum the voting scores of each valid candidate center word by weight. Take the valid candidate center word with the highest voting score as the final center word, and calculate the confidence of the final center word based on the distribution of the voting results. 2.The search term center word identification method based on multi-source voting mechanism according to claim 1, characterized in that, In step 1, obtaining the search terms to be processed includes obtaining the search terms entered by the user in real time and obtaining the search terms from the historical search log data stored in the system. 3.The search term center word identification method based on multi-source voting mechanism according to claim 1, characterized in that, The preprocessing in step 2 specifically includes: performing word segmentation, case normalization, special character cleaning, word standardization, and filtering on the search terms obtained in step 1 in sequence; the word standardization includes word form regularization, spelling correction, singular / plural regularization, and synonym expression regularization; the filtering refers to filtering words with no substantial semantic contribution from the search terms based on a preset thesaurus, including banned words, stop words, and words with low semantic weight; the words with low semantic weight are words included in the preset low semantic contribution thesaurus. 4.The search term center word identification method based on multi-source voting mechanism according to claim 1, characterized in that, In step 3, each heterogeneous information source runs independently in either parallel or asynchronous mode and outputs a set of candidate center words. In step 3, the information source based on vector semantic similarity vectorizes the search terms and information items, calculates the semantic similarity between them, selects the top K information items with the highest similarity, and extracts the keywords of the information items as candidate center words; the keywords refer to words extracted from the information items that can characterize the name and core attributes of the information items; In step 3, the information source based on the thesaurus analysis uses a constructed semantic knowledge base and rule base to match search terms with words in the knowledge base. When a search term matches a word in the semantic knowledge base, the word is determined as a candidate center word. The semantic knowledge base includes at least a general language thesaurus, an industry domain thesaurus, a category thesaurus, and an attribute thesaurus. In step 3, the semantic recognition information source based on model reasoning uses a computational model with semantic understanding capabilities to identify core words by analyzing the contribution of each word in the search term to expressing the overall search intent. Specifically, the process involves obtaining a quantitative score for the importance of each word in the search terms generated by the model, and then selecting the most important words as candidate center words based on a preset score threshold or ranking rule.

5. The method for identifying the center word of a search term based on a multi-source voting mechanism according to claim 4, characterized in that, In step 3, when extracting the candidate center words based on the information source of vector semantic similarity, the frequency of each candidate center word appearing in the first K information items is counted.

6. The method for identifying the center word of a search term based on a multi-source voting mechanism according to claim 5, characterized in that, In step 4, the comparison is an exact match judgment, which specifically includes: performing an exact match on each word in the candidate center word set in step 3, and judging whether the word is exactly the same as a word in the standardized search word set. If they are the same, they are retained; otherwise, they are discarded.

7. The method for identifying the center word of a search term based on a multi-source voting mechanism according to claim 6, characterized in that, In step 5, the voting weight is calculated and configured based on the historical recognition accuracy of each heterogeneous information source; Step 5 specifically includes: Step 501: Record the output state of each valid candidate center word in each heterogeneous information source, and calculate the basic voting value, where: For an information source based on vector semantic similarity, its basic voting value is equal to the ratio of the frequency of the effective candidate center word in the first K information items to K; For other heterogeneous information sources, if the valid candidate center word is output by the information source, the basic voting value is 1; otherwise, it is 0. Step 502: Based on the voting weights configured for each heterogeneous information source, the basic voting values ​​obtained by the effective candidate center words in Step 501 are weighted, and the voting score of each effective candidate center word is calculated. Step 503: Calculate the confidence level of the final center word based on the distribution of the voting results. The confidence level is related to the weighted voting score of the final center word, the weighted voting score of the second-ranked valid candidate center word, and the number of heterogeneous information sources from which the final center word is output.

8. In the method for identifying the search term center word based on a multi-source voting mechanism according to claim 7, in step 503, the confidence level of the center word is calculated by weighting according to the following formula: ； in, Conf represents the confidence level of the final headword; Vc represents the weighted voting score of the final headword; ΣVi represents the sum of the weighted voting scores of all valid candidate headwords; V2 represents the weighted voting score of the valid candidate keyword that ranks second in votes; and when hour, ; When there is no valid candidate center word ranked second Nc represents the number of heterogeneous information sources for the final output center word; N represents the total number of heterogeneous information sources participating in the vote; α, β, and γ are all weighting coefficients used to measure the share of votes received. Voting advantage and multi-source consistency rate The degree of influence of the three indicators on the final confidence level, and α+β+γ=1.

Images